GFS2: drop rindex glock to refresh rindex list
[linux-2.6.git] / fs / gfs2 / aops.c
1 /*
2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
3  * Copyright (C) 2004-2008 Red Hat, Inc.  All rights reserved.
4  *
5  * This copyrighted material is made available to anyone wishing to use,
6  * modify, copy, or redistribute it subject to the terms and conditions
7  * of the GNU General Public License version 2.
8  */
9
10 #include <linux/sched.h>
11 #include <linux/slab.h>
12 #include <linux/spinlock.h>
13 #include <linux/completion.h>
14 #include <linux/buffer_head.h>
15 #include <linux/pagemap.h>
16 #include <linux/pagevec.h>
17 #include <linux/mpage.h>
18 #include <linux/fs.h>
19 #include <linux/writeback.h>
20 #include <linux/swap.h>
21 #include <linux/gfs2_ondisk.h>
22 #include <linux/backing-dev.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
38
39 static void gfs2_page_add_databufs(struct gfs2_inode *ip, struct page *page,
40                                    unsigned int from, unsigned int to)
41 {
42         struct buffer_head *head = page_buffers(page);
43         unsigned int bsize = head->b_size;
44         struct buffer_head *bh;
45         unsigned int start, end;
46
47         for (bh = head, start = 0; bh != head || !start;
48              bh = bh->b_this_page, start = end) {
49                 end = start + bsize;
50                 if (end <= from || start >= to)
51                         continue;
52                 if (gfs2_is_jdata(ip))
53                         set_buffer_uptodate(bh);
54                 gfs2_trans_add_bh(ip->i_gl, bh, 0);
55         }
56 }
57
58 /**
59  * gfs2_get_block_noalloc - Fills in a buffer head with details about a block
60  * @inode: The inode
61  * @lblock: The block number to look up
62  * @bh_result: The buffer head to return the result in
63  * @create: Non-zero if we may add block to the file
64  *
65  * Returns: errno
66  */
67
68 static int gfs2_get_block_noalloc(struct inode *inode, sector_t lblock,
69                                   struct buffer_head *bh_result, int create)
70 {
71         int error;
72
73         error = gfs2_block_map(inode, lblock, bh_result, 0);
74         if (error)
75                 return error;
76         if (!buffer_mapped(bh_result))
77                 return -EIO;
78         return 0;
79 }
80
81 static int gfs2_get_block_direct(struct inode *inode, sector_t lblock,
82                                  struct buffer_head *bh_result, int create)
83 {
84         return gfs2_block_map(inode, lblock, bh_result, 0);
85 }
86
87 /**
88  * gfs2_writepage_common - Common bits of writepage
89  * @page: The page to be written
90  * @wbc: The writeback control
91  *
92  * Returns: 1 if writepage is ok, otherwise an error code or zero if no error.
93  */
94
95 static int gfs2_writepage_common(struct page *page,
96                                  struct writeback_control *wbc)
97 {
98         struct inode *inode = page->mapping->host;
99         struct gfs2_inode *ip = GFS2_I(inode);
100         struct gfs2_sbd *sdp = GFS2_SB(inode);
101         loff_t i_size = i_size_read(inode);
102         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
103         unsigned offset;
104
105         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
106                 goto out;
107         if (current->journal_info)
108                 goto redirty;
109         /* Is the page fully outside i_size? (truncate in progress) */
110         offset = i_size & (PAGE_CACHE_SIZE-1);
111         if (page->index > end_index || (page->index == end_index && !offset)) {
112                 page->mapping->a_ops->invalidatepage(page, 0);
113                 goto out;
114         }
115         return 1;
116 redirty:
117         redirty_page_for_writepage(wbc, page);
118 out:
119         unlock_page(page);
120         return 0;
121 }
122
123 /**
124  * gfs2_writeback_writepage - Write page for writeback mappings
125  * @page: The page
126  * @wbc: The writeback control
127  *
128  */
129
130 static int gfs2_writeback_writepage(struct page *page,
131                                     struct writeback_control *wbc)
132 {
133         int ret;
134
135         ret = gfs2_writepage_common(page, wbc);
136         if (ret <= 0)
137                 return ret;
138
139         ret = mpage_writepage(page, gfs2_get_block_noalloc, wbc);
140         if (ret == -EAGAIN)
141                 ret = block_write_full_page(page, gfs2_get_block_noalloc, wbc);
142         return ret;
143 }
144
145 /**
146  * gfs2_ordered_writepage - Write page for ordered data files
147  * @page: The page to write
148  * @wbc: The writeback control
149  *
150  */
151
152 static int gfs2_ordered_writepage(struct page *page,
153                                   struct writeback_control *wbc)
154 {
155         struct inode *inode = page->mapping->host;
156         struct gfs2_inode *ip = GFS2_I(inode);
157         int ret;
158
159         ret = gfs2_writepage_common(page, wbc);
160         if (ret <= 0)
161                 return ret;
162
163         if (!page_has_buffers(page)) {
164                 create_empty_buffers(page, inode->i_sb->s_blocksize,
165                                      (1 << BH_Dirty)|(1 << BH_Uptodate));
166         }
167         gfs2_page_add_databufs(ip, page, 0, inode->i_sb->s_blocksize-1);
168         return block_write_full_page(page, gfs2_get_block_noalloc, wbc);
169 }
170
171 /**
172  * __gfs2_jdata_writepage - The core of jdata writepage
173  * @page: The page to write
174  * @wbc: The writeback control
175  *
176  * This is shared between writepage and writepages and implements the
177  * core of the writepage operation. If a transaction is required then
178  * PageChecked will have been set and the transaction will have
179  * already been started before this is called.
180  */
181
182 static int __gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
183 {
184         struct inode *inode = page->mapping->host;
185         struct gfs2_inode *ip = GFS2_I(inode);
186         struct gfs2_sbd *sdp = GFS2_SB(inode);
187
188         if (PageChecked(page)) {
189                 ClearPageChecked(page);
190                 if (!page_has_buffers(page)) {
191                         create_empty_buffers(page, inode->i_sb->s_blocksize,
192                                              (1 << BH_Dirty)|(1 << BH_Uptodate));
193                 }
194                 gfs2_page_add_databufs(ip, page, 0, sdp->sd_vfs->s_blocksize-1);
195         }
196         return block_write_full_page(page, gfs2_get_block_noalloc, wbc);
197 }
198
199 /**
200  * gfs2_jdata_writepage - Write complete page
201  * @page: Page to write
202  *
203  * Returns: errno
204  *
205  */
206
207 static int gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
208 {
209         struct inode *inode = page->mapping->host;
210         struct gfs2_sbd *sdp = GFS2_SB(inode);
211         int ret;
212         int done_trans = 0;
213
214         if (PageChecked(page)) {
215                 if (wbc->sync_mode != WB_SYNC_ALL)
216                         goto out_ignore;
217                 ret = gfs2_trans_begin(sdp, RES_DINODE + 1, 0);
218                 if (ret)
219                         goto out_ignore;
220                 done_trans = 1;
221         }
222         ret = gfs2_writepage_common(page, wbc);
223         if (ret > 0)
224                 ret = __gfs2_jdata_writepage(page, wbc);
225         if (done_trans)
226                 gfs2_trans_end(sdp);
227         return ret;
228
229 out_ignore:
230         redirty_page_for_writepage(wbc, page);
231         unlock_page(page);
232         return 0;
233 }
234
235 /**
236  * gfs2_writeback_writepages - Write a bunch of dirty pages back to disk
237  * @mapping: The mapping to write
238  * @wbc: Write-back control
239  *
240  * For the data=writeback case we can already ignore buffer heads
241  * and write whole extents at once. This is a big reduction in the
242  * number of I/O requests we send and the bmap calls we make in this case.
243  */
244 static int gfs2_writeback_writepages(struct address_space *mapping,
245                                      struct writeback_control *wbc)
246 {
247         return mpage_writepages(mapping, wbc, gfs2_get_block_noalloc);
248 }
249
250 /**
251  * gfs2_write_jdata_pagevec - Write back a pagevec's worth of pages
252  * @mapping: The mapping
253  * @wbc: The writeback control
254  * @writepage: The writepage function to call for each page
255  * @pvec: The vector of pages
256  * @nr_pages: The number of pages to write
257  *
258  * Returns: non-zero if loop should terminate, zero otherwise
259  */
260
261 static int gfs2_write_jdata_pagevec(struct address_space *mapping,
262                                     struct writeback_control *wbc,
263                                     struct pagevec *pvec,
264                                     int nr_pages, pgoff_t end)
265 {
266         struct inode *inode = mapping->host;
267         struct gfs2_sbd *sdp = GFS2_SB(inode);
268         loff_t i_size = i_size_read(inode);
269         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
270         unsigned offset = i_size & (PAGE_CACHE_SIZE-1);
271         unsigned nrblocks = nr_pages * (PAGE_CACHE_SIZE/inode->i_sb->s_blocksize);
272         struct backing_dev_info *bdi = mapping->backing_dev_info;
273         int i;
274         int ret;
275
276         ret = gfs2_trans_begin(sdp, nrblocks, nrblocks);
277         if (ret < 0)
278                 return ret;
279
280         for(i = 0; i < nr_pages; i++) {
281                 struct page *page = pvec->pages[i];
282
283                 lock_page(page);
284
285                 if (unlikely(page->mapping != mapping)) {
286                         unlock_page(page);
287                         continue;
288                 }
289
290                 if (!wbc->range_cyclic && page->index > end) {
291                         ret = 1;
292                         unlock_page(page);
293                         continue;
294                 }
295
296                 if (wbc->sync_mode != WB_SYNC_NONE)
297                         wait_on_page_writeback(page);
298
299                 if (PageWriteback(page) ||
300                     !clear_page_dirty_for_io(page)) {
301                         unlock_page(page);
302                         continue;
303                 }
304
305                 /* Is the page fully outside i_size? (truncate in progress) */
306                 if (page->index > end_index || (page->index == end_index && !offset)) {
307                         page->mapping->a_ops->invalidatepage(page, 0);
308                         unlock_page(page);
309                         continue;
310                 }
311
312                 ret = __gfs2_jdata_writepage(page, wbc);
313
314                 if (ret || (--(wbc->nr_to_write) <= 0))
315                         ret = 1;
316                 if (wbc->nonblocking && bdi_write_congested(bdi)) {
317                         wbc->encountered_congestion = 1;
318                         ret = 1;
319                 }
320
321         }
322         gfs2_trans_end(sdp);
323         return ret;
324 }
325
326 /**
327  * gfs2_write_cache_jdata - Like write_cache_pages but different
328  * @mapping: The mapping to write
329  * @wbc: The writeback control
330  * @writepage: The writepage function to call
331  * @data: The data to pass to writepage
332  *
333  * The reason that we use our own function here is that we need to
334  * start transactions before we grab page locks. This allows us
335  * to get the ordering right.
336  */
337
338 static int gfs2_write_cache_jdata(struct address_space *mapping,
339                                   struct writeback_control *wbc)
340 {
341         struct backing_dev_info *bdi = mapping->backing_dev_info;
342         int ret = 0;
343         int done = 0;
344         struct pagevec pvec;
345         int nr_pages;
346         pgoff_t index;
347         pgoff_t end;
348         int scanned = 0;
349         int range_whole = 0;
350
351         if (wbc->nonblocking && bdi_write_congested(bdi)) {
352                 wbc->encountered_congestion = 1;
353                 return 0;
354         }
355
356         pagevec_init(&pvec, 0);
357         if (wbc->range_cyclic) {
358                 index = mapping->writeback_index; /* Start from prev offset */
359                 end = -1;
360         } else {
361                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
362                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
363                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
364                         range_whole = 1;
365                 scanned = 1;
366         }
367
368 retry:
369          while (!done && (index <= end) &&
370                 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
371                                                PAGECACHE_TAG_DIRTY,
372                                                min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
373                 scanned = 1;
374                 ret = gfs2_write_jdata_pagevec(mapping, wbc, &pvec, nr_pages, end);
375                 if (ret)
376                         done = 1;
377                 if (ret > 0)
378                         ret = 0;
379
380                 pagevec_release(&pvec);
381                 cond_resched();
382         }
383
384         if (!scanned && !done) {
385                 /*
386                  * We hit the last page and there is more work to be done: wrap
387                  * back to the start of the file
388                  */
389                 scanned = 1;
390                 index = 0;
391                 goto retry;
392         }
393
394         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
395                 mapping->writeback_index = index;
396         return ret;
397 }
398
399
400 /**
401  * gfs2_jdata_writepages - Write a bunch of dirty pages back to disk
402  * @mapping: The mapping to write
403  * @wbc: The writeback control
404  * 
405  */
406
407 static int gfs2_jdata_writepages(struct address_space *mapping,
408                                  struct writeback_control *wbc)
409 {
410         struct gfs2_inode *ip = GFS2_I(mapping->host);
411         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
412         int ret;
413
414         ret = gfs2_write_cache_jdata(mapping, wbc);
415         if (ret == 0 && wbc->sync_mode == WB_SYNC_ALL) {
416                 gfs2_log_flush(sdp, ip->i_gl);
417                 ret = gfs2_write_cache_jdata(mapping, wbc);
418         }
419         return ret;
420 }
421
422 /**
423  * stuffed_readpage - Fill in a Linux page with stuffed file data
424  * @ip: the inode
425  * @page: the page
426  *
427  * Returns: errno
428  */
429
430 static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
431 {
432         struct buffer_head *dibh;
433         void *kaddr;
434         int error;
435
436         /*
437          * Due to the order of unstuffing files and ->fault(), we can be
438          * asked for a zero page in the case of a stuffed file being extended,
439          * so we need to supply one here. It doesn't happen often.
440          */
441         if (unlikely(page->index)) {
442                 zero_user(page, 0, PAGE_CACHE_SIZE);
443                 SetPageUptodate(page);
444                 return 0;
445         }
446
447         error = gfs2_meta_inode_buffer(ip, &dibh);
448         if (error)
449                 return error;
450
451         kaddr = kmap_atomic(page, KM_USER0);
452         memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode),
453                ip->i_disksize);
454         memset(kaddr + ip->i_disksize, 0, PAGE_CACHE_SIZE - ip->i_disksize);
455         kunmap_atomic(kaddr, KM_USER0);
456         flush_dcache_page(page);
457         brelse(dibh);
458         SetPageUptodate(page);
459
460         return 0;
461 }
462
463
464 /**
465  * __gfs2_readpage - readpage
466  * @file: The file to read a page for
467  * @page: The page to read
468  *
469  * This is the core of gfs2's readpage. Its used by the internal file
470  * reading code as in that case we already hold the glock. Also its
471  * called by gfs2_readpage() once the required lock has been granted.
472  *
473  */
474
475 static int __gfs2_readpage(void *file, struct page *page)
476 {
477         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
478         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
479         int error;
480
481         if (gfs2_is_stuffed(ip)) {
482                 error = stuffed_readpage(ip, page);
483                 unlock_page(page);
484         } else {
485                 error = mpage_readpage(page, gfs2_block_map);
486         }
487
488         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
489                 return -EIO;
490
491         return error;
492 }
493
494 /**
495  * gfs2_readpage - read a page of a file
496  * @file: The file to read
497  * @page: The page of the file
498  *
499  * This deals with the locking required. We have to unlock and
500  * relock the page in order to get the locking in the right
501  * order.
502  */
503
504 static int gfs2_readpage(struct file *file, struct page *page)
505 {
506         struct address_space *mapping = page->mapping;
507         struct gfs2_inode *ip = GFS2_I(mapping->host);
508         struct gfs2_holder gh;
509         int error;
510
511         unlock_page(page);
512         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
513         error = gfs2_glock_nq(&gh);
514         if (unlikely(error))
515                 goto out;
516         error = AOP_TRUNCATED_PAGE;
517         lock_page(page);
518         if (page->mapping == mapping && !PageUptodate(page))
519                 error = __gfs2_readpage(file, page);
520         else
521                 unlock_page(page);
522         gfs2_glock_dq(&gh);
523 out:
524         gfs2_holder_uninit(&gh);
525         if (error && error != AOP_TRUNCATED_PAGE)
526                 lock_page(page);
527         return error;
528 }
529
530 /**
531  * gfs2_internal_read - read an internal file
532  * @ip: The gfs2 inode
533  * @ra_state: The readahead state (or NULL for no readahead)
534  * @buf: The buffer to fill
535  * @pos: The file position
536  * @size: The amount to read
537  *
538  */
539
540 int gfs2_internal_read(struct gfs2_inode *ip, struct file_ra_state *ra_state,
541                        char *buf, loff_t *pos, unsigned size)
542 {
543         struct address_space *mapping = ip->i_inode.i_mapping;
544         unsigned long index = *pos / PAGE_CACHE_SIZE;
545         unsigned offset = *pos & (PAGE_CACHE_SIZE - 1);
546         unsigned copied = 0;
547         unsigned amt;
548         struct page *page;
549         void *p;
550
551         do {
552                 amt = size - copied;
553                 if (offset + size > PAGE_CACHE_SIZE)
554                         amt = PAGE_CACHE_SIZE - offset;
555                 page = read_cache_page(mapping, index, __gfs2_readpage, NULL);
556                 if (IS_ERR(page))
557                         return PTR_ERR(page);
558                 p = kmap_atomic(page, KM_USER0);
559                 memcpy(buf + copied, p + offset, amt);
560                 kunmap_atomic(p, KM_USER0);
561                 mark_page_accessed(page);
562                 page_cache_release(page);
563                 copied += amt;
564                 index++;
565                 offset = 0;
566         } while(copied < size);
567         (*pos) += size;
568         return size;
569 }
570
571 /**
572  * gfs2_readpages - Read a bunch of pages at once
573  *
574  * Some notes:
575  * 1. This is only for readahead, so we can simply ignore any things
576  *    which are slightly inconvenient (such as locking conflicts between
577  *    the page lock and the glock) and return having done no I/O. Its
578  *    obviously not something we'd want to do on too regular a basis.
579  *    Any I/O we ignore at this time will be done via readpage later.
580  * 2. We don't handle stuffed files here we let readpage do the honours.
581  * 3. mpage_readpages() does most of the heavy lifting in the common case.
582  * 4. gfs2_block_map() is relied upon to set BH_Boundary in the right places.
583  */
584
585 static int gfs2_readpages(struct file *file, struct address_space *mapping,
586                           struct list_head *pages, unsigned nr_pages)
587 {
588         struct inode *inode = mapping->host;
589         struct gfs2_inode *ip = GFS2_I(inode);
590         struct gfs2_sbd *sdp = GFS2_SB(inode);
591         struct gfs2_holder gh;
592         int ret;
593
594         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, 0, &gh);
595         ret = gfs2_glock_nq(&gh);
596         if (unlikely(ret))
597                 goto out_uninit;
598         if (!gfs2_is_stuffed(ip))
599                 ret = mpage_readpages(mapping, pages, nr_pages, gfs2_block_map);
600         gfs2_glock_dq(&gh);
601 out_uninit:
602         gfs2_holder_uninit(&gh);
603         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
604                 ret = -EIO;
605         return ret;
606 }
607
608 /**
609  * gfs2_write_begin - Begin to write to a file
610  * @file: The file to write to
611  * @mapping: The mapping in which to write
612  * @pos: The file offset at which to start writing
613  * @len: Length of the write
614  * @flags: Various flags
615  * @pagep: Pointer to return the page
616  * @fsdata: Pointer to return fs data (unused by GFS2)
617  *
618  * Returns: errno
619  */
620
621 static int gfs2_write_begin(struct file *file, struct address_space *mapping,
622                             loff_t pos, unsigned len, unsigned flags,
623                             struct page **pagep, void **fsdata)
624 {
625         struct gfs2_inode *ip = GFS2_I(mapping->host);
626         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
627         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
628         unsigned int data_blocks = 0, ind_blocks = 0, rblocks;
629         int alloc_required;
630         int error = 0;
631         struct gfs2_alloc *al;
632         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
633         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
634         unsigned to = from + len;
635         struct page *page;
636
637         gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, 0, &ip->i_gh);
638         error = gfs2_glock_nq(&ip->i_gh);
639         if (unlikely(error))
640                 goto out_uninit;
641         if (&ip->i_inode == sdp->sd_rindex) {
642                 error = gfs2_glock_nq_init(m_ip->i_gl, LM_ST_EXCLUSIVE,
643                                            GL_NOCACHE, &m_ip->i_gh);
644                 if (unlikely(error)) {
645                         gfs2_glock_dq(&ip->i_gh);
646                         goto out_uninit;
647                 }
648         }
649
650         error = gfs2_write_alloc_required(ip, pos, len, &alloc_required);
651         if (error)
652                 goto out_unlock;
653
654         if (alloc_required || gfs2_is_jdata(ip))
655                 gfs2_write_calc_reserv(ip, len, &data_blocks, &ind_blocks);
656
657         if (alloc_required) {
658                 al = gfs2_alloc_get(ip);
659                 if (!al) {
660                         error = -ENOMEM;
661                         goto out_unlock;
662                 }
663
664                 error = gfs2_quota_lock_check(ip);
665                 if (error)
666                         goto out_alloc_put;
667
668                 al->al_requested = data_blocks + ind_blocks;
669                 error = gfs2_inplace_reserve(ip);
670                 if (error)
671                         goto out_qunlock;
672         }
673
674         rblocks = RES_DINODE + ind_blocks;
675         if (gfs2_is_jdata(ip))
676                 rblocks += data_blocks ? data_blocks : 1;
677         if (ind_blocks || data_blocks)
678                 rblocks += RES_STATFS + RES_QUOTA;
679         if (&ip->i_inode == sdp->sd_rindex)
680                 rblocks += 2 * RES_STATFS;
681
682         error = gfs2_trans_begin(sdp, rblocks,
683                                  PAGE_CACHE_SIZE/sdp->sd_sb.sb_bsize);
684         if (error)
685                 goto out_trans_fail;
686
687         error = -ENOMEM;
688         flags |= AOP_FLAG_NOFS;
689         page = grab_cache_page_write_begin(mapping, index, flags);
690         *pagep = page;
691         if (unlikely(!page))
692                 goto out_endtrans;
693
694         if (gfs2_is_stuffed(ip)) {
695                 error = 0;
696                 if (pos + len > sdp->sd_sb.sb_bsize - sizeof(struct gfs2_dinode)) {
697                         error = gfs2_unstuff_dinode(ip, page);
698                         if (error == 0)
699                                 goto prepare_write;
700                 } else if (!PageUptodate(page)) {
701                         error = stuffed_readpage(ip, page);
702                 }
703                 goto out;
704         }
705
706 prepare_write:
707         error = block_prepare_write(page, from, to, gfs2_block_map);
708 out:
709         if (error == 0)
710                 return 0;
711
712         page_cache_release(page);
713         if (pos + len > ip->i_inode.i_size)
714                 vmtruncate(&ip->i_inode, ip->i_inode.i_size);
715 out_endtrans:
716         gfs2_trans_end(sdp);
717 out_trans_fail:
718         if (alloc_required) {
719                 gfs2_inplace_release(ip);
720 out_qunlock:
721                 gfs2_quota_unlock(ip);
722 out_alloc_put:
723                 gfs2_alloc_put(ip);
724         }
725 out_unlock:
726         if (&ip->i_inode == sdp->sd_rindex) {
727                 gfs2_glock_dq(&m_ip->i_gh);
728                 gfs2_holder_uninit(&m_ip->i_gh);
729         }
730         gfs2_glock_dq(&ip->i_gh);
731 out_uninit:
732         gfs2_holder_uninit(&ip->i_gh);
733         return error;
734 }
735
736 /**
737  * adjust_fs_space - Adjusts the free space available due to gfs2_grow
738  * @inode: the rindex inode
739  */
740 static void adjust_fs_space(struct inode *inode)
741 {
742         struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
743         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
744         struct gfs2_inode *l_ip = GFS2_I(sdp->sd_sc_inode);
745         struct gfs2_statfs_change_host *m_sc = &sdp->sd_statfs_master;
746         struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
747         struct buffer_head *m_bh, *l_bh;
748         u64 fs_total, new_free;
749
750         /* Total up the file system space, according to the latest rindex. */
751         fs_total = gfs2_ri_total(sdp);
752         if (gfs2_meta_inode_buffer(m_ip, &m_bh) != 0)
753                 return;
754
755         spin_lock(&sdp->sd_statfs_spin);
756         gfs2_statfs_change_in(m_sc, m_bh->b_data +
757                               sizeof(struct gfs2_dinode));
758         if (fs_total > (m_sc->sc_total + l_sc->sc_total))
759                 new_free = fs_total - (m_sc->sc_total + l_sc->sc_total);
760         else
761                 new_free = 0;
762         spin_unlock(&sdp->sd_statfs_spin);
763         fs_warn(sdp, "File system extended by %llu blocks.\n",
764                 (unsigned long long)new_free);
765         gfs2_statfs_change(sdp, new_free, new_free, 0);
766
767         if (gfs2_meta_inode_buffer(l_ip, &l_bh) != 0)
768                 goto out;
769         update_statfs(sdp, m_bh, l_bh);
770         brelse(l_bh);
771 out:
772         brelse(m_bh);
773 }
774
775 /**
776  * gfs2_stuffed_write_end - Write end for stuffed files
777  * @inode: The inode
778  * @dibh: The buffer_head containing the on-disk inode
779  * @pos: The file position
780  * @len: The length of the write
781  * @copied: How much was actually copied by the VFS
782  * @page: The page
783  *
784  * This copies the data from the page into the inode block after
785  * the inode data structure itself.
786  *
787  * Returns: errno
788  */
789 static int gfs2_stuffed_write_end(struct inode *inode, struct buffer_head *dibh,
790                                   loff_t pos, unsigned len, unsigned copied,
791                                   struct page *page)
792 {
793         struct gfs2_inode *ip = GFS2_I(inode);
794         struct gfs2_sbd *sdp = GFS2_SB(inode);
795         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
796         u64 to = pos + copied;
797         void *kaddr;
798         unsigned char *buf = dibh->b_data + sizeof(struct gfs2_dinode);
799         struct gfs2_dinode *di = (struct gfs2_dinode *)dibh->b_data;
800
801         BUG_ON((pos + len) > (dibh->b_size - sizeof(struct gfs2_dinode)));
802         kaddr = kmap_atomic(page, KM_USER0);
803         memcpy(buf + pos, kaddr + pos, copied);
804         memset(kaddr + pos + copied, 0, len - copied);
805         flush_dcache_page(page);
806         kunmap_atomic(kaddr, KM_USER0);
807
808         if (!PageUptodate(page))
809                 SetPageUptodate(page);
810         unlock_page(page);
811         page_cache_release(page);
812
813         if (copied) {
814                 if (inode->i_size < to) {
815                         i_size_write(inode, to);
816                         ip->i_disksize = inode->i_size;
817                 }
818                 gfs2_dinode_out(ip, di);
819                 mark_inode_dirty(inode);
820         }
821
822         if (inode == sdp->sd_rindex) {
823                 adjust_fs_space(inode);
824                 ip->i_gh.gh_flags |= GL_NOCACHE;
825         }
826
827         brelse(dibh);
828         gfs2_trans_end(sdp);
829         if (inode == sdp->sd_rindex) {
830                 gfs2_glock_dq(&m_ip->i_gh);
831                 gfs2_holder_uninit(&m_ip->i_gh);
832         }
833         gfs2_glock_dq(&ip->i_gh);
834         gfs2_holder_uninit(&ip->i_gh);
835         return copied;
836 }
837
838 /**
839  * gfs2_write_end
840  * @file: The file to write to
841  * @mapping: The address space to write to
842  * @pos: The file position
843  * @len: The length of the data
844  * @copied:
845  * @page: The page that has been written
846  * @fsdata: The fsdata (unused in GFS2)
847  *
848  * The main write_end function for GFS2. We have a separate one for
849  * stuffed files as they are slightly different, otherwise we just
850  * put our locking around the VFS provided functions.
851  *
852  * Returns: errno
853  */
854
855 static int gfs2_write_end(struct file *file, struct address_space *mapping,
856                           loff_t pos, unsigned len, unsigned copied,
857                           struct page *page, void *fsdata)
858 {
859         struct inode *inode = page->mapping->host;
860         struct gfs2_inode *ip = GFS2_I(inode);
861         struct gfs2_sbd *sdp = GFS2_SB(inode);
862         struct gfs2_inode *m_ip = GFS2_I(sdp->sd_statfs_inode);
863         struct buffer_head *dibh;
864         struct gfs2_alloc *al = ip->i_alloc;
865         unsigned int from = pos & (PAGE_CACHE_SIZE - 1);
866         unsigned int to = from + len;
867         int ret;
868
869         BUG_ON(gfs2_glock_is_locked_by_me(ip->i_gl) == NULL);
870
871         ret = gfs2_meta_inode_buffer(ip, &dibh);
872         if (unlikely(ret)) {
873                 unlock_page(page);
874                 page_cache_release(page);
875                 goto failed;
876         }
877
878         gfs2_trans_add_bh(ip->i_gl, dibh, 1);
879
880         if (gfs2_is_stuffed(ip))
881                 return gfs2_stuffed_write_end(inode, dibh, pos, len, copied, page);
882
883         if (!gfs2_is_writeback(ip))
884                 gfs2_page_add_databufs(ip, page, from, to);
885
886         ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
887         if (ret > 0) {
888                 if (inode->i_size > ip->i_disksize)
889                         ip->i_disksize = inode->i_size;
890                 gfs2_dinode_out(ip, dibh->b_data);
891                 mark_inode_dirty(inode);
892         }
893
894         if (inode == sdp->sd_rindex) {
895                 adjust_fs_space(inode);
896                 ip->i_gh.gh_flags |= GL_NOCACHE;
897         }
898
899         brelse(dibh);
900         gfs2_trans_end(sdp);
901 failed:
902         if (al) {
903                 gfs2_inplace_release(ip);
904                 gfs2_quota_unlock(ip);
905                 gfs2_alloc_put(ip);
906         }
907         if (inode == sdp->sd_rindex) {
908                 gfs2_glock_dq(&m_ip->i_gh);
909                 gfs2_holder_uninit(&m_ip->i_gh);
910         }
911         gfs2_glock_dq(&ip->i_gh);
912         gfs2_holder_uninit(&ip->i_gh);
913         return ret;
914 }
915
916 /**
917  * gfs2_set_page_dirty - Page dirtying function
918  * @page: The page to dirty
919  *
920  * Returns: 1 if it dirtyed the page, or 0 otherwise
921  */
922  
923 static int gfs2_set_page_dirty(struct page *page)
924 {
925         SetPageChecked(page);
926         return __set_page_dirty_buffers(page);
927 }
928
929 /**
930  * gfs2_bmap - Block map function
931  * @mapping: Address space info
932  * @lblock: The block to map
933  *
934  * Returns: The disk address for the block or 0 on hole or error
935  */
936
937 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
938 {
939         struct gfs2_inode *ip = GFS2_I(mapping->host);
940         struct gfs2_holder i_gh;
941         sector_t dblock = 0;
942         int error;
943
944         error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
945         if (error)
946                 return 0;
947
948         if (!gfs2_is_stuffed(ip))
949                 dblock = generic_block_bmap(mapping, lblock, gfs2_block_map);
950
951         gfs2_glock_dq_uninit(&i_gh);
952
953         return dblock;
954 }
955
956 static void gfs2_discard(struct gfs2_sbd *sdp, struct buffer_head *bh)
957 {
958         struct gfs2_bufdata *bd;
959
960         lock_buffer(bh);
961         gfs2_log_lock(sdp);
962         clear_buffer_dirty(bh);
963         bd = bh->b_private;
964         if (bd) {
965                 if (!list_empty(&bd->bd_le.le_list) && !buffer_pinned(bh))
966                         list_del_init(&bd->bd_le.le_list);
967                 else
968                         gfs2_remove_from_journal(bh, current->journal_info, 0);
969         }
970         bh->b_bdev = NULL;
971         clear_buffer_mapped(bh);
972         clear_buffer_req(bh);
973         clear_buffer_new(bh);
974         gfs2_log_unlock(sdp);
975         unlock_buffer(bh);
976 }
977
978 static void gfs2_invalidatepage(struct page *page, unsigned long offset)
979 {
980         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
981         struct buffer_head *bh, *head;
982         unsigned long pos = 0;
983
984         BUG_ON(!PageLocked(page));
985         if (offset == 0)
986                 ClearPageChecked(page);
987         if (!page_has_buffers(page))
988                 goto out;
989
990         bh = head = page_buffers(page);
991         do {
992                 if (offset <= pos)
993                         gfs2_discard(sdp, bh);
994                 pos += bh->b_size;
995                 bh = bh->b_this_page;
996         } while (bh != head);
997 out:
998         if (offset == 0)
999                 try_to_release_page(page, 0);
1000 }
1001
1002 /**
1003  * gfs2_ok_for_dio - check that dio is valid on this file
1004  * @ip: The inode
1005  * @rw: READ or WRITE
1006  * @offset: The offset at which we are reading or writing
1007  *
1008  * Returns: 0 (to ignore the i/o request and thus fall back to buffered i/o)
1009  *          1 (to accept the i/o request)
1010  */
1011 static int gfs2_ok_for_dio(struct gfs2_inode *ip, int rw, loff_t offset)
1012 {
1013         /*
1014          * Should we return an error here? I can't see that O_DIRECT for
1015          * a stuffed file makes any sense. For now we'll silently fall
1016          * back to buffered I/O
1017          */
1018         if (gfs2_is_stuffed(ip))
1019                 return 0;
1020
1021         if (offset >= i_size_read(&ip->i_inode))
1022                 return 0;
1023         return 1;
1024 }
1025
1026
1027
1028 static ssize_t gfs2_direct_IO(int rw, struct kiocb *iocb,
1029                               const struct iovec *iov, loff_t offset,
1030                               unsigned long nr_segs)
1031 {
1032         struct file *file = iocb->ki_filp;
1033         struct inode *inode = file->f_mapping->host;
1034         struct gfs2_inode *ip = GFS2_I(inode);
1035         struct gfs2_holder gh;
1036         int rv;
1037
1038         /*
1039          * Deferred lock, even if its a write, since we do no allocation
1040          * on this path. All we need change is atime, and this lock mode
1041          * ensures that other nodes have flushed their buffered read caches
1042          * (i.e. their page cache entries for this inode). We do not,
1043          * unfortunately have the option of only flushing a range like
1044          * the VFS does.
1045          */
1046         gfs2_holder_init(ip->i_gl, LM_ST_DEFERRED, 0, &gh);
1047         rv = gfs2_glock_nq(&gh);
1048         if (rv)
1049                 return rv;
1050         rv = gfs2_ok_for_dio(ip, rw, offset);
1051         if (rv != 1)
1052                 goto out; /* dio not valid, fall back to buffered i/o */
1053
1054         rv = blockdev_direct_IO_no_locking(rw, iocb, inode, inode->i_sb->s_bdev,
1055                                            iov, offset, nr_segs,
1056                                            gfs2_get_block_direct, NULL);
1057 out:
1058         gfs2_glock_dq_m(1, &gh);
1059         gfs2_holder_uninit(&gh);
1060         return rv;
1061 }
1062
1063 /**
1064  * gfs2_releasepage - free the metadata associated with a page
1065  * @page: the page that's being released
1066  * @gfp_mask: passed from Linux VFS, ignored by us
1067  *
1068  * Call try_to_free_buffers() if the buffers in this page can be
1069  * released.
1070  *
1071  * Returns: 0
1072  */
1073
1074 int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
1075 {
1076         struct inode *aspace = page->mapping->host;
1077         struct gfs2_sbd *sdp = aspace->i_sb->s_fs_info;
1078         struct buffer_head *bh, *head;
1079         struct gfs2_bufdata *bd;
1080
1081         if (!page_has_buffers(page))
1082                 return 0;
1083
1084         gfs2_log_lock(sdp);
1085         head = bh = page_buffers(page);
1086         do {
1087                 if (atomic_read(&bh->b_count))
1088                         goto cannot_release;
1089                 bd = bh->b_private;
1090                 if (bd && bd->bd_ail)
1091                         goto cannot_release;
1092                 gfs2_assert_warn(sdp, !buffer_pinned(bh));
1093                 gfs2_assert_warn(sdp, !buffer_dirty(bh));
1094                 bh = bh->b_this_page;
1095         } while(bh != head);
1096         gfs2_log_unlock(sdp);
1097
1098         head = bh = page_buffers(page);
1099         do {
1100                 gfs2_log_lock(sdp);
1101                 bd = bh->b_private;
1102                 if (bd) {
1103                         gfs2_assert_warn(sdp, bd->bd_bh == bh);
1104                         gfs2_assert_warn(sdp, list_empty(&bd->bd_list_tr));
1105                         if (!list_empty(&bd->bd_le.le_list)) {
1106                                 if (!buffer_pinned(bh))
1107                                         list_del_init(&bd->bd_le.le_list);
1108                                 else
1109                                         bd = NULL;
1110                         }
1111                         if (bd)
1112                                 bd->bd_bh = NULL;
1113                         bh->b_private = NULL;
1114                 }
1115                 gfs2_log_unlock(sdp);
1116                 if (bd)
1117                         kmem_cache_free(gfs2_bufdata_cachep, bd);
1118
1119                 bh = bh->b_this_page;
1120         } while (bh != head);
1121
1122         return try_to_free_buffers(page);
1123 cannot_release:
1124         gfs2_log_unlock(sdp);
1125         return 0;
1126 }
1127
1128 static const struct address_space_operations gfs2_writeback_aops = {
1129         .writepage = gfs2_writeback_writepage,
1130         .writepages = gfs2_writeback_writepages,
1131         .readpage = gfs2_readpage,
1132         .readpages = gfs2_readpages,
1133         .sync_page = block_sync_page,
1134         .write_begin = gfs2_write_begin,
1135         .write_end = gfs2_write_end,
1136         .bmap = gfs2_bmap,
1137         .invalidatepage = gfs2_invalidatepage,
1138         .releasepage = gfs2_releasepage,
1139         .direct_IO = gfs2_direct_IO,
1140         .migratepage = buffer_migrate_page,
1141         .is_partially_uptodate = block_is_partially_uptodate,
1142         .error_remove_page = generic_error_remove_page,
1143 };
1144
1145 static const struct address_space_operations gfs2_ordered_aops = {
1146         .writepage = gfs2_ordered_writepage,
1147         .readpage = gfs2_readpage,
1148         .readpages = gfs2_readpages,
1149         .sync_page = block_sync_page,
1150         .write_begin = gfs2_write_begin,
1151         .write_end = gfs2_write_end,
1152         .set_page_dirty = gfs2_set_page_dirty,
1153         .bmap = gfs2_bmap,
1154         .invalidatepage = gfs2_invalidatepage,
1155         .releasepage = gfs2_releasepage,
1156         .direct_IO = gfs2_direct_IO,
1157         .migratepage = buffer_migrate_page,
1158         .is_partially_uptodate = block_is_partially_uptodate,
1159         .error_remove_page = generic_error_remove_page,
1160 };
1161
1162 static const struct address_space_operations gfs2_jdata_aops = {
1163         .writepage = gfs2_jdata_writepage,
1164         .writepages = gfs2_jdata_writepages,
1165         .readpage = gfs2_readpage,
1166         .readpages = gfs2_readpages,
1167         .sync_page = block_sync_page,
1168         .write_begin = gfs2_write_begin,
1169         .write_end = gfs2_write_end,
1170         .set_page_dirty = gfs2_set_page_dirty,
1171         .bmap = gfs2_bmap,
1172         .invalidatepage = gfs2_invalidatepage,
1173         .releasepage = gfs2_releasepage,
1174         .is_partially_uptodate = block_is_partially_uptodate,
1175         .error_remove_page = generic_error_remove_page,
1176 };
1177
1178 void gfs2_set_aops(struct inode *inode)
1179 {
1180         struct gfs2_inode *ip = GFS2_I(inode);
1181
1182         if (gfs2_is_writeback(ip))
1183                 inode->i_mapping->a_ops = &gfs2_writeback_aops;
1184         else if (gfs2_is_ordered(ip))
1185                 inode->i_mapping->a_ops = &gfs2_ordered_aops;
1186         else if (gfs2_is_jdata(ip))
1187                 inode->i_mapping->a_ops = &gfs2_jdata_aops;
1188         else
1189                 BUG();
1190 }
1191