Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
4 * Copyright (C) 2004-2008 Red Hat, Inc. All rights reserved.
5 */
6
7#include <linux/sched.h>
8#include <linux/slab.h>
9#include <linux/spinlock.h>
10#include <linux/completion.h>
11#include <linux/buffer_head.h>
12#include <linux/pagemap.h>
13#include <linux/pagevec.h>
14#include <linux/mpage.h>
15#include <linux/fs.h>
16#include <linux/writeback.h>
17#include <linux/swap.h>
18#include <linux/gfs2_ondisk.h>
19#include <linux/backing-dev.h>
20#include <linux/uio.h>
21#include <trace/events/writeback.h>
22#include <linux/sched/signal.h>
23
24#include "gfs2.h"
25#include "incore.h"
26#include "bmap.h"
27#include "glock.h"
28#include "inode.h"
29#include "log.h"
30#include "meta_io.h"
31#include "quota.h"
32#include "trans.h"
33#include "rgrp.h"
34#include "super.h"
35#include "util.h"
36#include "glops.h"
37#include "aops.h"
38
39
40void gfs2_page_add_databufs(struct gfs2_inode *ip, struct page *page,
41 unsigned int from, unsigned int len)
42{
43 struct buffer_head *head = page_buffers(page);
44 unsigned int bsize = head->b_size;
45 struct buffer_head *bh;
46 unsigned int to = from + len;
47 unsigned int start, end;
48
49 for (bh = head, start = 0; bh != head || !start;
50 bh = bh->b_this_page, start = end) {
51 end = start + bsize;
52 if (end <= from)
53 continue;
54 if (start >= to)
55 break;
56 set_buffer_uptodate(bh);
57 gfs2_trans_add_data(ip->i_gl, bh);
58 }
59}
60
61/**
62 * gfs2_get_block_noalloc - Fills in a buffer head with details about a block
63 * @inode: The inode
64 * @lblock: The block number to look up
65 * @bh_result: The buffer head to return the result in
66 * @create: Non-zero if we may add block to the file
67 *
68 * Returns: errno
69 */
70
71static int gfs2_get_block_noalloc(struct inode *inode, sector_t lblock,
72 struct buffer_head *bh_result, int create)
73{
74 int error;
75
76 error = gfs2_block_map(inode, lblock, bh_result, 0);
77 if (error)
78 return error;
79 if (!buffer_mapped(bh_result))
80 return -ENODATA;
81 return 0;
82}
83
84/**
85 * gfs2_write_jdata_page - gfs2 jdata-specific version of block_write_full_page
86 * @page: The page to write
87 * @wbc: The writeback control
88 *
89 * This is the same as calling block_write_full_page, but it also
90 * writes pages outside of i_size
91 */
92static int gfs2_write_jdata_page(struct page *page,
93 struct writeback_control *wbc)
94{
95 struct inode * const inode = page->mapping->host;
96 loff_t i_size = i_size_read(inode);
97 const pgoff_t end_index = i_size >> PAGE_SHIFT;
98 unsigned offset;
99
100 /*
101 * The page straddles i_size. It must be zeroed out on each and every
102 * writepage invocation because it may be mmapped. "A file is mapped
103 * in multiples of the page size. For a file that is not a multiple of
104 * the page size, the remaining memory is zeroed when mapped, and
105 * writes to that region are not written out to the file."
106 */
107 offset = i_size & (PAGE_SIZE - 1);
108 if (page->index == end_index && offset)
109 zero_user_segment(page, offset, PAGE_SIZE);
110
111 return __block_write_full_page(inode, page, gfs2_get_block_noalloc, wbc,
112 end_buffer_async_write);
113}
114
115/**
116 * __gfs2_jdata_writepage - The core of jdata writepage
117 * @page: The page to write
118 * @wbc: The writeback control
119 *
120 * This is shared between writepage and writepages and implements the
121 * core of the writepage operation. If a transaction is required then
122 * PageChecked will have been set and the transaction will have
123 * already been started before this is called.
124 */
125
126static int __gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
127{
128 struct inode *inode = page->mapping->host;
129 struct gfs2_inode *ip = GFS2_I(inode);
130 struct gfs2_sbd *sdp = GFS2_SB(inode);
131
132 if (PageChecked(page)) {
133 ClearPageChecked(page);
134 if (!page_has_buffers(page)) {
135 create_empty_buffers(page, inode->i_sb->s_blocksize,
136 BIT(BH_Dirty)|BIT(BH_Uptodate));
137 }
138 gfs2_page_add_databufs(ip, page, 0, sdp->sd_vfs->s_blocksize);
139 }
140 return gfs2_write_jdata_page(page, wbc);
141}
142
143/**
144 * gfs2_jdata_writepage - Write complete page
145 * @page: Page to write
146 * @wbc: The writeback control
147 *
148 * Returns: errno
149 *
150 */
151
152static int gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
153{
154 struct inode *inode = page->mapping->host;
155 struct gfs2_inode *ip = GFS2_I(inode);
156 struct gfs2_sbd *sdp = GFS2_SB(inode);
157
158 if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
159 goto out;
160 if (PageChecked(page) || current->journal_info)
161 goto out_ignore;
162 return __gfs2_jdata_writepage(page, wbc);
163
164out_ignore:
165 redirty_page_for_writepage(wbc, page);
166out:
167 unlock_page(page);
168 return 0;
169}
170
171/**
172 * gfs2_writepages - Write a bunch of dirty pages back to disk
173 * @mapping: The mapping to write
174 * @wbc: Write-back control
175 *
176 * Used for both ordered and writeback modes.
177 */
178static int gfs2_writepages(struct address_space *mapping,
179 struct writeback_control *wbc)
180{
181 struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
182 struct iomap_writepage_ctx wpc = { };
183 int ret;
184
185 /*
186 * Even if we didn't write any pages here, we might still be holding
187 * dirty pages in the ail. We forcibly flush the ail because we don't
188 * want balance_dirty_pages() to loop indefinitely trying to write out
189 * pages held in the ail that it can't find.
190 */
191 ret = iomap_writepages(mapping, wbc, &wpc, &gfs2_writeback_ops);
192 if (ret == 0)
193 set_bit(SDF_FORCE_AIL_FLUSH, &sdp->sd_flags);
194 return ret;
195}
196
197/**
198 * gfs2_write_jdata_pagevec - Write back a pagevec's worth of pages
199 * @mapping: The mapping
200 * @wbc: The writeback control
201 * @pvec: The vector of pages
202 * @nr_pages: The number of pages to write
203 * @done_index: Page index
204 *
205 * Returns: non-zero if loop should terminate, zero otherwise
206 */
207
208static int gfs2_write_jdata_pagevec(struct address_space *mapping,
209 struct writeback_control *wbc,
210 struct pagevec *pvec,
211 int nr_pages,
212 pgoff_t *done_index)
213{
214 struct inode *inode = mapping->host;
215 struct gfs2_sbd *sdp = GFS2_SB(inode);
216 unsigned nrblocks = nr_pages * (PAGE_SIZE >> inode->i_blkbits);
217 int i;
218 int ret;
219
220 ret = gfs2_trans_begin(sdp, nrblocks, nrblocks);
221 if (ret < 0)
222 return ret;
223
224 for(i = 0; i < nr_pages; i++) {
225 struct page *page = pvec->pages[i];
226
227 *done_index = page->index;
228
229 lock_page(page);
230
231 if (unlikely(page->mapping != mapping)) {
232continue_unlock:
233 unlock_page(page);
234 continue;
235 }
236
237 if (!PageDirty(page)) {
238 /* someone wrote it for us */
239 goto continue_unlock;
240 }
241
242 if (PageWriteback(page)) {
243 if (wbc->sync_mode != WB_SYNC_NONE)
244 wait_on_page_writeback(page);
245 else
246 goto continue_unlock;
247 }
248
249 BUG_ON(PageWriteback(page));
250 if (!clear_page_dirty_for_io(page))
251 goto continue_unlock;
252
253 trace_wbc_writepage(wbc, inode_to_bdi(inode));
254
255 ret = __gfs2_jdata_writepage(page, wbc);
256 if (unlikely(ret)) {
257 if (ret == AOP_WRITEPAGE_ACTIVATE) {
258 unlock_page(page);
259 ret = 0;
260 } else {
261
262 /*
263 * done_index is set past this page,
264 * so media errors will not choke
265 * background writeout for the entire
266 * file. This has consequences for
267 * range_cyclic semantics (ie. it may
268 * not be suitable for data integrity
269 * writeout).
270 */
271 *done_index = page->index + 1;
272 ret = 1;
273 break;
274 }
275 }
276
277 /*
278 * We stop writing back only if we are not doing
279 * integrity sync. In case of integrity sync we have to
280 * keep going until we have written all the pages
281 * we tagged for writeback prior to entering this loop.
282 */
283 if (--wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE) {
284 ret = 1;
285 break;
286 }
287
288 }
289 gfs2_trans_end(sdp);
290 return ret;
291}
292
293/**
294 * gfs2_write_cache_jdata - Like write_cache_pages but different
295 * @mapping: The mapping to write
296 * @wbc: The writeback control
297 *
298 * The reason that we use our own function here is that we need to
299 * start transactions before we grab page locks. This allows us
300 * to get the ordering right.
301 */
302
303static int gfs2_write_cache_jdata(struct address_space *mapping,
304 struct writeback_control *wbc)
305{
306 int ret = 0;
307 int done = 0;
308 struct pagevec pvec;
309 int nr_pages;
310 pgoff_t writeback_index;
311 pgoff_t index;
312 pgoff_t end;
313 pgoff_t done_index;
314 int cycled;
315 int range_whole = 0;
316 xa_mark_t tag;
317
318 pagevec_init(&pvec);
319 if (wbc->range_cyclic) {
320 writeback_index = mapping->writeback_index; /* prev offset */
321 index = writeback_index;
322 if (index == 0)
323 cycled = 1;
324 else
325 cycled = 0;
326 end = -1;
327 } else {
328 index = wbc->range_start >> PAGE_SHIFT;
329 end = wbc->range_end >> PAGE_SHIFT;
330 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
331 range_whole = 1;
332 cycled = 1; /* ignore range_cyclic tests */
333 }
334 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
335 tag = PAGECACHE_TAG_TOWRITE;
336 else
337 tag = PAGECACHE_TAG_DIRTY;
338
339retry:
340 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
341 tag_pages_for_writeback(mapping, index, end);
342 done_index = index;
343 while (!done && (index <= end)) {
344 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
345 tag);
346 if (nr_pages == 0)
347 break;
348
349 ret = gfs2_write_jdata_pagevec(mapping, wbc, &pvec, nr_pages, &done_index);
350 if (ret)
351 done = 1;
352 if (ret > 0)
353 ret = 0;
354 pagevec_release(&pvec);
355 cond_resched();
356 }
357
358 if (!cycled && !done) {
359 /*
360 * range_cyclic:
361 * We hit the last page and there is more work to be done: wrap
362 * back to the start of the file
363 */
364 cycled = 1;
365 index = 0;
366 end = writeback_index - 1;
367 goto retry;
368 }
369
370 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
371 mapping->writeback_index = done_index;
372
373 return ret;
374}
375
376
377/**
378 * gfs2_jdata_writepages - Write a bunch of dirty pages back to disk
379 * @mapping: The mapping to write
380 * @wbc: The writeback control
381 *
382 */
383
384static int gfs2_jdata_writepages(struct address_space *mapping,
385 struct writeback_control *wbc)
386{
387 struct gfs2_inode *ip = GFS2_I(mapping->host);
388 struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
389 int ret;
390
391 ret = gfs2_write_cache_jdata(mapping, wbc);
392 if (ret == 0 && wbc->sync_mode == WB_SYNC_ALL) {
393 gfs2_log_flush(sdp, ip->i_gl, GFS2_LOG_HEAD_FLUSH_NORMAL |
394 GFS2_LFC_JDATA_WPAGES);
395 ret = gfs2_write_cache_jdata(mapping, wbc);
396 }
397 return ret;
398}
399
400/**
401 * stuffed_readpage - Fill in a Linux page with stuffed file data
402 * @ip: the inode
403 * @page: the page
404 *
405 * Returns: errno
406 */
407static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
408{
409 struct buffer_head *dibh;
410 u64 dsize = i_size_read(&ip->i_inode);
411 void *kaddr;
412 int error;
413
414 /*
415 * Due to the order of unstuffing files and ->fault(), we can be
416 * asked for a zero page in the case of a stuffed file being extended,
417 * so we need to supply one here. It doesn't happen often.
418 */
419 if (unlikely(page->index)) {
420 zero_user(page, 0, PAGE_SIZE);
421 SetPageUptodate(page);
422 return 0;
423 }
424
425 error = gfs2_meta_inode_buffer(ip, &dibh);
426 if (error)
427 return error;
428
429 kaddr = kmap_atomic(page);
430 memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode), dsize);
431 memset(kaddr + dsize, 0, PAGE_SIZE - dsize);
432 kunmap_atomic(kaddr);
433 flush_dcache_page(page);
434 brelse(dibh);
435 SetPageUptodate(page);
436
437 return 0;
438}
439
440/**
441 * gfs2_read_folio - read a folio from a file
442 * @file: The file to read
443 * @folio: The folio in the file
444 */
445static int gfs2_read_folio(struct file *file, struct folio *folio)
446{
447 struct inode *inode = folio->mapping->host;
448 struct gfs2_inode *ip = GFS2_I(inode);
449 struct gfs2_sbd *sdp = GFS2_SB(inode);
450 int error;
451
452 if (!gfs2_is_jdata(ip) ||
453 (i_blocksize(inode) == PAGE_SIZE && !folio_buffers(folio))) {
454 error = iomap_read_folio(folio, &gfs2_iomap_ops);
455 } else if (gfs2_is_stuffed(ip)) {
456 error = stuffed_readpage(ip, &folio->page);
457 folio_unlock(folio);
458 } else {
459 error = mpage_read_folio(folio, gfs2_block_map);
460 }
461
462 if (unlikely(gfs2_withdrawn(sdp)))
463 return -EIO;
464
465 return error;
466}
467
468/**
469 * gfs2_internal_read - read an internal file
470 * @ip: The gfs2 inode
471 * @buf: The buffer to fill
472 * @pos: The file position
473 * @size: The amount to read
474 *
475 */
476
477int gfs2_internal_read(struct gfs2_inode *ip, char *buf, loff_t *pos,
478 unsigned size)
479{
480 struct address_space *mapping = ip->i_inode.i_mapping;
481 unsigned long index = *pos >> PAGE_SHIFT;
482 unsigned offset = *pos & (PAGE_SIZE - 1);
483 unsigned copied = 0;
484 unsigned amt;
485 struct page *page;
486 void *p;
487
488 do {
489 amt = size - copied;
490 if (offset + size > PAGE_SIZE)
491 amt = PAGE_SIZE - offset;
492 page = read_cache_page(mapping, index, gfs2_read_folio, NULL);
493 if (IS_ERR(page))
494 return PTR_ERR(page);
495 p = kmap_atomic(page);
496 memcpy(buf + copied, p + offset, amt);
497 kunmap_atomic(p);
498 put_page(page);
499 copied += amt;
500 index++;
501 offset = 0;
502 } while(copied < size);
503 (*pos) += size;
504 return size;
505}
506
507/**
508 * gfs2_readahead - Read a bunch of pages at once
509 * @rac: Read-ahead control structure
510 *
511 * Some notes:
512 * 1. This is only for readahead, so we can simply ignore any things
513 * which are slightly inconvenient (such as locking conflicts between
514 * the page lock and the glock) and return having done no I/O. Its
515 * obviously not something we'd want to do on too regular a basis.
516 * Any I/O we ignore at this time will be done via readpage later.
517 * 2. We don't handle stuffed files here we let readpage do the honours.
518 * 3. mpage_readahead() does most of the heavy lifting in the common case.
519 * 4. gfs2_block_map() is relied upon to set BH_Boundary in the right places.
520 */
521
522static void gfs2_readahead(struct readahead_control *rac)
523{
524 struct inode *inode = rac->mapping->host;
525 struct gfs2_inode *ip = GFS2_I(inode);
526
527 if (gfs2_is_stuffed(ip))
528 ;
529 else if (gfs2_is_jdata(ip))
530 mpage_readahead(rac, gfs2_block_map);
531 else
532 iomap_readahead(rac, &gfs2_iomap_ops);
533}
534
535/**
536 * adjust_fs_space - Adjusts the free space available due to gfs2_grow
537 * @inode: the rindex inode
538 */
539void adjust_fs_space(struct inode *inode)
540{
541 struct gfs2_sbd *sdp = GFS2_SB(inode);
542 struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
543 struct gfs2_statfs_change_host *m_sc = &sdp->sd_statfs_master;
544 struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
545 struct buffer_head *m_bh;
546 u64 fs_total, new_free;
547
548 if (gfs2_trans_begin(sdp, 2 * RES_STATFS, 0) != 0)
549 return;
550
551 /* Total up the file system space, according to the latest rindex. */
552 fs_total = gfs2_ri_total(sdp);
553 if (gfs2_meta_inode_buffer(m_ip, &m_bh) != 0)
554 goto out;
555
556 spin_lock(&sdp->sd_statfs_spin);
557 gfs2_statfs_change_in(m_sc, m_bh->b_data +
558 sizeof(struct gfs2_dinode));
559 if (fs_total > (m_sc->sc_total + l_sc->sc_total))
560 new_free = fs_total - (m_sc->sc_total + l_sc->sc_total);
561 else
562 new_free = 0;
563 spin_unlock(&sdp->sd_statfs_spin);
564 fs_warn(sdp, "File system extended by %llu blocks.\n",
565 (unsigned long long)new_free);
566 gfs2_statfs_change(sdp, new_free, new_free, 0);
567
568 update_statfs(sdp, m_bh);
569 brelse(m_bh);
570out:
571 sdp->sd_rindex_uptodate = 0;
572 gfs2_trans_end(sdp);
573}
574
575static bool jdata_dirty_folio(struct address_space *mapping,
576 struct folio *folio)
577{
578 if (current->journal_info)
579 folio_set_checked(folio);
580 return block_dirty_folio(mapping, folio);
581}
582
583/**
584 * gfs2_bmap - Block map function
585 * @mapping: Address space info
586 * @lblock: The block to map
587 *
588 * Returns: The disk address for the block or 0 on hole or error
589 */
590
591static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
592{
593 struct gfs2_inode *ip = GFS2_I(mapping->host);
594 struct gfs2_holder i_gh;
595 sector_t dblock = 0;
596 int error;
597
598 error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
599 if (error)
600 return 0;
601
602 if (!gfs2_is_stuffed(ip))
603 dblock = iomap_bmap(mapping, lblock, &gfs2_iomap_ops);
604
605 gfs2_glock_dq_uninit(&i_gh);
606
607 return dblock;
608}
609
610static void gfs2_discard(struct gfs2_sbd *sdp, struct buffer_head *bh)
611{
612 struct gfs2_bufdata *bd;
613
614 lock_buffer(bh);
615 gfs2_log_lock(sdp);
616 clear_buffer_dirty(bh);
617 bd = bh->b_private;
618 if (bd) {
619 if (!list_empty(&bd->bd_list) && !buffer_pinned(bh))
620 list_del_init(&bd->bd_list);
621 else {
622 spin_lock(&sdp->sd_ail_lock);
623 gfs2_remove_from_journal(bh, REMOVE_JDATA);
624 spin_unlock(&sdp->sd_ail_lock);
625 }
626 }
627 bh->b_bdev = NULL;
628 clear_buffer_mapped(bh);
629 clear_buffer_req(bh);
630 clear_buffer_new(bh);
631 gfs2_log_unlock(sdp);
632 unlock_buffer(bh);
633}
634
635static void gfs2_invalidate_folio(struct folio *folio, size_t offset,
636 size_t length)
637{
638 struct gfs2_sbd *sdp = GFS2_SB(folio->mapping->host);
639 size_t stop = offset + length;
640 int partial_page = (offset || length < folio_size(folio));
641 struct buffer_head *bh, *head;
642 unsigned long pos = 0;
643
644 BUG_ON(!folio_test_locked(folio));
645 if (!partial_page)
646 folio_clear_checked(folio);
647 head = folio_buffers(folio);
648 if (!head)
649 goto out;
650
651 bh = head;
652 do {
653 if (pos + bh->b_size > stop)
654 return;
655
656 if (offset <= pos)
657 gfs2_discard(sdp, bh);
658 pos += bh->b_size;
659 bh = bh->b_this_page;
660 } while (bh != head);
661out:
662 if (!partial_page)
663 filemap_release_folio(folio, 0);
664}
665
666/**
667 * gfs2_release_folio - free the metadata associated with a folio
668 * @folio: the folio that's being released
669 * @gfp_mask: passed from Linux VFS, ignored by us
670 *
671 * Calls try_to_free_buffers() to free the buffers and put the folio if the
672 * buffers can be released.
673 *
674 * Returns: true if the folio was put or else false
675 */
676
677bool gfs2_release_folio(struct folio *folio, gfp_t gfp_mask)
678{
679 struct address_space *mapping = folio->mapping;
680 struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
681 struct buffer_head *bh, *head;
682 struct gfs2_bufdata *bd;
683
684 head = folio_buffers(folio);
685 if (!head)
686 return false;
687
688 /*
689 * mm accommodates an old ext3 case where clean folios might
690 * not have had the dirty bit cleared. Thus, it can send actual
691 * dirty folios to ->release_folio() via shrink_active_list().
692 *
693 * As a workaround, we skip folios that contain dirty buffers
694 * below. Once ->release_folio isn't called on dirty folios
695 * anymore, we can warn on dirty buffers like we used to here
696 * again.
697 */
698
699 gfs2_log_lock(sdp);
700 bh = head;
701 do {
702 if (atomic_read(&bh->b_count))
703 goto cannot_release;
704 bd = bh->b_private;
705 if (bd && bd->bd_tr)
706 goto cannot_release;
707 if (buffer_dirty(bh) || WARN_ON(buffer_pinned(bh)))
708 goto cannot_release;
709 bh = bh->b_this_page;
710 } while (bh != head);
711
712 bh = head;
713 do {
714 bd = bh->b_private;
715 if (bd) {
716 gfs2_assert_warn(sdp, bd->bd_bh == bh);
717 bd->bd_bh = NULL;
718 bh->b_private = NULL;
719 /*
720 * The bd may still be queued as a revoke, in which
721 * case we must not dequeue nor free it.
722 */
723 if (!bd->bd_blkno && !list_empty(&bd->bd_list))
724 list_del_init(&bd->bd_list);
725 if (list_empty(&bd->bd_list))
726 kmem_cache_free(gfs2_bufdata_cachep, bd);
727 }
728
729 bh = bh->b_this_page;
730 } while (bh != head);
731 gfs2_log_unlock(sdp);
732
733 return try_to_free_buffers(folio);
734
735cannot_release:
736 gfs2_log_unlock(sdp);
737 return false;
738}
739
740static const struct address_space_operations gfs2_aops = {
741 .writepages = gfs2_writepages,
742 .read_folio = gfs2_read_folio,
743 .readahead = gfs2_readahead,
744 .dirty_folio = filemap_dirty_folio,
745 .release_folio = iomap_release_folio,
746 .invalidate_folio = iomap_invalidate_folio,
747 .bmap = gfs2_bmap,
748 .direct_IO = noop_direct_IO,
749 .migrate_folio = filemap_migrate_folio,
750 .is_partially_uptodate = iomap_is_partially_uptodate,
751 .error_remove_page = generic_error_remove_page,
752};
753
754static const struct address_space_operations gfs2_jdata_aops = {
755 .writepage = gfs2_jdata_writepage,
756 .writepages = gfs2_jdata_writepages,
757 .read_folio = gfs2_read_folio,
758 .readahead = gfs2_readahead,
759 .dirty_folio = jdata_dirty_folio,
760 .bmap = gfs2_bmap,
761 .invalidate_folio = gfs2_invalidate_folio,
762 .release_folio = gfs2_release_folio,
763 .is_partially_uptodate = block_is_partially_uptodate,
764 .error_remove_page = generic_error_remove_page,
765};
766
767void gfs2_set_aops(struct inode *inode)
768{
769 if (gfs2_is_jdata(GFS2_I(inode)))
770 inode->i_mapping->a_ops = &gfs2_jdata_aops;
771 else
772 inode->i_mapping->a_ops = &gfs2_aops;
773}
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * Copyright (C) Sistina Software, Inc. 1997-2003 All rights reserved.
4 * Copyright (C) 2004-2008 Red Hat, Inc. All rights reserved.
5 */
6
7#include <linux/sched.h>
8#include <linux/slab.h>
9#include <linux/spinlock.h>
10#include <linux/completion.h>
11#include <linux/buffer_head.h>
12#include <linux/pagemap.h>
13#include <linux/pagevec.h>
14#include <linux/mpage.h>
15#include <linux/fs.h>
16#include <linux/writeback.h>
17#include <linux/swap.h>
18#include <linux/gfs2_ondisk.h>
19#include <linux/backing-dev.h>
20#include <linux/uio.h>
21#include <trace/events/writeback.h>
22#include <linux/sched/signal.h>
23
24#include "gfs2.h"
25#include "incore.h"
26#include "bmap.h"
27#include "glock.h"
28#include "inode.h"
29#include "log.h"
30#include "meta_io.h"
31#include "quota.h"
32#include "trans.h"
33#include "rgrp.h"
34#include "super.h"
35#include "util.h"
36#include "glops.h"
37#include "aops.h"
38
39
40void gfs2_page_add_databufs(struct gfs2_inode *ip, struct page *page,
41 unsigned int from, unsigned int len)
42{
43 struct buffer_head *head = page_buffers(page);
44 unsigned int bsize = head->b_size;
45 struct buffer_head *bh;
46 unsigned int to = from + len;
47 unsigned int start, end;
48
49 for (bh = head, start = 0; bh != head || !start;
50 bh = bh->b_this_page, start = end) {
51 end = start + bsize;
52 if (end <= from)
53 continue;
54 if (start >= to)
55 break;
56 set_buffer_uptodate(bh);
57 gfs2_trans_add_data(ip->i_gl, bh);
58 }
59}
60
61/**
62 * gfs2_get_block_noalloc - Fills in a buffer head with details about a block
63 * @inode: The inode
64 * @lblock: The block number to look up
65 * @bh_result: The buffer head to return the result in
66 * @create: Non-zero if we may add block to the file
67 *
68 * Returns: errno
69 */
70
71static int gfs2_get_block_noalloc(struct inode *inode, sector_t lblock,
72 struct buffer_head *bh_result, int create)
73{
74 int error;
75
76 error = gfs2_block_map(inode, lblock, bh_result, 0);
77 if (error)
78 return error;
79 if (!buffer_mapped(bh_result))
80 return -EIO;
81 return 0;
82}
83
84/**
85 * gfs2_writepage - Write page for writeback mappings
86 * @page: The page
87 * @wbc: The writeback control
88 */
89static int gfs2_writepage(struct page *page, struct writeback_control *wbc)
90{
91 struct inode *inode = page->mapping->host;
92 struct gfs2_inode *ip = GFS2_I(inode);
93 struct gfs2_sbd *sdp = GFS2_SB(inode);
94 loff_t i_size = i_size_read(inode);
95 pgoff_t end_index = i_size >> PAGE_SHIFT;
96 unsigned offset;
97
98 if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
99 goto out;
100 if (current->journal_info)
101 goto redirty;
102 /* Is the page fully outside i_size? (truncate in progress) */
103 offset = i_size & (PAGE_SIZE-1);
104 if (page->index > end_index || (page->index == end_index && !offset)) {
105 page->mapping->a_ops->invalidatepage(page, 0, PAGE_SIZE);
106 goto out;
107 }
108
109 return nobh_writepage(page, gfs2_get_block_noalloc, wbc);
110
111redirty:
112 redirty_page_for_writepage(wbc, page);
113out:
114 unlock_page(page);
115 return 0;
116}
117
118/* This is the same as calling block_write_full_page, but it also
119 * writes pages outside of i_size
120 */
121static int gfs2_write_full_page(struct page *page, get_block_t *get_block,
122 struct writeback_control *wbc)
123{
124 struct inode * const inode = page->mapping->host;
125 loff_t i_size = i_size_read(inode);
126 const pgoff_t end_index = i_size >> PAGE_SHIFT;
127 unsigned offset;
128
129 /*
130 * The page straddles i_size. It must be zeroed out on each and every
131 * writepage invocation because it may be mmapped. "A file is mapped
132 * in multiples of the page size. For a file that is not a multiple of
133 * the page size, the remaining memory is zeroed when mapped, and
134 * writes to that region are not written out to the file."
135 */
136 offset = i_size & (PAGE_SIZE - 1);
137 if (page->index == end_index && offset)
138 zero_user_segment(page, offset, PAGE_SIZE);
139
140 return __block_write_full_page(inode, page, get_block, wbc,
141 end_buffer_async_write);
142}
143
144/**
145 * __gfs2_jdata_writepage - The core of jdata writepage
146 * @page: The page to write
147 * @wbc: The writeback control
148 *
149 * This is shared between writepage and writepages and implements the
150 * core of the writepage operation. If a transaction is required then
151 * PageChecked will have been set and the transaction will have
152 * already been started before this is called.
153 */
154
155static int __gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
156{
157 struct inode *inode = page->mapping->host;
158 struct gfs2_inode *ip = GFS2_I(inode);
159 struct gfs2_sbd *sdp = GFS2_SB(inode);
160
161 if (PageChecked(page)) {
162 ClearPageChecked(page);
163 if (!page_has_buffers(page)) {
164 create_empty_buffers(page, inode->i_sb->s_blocksize,
165 BIT(BH_Dirty)|BIT(BH_Uptodate));
166 }
167 gfs2_page_add_databufs(ip, page, 0, sdp->sd_vfs->s_blocksize);
168 }
169 return gfs2_write_full_page(page, gfs2_get_block_noalloc, wbc);
170}
171
172/**
173 * gfs2_jdata_writepage - Write complete page
174 * @page: Page to write
175 * @wbc: The writeback control
176 *
177 * Returns: errno
178 *
179 */
180
181static int gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
182{
183 struct inode *inode = page->mapping->host;
184 struct gfs2_inode *ip = GFS2_I(inode);
185 struct gfs2_sbd *sdp = GFS2_SB(inode);
186
187 if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
188 goto out;
189 if (PageChecked(page) || current->journal_info)
190 goto out_ignore;
191 return __gfs2_jdata_writepage(page, wbc);
192
193out_ignore:
194 redirty_page_for_writepage(wbc, page);
195out:
196 unlock_page(page);
197 return 0;
198}
199
200/**
201 * gfs2_writepages - Write a bunch of dirty pages back to disk
202 * @mapping: The mapping to write
203 * @wbc: Write-back control
204 *
205 * Used for both ordered and writeback modes.
206 */
207static int gfs2_writepages(struct address_space *mapping,
208 struct writeback_control *wbc)
209{
210 struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
211 int ret = mpage_writepages(mapping, wbc, gfs2_get_block_noalloc);
212
213 /*
214 * Even if we didn't write any pages here, we might still be holding
215 * dirty pages in the ail. We forcibly flush the ail because we don't
216 * want balance_dirty_pages() to loop indefinitely trying to write out
217 * pages held in the ail that it can't find.
218 */
219 if (ret == 0)
220 set_bit(SDF_FORCE_AIL_FLUSH, &sdp->sd_flags);
221
222 return ret;
223}
224
225/**
226 * gfs2_write_jdata_pagevec - Write back a pagevec's worth of pages
227 * @mapping: The mapping
228 * @wbc: The writeback control
229 * @pvec: The vector of pages
230 * @nr_pages: The number of pages to write
231 * @done_index: Page index
232 *
233 * Returns: non-zero if loop should terminate, zero otherwise
234 */
235
236static int gfs2_write_jdata_pagevec(struct address_space *mapping,
237 struct writeback_control *wbc,
238 struct pagevec *pvec,
239 int nr_pages,
240 pgoff_t *done_index)
241{
242 struct inode *inode = mapping->host;
243 struct gfs2_sbd *sdp = GFS2_SB(inode);
244 unsigned nrblocks = nr_pages * (PAGE_SIZE >> inode->i_blkbits);
245 int i;
246 int ret;
247
248 ret = gfs2_trans_begin(sdp, nrblocks, nrblocks);
249 if (ret < 0)
250 return ret;
251
252 for(i = 0; i < nr_pages; i++) {
253 struct page *page = pvec->pages[i];
254
255 *done_index = page->index;
256
257 lock_page(page);
258
259 if (unlikely(page->mapping != mapping)) {
260continue_unlock:
261 unlock_page(page);
262 continue;
263 }
264
265 if (!PageDirty(page)) {
266 /* someone wrote it for us */
267 goto continue_unlock;
268 }
269
270 if (PageWriteback(page)) {
271 if (wbc->sync_mode != WB_SYNC_NONE)
272 wait_on_page_writeback(page);
273 else
274 goto continue_unlock;
275 }
276
277 BUG_ON(PageWriteback(page));
278 if (!clear_page_dirty_for_io(page))
279 goto continue_unlock;
280
281 trace_wbc_writepage(wbc, inode_to_bdi(inode));
282
283 ret = __gfs2_jdata_writepage(page, wbc);
284 if (unlikely(ret)) {
285 if (ret == AOP_WRITEPAGE_ACTIVATE) {
286 unlock_page(page);
287 ret = 0;
288 } else {
289
290 /*
291 * done_index is set past this page,
292 * so media errors will not choke
293 * background writeout for the entire
294 * file. This has consequences for
295 * range_cyclic semantics (ie. it may
296 * not be suitable for data integrity
297 * writeout).
298 */
299 *done_index = page->index + 1;
300 ret = 1;
301 break;
302 }
303 }
304
305 /*
306 * We stop writing back only if we are not doing
307 * integrity sync. In case of integrity sync we have to
308 * keep going until we have written all the pages
309 * we tagged for writeback prior to entering this loop.
310 */
311 if (--wbc->nr_to_write <= 0 && wbc->sync_mode == WB_SYNC_NONE) {
312 ret = 1;
313 break;
314 }
315
316 }
317 gfs2_trans_end(sdp);
318 return ret;
319}
320
321/**
322 * gfs2_write_cache_jdata - Like write_cache_pages but different
323 * @mapping: The mapping to write
324 * @wbc: The writeback control
325 *
326 * The reason that we use our own function here is that we need to
327 * start transactions before we grab page locks. This allows us
328 * to get the ordering right.
329 */
330
331static int gfs2_write_cache_jdata(struct address_space *mapping,
332 struct writeback_control *wbc)
333{
334 int ret = 0;
335 int done = 0;
336 struct pagevec pvec;
337 int nr_pages;
338 pgoff_t writeback_index;
339 pgoff_t index;
340 pgoff_t end;
341 pgoff_t done_index;
342 int cycled;
343 int range_whole = 0;
344 xa_mark_t tag;
345
346 pagevec_init(&pvec);
347 if (wbc->range_cyclic) {
348 writeback_index = mapping->writeback_index; /* prev offset */
349 index = writeback_index;
350 if (index == 0)
351 cycled = 1;
352 else
353 cycled = 0;
354 end = -1;
355 } else {
356 index = wbc->range_start >> PAGE_SHIFT;
357 end = wbc->range_end >> PAGE_SHIFT;
358 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
359 range_whole = 1;
360 cycled = 1; /* ignore range_cyclic tests */
361 }
362 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
363 tag = PAGECACHE_TAG_TOWRITE;
364 else
365 tag = PAGECACHE_TAG_DIRTY;
366
367retry:
368 if (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)
369 tag_pages_for_writeback(mapping, index, end);
370 done_index = index;
371 while (!done && (index <= end)) {
372 nr_pages = pagevec_lookup_range_tag(&pvec, mapping, &index, end,
373 tag);
374 if (nr_pages == 0)
375 break;
376
377 ret = gfs2_write_jdata_pagevec(mapping, wbc, &pvec, nr_pages, &done_index);
378 if (ret)
379 done = 1;
380 if (ret > 0)
381 ret = 0;
382 pagevec_release(&pvec);
383 cond_resched();
384 }
385
386 if (!cycled && !done) {
387 /*
388 * range_cyclic:
389 * We hit the last page and there is more work to be done: wrap
390 * back to the start of the file
391 */
392 cycled = 1;
393 index = 0;
394 end = writeback_index - 1;
395 goto retry;
396 }
397
398 if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
399 mapping->writeback_index = done_index;
400
401 return ret;
402}
403
404
405/**
406 * gfs2_jdata_writepages - Write a bunch of dirty pages back to disk
407 * @mapping: The mapping to write
408 * @wbc: The writeback control
409 *
410 */
411
412static int gfs2_jdata_writepages(struct address_space *mapping,
413 struct writeback_control *wbc)
414{
415 struct gfs2_inode *ip = GFS2_I(mapping->host);
416 struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
417 int ret;
418
419 ret = gfs2_write_cache_jdata(mapping, wbc);
420 if (ret == 0 && wbc->sync_mode == WB_SYNC_ALL) {
421 gfs2_log_flush(sdp, ip->i_gl, GFS2_LOG_HEAD_FLUSH_NORMAL |
422 GFS2_LFC_JDATA_WPAGES);
423 ret = gfs2_write_cache_jdata(mapping, wbc);
424 }
425 return ret;
426}
427
428/**
429 * stuffed_readpage - Fill in a Linux page with stuffed file data
430 * @ip: the inode
431 * @page: the page
432 *
433 * Returns: errno
434 */
435static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
436{
437 struct buffer_head *dibh;
438 u64 dsize = i_size_read(&ip->i_inode);
439 void *kaddr;
440 int error;
441
442 /*
443 * Due to the order of unstuffing files and ->fault(), we can be
444 * asked for a zero page in the case of a stuffed file being extended,
445 * so we need to supply one here. It doesn't happen often.
446 */
447 if (unlikely(page->index)) {
448 zero_user(page, 0, PAGE_SIZE);
449 SetPageUptodate(page);
450 return 0;
451 }
452
453 error = gfs2_meta_inode_buffer(ip, &dibh);
454 if (error)
455 return error;
456
457 kaddr = kmap_atomic(page);
458 if (dsize > gfs2_max_stuffed_size(ip))
459 dsize = gfs2_max_stuffed_size(ip);
460 memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode), dsize);
461 memset(kaddr + dsize, 0, PAGE_SIZE - dsize);
462 kunmap_atomic(kaddr);
463 flush_dcache_page(page);
464 brelse(dibh);
465 SetPageUptodate(page);
466
467 return 0;
468}
469
470
471static int __gfs2_readpage(void *file, struct page *page)
472{
473 struct gfs2_inode *ip = GFS2_I(page->mapping->host);
474 struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
475 int error;
476
477 if (i_blocksize(page->mapping->host) == PAGE_SIZE &&
478 !page_has_buffers(page)) {
479 error = iomap_readpage(page, &gfs2_iomap_ops);
480 } else if (gfs2_is_stuffed(ip)) {
481 error = stuffed_readpage(ip, page);
482 unlock_page(page);
483 } else {
484 error = mpage_readpage(page, gfs2_block_map);
485 }
486
487 if (unlikely(gfs2_withdrawn(sdp)))
488 return -EIO;
489
490 return error;
491}
492
493/**
494 * gfs2_readpage - read a page of a file
495 * @file: The file to read
496 * @page: The page of the file
497 */
498
499static int gfs2_readpage(struct file *file, struct page *page)
500{
501 return __gfs2_readpage(file, page);
502}
503
504/**
505 * gfs2_internal_read - read an internal file
506 * @ip: The gfs2 inode
507 * @buf: The buffer to fill
508 * @pos: The file position
509 * @size: The amount to read
510 *
511 */
512
513int gfs2_internal_read(struct gfs2_inode *ip, char *buf, loff_t *pos,
514 unsigned size)
515{
516 struct address_space *mapping = ip->i_inode.i_mapping;
517 unsigned long index = *pos >> PAGE_SHIFT;
518 unsigned offset = *pos & (PAGE_SIZE - 1);
519 unsigned copied = 0;
520 unsigned amt;
521 struct page *page;
522 void *p;
523
524 do {
525 amt = size - copied;
526 if (offset + size > PAGE_SIZE)
527 amt = PAGE_SIZE - offset;
528 page = read_cache_page(mapping, index, __gfs2_readpage, NULL);
529 if (IS_ERR(page))
530 return PTR_ERR(page);
531 p = kmap_atomic(page);
532 memcpy(buf + copied, p + offset, amt);
533 kunmap_atomic(p);
534 put_page(page);
535 copied += amt;
536 index++;
537 offset = 0;
538 } while(copied < size);
539 (*pos) += size;
540 return size;
541}
542
543/**
544 * gfs2_readahead - Read a bunch of pages at once
545 * @file: The file to read from
546 * @mapping: Address space info
547 * @pages: List of pages to read
548 * @nr_pages: Number of pages to read
549 *
550 * Some notes:
551 * 1. This is only for readahead, so we can simply ignore any things
552 * which are slightly inconvenient (such as locking conflicts between
553 * the page lock and the glock) and return having done no I/O. Its
554 * obviously not something we'd want to do on too regular a basis.
555 * Any I/O we ignore at this time will be done via readpage later.
556 * 2. We don't handle stuffed files here we let readpage do the honours.
557 * 3. mpage_readahead() does most of the heavy lifting in the common case.
558 * 4. gfs2_block_map() is relied upon to set BH_Boundary in the right places.
559 */
560
561static void gfs2_readahead(struct readahead_control *rac)
562{
563 struct inode *inode = rac->mapping->host;
564 struct gfs2_inode *ip = GFS2_I(inode);
565
566 if (!gfs2_is_stuffed(ip))
567 mpage_readahead(rac, gfs2_block_map);
568}
569
570/**
571 * adjust_fs_space - Adjusts the free space available due to gfs2_grow
572 * @inode: the rindex inode
573 */
574void adjust_fs_space(struct inode *inode)
575{
576 struct gfs2_sbd *sdp = GFS2_SB(inode);
577 struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
578 struct gfs2_inode *l_ip = GFS2_I(sdp->sd_sc_inode);
579 struct gfs2_statfs_change_host *m_sc = &sdp->sd_statfs_master;
580 struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
581 struct buffer_head *m_bh, *l_bh;
582 u64 fs_total, new_free;
583
584 if (gfs2_trans_begin(sdp, 2 * RES_STATFS, 0) != 0)
585 return;
586
587 /* Total up the file system space, according to the latest rindex. */
588 fs_total = gfs2_ri_total(sdp);
589 if (gfs2_meta_inode_buffer(m_ip, &m_bh) != 0)
590 goto out;
591
592 spin_lock(&sdp->sd_statfs_spin);
593 gfs2_statfs_change_in(m_sc, m_bh->b_data +
594 sizeof(struct gfs2_dinode));
595 if (fs_total > (m_sc->sc_total + l_sc->sc_total))
596 new_free = fs_total - (m_sc->sc_total + l_sc->sc_total);
597 else
598 new_free = 0;
599 spin_unlock(&sdp->sd_statfs_spin);
600 fs_warn(sdp, "File system extended by %llu blocks.\n",
601 (unsigned long long)new_free);
602 gfs2_statfs_change(sdp, new_free, new_free, 0);
603
604 if (gfs2_meta_inode_buffer(l_ip, &l_bh) != 0)
605 goto out2;
606 update_statfs(sdp, m_bh, l_bh);
607 brelse(l_bh);
608out2:
609 brelse(m_bh);
610out:
611 sdp->sd_rindex_uptodate = 0;
612 gfs2_trans_end(sdp);
613}
614
615/**
616 * jdata_set_page_dirty - Page dirtying function
617 * @page: The page to dirty
618 *
619 * Returns: 1 if it dirtyed the page, or 0 otherwise
620 */
621
622static int jdata_set_page_dirty(struct page *page)
623{
624 SetPageChecked(page);
625 return __set_page_dirty_buffers(page);
626}
627
628/**
629 * gfs2_bmap - Block map function
630 * @mapping: Address space info
631 * @lblock: The block to map
632 *
633 * Returns: The disk address for the block or 0 on hole or error
634 */
635
636static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
637{
638 struct gfs2_inode *ip = GFS2_I(mapping->host);
639 struct gfs2_holder i_gh;
640 sector_t dblock = 0;
641 int error;
642
643 error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
644 if (error)
645 return 0;
646
647 if (!gfs2_is_stuffed(ip))
648 dblock = iomap_bmap(mapping, lblock, &gfs2_iomap_ops);
649
650 gfs2_glock_dq_uninit(&i_gh);
651
652 return dblock;
653}
654
655static void gfs2_discard(struct gfs2_sbd *sdp, struct buffer_head *bh)
656{
657 struct gfs2_bufdata *bd;
658
659 lock_buffer(bh);
660 gfs2_log_lock(sdp);
661 clear_buffer_dirty(bh);
662 bd = bh->b_private;
663 if (bd) {
664 if (!list_empty(&bd->bd_list) && !buffer_pinned(bh))
665 list_del_init(&bd->bd_list);
666 else
667 gfs2_remove_from_journal(bh, REMOVE_JDATA);
668 }
669 bh->b_bdev = NULL;
670 clear_buffer_mapped(bh);
671 clear_buffer_req(bh);
672 clear_buffer_new(bh);
673 gfs2_log_unlock(sdp);
674 unlock_buffer(bh);
675}
676
677static void gfs2_invalidatepage(struct page *page, unsigned int offset,
678 unsigned int length)
679{
680 struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
681 unsigned int stop = offset + length;
682 int partial_page = (offset || length < PAGE_SIZE);
683 struct buffer_head *bh, *head;
684 unsigned long pos = 0;
685
686 BUG_ON(!PageLocked(page));
687 if (!partial_page)
688 ClearPageChecked(page);
689 if (!page_has_buffers(page))
690 goto out;
691
692 bh = head = page_buffers(page);
693 do {
694 if (pos + bh->b_size > stop)
695 return;
696
697 if (offset <= pos)
698 gfs2_discard(sdp, bh);
699 pos += bh->b_size;
700 bh = bh->b_this_page;
701 } while (bh != head);
702out:
703 if (!partial_page)
704 try_to_release_page(page, 0);
705}
706
707/**
708 * gfs2_releasepage - free the metadata associated with a page
709 * @page: the page that's being released
710 * @gfp_mask: passed from Linux VFS, ignored by us
711 *
712 * Calls try_to_free_buffers() to free the buffers and put the page if the
713 * buffers can be released.
714 *
715 * Returns: 1 if the page was put or else 0
716 */
717
718int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
719{
720 struct address_space *mapping = page->mapping;
721 struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
722 struct buffer_head *bh, *head;
723 struct gfs2_bufdata *bd;
724
725 if (!page_has_buffers(page))
726 return 0;
727
728 /*
729 * From xfs_vm_releasepage: mm accommodates an old ext3 case where
730 * clean pages might not have had the dirty bit cleared. Thus, it can
731 * send actual dirty pages to ->releasepage() via shrink_active_list().
732 *
733 * As a workaround, we skip pages that contain dirty buffers below.
734 * Once ->releasepage isn't called on dirty pages anymore, we can warn
735 * on dirty buffers like we used to here again.
736 */
737
738 gfs2_log_lock(sdp);
739 spin_lock(&sdp->sd_ail_lock);
740 head = bh = page_buffers(page);
741 do {
742 if (atomic_read(&bh->b_count))
743 goto cannot_release;
744 bd = bh->b_private;
745 if (bd && bd->bd_tr)
746 goto cannot_release;
747 if (buffer_dirty(bh) || WARN_ON(buffer_pinned(bh)))
748 goto cannot_release;
749 bh = bh->b_this_page;
750 } while(bh != head);
751 spin_unlock(&sdp->sd_ail_lock);
752
753 head = bh = page_buffers(page);
754 do {
755 bd = bh->b_private;
756 if (bd) {
757 gfs2_assert_warn(sdp, bd->bd_bh == bh);
758 bd->bd_bh = NULL;
759 bh->b_private = NULL;
760 /*
761 * The bd may still be queued as a revoke, in which
762 * case we must not dequeue nor free it.
763 */
764 if (!bd->bd_blkno && !list_empty(&bd->bd_list))
765 list_del_init(&bd->bd_list);
766 if (list_empty(&bd->bd_list))
767 kmem_cache_free(gfs2_bufdata_cachep, bd);
768 }
769
770 bh = bh->b_this_page;
771 } while (bh != head);
772 gfs2_log_unlock(sdp);
773
774 return try_to_free_buffers(page);
775
776cannot_release:
777 spin_unlock(&sdp->sd_ail_lock);
778 gfs2_log_unlock(sdp);
779 return 0;
780}
781
782static const struct address_space_operations gfs2_aops = {
783 .writepage = gfs2_writepage,
784 .writepages = gfs2_writepages,
785 .readpage = gfs2_readpage,
786 .readahead = gfs2_readahead,
787 .bmap = gfs2_bmap,
788 .invalidatepage = gfs2_invalidatepage,
789 .releasepage = gfs2_releasepage,
790 .direct_IO = noop_direct_IO,
791 .migratepage = buffer_migrate_page,
792 .is_partially_uptodate = block_is_partially_uptodate,
793 .error_remove_page = generic_error_remove_page,
794};
795
796static const struct address_space_operations gfs2_jdata_aops = {
797 .writepage = gfs2_jdata_writepage,
798 .writepages = gfs2_jdata_writepages,
799 .readpage = gfs2_readpage,
800 .readahead = gfs2_readahead,
801 .set_page_dirty = jdata_set_page_dirty,
802 .bmap = gfs2_bmap,
803 .invalidatepage = gfs2_invalidatepage,
804 .releasepage = gfs2_releasepage,
805 .is_partially_uptodate = block_is_partially_uptodate,
806 .error_remove_page = generic_error_remove_page,
807};
808
809void gfs2_set_aops(struct inode *inode)
810{
811 if (gfs2_is_jdata(GFS2_I(inode)))
812 inode->i_mapping->a_ops = &gfs2_jdata_aops;
813 else
814 inode->i_mapping->a_ops = &gfs2_aops;
815}