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_shared.h"
20 #include "xfs_format.h"
21 #include "xfs_log_format.h"
22 #include "xfs_trans_resv.h"
23 #include "xfs_mount.h"
24 #include "xfs_inode.h"
25 #include "xfs_trans.h"
26 #include "xfs_inode_item.h"
27 #include "xfs_alloc.h"
28 #include "xfs_error.h"
29 #include "xfs_iomap.h"
30 #include "xfs_trace.h"
31 #include "xfs_bmap.h"
32 #include "xfs_bmap_util.h"
33 #include "xfs_bmap_btree.h"
34 #include <linux/gfp.h>
35 #include <linux/mpage.h>
36 #include <linux/pagevec.h>
37 #include <linux/writeback.h>
38
39 void
40 xfs_count_page_state(
41 struct page *page,
42 int *delalloc,
43 int *unwritten)
44 {
45 struct buffer_head *bh, *head;
46
47 *delalloc = *unwritten = 0;
48
49 bh = head = page_buffers(page);
50 do {
51 if (buffer_unwritten(bh))
52 (*unwritten) = 1;
53 else if (buffer_delay(bh))
54 (*delalloc) = 1;
55 } while ((bh = bh->b_this_page) != head);
56 }
57
58 STATIC struct block_device *
59 xfs_find_bdev_for_inode(
60 struct inode *inode)
61 {
62 struct xfs_inode *ip = XFS_I(inode);
63 struct xfs_mount *mp = ip->i_mount;
64
65 if (XFS_IS_REALTIME_INODE(ip))
66 return mp->m_rtdev_targp->bt_bdev;
67 else
68 return mp->m_ddev_targp->bt_bdev;
69 }
70
71 /*
72 * We're now finished for good with this ioend structure.
73 * Update the page state via the associated buffer_heads,
74 * release holds on the inode and bio, and finally free
75 * up memory. Do not use the ioend after this.
76 */
77 STATIC void
78 xfs_destroy_ioend(
79 xfs_ioend_t *ioend)
80 {
81 struct buffer_head *bh, *next;
82
83 for (bh = ioend->io_buffer_head; bh; bh = next) {
84 next = bh->b_private;
85 bh->b_end_io(bh, !ioend->io_error);
86 }
87
88 mempool_free(ioend, xfs_ioend_pool);
89 }
90
91 /*
92 * Fast and loose check if this write could update the on-disk inode size.
93 */
94 static inline bool xfs_ioend_is_append(struct xfs_ioend *ioend)
95 {
96 return ioend->io_offset + ioend->io_size >
97 XFS_I(ioend->io_inode)->i_d.di_size;
98 }
99
100 STATIC int
101 xfs_setfilesize_trans_alloc(
102 struct xfs_ioend *ioend)
103 {
104 struct xfs_mount *mp = XFS_I(ioend->io_inode)->i_mount;
105 struct xfs_trans *tp;
106 int error;
107
108 tp = xfs_trans_alloc(mp, XFS_TRANS_FSYNC_TS);
109
110 error = xfs_trans_reserve(tp, &M_RES(mp)->tr_fsyncts, 0, 0);
111 if (error) {
112 xfs_trans_cancel(tp);
113 return error;
114 }
115
116 ioend->io_append_trans = tp;
117
118 /*
119 * We may pass freeze protection with a transaction. So tell lockdep
120 * we released it.
121 */
122 rwsem_release(&ioend->io_inode->i_sb->s_writers.lock_map[SB_FREEZE_FS-1],
123 1, _THIS_IP_);
124 /*
125 * We hand off the transaction to the completion thread now, so
126 * clear the flag here.
127 */
128 current_restore_flags_nested(&tp->t_pflags, PF_FSTRANS);
129 return 0;
130 }
131
132 /*
133 * Update on-disk file size now that data has been written to disk.
134 */
135 STATIC int
136 xfs_setfilesize(
137 struct xfs_inode *ip,
138 struct xfs_trans *tp,
139 xfs_off_t offset,
140 size_t size)
141 {
142 xfs_fsize_t isize;
143
144 xfs_ilock(ip, XFS_ILOCK_EXCL);
145 isize = xfs_new_eof(ip, offset + size);
146 if (!isize) {
147 xfs_iunlock(ip, XFS_ILOCK_EXCL);
148 xfs_trans_cancel(tp);
149 return 0;
150 }
151
152 trace_xfs_setfilesize(ip, offset, size);
153
154 ip->i_d.di_size = isize;
155 xfs_trans_ijoin(tp, ip, XFS_ILOCK_EXCL);
156 xfs_trans_log_inode(tp, ip, XFS_ILOG_CORE);
157
158 return xfs_trans_commit(tp);
159 }
160
161 STATIC int
162 xfs_setfilesize_ioend(
163 struct xfs_ioend *ioend)
164 {
165 struct xfs_inode *ip = XFS_I(ioend->io_inode);
166 struct xfs_trans *tp = ioend->io_append_trans;
167
168 /*
169 * The transaction may have been allocated in the I/O submission thread,
170 * thus we need to mark ourselves as being in a transaction manually.
171 * Similarly for freeze protection.
172 */
173 current_set_flags_nested(&tp->t_pflags, PF_FSTRANS);
174 rwsem_acquire_read(&VFS_I(ip)->i_sb->s_writers.lock_map[SB_FREEZE_FS-1],
175 0, 1, _THIS_IP_);
176
177 return xfs_setfilesize(ip, tp, ioend->io_offset, ioend->io_size);
178 }
179
180 /*
181 * Schedule IO completion handling on the final put of an ioend.
182 *
183 * If there is no work to do we might as well call it a day and free the
184 * ioend right now.
185 */
186 STATIC void
187 xfs_finish_ioend(
188 struct xfs_ioend *ioend)
189 {
190 if (atomic_dec_and_test(&ioend->io_remaining)) {
191 struct xfs_mount *mp = XFS_I(ioend->io_inode)->i_mount;
192
193 if (ioend->io_type == XFS_IO_UNWRITTEN)
194 queue_work(mp->m_unwritten_workqueue, &ioend->io_work);
195 else if (ioend->io_append_trans)
196 queue_work(mp->m_data_workqueue, &ioend->io_work);
197 else
198 xfs_destroy_ioend(ioend);
199 }
200 }
201
202 /*
203 * IO write completion.
204 */
205 STATIC void
206 xfs_end_io(
207 struct work_struct *work)
208 {
209 xfs_ioend_t *ioend = container_of(work, xfs_ioend_t, io_work);
210 struct xfs_inode *ip = XFS_I(ioend->io_inode);
211 int error = 0;
212
213 if (XFS_FORCED_SHUTDOWN(ip->i_mount)) {
214 ioend->io_error = -EIO;
215 goto done;
216 }
217 if (ioend->io_error)
218 goto done;
219
220 /*
221 * For unwritten extents we need to issue transactions to convert a
222 * range to normal written extens after the data I/O has finished.
223 */
224 if (ioend->io_type == XFS_IO_UNWRITTEN) {
225 error = xfs_iomap_write_unwritten(ip, ioend->io_offset,
226 ioend->io_size);
227 } else if (ioend->io_append_trans) {
228 error = xfs_setfilesize_ioend(ioend);
229 } else {
230 ASSERT(!xfs_ioend_is_append(ioend));
231 }
232
233 done:
234 if (error)
235 ioend->io_error = error;
236 xfs_destroy_ioend(ioend);
237 }
238
239 /*
240 * Allocate and initialise an IO completion structure.
241 * We need to track unwritten extent write completion here initially.
242 * We'll need to extend this for updating the ondisk inode size later
243 * (vs. incore size).
244 */
245 STATIC xfs_ioend_t *
246 xfs_alloc_ioend(
247 struct inode *inode,
248 unsigned int type)
249 {
250 xfs_ioend_t *ioend;
251
252 ioend = mempool_alloc(xfs_ioend_pool, GFP_NOFS);
253
254 /*
255 * Set the count to 1 initially, which will prevent an I/O
256 * completion callback from happening before we have started
257 * all the I/O from calling the completion routine too early.
258 */
259 atomic_set(&ioend->io_remaining, 1);
260 ioend->io_error = 0;
261 ioend->io_list = NULL;
262 ioend->io_type = type;
263 ioend->io_inode = inode;
264 ioend->io_buffer_head = NULL;
265 ioend->io_buffer_tail = NULL;
266 ioend->io_offset = 0;
267 ioend->io_size = 0;
268 ioend->io_append_trans = NULL;
269
270 INIT_WORK(&ioend->io_work, xfs_end_io);
271 return ioend;
272 }
273
274 STATIC int
275 xfs_map_blocks(
276 struct inode *inode,
277 loff_t offset,
278 struct xfs_bmbt_irec *imap,
279 int type,
280 int nonblocking)
281 {
282 struct xfs_inode *ip = XFS_I(inode);
283 struct xfs_mount *mp = ip->i_mount;
284 ssize_t count = 1 << inode->i_blkbits;
285 xfs_fileoff_t offset_fsb, end_fsb;
286 int error = 0;
287 int bmapi_flags = XFS_BMAPI_ENTIRE;
288 int nimaps = 1;
289
290 if (XFS_FORCED_SHUTDOWN(mp))
291 return -EIO;
292
293 if (type == XFS_IO_UNWRITTEN)
294 bmapi_flags |= XFS_BMAPI_IGSTATE;
295
296 if (!xfs_ilock_nowait(ip, XFS_ILOCK_SHARED)) {
297 if (nonblocking)
298 return -EAGAIN;
299 xfs_ilock(ip, XFS_ILOCK_SHARED);
300 }
301
302 ASSERT(ip->i_d.di_format != XFS_DINODE_FMT_BTREE ||
303 (ip->i_df.if_flags & XFS_IFEXTENTS));
304 ASSERT(offset <= mp->m_super->s_maxbytes);
305
306 if (offset + count > mp->m_super->s_maxbytes)
307 count = mp->m_super->s_maxbytes - offset;
308 end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + count);
309 offset_fsb = XFS_B_TO_FSBT(mp, offset);
310 error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
311 imap, &nimaps, bmapi_flags);
312 xfs_iunlock(ip, XFS_ILOCK_SHARED);
313
314 if (error)
315 return error;
316
317 if (type == XFS_IO_DELALLOC &&
318 (!nimaps || isnullstartblock(imap->br_startblock))) {
319 error = xfs_iomap_write_allocate(ip, offset, imap);
320 if (!error)
321 trace_xfs_map_blocks_alloc(ip, offset, count, type, imap);
322 return error;
323 }
324
325 #ifdef DEBUG
326 if (type == XFS_IO_UNWRITTEN) {
327 ASSERT(nimaps);
328 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
329 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
330 }
331 #endif
332 if (nimaps)
333 trace_xfs_map_blocks_found(ip, offset, count, type, imap);
334 return 0;
335 }
336
337 STATIC int
338 xfs_imap_valid(
339 struct inode *inode,
340 struct xfs_bmbt_irec *imap,
341 xfs_off_t offset)
342 {
343 offset >>= inode->i_blkbits;
344
345 return offset >= imap->br_startoff &&
346 offset < imap->br_startoff + imap->br_blockcount;
347 }
348
349 /*
350 * BIO completion handler for buffered IO.
351 */
352 STATIC void
353 xfs_end_bio(
354 struct bio *bio,
355 int error)
356 {
357 xfs_ioend_t *ioend = bio->bi_private;
358
359 if (!ioend->io_error && !test_bit(BIO_UPTODATE, &bio->bi_flags))
360 ioend->io_error = error;
361
362 /* Toss bio and pass work off to an xfsdatad thread */
363 bio->bi_private = NULL;
364 bio->bi_end_io = NULL;
365 bio_put(bio);
366
367 xfs_finish_ioend(ioend);
368 }
369
370 STATIC void
371 xfs_submit_ioend_bio(
372 struct writeback_control *wbc,
373 xfs_ioend_t *ioend,
374 struct bio *bio)
375 {
376 atomic_inc(&ioend->io_remaining);
377 bio->bi_private = ioend;
378 bio->bi_end_io = xfs_end_bio;
379 submit_bio(wbc->sync_mode == WB_SYNC_ALL ? WRITE_SYNC : WRITE, bio);
380 }
381
382 STATIC struct bio *
383 xfs_alloc_ioend_bio(
384 struct buffer_head *bh)
385 {
386 int nvecs = bio_get_nr_vecs(bh->b_bdev);
387 struct bio *bio = bio_alloc(GFP_NOIO, nvecs);
388
389 ASSERT(bio->bi_private == NULL);
390 bio->bi_iter.bi_sector = bh->b_blocknr * (bh->b_size >> 9);
391 bio->bi_bdev = bh->b_bdev;
392 return bio;
393 }
394
395 STATIC void
396 xfs_start_buffer_writeback(
397 struct buffer_head *bh)
398 {
399 ASSERT(buffer_mapped(bh));
400 ASSERT(buffer_locked(bh));
401 ASSERT(!buffer_delay(bh));
402 ASSERT(!buffer_unwritten(bh));
403
404 mark_buffer_async_write(bh);
405 set_buffer_uptodate(bh);
406 clear_buffer_dirty(bh);
407 }
408
409 STATIC void
410 xfs_start_page_writeback(
411 struct page *page,
412 int clear_dirty,
413 int buffers)
414 {
415 ASSERT(PageLocked(page));
416 ASSERT(!PageWriteback(page));
417
418 /*
419 * if the page was not fully cleaned, we need to ensure that the higher
420 * layers come back to it correctly. That means we need to keep the page
421 * dirty, and for WB_SYNC_ALL writeback we need to ensure the
422 * PAGECACHE_TAG_TOWRITE index mark is not removed so another attempt to
423 * write this page in this writeback sweep will be made.
424 */
425 if (clear_dirty) {
426 clear_page_dirty_for_io(page);
427 set_page_writeback(page);
428 } else
429 set_page_writeback_keepwrite(page);
430
431 unlock_page(page);
432
433 /* If no buffers on the page are to be written, finish it here */
434 if (!buffers)
435 end_page_writeback(page);
436 }
437
438 static inline int xfs_bio_add_buffer(struct bio *bio, struct buffer_head *bh)
439 {
440 return bio_add_page(bio, bh->b_page, bh->b_size, bh_offset(bh));
441 }
442
443 /*
444 * Submit all of the bios for all of the ioends we have saved up, covering the
445 * initial writepage page and also any probed pages.
446 *
447 * Because we may have multiple ioends spanning a page, we need to start
448 * writeback on all the buffers before we submit them for I/O. If we mark the
449 * buffers as we got, then we can end up with a page that only has buffers
450 * marked async write and I/O complete on can occur before we mark the other
451 * buffers async write.
452 *
453 * The end result of this is that we trip a bug in end_page_writeback() because
454 * we call it twice for the one page as the code in end_buffer_async_write()
455 * assumes that all buffers on the page are started at the same time.
456 *
457 * The fix is two passes across the ioend list - one to start writeback on the
458 * buffer_heads, and then submit them for I/O on the second pass.
459 *
460 * If @fail is non-zero, it means that we have a situation where some part of
461 * the submission process has failed after we have marked paged for writeback
462 * and unlocked them. In this situation, we need to fail the ioend chain rather
463 * than submit it to IO. This typically only happens on a filesystem shutdown.
464 */
465 STATIC void
466 xfs_submit_ioend(
467 struct writeback_control *wbc,
468 xfs_ioend_t *ioend,
469 int fail)
470 {
471 xfs_ioend_t *head = ioend;
472 xfs_ioend_t *next;
473 struct buffer_head *bh;
474 struct bio *bio;
475 sector_t lastblock = 0;
476
477 /* Pass 1 - start writeback */
478 do {
479 next = ioend->io_list;
480 for (bh = ioend->io_buffer_head; bh; bh = bh->b_private)
481 xfs_start_buffer_writeback(bh);
482 } while ((ioend = next) != NULL);
483
484 /* Pass 2 - submit I/O */
485 ioend = head;
486 do {
487 next = ioend->io_list;
488 bio = NULL;
489
490 /*
491 * If we are failing the IO now, just mark the ioend with an
492 * error and finish it. This will run IO completion immediately
493 * as there is only one reference to the ioend at this point in
494 * time.
495 */
496 if (fail) {
497 ioend->io_error = fail;
498 xfs_finish_ioend(ioend);
499 continue;
500 }
501
502 for (bh = ioend->io_buffer_head; bh; bh = bh->b_private) {
503
504 if (!bio) {
505 retry:
506 bio = xfs_alloc_ioend_bio(bh);
507 } else if (bh->b_blocknr != lastblock + 1) {
508 xfs_submit_ioend_bio(wbc, ioend, bio);
509 goto retry;
510 }
511
512 if (xfs_bio_add_buffer(bio, bh) != bh->b_size) {
513 xfs_submit_ioend_bio(wbc, ioend, bio);
514 goto retry;
515 }
516
517 lastblock = bh->b_blocknr;
518 }
519 if (bio)
520 xfs_submit_ioend_bio(wbc, ioend, bio);
521 xfs_finish_ioend(ioend);
522 } while ((ioend = next) != NULL);
523 }
524
525 /*
526 * Cancel submission of all buffer_heads so far in this endio.
527 * Toss the endio too. Only ever called for the initial page
528 * in a writepage request, so only ever one page.
529 */
530 STATIC void
531 xfs_cancel_ioend(
532 xfs_ioend_t *ioend)
533 {
534 xfs_ioend_t *next;
535 struct buffer_head *bh, *next_bh;
536
537 do {
538 next = ioend->io_list;
539 bh = ioend->io_buffer_head;
540 do {
541 next_bh = bh->b_private;
542 clear_buffer_async_write(bh);
543 /*
544 * The unwritten flag is cleared when added to the
545 * ioend. We're not submitting for I/O so mark the
546 * buffer unwritten again for next time around.
547 */
548 if (ioend->io_type == XFS_IO_UNWRITTEN)
549 set_buffer_unwritten(bh);
550 unlock_buffer(bh);
551 } while ((bh = next_bh) != NULL);
552
553 mempool_free(ioend, xfs_ioend_pool);
554 } while ((ioend = next) != NULL);
555 }
556
557 /*
558 * Test to see if we've been building up a completion structure for
559 * earlier buffers -- if so, we try to append to this ioend if we
560 * can, otherwise we finish off any current ioend and start another.
561 * Return true if we've finished the given ioend.
562 */
563 STATIC void
564 xfs_add_to_ioend(
565 struct inode *inode,
566 struct buffer_head *bh,
567 xfs_off_t offset,
568 unsigned int type,
569 xfs_ioend_t **result,
570 int need_ioend)
571 {
572 xfs_ioend_t *ioend = *result;
573
574 if (!ioend || need_ioend || type != ioend->io_type) {
575 xfs_ioend_t *previous = *result;
576
577 ioend = xfs_alloc_ioend(inode, type);
578 ioend->io_offset = offset;
579 ioend->io_buffer_head = bh;
580 ioend->io_buffer_tail = bh;
581 if (previous)
582 previous->io_list = ioend;
583 *result = ioend;
584 } else {
585 ioend->io_buffer_tail->b_private = bh;
586 ioend->io_buffer_tail = bh;
587 }
588
589 bh->b_private = NULL;
590 ioend->io_size += bh->b_size;
591 }
592
593 STATIC void
594 xfs_map_buffer(
595 struct inode *inode,
596 struct buffer_head *bh,
597 struct xfs_bmbt_irec *imap,
598 xfs_off_t offset)
599 {
600 sector_t bn;
601 struct xfs_mount *m = XFS_I(inode)->i_mount;
602 xfs_off_t iomap_offset = XFS_FSB_TO_B(m, imap->br_startoff);
603 xfs_daddr_t iomap_bn = xfs_fsb_to_db(XFS_I(inode), imap->br_startblock);
604
605 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
606 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
607
608 bn = (iomap_bn >> (inode->i_blkbits - BBSHIFT)) +
609 ((offset - iomap_offset) >> inode->i_blkbits);
610
611 ASSERT(bn || XFS_IS_REALTIME_INODE(XFS_I(inode)));
612
613 bh->b_blocknr = bn;
614 set_buffer_mapped(bh);
615 }
616
617 STATIC void
618 xfs_map_at_offset(
619 struct inode *inode,
620 struct buffer_head *bh,
621 struct xfs_bmbt_irec *imap,
622 xfs_off_t offset)
623 {
624 ASSERT(imap->br_startblock != HOLESTARTBLOCK);
625 ASSERT(imap->br_startblock != DELAYSTARTBLOCK);
626
627 xfs_map_buffer(inode, bh, imap, offset);
628 set_buffer_mapped(bh);
629 clear_buffer_delay(bh);
630 clear_buffer_unwritten(bh);
631 }
632
633 /*
634 * Test if a given page contains at least one buffer of a given @type.
635 * If @check_all_buffers is true, then we walk all the buffers in the page to
636 * try to find one of the type passed in. If it is not set, then the caller only
637 * needs to check the first buffer on the page for a match.
638 */
639 STATIC bool
640 xfs_check_page_type(
641 struct page *page,
642 unsigned int type,
643 bool check_all_buffers)
644 {
645 struct buffer_head *bh;
646 struct buffer_head *head;
647
648 if (PageWriteback(page))
649 return false;
650 if (!page->mapping)
651 return false;
652 if (!page_has_buffers(page))
653 return false;
654
655 bh = head = page_buffers(page);
656 do {
657 if (buffer_unwritten(bh)) {
658 if (type == XFS_IO_UNWRITTEN)
659 return true;
660 } else if (buffer_delay(bh)) {
661 if (type == XFS_IO_DELALLOC)
662 return true;
663 } else if (buffer_dirty(bh) && buffer_mapped(bh)) {
664 if (type == XFS_IO_OVERWRITE)
665 return true;
666 }
667
668 /* If we are only checking the first buffer, we are done now. */
669 if (!check_all_buffers)
670 break;
671 } while ((bh = bh->b_this_page) != head);
672
673 return false;
674 }
675
676 /*
677 * Allocate & map buffers for page given the extent map. Write it out.
678 * except for the original page of a writepage, this is called on
679 * delalloc/unwritten pages only, for the original page it is possible
680 * that the page has no mapping at all.
681 */
682 STATIC int
683 xfs_convert_page(
684 struct inode *inode,
685 struct page *page,
686 loff_t tindex,
687 struct xfs_bmbt_irec *imap,
688 xfs_ioend_t **ioendp,
689 struct writeback_control *wbc)
690 {
691 struct buffer_head *bh, *head;
692 xfs_off_t end_offset;
693 unsigned long p_offset;
694 unsigned int type;
695 int len, page_dirty;
696 int count = 0, done = 0, uptodate = 1;
697 xfs_off_t offset = page_offset(page);
698
699 if (page->index != tindex)
700 goto fail;
701 if (!trylock_page(page))
702 goto fail;
703 if (PageWriteback(page))
704 goto fail_unlock_page;
705 if (page->mapping != inode->i_mapping)
706 goto fail_unlock_page;
707 if (!xfs_check_page_type(page, (*ioendp)->io_type, false))
708 goto fail_unlock_page;
709
710 /*
711 * page_dirty is initially a count of buffers on the page before
712 * EOF and is decremented as we move each into a cleanable state.
713 *
714 * Derivation:
715 *
716 * End offset is the highest offset that this page should represent.
717 * If we are on the last page, (end_offset & (PAGE_CACHE_SIZE - 1))
718 * will evaluate non-zero and be less than PAGE_CACHE_SIZE and
719 * hence give us the correct page_dirty count. On any other page,
720 * it will be zero and in that case we need page_dirty to be the
721 * count of buffers on the page.
722 */
723 end_offset = min_t(unsigned long long,
724 (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT,
725 i_size_read(inode));
726
727 /*
728 * If the current map does not span the entire page we are about to try
729 * to write, then give up. The only way we can write a page that spans
730 * multiple mappings in a single writeback iteration is via the
731 * xfs_vm_writepage() function. Data integrity writeback requires the
732 * entire page to be written in a single attempt, otherwise the part of
733 * the page we don't write here doesn't get written as part of the data
734 * integrity sync.
735 *
736 * For normal writeback, we also don't attempt to write partial pages
737 * here as it simply means that write_cache_pages() will see it under
738 * writeback and ignore the page until some point in the future, at
739 * which time this will be the only page in the file that needs
740 * writeback. Hence for more optimal IO patterns, we should always
741 * avoid partial page writeback due to multiple mappings on a page here.
742 */
743 if (!xfs_imap_valid(inode, imap, end_offset))
744 goto fail_unlock_page;
745
746 len = 1 << inode->i_blkbits;
747 p_offset = min_t(unsigned long, end_offset & (PAGE_CACHE_SIZE - 1),
748 PAGE_CACHE_SIZE);
749 p_offset = p_offset ? roundup(p_offset, len) : PAGE_CACHE_SIZE;
750 page_dirty = p_offset / len;
751
752 /*
753 * The moment we find a buffer that doesn't match our current type
754 * specification or can't be written, abort the loop and start
755 * writeback. As per the above xfs_imap_valid() check, only
756 * xfs_vm_writepage() can handle partial page writeback fully - we are
757 * limited here to the buffers that are contiguous with the current
758 * ioend, and hence a buffer we can't write breaks that contiguity and
759 * we have to defer the rest of the IO to xfs_vm_writepage().
760 */
761 bh = head = page_buffers(page);
762 do {
763 if (offset >= end_offset)
764 break;
765 if (!buffer_uptodate(bh))
766 uptodate = 0;
767 if (!(PageUptodate(page) || buffer_uptodate(bh))) {
768 done = 1;
769 break;
770 }
771
772 if (buffer_unwritten(bh) || buffer_delay(bh) ||
773 buffer_mapped(bh)) {
774 if (buffer_unwritten(bh))
775 type = XFS_IO_UNWRITTEN;
776 else if (buffer_delay(bh))
777 type = XFS_IO_DELALLOC;
778 else
779 type = XFS_IO_OVERWRITE;
780
781 /*
782 * imap should always be valid because of the above
783 * partial page end_offset check on the imap.
784 */
785 ASSERT(xfs_imap_valid(inode, imap, offset));
786
787 lock_buffer(bh);
788 if (type != XFS_IO_OVERWRITE)
789 xfs_map_at_offset(inode, bh, imap, offset);
790 xfs_add_to_ioend(inode, bh, offset, type,
791 ioendp, done);
792
793 page_dirty--;
794 count++;
795 } else {
796 done = 1;
797 break;
798 }
799 } while (offset += len, (bh = bh->b_this_page) != head);
800
801 if (uptodate && bh == head)
802 SetPageUptodate(page);
803
804 if (count) {
805 if (--wbc->nr_to_write <= 0 &&
806 wbc->sync_mode == WB_SYNC_NONE)
807 done = 1;
808 }
809 xfs_start_page_writeback(page, !page_dirty, count);
810
811 return done;
812 fail_unlock_page:
813 unlock_page(page);
814 fail:
815 return 1;
816 }
817
818 /*
819 * Convert & write out a cluster of pages in the same extent as defined
820 * by mp and following the start page.
821 */
822 STATIC void
823 xfs_cluster_write(
824 struct inode *inode,
825 pgoff_t tindex,
826 struct xfs_bmbt_irec *imap,
827 xfs_ioend_t **ioendp,
828 struct writeback_control *wbc,
829 pgoff_t tlast)
830 {
831 struct pagevec pvec;
832 int done = 0, i;
833
834 pagevec_init(&pvec, 0);
835 while (!done && tindex <= tlast) {
836 unsigned len = min_t(pgoff_t, PAGEVEC_SIZE, tlast - tindex + 1);
837
838 if (!pagevec_lookup(&pvec, inode->i_mapping, tindex, len))
839 break;
840
841 for (i = 0; i < pagevec_count(&pvec); i++) {
842 done = xfs_convert_page(inode, pvec.pages[i], tindex++,
843 imap, ioendp, wbc);
844 if (done)
845 break;
846 }
847
848 pagevec_release(&pvec);
849 cond_resched();
850 }
851 }
852
853 STATIC void
854 xfs_vm_invalidatepage(
855 struct page *page,
856 unsigned int offset,
857 unsigned int length)
858 {
859 trace_xfs_invalidatepage(page->mapping->host, page, offset,
860 length);
861 block_invalidatepage(page, offset, length);
862 }
863
864 /*
865 * If the page has delalloc buffers on it, we need to punch them out before we
866 * invalidate the page. If we don't, we leave a stale delalloc mapping on the
867 * inode that can trip a BUG() in xfs_get_blocks() later on if a direct IO read
868 * is done on that same region - the delalloc extent is returned when none is
869 * supposed to be there.
870 *
871 * We prevent this by truncating away the delalloc regions on the page before
872 * invalidating it. Because they are delalloc, we can do this without needing a
873 * transaction. Indeed - if we get ENOSPC errors, we have to be able to do this
874 * truncation without a transaction as there is no space left for block
875 * reservation (typically why we see a ENOSPC in writeback).
876 *
877 * This is not a performance critical path, so for now just do the punching a
878 * buffer head at a time.
879 */
880 STATIC void
881 xfs_aops_discard_page(
882 struct page *page)
883 {
884 struct inode *inode = page->mapping->host;
885 struct xfs_inode *ip = XFS_I(inode);
886 struct buffer_head *bh, *head;
887 loff_t offset = page_offset(page);
888
889 if (!xfs_check_page_type(page, XFS_IO_DELALLOC, true))
890 goto out_invalidate;
891
892 if (XFS_FORCED_SHUTDOWN(ip->i_mount))
893 goto out_invalidate;
894
895 xfs_alert(ip->i_mount,
896 "page discard on page %p, inode 0x%llx, offset %llu.",
897 page, ip->i_ino, offset);
898
899 xfs_ilock(ip, XFS_ILOCK_EXCL);
900 bh = head = page_buffers(page);
901 do {
902 int error;
903 xfs_fileoff_t start_fsb;
904
905 if (!buffer_delay(bh))
906 goto next_buffer;
907
908 start_fsb = XFS_B_TO_FSBT(ip->i_mount, offset);
909 error = xfs_bmap_punch_delalloc_range(ip, start_fsb, 1);
910 if (error) {
911 /* something screwed, just bail */
912 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
913 xfs_alert(ip->i_mount,
914 "page discard unable to remove delalloc mapping.");
915 }
916 break;
917 }
918 next_buffer:
919 offset += 1 << inode->i_blkbits;
920
921 } while ((bh = bh->b_this_page) != head);
922
923 xfs_iunlock(ip, XFS_ILOCK_EXCL);
924 out_invalidate:
925 xfs_vm_invalidatepage(page, 0, PAGE_CACHE_SIZE);
926 return;
927 }
928
929 /*
930 * Write out a dirty page.
931 *
932 * For delalloc space on the page we need to allocate space and flush it.
933 * For unwritten space on the page we need to start the conversion to
934 * regular allocated space.
935 * For any other dirty buffer heads on the page we should flush them.
936 */
937 STATIC int
938 xfs_vm_writepage(
939 struct page *page,
940 struct writeback_control *wbc)
941 {
942 struct inode *inode = page->mapping->host;
943 struct buffer_head *bh, *head;
944 struct xfs_bmbt_irec imap;
945 xfs_ioend_t *ioend = NULL, *iohead = NULL;
946 loff_t offset;
947 unsigned int type;
948 __uint64_t end_offset;
949 pgoff_t end_index, last_index;
950 ssize_t len;
951 int err, imap_valid = 0, uptodate = 1;
952 int count = 0;
953 int nonblocking = 0;
954
955 trace_xfs_writepage(inode, page, 0, 0);
956
957 ASSERT(page_has_buffers(page));
958
959 /*
960 * Refuse to write the page out if we are called from reclaim context.
961 *
962 * This avoids stack overflows when called from deeply used stacks in
963 * random callers for direct reclaim or memcg reclaim. We explicitly
964 * allow reclaim from kswapd as the stack usage there is relatively low.
965 *
966 * This should never happen except in the case of a VM regression so
967 * warn about it.
968 */
969 if (WARN_ON_ONCE((current->flags & (PF_MEMALLOC|PF_KSWAPD)) ==
970 PF_MEMALLOC))
971 goto redirty;
972
973 /*
974 * Given that we do not allow direct reclaim to call us, we should
975 * never be called while in a filesystem transaction.
976 */
977 if (WARN_ON_ONCE(current->flags & PF_FSTRANS))
978 goto redirty;
979
980 /* Is this page beyond the end of the file? */
981 offset = i_size_read(inode);
982 end_index = offset >> PAGE_CACHE_SHIFT;
983 last_index = (offset - 1) >> PAGE_CACHE_SHIFT;
984
985 /*
986 * The page index is less than the end_index, adjust the end_offset
987 * to the highest offset that this page should represent.
988 * -----------------------------------------------------
989 * | file mapping | <EOF> |
990 * -----------------------------------------------------
991 * | Page ... | Page N-2 | Page N-1 | Page N | |
992 * ^--------------------------------^----------|--------
993 * | desired writeback range | see else |
994 * ---------------------------------^------------------|
995 */
996 if (page->index < end_index)
997 end_offset = (xfs_off_t)(page->index + 1) << PAGE_CACHE_SHIFT;
998 else {
999 /*
1000 * Check whether the page to write out is beyond or straddles
1001 * i_size or not.
1002 * -------------------------------------------------------
1003 * | file mapping | <EOF> |
1004 * -------------------------------------------------------
1005 * | Page ... | Page N-2 | Page N-1 | Page N | Beyond |
1006 * ^--------------------------------^-----------|---------
1007 * | | Straddles |
1008 * ---------------------------------^-----------|--------|
1009 */
1010 unsigned offset_into_page = offset & (PAGE_CACHE_SIZE - 1);
1011
1012 /*
1013 * Skip the page if it is fully outside i_size, e.g. due to a
1014 * truncate operation that is in progress. We must redirty the
1015 * page so that reclaim stops reclaiming it. Otherwise
1016 * xfs_vm_releasepage() is called on it and gets confused.
1017 *
1018 * Note that the end_index is unsigned long, it would overflow
1019 * if the given offset is greater than 16TB on 32-bit system
1020 * and if we do check the page is fully outside i_size or not
1021 * via "if (page->index >= end_index + 1)" as "end_index + 1"
1022 * will be evaluated to 0. Hence this page will be redirtied
1023 * and be written out repeatedly which would result in an
1024 * infinite loop, the user program that perform this operation
1025 * will hang. Instead, we can verify this situation by checking
1026 * if the page to write is totally beyond the i_size or if it's
1027 * offset is just equal to the EOF.
1028 */
1029 if (page->index > end_index ||
1030 (page->index == end_index && offset_into_page == 0))
1031 goto redirty;
1032
1033 /*
1034 * The page straddles i_size. It must be zeroed out on each
1035 * and every writepage invocation because it may be mmapped.
1036 * "A file is mapped in multiples of the page size. For a file
1037 * that is not a multiple of the page size, the remaining
1038 * memory is zeroed when mapped, and writes to that region are
1039 * not written out to the file."
1040 */
1041 zero_user_segment(page, offset_into_page, PAGE_CACHE_SIZE);
1042
1043 /* Adjust the end_offset to the end of file */
1044 end_offset = offset;
1045 }
1046
1047 len = 1 << inode->i_blkbits;
1048
1049 bh = head = page_buffers(page);
1050 offset = page_offset(page);
1051 type = XFS_IO_OVERWRITE;
1052
1053 if (wbc->sync_mode == WB_SYNC_NONE)
1054 nonblocking = 1;
1055
1056 do {
1057 int new_ioend = 0;
1058
1059 if (offset >= end_offset)
1060 break;
1061 if (!buffer_uptodate(bh))
1062 uptodate = 0;
1063
1064 /*
1065 * set_page_dirty dirties all buffers in a page, independent
1066 * of their state. The dirty state however is entirely
1067 * meaningless for holes (!mapped && uptodate), so skip
1068 * buffers covering holes here.
1069 */
1070 if (!buffer_mapped(bh) && buffer_uptodate(bh)) {
1071 imap_valid = 0;
1072 continue;
1073 }
1074
1075 if (buffer_unwritten(bh)) {
1076 if (type != XFS_IO_UNWRITTEN) {
1077 type = XFS_IO_UNWRITTEN;
1078 imap_valid = 0;
1079 }
1080 } else if (buffer_delay(bh)) {
1081 if (type != XFS_IO_DELALLOC) {
1082 type = XFS_IO_DELALLOC;
1083 imap_valid = 0;
1084 }
1085 } else if (buffer_uptodate(bh)) {
1086 if (type != XFS_IO_OVERWRITE) {
1087 type = XFS_IO_OVERWRITE;
1088 imap_valid = 0;
1089 }
1090 } else {
1091 if (PageUptodate(page))
1092 ASSERT(buffer_mapped(bh));
1093 /*
1094 * This buffer is not uptodate and will not be
1095 * written to disk. Ensure that we will put any
1096 * subsequent writeable buffers into a new
1097 * ioend.
1098 */
1099 imap_valid = 0;
1100 continue;
1101 }
1102
1103 if (imap_valid)
1104 imap_valid = xfs_imap_valid(inode, &imap, offset);
1105 if (!imap_valid) {
1106 /*
1107 * If we didn't have a valid mapping then we need to
1108 * put the new mapping into a separate ioend structure.
1109 * This ensures non-contiguous extents always have
1110 * separate ioends, which is particularly important
1111 * for unwritten extent conversion at I/O completion
1112 * time.
1113 */
1114 new_ioend = 1;
1115 err = xfs_map_blocks(inode, offset, &imap, type,
1116 nonblocking);
1117 if (err)
1118 goto error;
1119 imap_valid = xfs_imap_valid(inode, &imap, offset);
1120 }
1121 if (imap_valid) {
1122 lock_buffer(bh);
1123 if (type != XFS_IO_OVERWRITE)
1124 xfs_map_at_offset(inode, bh, &imap, offset);
1125 xfs_add_to_ioend(inode, bh, offset, type, &ioend,
1126 new_ioend);
1127 count++;
1128 }
1129
1130 if (!iohead)
1131 iohead = ioend;
1132
1133 } while (offset += len, ((bh = bh->b_this_page) != head));
1134
1135 if (uptodate && bh == head)
1136 SetPageUptodate(page);
1137
1138 xfs_start_page_writeback(page, 1, count);
1139
1140 /* if there is no IO to be submitted for this page, we are done */
1141 if (!ioend)
1142 return 0;
1143
1144 ASSERT(iohead);
1145
1146 /*
1147 * Any errors from this point onwards need tobe reported through the IO
1148 * completion path as we have marked the initial page as under writeback
1149 * and unlocked it.
1150 */
1151 if (imap_valid) {
1152 xfs_off_t end_index;
1153
1154 end_index = imap.br_startoff + imap.br_blockcount;
1155
1156 /* to bytes */
1157 end_index <<= inode->i_blkbits;
1158
1159 /* to pages */
1160 end_index = (end_index - 1) >> PAGE_CACHE_SHIFT;
1161
1162 /* check against file size */
1163 if (end_index > last_index)
1164 end_index = last_index;
1165
1166 xfs_cluster_write(inode, page->index + 1, &imap, &ioend,
1167 wbc, end_index);
1168 }
1169
1170
1171 /*
1172 * Reserve log space if we might write beyond the on-disk inode size.
1173 */
1174 err = 0;
1175 if (ioend->io_type != XFS_IO_UNWRITTEN && xfs_ioend_is_append(ioend))
1176 err = xfs_setfilesize_trans_alloc(ioend);
1177
1178 xfs_submit_ioend(wbc, iohead, err);
1179
1180 return 0;
1181
1182 error:
1183 if (iohead)
1184 xfs_cancel_ioend(iohead);
1185
1186 if (err == -EAGAIN)
1187 goto redirty;
1188
1189 xfs_aops_discard_page(page);
1190 ClearPageUptodate(page);
1191 unlock_page(page);
1192 return err;
1193
1194 redirty:
1195 redirty_page_for_writepage(wbc, page);
1196 unlock_page(page);
1197 return 0;
1198 }
1199
1200 STATIC int
1201 xfs_vm_writepages(
1202 struct address_space *mapping,
1203 struct writeback_control *wbc)
1204 {
1205 xfs_iflags_clear(XFS_I(mapping->host), XFS_ITRUNCATED);
1206 return generic_writepages(mapping, wbc);
1207 }
1208
1209 /*
1210 * Called to move a page into cleanable state - and from there
1211 * to be released. The page should already be clean. We always
1212 * have buffer heads in this call.
1213 *
1214 * Returns 1 if the page is ok to release, 0 otherwise.
1215 */
1216 STATIC int
1217 xfs_vm_releasepage(
1218 struct page *page,
1219 gfp_t gfp_mask)
1220 {
1221 int delalloc, unwritten;
1222
1223 trace_xfs_releasepage(page->mapping->host, page, 0, 0);
1224
1225 xfs_count_page_state(page, &delalloc, &unwritten);
1226
1227 if (WARN_ON_ONCE(delalloc))
1228 return 0;
1229 if (WARN_ON_ONCE(unwritten))
1230 return 0;
1231
1232 return try_to_free_buffers(page);
1233 }
1234
1235 /*
1236 * When we map a DIO buffer, we may need to attach an ioend that describes the
1237 * type of write IO we are doing. This passes to the completion function the
1238 * operations it needs to perform. If the mapping is for an overwrite wholly
1239 * within the EOF then we don't need an ioend and so we don't allocate one.
1240 * This avoids the unnecessary overhead of allocating and freeing ioends for
1241 * workloads that don't require transactions on IO completion.
1242 *
1243 * If we get multiple mappings in a single IO, we might be mapping different
1244 * types. But because the direct IO can only have a single private pointer, we
1245 * need to ensure that:
1246 *
1247 * a) i) the ioend spans the entire region of unwritten mappings; or
1248 * ii) the ioend spans all the mappings that cross or are beyond EOF; and
1249 * b) if it contains unwritten extents, it is *permanently* marked as such
1250 *
1251 * We could do this by chaining ioends like buffered IO does, but we only
1252 * actually get one IO completion callback from the direct IO, and that spans
1253 * the entire IO regardless of how many mappings and IOs are needed to complete
1254 * the DIO. There is only going to be one reference to the ioend and its life
1255 * cycle is constrained by the DIO completion code. hence we don't need
1256 * reference counting here.
1257 */
1258 static void
1259 xfs_map_direct(
1260 struct inode *inode,
1261 struct buffer_head *bh_result,
1262 struct xfs_bmbt_irec *imap,
1263 xfs_off_t offset)
1264 {
1265 struct xfs_ioend *ioend;
1266 xfs_off_t size = bh_result->b_size;
1267 int type;
1268
1269 if (ISUNWRITTEN(imap))
1270 type = XFS_IO_UNWRITTEN;
1271 else
1272 type = XFS_IO_OVERWRITE;
1273
1274 trace_xfs_gbmap_direct(XFS_I(inode), offset, size, type, imap);
1275
1276 if (bh_result->b_private) {
1277 ioend = bh_result->b_private;
1278 ASSERT(ioend->io_size > 0);
1279 ASSERT(offset >= ioend->io_offset);
1280 if (offset + size > ioend->io_offset + ioend->io_size)
1281 ioend->io_size = offset - ioend->io_offset + size;
1282
1283 if (type == XFS_IO_UNWRITTEN && type != ioend->io_type)
1284 ioend->io_type = XFS_IO_UNWRITTEN;
1285
1286 trace_xfs_gbmap_direct_update(XFS_I(inode), ioend->io_offset,
1287 ioend->io_size, ioend->io_type,
1288 imap);
1289 } else if (type == XFS_IO_UNWRITTEN ||
1290 offset + size > i_size_read(inode)) {
1291 ioend = xfs_alloc_ioend(inode, type);
1292 ioend->io_offset = offset;
1293 ioend->io_size = size;
1294
1295 bh_result->b_private = ioend;
1296 set_buffer_defer_completion(bh_result);
1297
1298 trace_xfs_gbmap_direct_new(XFS_I(inode), offset, size, type,
1299 imap);
1300 } else {
1301 trace_xfs_gbmap_direct_none(XFS_I(inode), offset, size, type,
1302 imap);
1303 }
1304 }
1305
1306 /*
1307 * If this is O_DIRECT or the mpage code calling tell them how large the mapping
1308 * is, so that we can avoid repeated get_blocks calls.
1309 *
1310 * If the mapping spans EOF, then we have to break the mapping up as the mapping
1311 * for blocks beyond EOF must be marked new so that sub block regions can be
1312 * correctly zeroed. We can't do this for mappings within EOF unless the mapping
1313 * was just allocated or is unwritten, otherwise the callers would overwrite
1314 * existing data with zeros. Hence we have to split the mapping into a range up
1315 * to and including EOF, and a second mapping for beyond EOF.
1316 */
1317 static void
1318 xfs_map_trim_size(
1319 struct inode *inode,
1320 sector_t iblock,
1321 struct buffer_head *bh_result,
1322 struct xfs_bmbt_irec *imap,
1323 xfs_off_t offset,
1324 ssize_t size)
1325 {
1326 xfs_off_t mapping_size;
1327
1328 mapping_size = imap->br_startoff + imap->br_blockcount - iblock;
1329 mapping_size <<= inode->i_blkbits;
1330
1331 ASSERT(mapping_size > 0);
1332 if (mapping_size > size)
1333 mapping_size = size;
1334 if (offset < i_size_read(inode) &&
1335 offset + mapping_size >= i_size_read(inode)) {
1336 /* limit mapping to block that spans EOF */
1337 mapping_size = roundup_64(i_size_read(inode) - offset,
1338 1 << inode->i_blkbits);
1339 }
1340 if (mapping_size > LONG_MAX)
1341 mapping_size = LONG_MAX;
1342
1343 bh_result->b_size = mapping_size;
1344 }
1345
1346 STATIC int
1347 __xfs_get_blocks(
1348 struct inode *inode,
1349 sector_t iblock,
1350 struct buffer_head *bh_result,
1351 int create,
1352 bool direct)
1353 {
1354 struct xfs_inode *ip = XFS_I(inode);
1355 struct xfs_mount *mp = ip->i_mount;
1356 xfs_fileoff_t offset_fsb, end_fsb;
1357 int error = 0;
1358 int lockmode = 0;
1359 struct xfs_bmbt_irec imap;
1360 int nimaps = 1;
1361 xfs_off_t offset;
1362 ssize_t size;
1363 int new = 0;
1364
1365 if (XFS_FORCED_SHUTDOWN(mp))
1366 return -EIO;
1367
1368 offset = (xfs_off_t)iblock << inode->i_blkbits;
1369 ASSERT(bh_result->b_size >= (1 << inode->i_blkbits));
1370 size = bh_result->b_size;
1371
1372 if (!create && direct && offset >= i_size_read(inode))
1373 return 0;
1374
1375 /*
1376 * Direct I/O is usually done on preallocated files, so try getting
1377 * a block mapping without an exclusive lock first. For buffered
1378 * writes we already have the exclusive iolock anyway, so avoiding
1379 * a lock roundtrip here by taking the ilock exclusive from the
1380 * beginning is a useful micro optimization.
1381 */
1382 if (create && !direct) {
1383 lockmode = XFS_ILOCK_EXCL;
1384 xfs_ilock(ip, lockmode);
1385 } else {
1386 lockmode = xfs_ilock_data_map_shared(ip);
1387 }
1388
1389 ASSERT(offset <= mp->m_super->s_maxbytes);
1390 if (offset + size > mp->m_super->s_maxbytes)
1391 size = mp->m_super->s_maxbytes - offset;
1392 end_fsb = XFS_B_TO_FSB(mp, (xfs_ufsize_t)offset + size);
1393 offset_fsb = XFS_B_TO_FSBT(mp, offset);
1394
1395 error = xfs_bmapi_read(ip, offset_fsb, end_fsb - offset_fsb,
1396 &imap, &nimaps, XFS_BMAPI_ENTIRE);
1397 if (error)
1398 goto out_unlock;
1399
1400 if (create &&
1401 (!nimaps ||
1402 (imap.br_startblock == HOLESTARTBLOCK ||
1403 imap.br_startblock == DELAYSTARTBLOCK))) {
1404 if (direct || xfs_get_extsz_hint(ip)) {
1405 /*
1406 * Drop the ilock in preparation for starting the block
1407 * allocation transaction. It will be retaken
1408 * exclusively inside xfs_iomap_write_direct for the
1409 * actual allocation.
1410 */
1411 xfs_iunlock(ip, lockmode);
1412 error = xfs_iomap_write_direct(ip, offset, size,
1413 &imap, nimaps);
1414 if (error)
1415 return error;
1416 new = 1;
1417
1418 } else {
1419 /*
1420 * Delalloc reservations do not require a transaction,
1421 * we can go on without dropping the lock here. If we
1422 * are allocating a new delalloc block, make sure that
1423 * we set the new flag so that we mark the buffer new so
1424 * that we know that it is newly allocated if the write
1425 * fails.
1426 */
1427 if (nimaps && imap.br_startblock == HOLESTARTBLOCK)
1428 new = 1;
1429 error = xfs_iomap_write_delay(ip, offset, size, &imap);
1430 if (error)
1431 goto out_unlock;
1432
1433 xfs_iunlock(ip, lockmode);
1434 }
1435 trace_xfs_get_blocks_alloc(ip, offset, size,
1436 ISUNWRITTEN(&imap) ? XFS_IO_UNWRITTEN
1437 : XFS_IO_DELALLOC, &imap);
1438 } else if (nimaps) {
1439 trace_xfs_get_blocks_found(ip, offset, size,
1440 ISUNWRITTEN(&imap) ? XFS_IO_UNWRITTEN
1441 : XFS_IO_OVERWRITE, &imap);
1442 xfs_iunlock(ip, lockmode);
1443 } else {
1444 trace_xfs_get_blocks_notfound(ip, offset, size);
1445 goto out_unlock;
1446 }
1447
1448 /* trim mapping down to size requested */
1449 if (direct || size > (1 << inode->i_blkbits))
1450 xfs_map_trim_size(inode, iblock, bh_result,
1451 &imap, offset, size);
1452
1453 /*
1454 * For unwritten extents do not report a disk address in the buffered
1455 * read case (treat as if we're reading into a hole).
1456 */
1457 if (imap.br_startblock != HOLESTARTBLOCK &&
1458 imap.br_startblock != DELAYSTARTBLOCK &&
1459 (create || !ISUNWRITTEN(&imap))) {
1460 xfs_map_buffer(inode, bh_result, &imap, offset);
1461 if (ISUNWRITTEN(&imap))
1462 set_buffer_unwritten(bh_result);
1463 /* direct IO needs special help */
1464 if (create && direct)
1465 xfs_map_direct(inode, bh_result, &imap, offset);
1466 }
1467
1468 /*
1469 * If this is a realtime file, data may be on a different device.
1470 * to that pointed to from the buffer_head b_bdev currently.
1471 */
1472 bh_result->b_bdev = xfs_find_bdev_for_inode(inode);
1473
1474 /*
1475 * If we previously allocated a block out beyond eof and we are now
1476 * coming back to use it then we will need to flag it as new even if it
1477 * has a disk address.
1478 *
1479 * With sub-block writes into unwritten extents we also need to mark
1480 * the buffer as new so that the unwritten parts of the buffer gets
1481 * correctly zeroed.
1482 */
1483 if (create &&
1484 ((!buffer_mapped(bh_result) && !buffer_uptodate(bh_result)) ||
1485 (offset >= i_size_read(inode)) ||
1486 (new || ISUNWRITTEN(&imap))))
1487 set_buffer_new(bh_result);
1488
1489 if (imap.br_startblock == DELAYSTARTBLOCK) {
1490 BUG_ON(direct);
1491 if (create) {
1492 set_buffer_uptodate(bh_result);
1493 set_buffer_mapped(bh_result);
1494 set_buffer_delay(bh_result);
1495 }
1496 }
1497
1498 return 0;
1499
1500 out_unlock:
1501 xfs_iunlock(ip, lockmode);
1502 return error;
1503 }
1504
1505 int
1506 xfs_get_blocks(
1507 struct inode *inode,
1508 sector_t iblock,
1509 struct buffer_head *bh_result,
1510 int create)
1511 {
1512 return __xfs_get_blocks(inode, iblock, bh_result, create, false);
1513 }
1514
1515 int
1516 xfs_get_blocks_direct(
1517 struct inode *inode,
1518 sector_t iblock,
1519 struct buffer_head *bh_result,
1520 int create)
1521 {
1522 return __xfs_get_blocks(inode, iblock, bh_result, create, true);
1523 }
1524
1525 static void
1526 __xfs_end_io_direct_write(
1527 struct inode *inode,
1528 struct xfs_ioend *ioend,
1529 loff_t offset,
1530 ssize_t size)
1531 {
1532 struct xfs_mount *mp = XFS_I(inode)->i_mount;
1533
1534 if (XFS_FORCED_SHUTDOWN(mp) || ioend->io_error)
1535 goto out_end_io;
1536
1537 /*
1538 * dio completion end_io functions are only called on writes if more
1539 * than 0 bytes was written.
1540 */
1541 ASSERT(size > 0);
1542
1543 /*
1544 * The ioend only maps whole blocks, while the IO may be sector aligned.
1545 * Hence the ioend offset/size may not match the IO offset/size exactly.
1546 * Because we don't map overwrites within EOF into the ioend, the offset
1547 * may not match, but only if the endio spans EOF. Either way, write
1548 * the IO sizes into the ioend so that completion processing does the
1549 * right thing.
1550 */
1551 ASSERT(offset + size <= ioend->io_offset + ioend->io_size);
1552 ioend->io_size = size;
1553 ioend->io_offset = offset;
1554
1555 /*
1556 * The ioend tells us whether we are doing unwritten extent conversion
1557 * or an append transaction that updates the on-disk file size. These
1558 * cases are the only cases where we should *potentially* be needing
1559 * to update the VFS inode size.
1560 *
1561 * We need to update the in-core inode size here so that we don't end up
1562 * with the on-disk inode size being outside the in-core inode size. We
1563 * have no other method of updating EOF for AIO, so always do it here
1564 * if necessary.
1565 *
1566 * We need to lock the test/set EOF update as we can be racing with
1567 * other IO completions here to update the EOF. Failing to serialise
1568 * here can result in EOF moving backwards and Bad Things Happen when
1569 * that occurs.
1570 */
1571 spin_lock(&XFS_I(inode)->i_flags_lock);
1572 if (offset + size > i_size_read(inode))
1573 i_size_write(inode, offset + size);
1574 spin_unlock(&XFS_I(inode)->i_flags_lock);
1575
1576 /*
1577 * If we are doing an append IO that needs to update the EOF on disk,
1578 * do the transaction reserve now so we can use common end io
1579 * processing. Stashing the error (if there is one) in the ioend will
1580 * result in the ioend processing passing on the error if it is
1581 * possible as we can't return it from here.
1582 */
1583 if (ioend->io_type == XFS_IO_OVERWRITE)
1584 ioend->io_error = xfs_setfilesize_trans_alloc(ioend);
1585
1586 out_end_io:
1587 xfs_end_io(&ioend->io_work);
1588 return;
1589 }
1590
1591 /*
1592 * Complete a direct I/O write request.
1593 *
1594 * The ioend structure is passed from __xfs_get_blocks() to tell us what to do.
1595 * If no ioend exists (i.e. @private == NULL) then the write IO is an overwrite
1596 * wholly within the EOF and so there is nothing for us to do. Note that in this
1597 * case the completion can be called in interrupt context, whereas if we have an
1598 * ioend we will always be called in task context (i.e. from a workqueue).
1599 */
1600 STATIC void
1601 xfs_end_io_direct_write(
1602 struct kiocb *iocb,
1603 loff_t offset,
1604 ssize_t size,
1605 void *private)
1606 {
1607 struct inode *inode = file_inode(iocb->ki_filp);
1608 struct xfs_ioend *ioend = private;
1609
1610 trace_xfs_gbmap_direct_endio(XFS_I(inode), offset, size,
1611 ioend ? ioend->io_type : 0, NULL);
1612
1613 if (!ioend) {
1614 ASSERT(offset + size <= i_size_read(inode));
1615 return;
1616 }
1617
1618 __xfs_end_io_direct_write(inode, ioend, offset, size);
1619 }
1620
1621 /*
1622 * For DAX we need a mapping buffer callback for unwritten extent conversion
1623 * when page faults allocate blocks and then zero them. Note that in this
1624 * case the mapping indicated by the ioend may extend beyond EOF. We most
1625 * definitely do not want to extend EOF here, so we trim back the ioend size to
1626 * EOF.
1627 */
1628 #ifdef CONFIG_FS_DAX
1629 void
1630 xfs_end_io_dax_write(
1631 struct buffer_head *bh,
1632 int uptodate)
1633 {
1634 struct xfs_ioend *ioend = bh->b_private;
1635 struct inode *inode = ioend->io_inode;
1636 ssize_t size = ioend->io_size;
1637
1638 ASSERT(IS_DAX(ioend->io_inode));
1639
1640 /* if there was an error zeroing, then don't convert it */
1641 if (!uptodate)
1642 ioend->io_error = -EIO;
1643
1644 /*
1645 * Trim update to EOF, so we don't extend EOF during unwritten extent
1646 * conversion of partial EOF blocks.
1647 */
1648 spin_lock(&XFS_I(inode)->i_flags_lock);
1649 if (ioend->io_offset + size > i_size_read(inode))
1650 size = i_size_read(inode) - ioend->io_offset;
1651 spin_unlock(&XFS_I(inode)->i_flags_lock);
1652
1653 __xfs_end_io_direct_write(inode, ioend, ioend->io_offset, size);
1654
1655 }
1656 #else
1657 void xfs_end_io_dax_write(struct buffer_head *bh, int uptodate) { }
1658 #endif
1659
1660 static inline ssize_t
1661 xfs_vm_do_dio(
1662 struct inode *inode,
1663 struct kiocb *iocb,
1664 struct iov_iter *iter,
1665 loff_t offset,
1666 void (*endio)(struct kiocb *iocb,
1667 loff_t offset,
1668 ssize_t size,
1669 void *private),
1670 int flags)
1671 {
1672 struct block_device *bdev;
1673
1674 if (IS_DAX(inode))
1675 return dax_do_io(iocb, inode, iter, offset,
1676 xfs_get_blocks_direct, endio, 0);
1677
1678 bdev = xfs_find_bdev_for_inode(inode);
1679 return __blockdev_direct_IO(iocb, inode, bdev, iter, offset,
1680 xfs_get_blocks_direct, endio, NULL, flags);
1681 }
1682
1683 STATIC ssize_t
1684 xfs_vm_direct_IO(
1685 struct kiocb *iocb,
1686 struct iov_iter *iter,
1687 loff_t offset)
1688 {
1689 struct inode *inode = iocb->ki_filp->f_mapping->host;
1690
1691 if (iov_iter_rw(iter) == WRITE)
1692 return xfs_vm_do_dio(inode, iocb, iter, offset,
1693 xfs_end_io_direct_write, DIO_ASYNC_EXTEND);
1694 return xfs_vm_do_dio(inode, iocb, iter, offset, NULL, 0);
1695 }
1696
1697 /*
1698 * Punch out the delalloc blocks we have already allocated.
1699 *
1700 * Don't bother with xfs_setattr given that nothing can have made it to disk yet
1701 * as the page is still locked at this point.
1702 */
1703 STATIC void
1704 xfs_vm_kill_delalloc_range(
1705 struct inode *inode,
1706 loff_t start,
1707 loff_t end)
1708 {
1709 struct xfs_inode *ip = XFS_I(inode);
1710 xfs_fileoff_t start_fsb;
1711 xfs_fileoff_t end_fsb;
1712 int error;
1713
1714 start_fsb = XFS_B_TO_FSB(ip->i_mount, start);
1715 end_fsb = XFS_B_TO_FSB(ip->i_mount, end);
1716 if (end_fsb <= start_fsb)
1717 return;
1718
1719 xfs_ilock(ip, XFS_ILOCK_EXCL);
1720 error = xfs_bmap_punch_delalloc_range(ip, start_fsb,
1721 end_fsb - start_fsb);
1722 if (error) {
1723 /* something screwed, just bail */
1724 if (!XFS_FORCED_SHUTDOWN(ip->i_mount)) {
1725 xfs_alert(ip->i_mount,
1726 "xfs_vm_write_failed: unable to clean up ino %lld",
1727 ip->i_ino);
1728 }
1729 }
1730 xfs_iunlock(ip, XFS_ILOCK_EXCL);
1731 }
1732
1733 STATIC void
1734 xfs_vm_write_failed(
1735 struct inode *inode,
1736 struct page *page,
1737 loff_t pos,
1738 unsigned len)
1739 {
1740 loff_t block_offset;
1741 loff_t block_start;
1742 loff_t block_end;
1743 loff_t from = pos & (PAGE_CACHE_SIZE - 1);
1744 loff_t to = from + len;
1745 struct buffer_head *bh, *head;
1746
1747 /*
1748 * The request pos offset might be 32 or 64 bit, this is all fine
1749 * on 64-bit platform. However, for 64-bit pos request on 32-bit
1750 * platform, the high 32-bit will be masked off if we evaluate the
1751 * block_offset via (pos & PAGE_MASK) because the PAGE_MASK is
1752 * 0xfffff000 as an unsigned long, hence the result is incorrect
1753 * which could cause the following ASSERT failed in most cases.
1754 * In order to avoid this, we can evaluate the block_offset of the
1755 * start of the page by using shifts rather than masks the mismatch
1756 * problem.
1757 */
1758 block_offset = (pos >> PAGE_CACHE_SHIFT) << PAGE_CACHE_SHIFT;
1759
1760 ASSERT(block_offset + from == pos);
1761
1762 head = page_buffers(page);
1763 block_start = 0;
1764 for (bh = head; bh != head || !block_start;
1765 bh = bh->b_this_page, block_start = block_end,
1766 block_offset += bh->b_size) {
1767 block_end = block_start + bh->b_size;
1768
1769 /* skip buffers before the write */
1770 if (block_end <= from)
1771 continue;
1772
1773 /* if the buffer is after the write, we're done */
1774 if (block_start >= to)
1775 break;
1776
1777 if (!buffer_delay(bh))
1778 continue;
1779
1780 if (!buffer_new(bh) && block_offset < i_size_read(inode))
1781 continue;
1782
1783 xfs_vm_kill_delalloc_range(inode, block_offset,
1784 block_offset + bh->b_size);
1785
1786 /*
1787 * This buffer does not contain data anymore. make sure anyone
1788 * who finds it knows that for certain.
1789 */
1790 clear_buffer_delay(bh);
1791 clear_buffer_uptodate(bh);
1792 clear_buffer_mapped(bh);
1793 clear_buffer_new(bh);
1794 clear_buffer_dirty(bh);
1795 }
1796
1797 }
1798
1799 /*
1800 * This used to call block_write_begin(), but it unlocks and releases the page
1801 * on error, and we need that page to be able to punch stale delalloc blocks out
1802 * on failure. hence we copy-n-waste it here and call xfs_vm_write_failed() at
1803 * the appropriate point.
1804 */
1805 STATIC int
1806 xfs_vm_write_begin(
1807 struct file *file,
1808 struct address_space *mapping,
1809 loff_t pos,
1810 unsigned len,
1811 unsigned flags,
1812 struct page **pagep,
1813 void **fsdata)
1814 {
1815 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
1816 struct page *page;
1817 int status;
1818
1819 ASSERT(len <= PAGE_CACHE_SIZE);
1820
1821 page = grab_cache_page_write_begin(mapping, index, flags);
1822 if (!page)
1823 return -ENOMEM;
1824
1825 status = __block_write_begin(page, pos, len, xfs_get_blocks);
1826 if (unlikely(status)) {
1827 struct inode *inode = mapping->host;
1828 size_t isize = i_size_read(inode);
1829
1830 xfs_vm_write_failed(inode, page, pos, len);
1831 unlock_page(page);
1832
1833 /*
1834 * If the write is beyond EOF, we only want to kill blocks
1835 * allocated in this write, not blocks that were previously
1836 * written successfully.
1837 */
1838 if (pos + len > isize) {
1839 ssize_t start = max_t(ssize_t, pos, isize);
1840
1841 truncate_pagecache_range(inode, start, pos + len);
1842 }
1843
1844 page_cache_release(page);
1845 page = NULL;
1846 }
1847
1848 *pagep = page;
1849 return status;
1850 }
1851
1852 /*
1853 * On failure, we only need to kill delalloc blocks beyond EOF in the range of
1854 * this specific write because they will never be written. Previous writes
1855 * beyond EOF where block allocation succeeded do not need to be trashed, so
1856 * only new blocks from this write should be trashed. For blocks within
1857 * EOF, generic_write_end() zeros them so they are safe to leave alone and be
1858 * written with all the other valid data.
1859 */
1860 STATIC int
1861 xfs_vm_write_end(
1862 struct file *file,
1863 struct address_space *mapping,
1864 loff_t pos,
1865 unsigned len,
1866 unsigned copied,
1867 struct page *page,
1868 void *fsdata)
1869 {
1870 int ret;
1871
1872 ASSERT(len <= PAGE_CACHE_SIZE);
1873
1874 ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
1875 if (unlikely(ret < len)) {
1876 struct inode *inode = mapping->host;
1877 size_t isize = i_size_read(inode);
1878 loff_t to = pos + len;
1879
1880 if (to > isize) {
1881 /* only kill blocks in this write beyond EOF */
1882 if (pos > isize)
1883 isize = pos;
1884 xfs_vm_kill_delalloc_range(inode, isize, to);
1885 truncate_pagecache_range(inode, isize, to);
1886 }
1887 }
1888 return ret;
1889 }
1890
1891 STATIC sector_t
1892 xfs_vm_bmap(
1893 struct address_space *mapping,
1894 sector_t block)
1895 {
1896 struct inode *inode = (struct inode *)mapping->host;
1897 struct xfs_inode *ip = XFS_I(inode);
1898
1899 trace_xfs_vm_bmap(XFS_I(inode));
1900 xfs_ilock(ip, XFS_IOLOCK_SHARED);
1901 filemap_write_and_wait(mapping);
1902 xfs_iunlock(ip, XFS_IOLOCK_SHARED);
1903 return generic_block_bmap(mapping, block, xfs_get_blocks);
1904 }
1905
1906 STATIC int
1907 xfs_vm_readpage(
1908 struct file *unused,
1909 struct page *page)
1910 {
1911 return mpage_readpage(page, xfs_get_blocks);
1912 }
1913
1914 STATIC int
1915 xfs_vm_readpages(
1916 struct file *unused,
1917 struct address_space *mapping,
1918 struct list_head *pages,
1919 unsigned nr_pages)
1920 {
1921 return mpage_readpages(mapping, pages, nr_pages, xfs_get_blocks);
1922 }
1923
1924 /*
1925 * This is basically a copy of __set_page_dirty_buffers() with one
1926 * small tweak: buffers beyond EOF do not get marked dirty. If we mark them
1927 * dirty, we'll never be able to clean them because we don't write buffers
1928 * beyond EOF, and that means we can't invalidate pages that span EOF
1929 * that have been marked dirty. Further, the dirty state can leak into
1930 * the file interior if the file is extended, resulting in all sorts of
1931 * bad things happening as the state does not match the underlying data.
1932 *
1933 * XXX: this really indicates that bufferheads in XFS need to die. Warts like
1934 * this only exist because of bufferheads and how the generic code manages them.
1935 */
1936 STATIC int
1937 xfs_vm_set_page_dirty(
1938 struct page *page)
1939 {
1940 struct address_space *mapping = page->mapping;
1941 struct inode *inode = mapping->host;
1942 loff_t end_offset;
1943 loff_t offset;
1944 int newly_dirty;
1945 struct mem_cgroup *memcg;
1946
1947 if (unlikely(!mapping))
1948 return !TestSetPageDirty(page);
1949
1950 end_offset = i_size_read(inode);
1951 offset = page_offset(page);
1952
1953 spin_lock(&mapping->private_lock);
1954 if (page_has_buffers(page)) {
1955 struct buffer_head *head = page_buffers(page);
1956 struct buffer_head *bh = head;
1957
1958 do {
1959 if (offset < end_offset)
1960 set_buffer_dirty(bh);
1961 bh = bh->b_this_page;
1962 offset += 1 << inode->i_blkbits;
1963 } while (bh != head);
1964 }
1965 /*
1966 * Use mem_group_begin_page_stat() to keep PageDirty synchronized with
1967 * per-memcg dirty page counters.
1968 */
1969 memcg = mem_cgroup_begin_page_stat(page);
1970 newly_dirty = !TestSetPageDirty(page);
1971 spin_unlock(&mapping->private_lock);
1972
1973 if (newly_dirty) {
1974 /* sigh - __set_page_dirty() is static, so copy it here, too */
1975 unsigned long flags;
1976
1977 spin_lock_irqsave(&mapping->tree_lock, flags);
1978 if (page->mapping) { /* Race with truncate? */
1979 WARN_ON_ONCE(!PageUptodate(page));
1980 account_page_dirtied(page, mapping, memcg);
1981 radix_tree_tag_set(&mapping->page_tree,
1982 page_index(page), PAGECACHE_TAG_DIRTY);
1983 }
1984 spin_unlock_irqrestore(&mapping->tree_lock, flags);
1985 }
1986 mem_cgroup_end_page_stat(memcg);
1987 if (newly_dirty)
1988 __mark_inode_dirty(mapping->host, I_DIRTY_PAGES);
1989 return newly_dirty;
1990 }
1991
1992 const struct address_space_operations xfs_address_space_operations = {
1993 .readpage = xfs_vm_readpage,
1994 .readpages = xfs_vm_readpages,
1995 .writepage = xfs_vm_writepage,
1996 .writepages = xfs_vm_writepages,
1997 .set_page_dirty = xfs_vm_set_page_dirty,
1998 .releasepage = xfs_vm_releasepage,
1999 .invalidatepage = xfs_vm_invalidatepage,
2000 .write_begin = xfs_vm_write_begin,
2001 .write_end = xfs_vm_write_end,
2002 .bmap = xfs_vm_bmap,
2003 .direct_IO = xfs_vm_direct_IO,
2004 .migratepage = buffer_migrate_page,
2005 .is_partially_uptodate = block_is_partially_uptodate,
2006 .error_remove_page = generic_error_remove_page,
2007 };
2008
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.