xref: /openbmc/linux/fs/gfs2/aops.c (revision f17f06a0)
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 
40 void 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 
71 static 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  */
89 static 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 
111 redirty:
112 	redirty_page_for_writepage(wbc, page);
113 out:
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  */
121 static 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 
155 static 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 
181 static 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 
193 out_ignore:
194 	redirty_page_for_writepage(wbc, page);
195 out:
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  */
207 static 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 
236 static 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)) {
260 continue_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 
331 static 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 uninitialized_var(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 
367 retry:
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 
412 static 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  */
435 static 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 
471 /**
472  * __gfs2_readpage - readpage
473  * @file: The file to read a page for
474  * @page: The page to read
475  *
476  * This is the core of gfs2's readpage. It's used by the internal file
477  * reading code as in that case we already hold the glock. Also it's
478  * called by gfs2_readpage() once the required lock has been granted.
479  */
480 
481 static int __gfs2_readpage(void *file, struct page *page)
482 {
483 	struct gfs2_inode *ip = GFS2_I(page->mapping->host);
484 	struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
485 
486 	int error;
487 
488 	if (i_blocksize(page->mapping->host) == PAGE_SIZE &&
489 	    !page_has_buffers(page)) {
490 		error = iomap_readpage(page, &gfs2_iomap_ops);
491 	} else if (gfs2_is_stuffed(ip)) {
492 		error = stuffed_readpage(ip, page);
493 		unlock_page(page);
494 	} else {
495 		error = mpage_readpage(page, gfs2_block_map);
496 	}
497 
498 	if (unlikely(gfs2_withdrawn(sdp)))
499 		return -EIO;
500 
501 	return error;
502 }
503 
504 /**
505  * gfs2_readpage - read a page of a file
506  * @file: The file to read
507  * @page: The page of the file
508  *
509  * This deals with the locking required. We have to unlock and
510  * relock the page in order to get the locking in the right
511  * order.
512  */
513 
514 static int gfs2_readpage(struct file *file, struct page *page)
515 {
516 	struct address_space *mapping = page->mapping;
517 	struct gfs2_inode *ip = GFS2_I(mapping->host);
518 	struct gfs2_holder gh;
519 	int error;
520 
521 	unlock_page(page);
522 	gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
523 	error = gfs2_glock_nq(&gh);
524 	if (unlikely(error))
525 		goto out;
526 	error = AOP_TRUNCATED_PAGE;
527 	lock_page(page);
528 	if (page->mapping == mapping && !PageUptodate(page))
529 		error = __gfs2_readpage(file, page);
530 	else
531 		unlock_page(page);
532 	gfs2_glock_dq(&gh);
533 out:
534 	gfs2_holder_uninit(&gh);
535 	if (error && error != AOP_TRUNCATED_PAGE)
536 		lock_page(page);
537 	return error;
538 }
539 
540 /**
541  * gfs2_internal_read - read an internal file
542  * @ip: The gfs2 inode
543  * @buf: The buffer to fill
544  * @pos: The file position
545  * @size: The amount to read
546  *
547  */
548 
549 int gfs2_internal_read(struct gfs2_inode *ip, char *buf, loff_t *pos,
550                        unsigned size)
551 {
552 	struct address_space *mapping = ip->i_inode.i_mapping;
553 	unsigned long index = *pos >> PAGE_SHIFT;
554 	unsigned offset = *pos & (PAGE_SIZE - 1);
555 	unsigned copied = 0;
556 	unsigned amt;
557 	struct page *page;
558 	void *p;
559 
560 	do {
561 		amt = size - copied;
562 		if (offset + size > PAGE_SIZE)
563 			amt = PAGE_SIZE - offset;
564 		page = read_cache_page(mapping, index, __gfs2_readpage, NULL);
565 		if (IS_ERR(page))
566 			return PTR_ERR(page);
567 		p = kmap_atomic(page);
568 		memcpy(buf + copied, p + offset, amt);
569 		kunmap_atomic(p);
570 		put_page(page);
571 		copied += amt;
572 		index++;
573 		offset = 0;
574 	} while(copied < size);
575 	(*pos) += size;
576 	return size;
577 }
578 
579 /**
580  * gfs2_readpages - Read a bunch of pages at once
581  * @file: The file to read from
582  * @mapping: Address space info
583  * @pages: List of pages to read
584  * @nr_pages: Number of pages to read
585  *
586  * Some notes:
587  * 1. This is only for readahead, so we can simply ignore any things
588  *    which are slightly inconvenient (such as locking conflicts between
589  *    the page lock and the glock) and return having done no I/O. Its
590  *    obviously not something we'd want to do on too regular a basis.
591  *    Any I/O we ignore at this time will be done via readpage later.
592  * 2. We don't handle stuffed files here we let readpage do the honours.
593  * 3. mpage_readpages() does most of the heavy lifting in the common case.
594  * 4. gfs2_block_map() is relied upon to set BH_Boundary in the right places.
595  */
596 
597 static int gfs2_readpages(struct file *file, struct address_space *mapping,
598 			  struct list_head *pages, unsigned nr_pages)
599 {
600 	struct inode *inode = mapping->host;
601 	struct gfs2_inode *ip = GFS2_I(inode);
602 	struct gfs2_sbd *sdp = GFS2_SB(inode);
603 	struct gfs2_holder gh;
604 	int ret;
605 
606 	gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
607 	ret = gfs2_glock_nq(&gh);
608 	if (unlikely(ret))
609 		goto out_uninit;
610 	if (!gfs2_is_stuffed(ip))
611 		ret = mpage_readpages(mapping, pages, nr_pages, gfs2_block_map);
612 	gfs2_glock_dq(&gh);
613 out_uninit:
614 	gfs2_holder_uninit(&gh);
615 	if (unlikely(gfs2_withdrawn(sdp)))
616 		ret = -EIO;
617 	return ret;
618 }
619 
620 /**
621  * adjust_fs_space - Adjusts the free space available due to gfs2_grow
622  * @inode: the rindex inode
623  */
624 void adjust_fs_space(struct inode *inode)
625 {
626 	struct gfs2_sbd *sdp = GFS2_SB(inode);
627 	struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
628 	struct gfs2_inode *l_ip = GFS2_I(sdp->sd_sc_inode);
629 	struct gfs2_statfs_change_host *m_sc = &sdp->sd_statfs_master;
630 	struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
631 	struct buffer_head *m_bh, *l_bh;
632 	u64 fs_total, new_free;
633 
634 	if (gfs2_trans_begin(sdp, 2 * RES_STATFS, 0) != 0)
635 		return;
636 
637 	/* Total up the file system space, according to the latest rindex. */
638 	fs_total = gfs2_ri_total(sdp);
639 	if (gfs2_meta_inode_buffer(m_ip, &m_bh) != 0)
640 		goto out;
641 
642 	spin_lock(&sdp->sd_statfs_spin);
643 	gfs2_statfs_change_in(m_sc, m_bh->b_data +
644 			      sizeof(struct gfs2_dinode));
645 	if (fs_total > (m_sc->sc_total + l_sc->sc_total))
646 		new_free = fs_total - (m_sc->sc_total + l_sc->sc_total);
647 	else
648 		new_free = 0;
649 	spin_unlock(&sdp->sd_statfs_spin);
650 	fs_warn(sdp, "File system extended by %llu blocks.\n",
651 		(unsigned long long)new_free);
652 	gfs2_statfs_change(sdp, new_free, new_free, 0);
653 
654 	if (gfs2_meta_inode_buffer(l_ip, &l_bh) != 0)
655 		goto out2;
656 	update_statfs(sdp, m_bh, l_bh);
657 	brelse(l_bh);
658 out2:
659 	brelse(m_bh);
660 out:
661 	sdp->sd_rindex_uptodate = 0;
662 	gfs2_trans_end(sdp);
663 }
664 
665 /**
666  * jdata_set_page_dirty - Page dirtying function
667  * @page: The page to dirty
668  *
669  * Returns: 1 if it dirtyed the page, or 0 otherwise
670  */
671 
672 static int jdata_set_page_dirty(struct page *page)
673 {
674 	SetPageChecked(page);
675 	return __set_page_dirty_buffers(page);
676 }
677 
678 /**
679  * gfs2_bmap - Block map function
680  * @mapping: Address space info
681  * @lblock: The block to map
682  *
683  * Returns: The disk address for the block or 0 on hole or error
684  */
685 
686 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
687 {
688 	struct gfs2_inode *ip = GFS2_I(mapping->host);
689 	struct gfs2_holder i_gh;
690 	sector_t dblock = 0;
691 	int error;
692 
693 	error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
694 	if (error)
695 		return 0;
696 
697 	if (!gfs2_is_stuffed(ip))
698 		dblock = iomap_bmap(mapping, lblock, &gfs2_iomap_ops);
699 
700 	gfs2_glock_dq_uninit(&i_gh);
701 
702 	return dblock;
703 }
704 
705 static void gfs2_discard(struct gfs2_sbd *sdp, struct buffer_head *bh)
706 {
707 	struct gfs2_bufdata *bd;
708 
709 	lock_buffer(bh);
710 	gfs2_log_lock(sdp);
711 	clear_buffer_dirty(bh);
712 	bd = bh->b_private;
713 	if (bd) {
714 		if (!list_empty(&bd->bd_list) && !buffer_pinned(bh))
715 			list_del_init(&bd->bd_list);
716 		else
717 			gfs2_remove_from_journal(bh, REMOVE_JDATA);
718 	}
719 	bh->b_bdev = NULL;
720 	clear_buffer_mapped(bh);
721 	clear_buffer_req(bh);
722 	clear_buffer_new(bh);
723 	gfs2_log_unlock(sdp);
724 	unlock_buffer(bh);
725 }
726 
727 static void gfs2_invalidatepage(struct page *page, unsigned int offset,
728 				unsigned int length)
729 {
730 	struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
731 	unsigned int stop = offset + length;
732 	int partial_page = (offset || length < PAGE_SIZE);
733 	struct buffer_head *bh, *head;
734 	unsigned long pos = 0;
735 
736 	BUG_ON(!PageLocked(page));
737 	if (!partial_page)
738 		ClearPageChecked(page);
739 	if (!page_has_buffers(page))
740 		goto out;
741 
742 	bh = head = page_buffers(page);
743 	do {
744 		if (pos + bh->b_size > stop)
745 			return;
746 
747 		if (offset <= pos)
748 			gfs2_discard(sdp, bh);
749 		pos += bh->b_size;
750 		bh = bh->b_this_page;
751 	} while (bh != head);
752 out:
753 	if (!partial_page)
754 		try_to_release_page(page, 0);
755 }
756 
757 /**
758  * gfs2_releasepage - free the metadata associated with a page
759  * @page: the page that's being released
760  * @gfp_mask: passed from Linux VFS, ignored by us
761  *
762  * Calls try_to_free_buffers() to free the buffers and put the page if the
763  * buffers can be released.
764  *
765  * Returns: 1 if the page was put or else 0
766  */
767 
768 int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
769 {
770 	struct address_space *mapping = page->mapping;
771 	struct gfs2_sbd *sdp = gfs2_mapping2sbd(mapping);
772 	struct buffer_head *bh, *head;
773 	struct gfs2_bufdata *bd;
774 
775 	if (!page_has_buffers(page))
776 		return 0;
777 
778 	/*
779 	 * From xfs_vm_releasepage: mm accommodates an old ext3 case where
780 	 * clean pages might not have had the dirty bit cleared.  Thus, it can
781 	 * send actual dirty pages to ->releasepage() via shrink_active_list().
782 	 *
783 	 * As a workaround, we skip pages that contain dirty buffers below.
784 	 * Once ->releasepage isn't called on dirty pages anymore, we can warn
785 	 * on dirty buffers like we used to here again.
786 	 */
787 
788 	gfs2_log_lock(sdp);
789 	spin_lock(&sdp->sd_ail_lock);
790 	head = bh = page_buffers(page);
791 	do {
792 		if (atomic_read(&bh->b_count))
793 			goto cannot_release;
794 		bd = bh->b_private;
795 		if (bd && bd->bd_tr)
796 			goto cannot_release;
797 		if (buffer_dirty(bh) || WARN_ON(buffer_pinned(bh)))
798 			goto cannot_release;
799 		bh = bh->b_this_page;
800 	} while(bh != head);
801 	spin_unlock(&sdp->sd_ail_lock);
802 
803 	head = bh = page_buffers(page);
804 	do {
805 		bd = bh->b_private;
806 		if (bd) {
807 			gfs2_assert_warn(sdp, bd->bd_bh == bh);
808 			if (!list_empty(&bd->bd_list))
809 				list_del_init(&bd->bd_list);
810 			bd->bd_bh = NULL;
811 			bh->b_private = NULL;
812 			kmem_cache_free(gfs2_bufdata_cachep, bd);
813 		}
814 
815 		bh = bh->b_this_page;
816 	} while (bh != head);
817 	gfs2_log_unlock(sdp);
818 
819 	return try_to_free_buffers(page);
820 
821 cannot_release:
822 	spin_unlock(&sdp->sd_ail_lock);
823 	gfs2_log_unlock(sdp);
824 	return 0;
825 }
826 
827 static const struct address_space_operations gfs2_aops = {
828 	.writepage = gfs2_writepage,
829 	.writepages = gfs2_writepages,
830 	.readpage = gfs2_readpage,
831 	.readpages = gfs2_readpages,
832 	.bmap = gfs2_bmap,
833 	.invalidatepage = gfs2_invalidatepage,
834 	.releasepage = gfs2_releasepage,
835 	.direct_IO = noop_direct_IO,
836 	.migratepage = buffer_migrate_page,
837 	.is_partially_uptodate = block_is_partially_uptodate,
838 	.error_remove_page = generic_error_remove_page,
839 };
840 
841 static const struct address_space_operations gfs2_jdata_aops = {
842 	.writepage = gfs2_jdata_writepage,
843 	.writepages = gfs2_jdata_writepages,
844 	.readpage = gfs2_readpage,
845 	.readpages = gfs2_readpages,
846 	.set_page_dirty = jdata_set_page_dirty,
847 	.bmap = gfs2_bmap,
848 	.invalidatepage = gfs2_invalidatepage,
849 	.releasepage = gfs2_releasepage,
850 	.is_partially_uptodate = block_is_partially_uptodate,
851 	.error_remove_page = generic_error_remove_page,
852 };
853 
854 void gfs2_set_aops(struct inode *inode)
855 {
856 	if (gfs2_is_jdata(GFS2_I(inode)))
857 		inode->i_mapping->a_ops = &gfs2_jdata_aops;
858 	else
859 		inode->i_mapping->a_ops = &gfs2_aops;
860 }
861