[GFS2] Reduce inode size by moving i_alloc out of line
[linux-2.6.git] / fs / gfs2 / ops_address.c
1 /*
2  * Copyright (C) Sistina Software, Inc.  1997-2003 All rights reserved.
3  * Copyright (C) 2004-2007 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/lm_interface.h>
23 #include <linux/backing-dev.h>
24 #include <linux/pagevec.h>
25
26 #include "gfs2.h"
27 #include "incore.h"
28 #include "bmap.h"
29 #include "glock.h"
30 #include "inode.h"
31 #include "log.h"
32 #include "meta_io.h"
33 #include "ops_address.h"
34 #include "quota.h"
35 #include "trans.h"
36 #include "rgrp.h"
37 #include "super.h"
38 #include "util.h"
39 #include "glops.h"
40
41
42 static void gfs2_page_add_databufs(struct gfs2_inode *ip, struct page *page,
43                                    unsigned int from, unsigned int to)
44 {
45         struct buffer_head *head = page_buffers(page);
46         unsigned int bsize = head->b_size;
47         struct buffer_head *bh;
48         unsigned int start, end;
49
50         for (bh = head, start = 0; bh != head || !start;
51              bh = bh->b_this_page, start = end) {
52                 end = start + bsize;
53                 if (end <= from || start >= to)
54                         continue;
55                 if (gfs2_is_jdata(ip))
56                         set_buffer_uptodate(bh);
57                 gfs2_trans_add_bh(ip->i_gl, bh, 0);
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 static int gfs2_get_block_direct(struct inode *inode, sector_t lblock,
85                                  struct buffer_head *bh_result, int create)
86 {
87         return gfs2_block_map(inode, lblock, bh_result, 0);
88 }
89
90 /**
91  * gfs2_writepage_common - Common bits of writepage
92  * @page: The page to be written
93  * @wbc: The writeback control
94  *
95  * Returns: 1 if writepage is ok, otherwise an error code or zero if no error.
96  */
97
98 static int gfs2_writepage_common(struct page *page,
99                                  struct writeback_control *wbc)
100 {
101         struct inode *inode = page->mapping->host;
102         struct gfs2_inode *ip = GFS2_I(inode);
103         struct gfs2_sbd *sdp = GFS2_SB(inode);
104         loff_t i_size = i_size_read(inode);
105         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
106         unsigned offset;
107         int ret = -EIO;
108
109         if (gfs2_assert_withdraw(sdp, gfs2_glock_is_held_excl(ip->i_gl)))
110                 goto out;
111         ret = 0;
112         if (current->journal_info)
113                 goto redirty;
114         /* Is the page fully outside i_size? (truncate in progress) */
115         offset = i_size & (PAGE_CACHE_SIZE-1);
116         if (page->index > end_index || (page->index == end_index && !offset)) {
117                 page->mapping->a_ops->invalidatepage(page, 0);
118                 goto out;
119         }
120         return 1;
121 redirty:
122         redirty_page_for_writepage(wbc, page);
123 out:
124         unlock_page(page);
125         return 0;
126 }
127
128 /**
129  * gfs2_writeback_writepage - Write page for writeback mappings
130  * @page: The page
131  * @wbc: The writeback control
132  *
133  */
134
135 static int gfs2_writeback_writepage(struct page *page,
136                                     struct writeback_control *wbc)
137 {
138         int ret;
139
140         ret = gfs2_writepage_common(page, wbc);
141         if (ret <= 0)
142                 return ret;
143
144         ret = mpage_writepage(page, gfs2_get_block_noalloc, wbc);
145         if (ret == -EAGAIN)
146                 ret = block_write_full_page(page, gfs2_get_block_noalloc, wbc);
147         return ret;
148 }
149
150 /**
151  * gfs2_ordered_writepage - Write page for ordered data files
152  * @page: The page to write
153  * @wbc: The writeback control
154  *
155  */
156
157 static int gfs2_ordered_writepage(struct page *page,
158                                   struct writeback_control *wbc)
159 {
160         struct inode *inode = page->mapping->host;
161         struct gfs2_inode *ip = GFS2_I(inode);
162         int ret;
163
164         ret = gfs2_writepage_common(page, wbc);
165         if (ret <= 0)
166                 return ret;
167
168         if (!page_has_buffers(page)) {
169                 create_empty_buffers(page, inode->i_sb->s_blocksize,
170                                      (1 << BH_Dirty)|(1 << BH_Uptodate));
171         }
172         gfs2_page_add_databufs(ip, page, 0, inode->i_sb->s_blocksize-1);
173         return block_write_full_page(page, gfs2_get_block_noalloc, wbc);
174 }
175
176 /**
177  * __gfs2_jdata_writepage - The core of jdata writepage
178  * @page: The page to write
179  * @wbc: The writeback control
180  *
181  * This is shared between writepage and writepages and implements the
182  * core of the writepage operation. If a transaction is required then
183  * PageChecked will have been set and the transaction will have
184  * already been started before this is called.
185  */
186
187 static int __gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
188 {
189         struct inode *inode = page->mapping->host;
190         struct gfs2_inode *ip = GFS2_I(inode);
191         struct gfs2_sbd *sdp = GFS2_SB(inode);
192
193         if (PageChecked(page)) {
194                 ClearPageChecked(page);
195                 if (!page_has_buffers(page)) {
196                         create_empty_buffers(page, inode->i_sb->s_blocksize,
197                                              (1 << BH_Dirty)|(1 << BH_Uptodate));
198                 }
199                 gfs2_page_add_databufs(ip, page, 0, sdp->sd_vfs->s_blocksize-1);
200         }
201         return block_write_full_page(page, gfs2_get_block_noalloc, wbc);
202 }
203
204 /**
205  * gfs2_jdata_writepage - Write complete page
206  * @page: Page to write
207  *
208  * Returns: errno
209  *
210  */
211
212 static int gfs2_jdata_writepage(struct page *page, struct writeback_control *wbc)
213 {
214         struct inode *inode = page->mapping->host;
215         struct gfs2_sbd *sdp = GFS2_SB(inode);
216         int error;
217         int done_trans = 0;
218
219         error = gfs2_writepage_common(page, wbc);
220         if (error <= 0)
221                 return error;
222
223         if (PageChecked(page)) {
224                 if (wbc->sync_mode != WB_SYNC_ALL)
225                         goto out_ignore;
226                 error = gfs2_trans_begin(sdp, RES_DINODE + 1, 0);
227                 if (error)
228                         goto out_ignore;
229                 done_trans = 1;
230         }
231         error = __gfs2_jdata_writepage(page, wbc);
232         if (done_trans)
233                 gfs2_trans_end(sdp);
234         return error;
235
236 out_ignore:
237         redirty_page_for_writepage(wbc, page);
238         unlock_page(page);
239         return 0;
240 }
241
242 /**
243  * gfs2_writeback_writepages - Write a bunch of dirty pages back to disk
244  * @mapping: The mapping to write
245  * @wbc: Write-back control
246  *
247  * For the data=writeback case we can already ignore buffer heads
248  * and write whole extents at once. This is a big reduction in the
249  * number of I/O requests we send and the bmap calls we make in this case.
250  */
251 static int gfs2_writeback_writepages(struct address_space *mapping,
252                                      struct writeback_control *wbc)
253 {
254         return mpage_writepages(mapping, wbc, gfs2_get_block_noalloc);
255 }
256
257 /**
258  * gfs2_write_jdata_pagevec - Write back a pagevec's worth of pages
259  * @mapping: The mapping
260  * @wbc: The writeback control
261  * @writepage: The writepage function to call for each page
262  * @pvec: The vector of pages
263  * @nr_pages: The number of pages to write
264  *
265  * Returns: non-zero if loop should terminate, zero otherwise
266  */
267
268 static int gfs2_write_jdata_pagevec(struct address_space *mapping,
269                                     struct writeback_control *wbc,
270                                     struct pagevec *pvec,
271                                     int nr_pages, pgoff_t end)
272 {
273         struct inode *inode = mapping->host;
274         struct gfs2_sbd *sdp = GFS2_SB(inode);
275         loff_t i_size = i_size_read(inode);
276         pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
277         unsigned offset = i_size & (PAGE_CACHE_SIZE-1);
278         unsigned nrblocks = nr_pages * (PAGE_CACHE_SIZE/inode->i_sb->s_blocksize);
279         struct backing_dev_info *bdi = mapping->backing_dev_info;
280         int i;
281         int ret;
282
283         ret = gfs2_trans_begin(sdp, nrblocks, 0);
284         if (ret < 0)
285                 return ret;
286
287         for(i = 0; i < nr_pages; i++) {
288                 struct page *page = pvec->pages[i];
289
290                 lock_page(page);
291
292                 if (unlikely(page->mapping != mapping)) {
293                         unlock_page(page);
294                         continue;
295                 }
296
297                 if (!wbc->range_cyclic && page->index > end) {
298                         ret = 1;
299                         unlock_page(page);
300                         continue;
301                 }
302
303                 if (wbc->sync_mode != WB_SYNC_NONE)
304                         wait_on_page_writeback(page);
305
306                 if (PageWriteback(page) ||
307                     !clear_page_dirty_for_io(page)) {
308                         unlock_page(page);
309                         continue;
310                 }
311
312                 /* Is the page fully outside i_size? (truncate in progress) */
313                 if (page->index > end_index || (page->index == end_index && !offset)) {
314                         page->mapping->a_ops->invalidatepage(page, 0);
315                         unlock_page(page);
316                         continue;
317                 }
318
319                 ret = __gfs2_jdata_writepage(page, wbc);
320
321                 if (ret || (--(wbc->nr_to_write) <= 0))
322                         ret = 1;
323                 if (wbc->nonblocking && bdi_write_congested(bdi)) {
324                         wbc->encountered_congestion = 1;
325                         ret = 1;
326                 }
327
328         }
329         gfs2_trans_end(sdp);
330         return ret;
331 }
332
333 /**
334  * gfs2_write_cache_jdata - Like write_cache_pages but different
335  * @mapping: The mapping to write
336  * @wbc: The writeback control
337  * @writepage: The writepage function to call
338  * @data: The data to pass to writepage
339  *
340  * The reason that we use our own function here is that we need to
341  * start transactions before we grab page locks. This allows us
342  * to get the ordering right.
343  */
344
345 static int gfs2_write_cache_jdata(struct address_space *mapping,
346                                   struct writeback_control *wbc)
347 {
348         struct backing_dev_info *bdi = mapping->backing_dev_info;
349         int ret = 0;
350         int done = 0;
351         struct pagevec pvec;
352         int nr_pages;
353         pgoff_t index;
354         pgoff_t end;
355         int scanned = 0;
356         int range_whole = 0;
357
358         if (wbc->nonblocking && bdi_write_congested(bdi)) {
359                 wbc->encountered_congestion = 1;
360                 return 0;
361         }
362
363         pagevec_init(&pvec, 0);
364         if (wbc->range_cyclic) {
365                 index = mapping->writeback_index; /* Start from prev offset */
366                 end = -1;
367         } else {
368                 index = wbc->range_start >> PAGE_CACHE_SHIFT;
369                 end = wbc->range_end >> PAGE_CACHE_SHIFT;
370                 if (wbc->range_start == 0 && wbc->range_end == LLONG_MAX)
371                         range_whole = 1;
372                 scanned = 1;
373         }
374
375 retry:
376          while (!done && (index <= end) &&
377                 (nr_pages = pagevec_lookup_tag(&pvec, mapping, &index,
378                                                PAGECACHE_TAG_DIRTY,
379                                                min(end - index, (pgoff_t)PAGEVEC_SIZE-1) + 1))) {
380                 scanned = 1;
381                 ret = gfs2_write_jdata_pagevec(mapping, wbc, &pvec, nr_pages, end);
382                 if (ret)
383                         done = 1;
384                 if (ret > 0)
385                         ret = 0;
386
387                 pagevec_release(&pvec);
388                 cond_resched();
389         }
390
391         if (!scanned && !done) {
392                 /*
393                  * We hit the last page and there is more work to be done: wrap
394                  * back to the start of the file
395                  */
396                 scanned = 1;
397                 index = 0;
398                 goto retry;
399         }
400
401         if (wbc->range_cyclic || (range_whole && wbc->nr_to_write > 0))
402                 mapping->writeback_index = index;
403         return ret;
404 }
405
406
407 /**
408  * gfs2_jdata_writepages - Write a bunch of dirty pages back to disk
409  * @mapping: The mapping to write
410  * @wbc: The writeback control
411  * 
412  */
413
414 static int gfs2_jdata_writepages(struct address_space *mapping,
415                                  struct writeback_control *wbc)
416 {
417         struct gfs2_inode *ip = GFS2_I(mapping->host);
418         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
419         int ret;
420
421         ret = gfs2_write_cache_jdata(mapping, wbc);
422         if (ret == 0 && wbc->sync_mode == WB_SYNC_ALL) {
423                 gfs2_log_flush(sdp, ip->i_gl);
424                 ret = gfs2_write_cache_jdata(mapping, wbc);
425         }
426         return ret;
427 }
428
429 /**
430  * stuffed_readpage - Fill in a Linux page with stuffed file data
431  * @ip: the inode
432  * @page: the page
433  *
434  * Returns: errno
435  */
436
437 static int stuffed_readpage(struct gfs2_inode *ip, struct page *page)
438 {
439         struct buffer_head *dibh;
440         void *kaddr;
441         int error;
442
443         /*
444          * Due to the order of unstuffing files and ->nopage(), we can be
445          * asked for a zero page in the case of a stuffed file being extended,
446          * so we need to supply one here. It doesn't happen often.
447          */
448         if (unlikely(page->index)) {
449                 zero_user_page(page, 0, PAGE_CACHE_SIZE, KM_USER0);
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, KM_USER0);
458         memcpy(kaddr, dibh->b_data + sizeof(struct gfs2_dinode),
459                ip->i_di.di_size);
460         memset(kaddr + ip->i_di.di_size, 0, PAGE_CACHE_SIZE - ip->i_di.di_size);
461         kunmap_atomic(kaddr, KM_USER0);
462         flush_dcache_page(page);
463         brelse(dibh);
464         SetPageUptodate(page);
465
466         return 0;
467 }
468
469
470 /**
471  * __gfs2_readpage - readpage
472  * @file: The file to read a page for
473  * @page: The page to read
474  *
475  * This is the core of gfs2's readpage. Its used by the internal file
476  * reading code as in that case we already hold the glock. Also its
477  * called by gfs2_readpage() once the required lock has been granted.
478  *
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         int error;
486
487         if (gfs2_is_stuffed(ip)) {
488                 error = stuffed_readpage(ip, page);
489                 unlock_page(page);
490         } else {
491                 error = mpage_readpage(page, gfs2_block_map);
492         }
493
494         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
495                 return -EIO;
496
497         return error;
498 }
499
500 /**
501  * gfs2_readpage - read a page of a file
502  * @file: The file to read
503  * @page: The page of the file
504  *
505  * This deals with the locking required. We use a trylock in order to
506  * avoid the page lock / glock ordering problems returning AOP_TRUNCATED_PAGE
507  * in the event that we are unable to get the lock.
508  */
509
510 static int gfs2_readpage(struct file *file, struct page *page)
511 {
512         struct gfs2_inode *ip = GFS2_I(page->mapping->host);
513         struct gfs2_holder gh;
514         int error;
515
516         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, GL_ATIME|LM_FLAG_TRY_1CB, &gh);
517         error = gfs2_glock_nq_atime(&gh);
518         if (unlikely(error)) {
519                 unlock_page(page);
520                 goto out;
521         }
522         error = __gfs2_readpage(file, page);
523         gfs2_glock_dq(&gh);
524 out:
525         gfs2_holder_uninit(&gh);
526         if (error == GLR_TRYFAILED) {
527                 yield();
528                 return AOP_TRUNCATED_PAGE;
529         }
530         return error;
531 }
532
533 /**
534  * gfs2_internal_read - read an internal file
535  * @ip: The gfs2 inode
536  * @ra_state: The readahead state (or NULL for no readahead)
537  * @buf: The buffer to fill
538  * @pos: The file position
539  * @size: The amount to read
540  *
541  */
542
543 int gfs2_internal_read(struct gfs2_inode *ip, struct file_ra_state *ra_state,
544                        char *buf, loff_t *pos, unsigned size)
545 {
546         struct address_space *mapping = ip->i_inode.i_mapping;
547         unsigned long index = *pos / PAGE_CACHE_SIZE;
548         unsigned offset = *pos & (PAGE_CACHE_SIZE - 1);
549         unsigned copied = 0;
550         unsigned amt;
551         struct page *page;
552         void *p;
553
554         do {
555                 amt = size - copied;
556                 if (offset + size > PAGE_CACHE_SIZE)
557                         amt = PAGE_CACHE_SIZE - offset;
558                 page = read_cache_page(mapping, index, __gfs2_readpage, NULL);
559                 if (IS_ERR(page))
560                         return PTR_ERR(page);
561                 p = kmap_atomic(page, KM_USER0);
562                 memcpy(buf + copied, p + offset, amt);
563                 kunmap_atomic(p, KM_USER0);
564                 mark_page_accessed(page);
565                 page_cache_release(page);
566                 copied += amt;
567                 index++;
568                 offset = 0;
569         } while(copied < size);
570         (*pos) += size;
571         return size;
572 }
573
574 /**
575  * gfs2_readpages - Read a bunch of pages at once
576  *
577  * Some notes:
578  * 1. This is only for readahead, so we can simply ignore any things
579  *    which are slightly inconvenient (such as locking conflicts between
580  *    the page lock and the glock) and return having done no I/O. Its
581  *    obviously not something we'd want to do on too regular a basis.
582  *    Any I/O we ignore at this time will be done via readpage later.
583  * 2. We don't handle stuffed files here we let readpage do the honours.
584  * 3. mpage_readpages() does most of the heavy lifting in the common case.
585  * 4. gfs2_block_map() is relied upon to set BH_Boundary in the right places.
586  */
587
588 static int gfs2_readpages(struct file *file, struct address_space *mapping,
589                           struct list_head *pages, unsigned nr_pages)
590 {
591         struct inode *inode = mapping->host;
592         struct gfs2_inode *ip = GFS2_I(inode);
593         struct gfs2_sbd *sdp = GFS2_SB(inode);
594         struct gfs2_holder gh;
595         int ret;
596
597         gfs2_holder_init(ip->i_gl, LM_ST_SHARED, GL_ATIME, &gh);
598         ret = gfs2_glock_nq_atime(&gh);
599         if (unlikely(ret))
600                 goto out_uninit;
601         if (!gfs2_is_stuffed(ip))
602                 ret = mpage_readpages(mapping, pages, nr_pages, gfs2_block_map);
603         gfs2_glock_dq(&gh);
604 out_uninit:
605         gfs2_holder_uninit(&gh);
606         if (unlikely(test_bit(SDF_SHUTDOWN, &sdp->sd_flags)))
607                 ret = -EIO;
608         return ret;
609 }
610
611 /**
612  * gfs2_write_begin - Begin to write to a file
613  * @file: The file to write to
614  * @mapping: The mapping in which to write
615  * @pos: The file offset at which to start writing
616  * @len: Length of the write
617  * @flags: Various flags
618  * @pagep: Pointer to return the page
619  * @fsdata: Pointer to return fs data (unused by GFS2)
620  *
621  * Returns: errno
622  */
623
624 static int gfs2_write_begin(struct file *file, struct address_space *mapping,
625                             loff_t pos, unsigned len, unsigned flags,
626                             struct page **pagep, void **fsdata)
627 {
628         struct gfs2_inode *ip = GFS2_I(mapping->host);
629         struct gfs2_sbd *sdp = GFS2_SB(mapping->host);
630         unsigned int data_blocks, ind_blocks, rblocks;
631         int alloc_required;
632         int error = 0;
633         struct gfs2_alloc *al;
634         pgoff_t index = pos >> PAGE_CACHE_SHIFT;
635         unsigned from = pos & (PAGE_CACHE_SIZE - 1);
636         unsigned to = from + len;
637         struct page *page;
638
639         gfs2_holder_init(ip->i_gl, LM_ST_EXCLUSIVE, GL_ATIME, &ip->i_gh);
640         error = gfs2_glock_nq_atime(&ip->i_gh);
641         if (unlikely(error))
642                 goto out_uninit;
643
644         gfs2_write_calc_reserv(ip, len, &data_blocks, &ind_blocks);
645         error = gfs2_write_alloc_required(ip, pos, len, &alloc_required);
646         if (error)
647                 goto out_unlock;
648
649         if (alloc_required) {
650                 al = gfs2_alloc_get(ip);
651
652                 error = gfs2_quota_lock(ip, NO_QUOTA_CHANGE, NO_QUOTA_CHANGE);
653                 if (error)
654                         goto out_alloc_put;
655
656                 error = gfs2_quota_check(ip, ip->i_inode.i_uid, ip->i_inode.i_gid);
657                 if (error)
658                         goto out_qunlock;
659
660                 al->al_requested = data_blocks + ind_blocks;
661                 error = gfs2_inplace_reserve(ip);
662                 if (error)
663                         goto out_qunlock;
664         }
665
666         rblocks = RES_DINODE + ind_blocks;
667         if (gfs2_is_jdata(ip))
668                 rblocks += data_blocks ? data_blocks : 1;
669         if (ind_blocks || data_blocks)
670                 rblocks += RES_STATFS + RES_QUOTA;
671
672         error = gfs2_trans_begin(sdp, rblocks,
673                                  PAGE_CACHE_SIZE/sdp->sd_sb.sb_bsize);
674         if (error)
675                 goto out_trans_fail;
676
677         error = -ENOMEM;
678         page = __grab_cache_page(mapping, index);
679         *pagep = page;
680         if (unlikely(!page))
681                 goto out_endtrans;
682
683         if (gfs2_is_stuffed(ip)) {
684                 error = 0;
685                 if (pos + len > sdp->sd_sb.sb_bsize - sizeof(struct gfs2_dinode)) {
686                         error = gfs2_unstuff_dinode(ip, page);
687                         if (error == 0)
688                                 goto prepare_write;
689                 } else if (!PageUptodate(page)) {
690                         error = stuffed_readpage(ip, page);
691                 }
692                 goto out;
693         }
694
695 prepare_write:
696         error = block_prepare_write(page, from, to, gfs2_block_map);
697 out:
698         if (error == 0)
699                 return 0;
700
701         page_cache_release(page);
702         if (pos + len > ip->i_inode.i_size)
703                 vmtruncate(&ip->i_inode, ip->i_inode.i_size);
704 out_endtrans:
705         gfs2_trans_end(sdp);
706 out_trans_fail:
707         if (alloc_required) {
708                 gfs2_inplace_release(ip);
709 out_qunlock:
710                 gfs2_quota_unlock(ip);
711 out_alloc_put:
712                 gfs2_alloc_put(ip);
713         }
714 out_unlock:
715         gfs2_glock_dq(&ip->i_gh);
716 out_uninit:
717         gfs2_holder_uninit(&ip->i_gh);
718         return error;
719 }
720
721 /**
722  * adjust_fs_space - Adjusts the free space available due to gfs2_grow
723  * @inode: the rindex inode
724  */
725 static void adjust_fs_space(struct inode *inode)
726 {
727         struct gfs2_sbd *sdp = inode->i_sb->s_fs_info;
728         struct gfs2_statfs_change_host *m_sc = &sdp->sd_statfs_master;
729         struct gfs2_statfs_change_host *l_sc = &sdp->sd_statfs_local;
730         u64 fs_total, new_free;
731
732         /* Total up the file system space, according to the latest rindex. */
733         fs_total = gfs2_ri_total(sdp);
734
735         spin_lock(&sdp->sd_statfs_spin);
736         if (fs_total > (m_sc->sc_total + l_sc->sc_total))
737                 new_free = fs_total - (m_sc->sc_total + l_sc->sc_total);
738         else
739                 new_free = 0;
740         spin_unlock(&sdp->sd_statfs_spin);
741         fs_warn(sdp, "File system extended by %llu blocks.\n",
742                 (unsigned long long)new_free);
743         gfs2_statfs_change(sdp, new_free, new_free, 0);
744 }
745
746 /**
747  * gfs2_stuffed_write_end - Write end for stuffed files
748  * @inode: The inode
749  * @dibh: The buffer_head containing the on-disk inode
750  * @pos: The file position
751  * @len: The length of the write
752  * @copied: How much was actually copied by the VFS
753  * @page: The page
754  *
755  * This copies the data from the page into the inode block after
756  * the inode data structure itself.
757  *
758  * Returns: errno
759  */
760 static int gfs2_stuffed_write_end(struct inode *inode, struct buffer_head *dibh,
761                                   loff_t pos, unsigned len, unsigned copied,
762                                   struct page *page)
763 {
764         struct gfs2_inode *ip = GFS2_I(inode);
765         struct gfs2_sbd *sdp = GFS2_SB(inode);
766         u64 to = pos + copied;
767         void *kaddr;
768         unsigned char *buf = dibh->b_data + sizeof(struct gfs2_dinode);
769         struct gfs2_dinode *di = (struct gfs2_dinode *)dibh->b_data;
770
771         BUG_ON((pos + len) > (dibh->b_size - sizeof(struct gfs2_dinode)));
772         kaddr = kmap_atomic(page, KM_USER0);
773         memcpy(buf + pos, kaddr + pos, copied);
774         memset(kaddr + pos + copied, 0, len - copied);
775         flush_dcache_page(page);
776         kunmap_atomic(kaddr, KM_USER0);
777
778         if (!PageUptodate(page))
779                 SetPageUptodate(page);
780         unlock_page(page);
781         page_cache_release(page);
782
783         if (inode->i_size < to) {
784                 i_size_write(inode, to);
785                 ip->i_di.di_size = inode->i_size;
786                 di->di_size = cpu_to_be64(inode->i_size);
787                 mark_inode_dirty(inode);
788         }
789
790         if (inode == sdp->sd_rindex)
791                 adjust_fs_space(inode);
792
793         brelse(dibh);
794         gfs2_trans_end(sdp);
795         gfs2_glock_dq(&ip->i_gh);
796         gfs2_holder_uninit(&ip->i_gh);
797         return copied;
798 }
799
800 /**
801  * gfs2_write_end
802  * @file: The file to write to
803  * @mapping: The address space to write to
804  * @pos: The file position
805  * @len: The length of the data
806  * @copied:
807  * @page: The page that has been written
808  * @fsdata: The fsdata (unused in GFS2)
809  *
810  * The main write_end function for GFS2. We have a separate one for
811  * stuffed files as they are slightly different, otherwise we just
812  * put our locking around the VFS provided functions.
813  *
814  * Returns: errno
815  */
816
817 static int gfs2_write_end(struct file *file, struct address_space *mapping,
818                           loff_t pos, unsigned len, unsigned copied,
819                           struct page *page, void *fsdata)
820 {
821         struct inode *inode = page->mapping->host;
822         struct gfs2_inode *ip = GFS2_I(inode);
823         struct gfs2_sbd *sdp = GFS2_SB(inode);
824         struct buffer_head *dibh;
825         struct gfs2_alloc *al = ip->i_alloc;
826         struct gfs2_dinode *di;
827         unsigned int from = pos & (PAGE_CACHE_SIZE - 1);
828         unsigned int to = from + len;
829         int ret;
830
831         BUG_ON(gfs2_glock_is_locked_by_me(ip->i_gl) == 0);
832
833         ret = gfs2_meta_inode_buffer(ip, &dibh);
834         if (unlikely(ret)) {
835                 unlock_page(page);
836                 page_cache_release(page);
837                 goto failed;
838         }
839
840         gfs2_trans_add_bh(ip->i_gl, dibh, 1);
841
842         if (gfs2_is_stuffed(ip))
843                 return gfs2_stuffed_write_end(inode, dibh, pos, len, copied, page);
844
845         if (!gfs2_is_writeback(ip))
846                 gfs2_page_add_databufs(ip, page, from, to);
847
848         ret = generic_write_end(file, mapping, pos, len, copied, page, fsdata);
849
850         if (likely(ret >= 0) && (inode->i_size > ip->i_di.di_size)) {
851                 di = (struct gfs2_dinode *)dibh->b_data;
852                 ip->i_di.di_size = inode->i_size;
853                 di->di_size = cpu_to_be64(inode->i_size);
854                 mark_inode_dirty(inode);
855         }
856
857         if (inode == sdp->sd_rindex)
858                 adjust_fs_space(inode);
859
860         brelse(dibh);
861         gfs2_trans_end(sdp);
862 failed:
863         if (al) {
864                 gfs2_inplace_release(ip);
865                 gfs2_quota_unlock(ip);
866                 gfs2_alloc_put(ip);
867         }
868         gfs2_glock_dq(&ip->i_gh);
869         gfs2_holder_uninit(&ip->i_gh);
870         return ret;
871 }
872
873 /**
874  * gfs2_set_page_dirty - Page dirtying function
875  * @page: The page to dirty
876  *
877  * Returns: 1 if it dirtyed the page, or 0 otherwise
878  */
879  
880 static int gfs2_set_page_dirty(struct page *page)
881 {
882         SetPageChecked(page);
883         return __set_page_dirty_buffers(page);
884 }
885
886 /**
887  * gfs2_bmap - Block map function
888  * @mapping: Address space info
889  * @lblock: The block to map
890  *
891  * Returns: The disk address for the block or 0 on hole or error
892  */
893
894 static sector_t gfs2_bmap(struct address_space *mapping, sector_t lblock)
895 {
896         struct gfs2_inode *ip = GFS2_I(mapping->host);
897         struct gfs2_holder i_gh;
898         sector_t dblock = 0;
899         int error;
900
901         error = gfs2_glock_nq_init(ip->i_gl, LM_ST_SHARED, LM_FLAG_ANY, &i_gh);
902         if (error)
903                 return 0;
904
905         if (!gfs2_is_stuffed(ip))
906                 dblock = generic_block_bmap(mapping, lblock, gfs2_block_map);
907
908         gfs2_glock_dq_uninit(&i_gh);
909
910         return dblock;
911 }
912
913 static void gfs2_discard(struct gfs2_sbd *sdp, struct buffer_head *bh)
914 {
915         struct gfs2_bufdata *bd;
916
917         lock_buffer(bh);
918         gfs2_log_lock(sdp);
919         clear_buffer_dirty(bh);
920         bd = bh->b_private;
921         if (bd) {
922                 if (!list_empty(&bd->bd_le.le_list) && !buffer_pinned(bh))
923                         list_del_init(&bd->bd_le.le_list);
924                 else
925                         gfs2_remove_from_journal(bh, current->journal_info, 0);
926         }
927         bh->b_bdev = NULL;
928         clear_buffer_mapped(bh);
929         clear_buffer_req(bh);
930         clear_buffer_new(bh);
931         gfs2_log_unlock(sdp);
932         unlock_buffer(bh);
933 }
934
935 static void gfs2_invalidatepage(struct page *page, unsigned long offset)
936 {
937         struct gfs2_sbd *sdp = GFS2_SB(page->mapping->host);
938         struct buffer_head *bh, *head;
939         unsigned long pos = 0;
940
941         BUG_ON(!PageLocked(page));
942         if (offset == 0)
943                 ClearPageChecked(page);
944         if (!page_has_buffers(page))
945                 goto out;
946
947         bh = head = page_buffers(page);
948         do {
949                 if (offset <= pos)
950                         gfs2_discard(sdp, bh);
951                 pos += bh->b_size;
952                 bh = bh->b_this_page;
953         } while (bh != head);
954 out:
955         if (offset == 0)
956                 try_to_release_page(page, 0);
957 }
958
959 /**
960  * gfs2_ok_for_dio - check that dio is valid on this file
961  * @ip: The inode
962  * @rw: READ or WRITE
963  * @offset: The offset at which we are reading or writing
964  *
965  * Returns: 0 (to ignore the i/o request and thus fall back to buffered i/o)
966  *          1 (to accept the i/o request)
967  */
968 static int gfs2_ok_for_dio(struct gfs2_inode *ip, int rw, loff_t offset)
969 {
970         /*
971          * Should we return an error here? I can't see that O_DIRECT for
972          * a stuffed file makes any sense. For now we'll silently fall
973          * back to buffered I/O
974          */
975         if (gfs2_is_stuffed(ip))
976                 return 0;
977
978         if (offset > i_size_read(&ip->i_inode))
979                 return 0;
980         return 1;
981 }
982
983
984
985 static ssize_t gfs2_direct_IO(int rw, struct kiocb *iocb,
986                               const struct iovec *iov, loff_t offset,
987                               unsigned long nr_segs)
988 {
989         struct file *file = iocb->ki_filp;
990         struct inode *inode = file->f_mapping->host;
991         struct gfs2_inode *ip = GFS2_I(inode);
992         struct gfs2_holder gh;
993         int rv;
994
995         /*
996          * Deferred lock, even if its a write, since we do no allocation
997          * on this path. All we need change is atime, and this lock mode
998          * ensures that other nodes have flushed their buffered read caches
999          * (i.e. their page cache entries for this inode). We do not,
1000          * unfortunately have the option of only flushing a range like
1001          * the VFS does.
1002          */
1003         gfs2_holder_init(ip->i_gl, LM_ST_DEFERRED, GL_ATIME, &gh);
1004         rv = gfs2_glock_nq_atime(&gh);
1005         if (rv)
1006                 return rv;
1007         rv = gfs2_ok_for_dio(ip, rw, offset);
1008         if (rv != 1)
1009                 goto out; /* dio not valid, fall back to buffered i/o */
1010
1011         rv = blockdev_direct_IO_no_locking(rw, iocb, inode, inode->i_sb->s_bdev,
1012                                            iov, offset, nr_segs,
1013                                            gfs2_get_block_direct, NULL);
1014 out:
1015         gfs2_glock_dq_m(1, &gh);
1016         gfs2_holder_uninit(&gh);
1017         return rv;
1018 }
1019
1020 /**
1021  * gfs2_releasepage - free the metadata associated with a page
1022  * @page: the page that's being released
1023  * @gfp_mask: passed from Linux VFS, ignored by us
1024  *
1025  * Call try_to_free_buffers() if the buffers in this page can be
1026  * released.
1027  *
1028  * Returns: 0
1029  */
1030
1031 int gfs2_releasepage(struct page *page, gfp_t gfp_mask)
1032 {
1033         struct inode *aspace = page->mapping->host;
1034         struct gfs2_sbd *sdp = aspace->i_sb->s_fs_info;
1035         struct buffer_head *bh, *head;
1036         struct gfs2_bufdata *bd;
1037
1038         if (!page_has_buffers(page))
1039                 return 0;
1040
1041         gfs2_log_lock(sdp);
1042         head = bh = page_buffers(page);
1043         do {
1044                 if (atomic_read(&bh->b_count))
1045                         goto cannot_release;
1046                 bd = bh->b_private;
1047                 if (bd && bd->bd_ail)
1048                         goto cannot_release;
1049                 gfs2_assert_warn(sdp, !buffer_pinned(bh));
1050                 gfs2_assert_warn(sdp, !buffer_dirty(bh));
1051                 bh = bh->b_this_page;
1052         } while(bh != head);
1053         gfs2_log_unlock(sdp);
1054
1055         head = bh = page_buffers(page);
1056         do {
1057                 gfs2_log_lock(sdp);
1058                 bd = bh->b_private;
1059                 if (bd) {
1060                         gfs2_assert_warn(sdp, bd->bd_bh == bh);
1061                         gfs2_assert_warn(sdp, list_empty(&bd->bd_list_tr));
1062                         if (!list_empty(&bd->bd_le.le_list)) {
1063                                 if (!buffer_pinned(bh))
1064                                         list_del_init(&bd->bd_le.le_list);
1065                                 else
1066                                         bd = NULL;
1067                         }
1068                         if (bd)
1069                                 bd->bd_bh = NULL;
1070                         bh->b_private = NULL;
1071                 }
1072                 gfs2_log_unlock(sdp);
1073                 if (bd)
1074                         kmem_cache_free(gfs2_bufdata_cachep, bd);
1075
1076                 bh = bh->b_this_page;
1077         } while (bh != head);
1078
1079         return try_to_free_buffers(page);
1080 cannot_release:
1081         gfs2_log_unlock(sdp);
1082         return 0;
1083 }
1084
1085 static const struct address_space_operations gfs2_writeback_aops = {
1086         .writepage = gfs2_writeback_writepage,
1087         .writepages = gfs2_writeback_writepages,
1088         .readpage = gfs2_readpage,
1089         .readpages = gfs2_readpages,
1090         .sync_page = block_sync_page,
1091         .write_begin = gfs2_write_begin,
1092         .write_end = gfs2_write_end,
1093         .bmap = gfs2_bmap,
1094         .invalidatepage = gfs2_invalidatepage,
1095         .releasepage = gfs2_releasepage,
1096         .direct_IO = gfs2_direct_IO,
1097         .migratepage = buffer_migrate_page,
1098 };
1099
1100 static const struct address_space_operations gfs2_ordered_aops = {
1101         .writepage = gfs2_ordered_writepage,
1102         .readpage = gfs2_readpage,
1103         .readpages = gfs2_readpages,
1104         .sync_page = block_sync_page,
1105         .write_begin = gfs2_write_begin,
1106         .write_end = gfs2_write_end,
1107         .set_page_dirty = gfs2_set_page_dirty,
1108         .bmap = gfs2_bmap,
1109         .invalidatepage = gfs2_invalidatepage,
1110         .releasepage = gfs2_releasepage,
1111         .direct_IO = gfs2_direct_IO,
1112         .migratepage = buffer_migrate_page,
1113 };
1114
1115 static const struct address_space_operations gfs2_jdata_aops = {
1116         .writepage = gfs2_jdata_writepage,
1117         .writepages = gfs2_jdata_writepages,
1118         .readpage = gfs2_readpage,
1119         .readpages = gfs2_readpages,
1120         .sync_page = block_sync_page,
1121         .write_begin = gfs2_write_begin,
1122         .write_end = gfs2_write_end,
1123         .set_page_dirty = gfs2_set_page_dirty,
1124         .bmap = gfs2_bmap,
1125         .invalidatepage = gfs2_invalidatepage,
1126         .releasepage = gfs2_releasepage,
1127 };
1128
1129 void gfs2_set_aops(struct inode *inode)
1130 {
1131         struct gfs2_inode *ip = GFS2_I(inode);
1132
1133         if (gfs2_is_writeback(ip))
1134                 inode->i_mapping->a_ops = &gfs2_writeback_aops;
1135         else if (gfs2_is_ordered(ip))
1136                 inode->i_mapping->a_ops = &gfs2_ordered_aops;
1137         else if (gfs2_is_jdata(ip))
1138                 inode->i_mapping->a_ops = &gfs2_jdata_aops;
1139         else
1140                 BUG();
1141 }
1142