Merge branch 'for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jmorris...
[linux-2.6] / fs / ocfs2 / aops.c
1 /* -*- mode: c; c-basic-offset: 8; -*-
2  * vim: noexpandtab sw=8 ts=8 sts=0:
3  *
4  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
5  *
6  * This program is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public
8  * License as published by the Free Software Foundation; either
9  * version 2 of the License, or (at your option) any later version.
10  *
11  * This program is distributed in the hope that it will be useful,
12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
14  * General Public License for more details.
15  *
16  * You should have received a copy of the GNU General Public
17  * License along with this program; if not, write to the
18  * Free Software Foundation, Inc., 59 Temple Place - Suite 330,
19  * Boston, MA 021110-1307, USA.
20  */
21
22 #include <linux/fs.h>
23 #include <linux/slab.h>
24 #include <linux/highmem.h>
25 #include <linux/pagemap.h>
26 #include <asm/byteorder.h>
27 #include <linux/swap.h>
28 #include <linux/pipe_fs_i.h>
29
30 #define MLOG_MASK_PREFIX ML_FILE_IO
31 #include <cluster/masklog.h>
32
33 #include "ocfs2.h"
34
35 #include "alloc.h"
36 #include "aops.h"
37 #include "dlmglue.h"
38 #include "extent_map.h"
39 #include "file.h"
40 #include "inode.h"
41 #include "journal.h"
42 #include "suballoc.h"
43 #include "super.h"
44 #include "symlink.h"
45
46 #include "buffer_head_io.h"
47
48 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
49                                    struct buffer_head *bh_result, int create)
50 {
51         int err = -EIO;
52         int status;
53         struct ocfs2_dinode *fe = NULL;
54         struct buffer_head *bh = NULL;
55         struct buffer_head *buffer_cache_bh = NULL;
56         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
57         void *kaddr;
58
59         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
60                    (unsigned long long)iblock, bh_result, create);
61
62         BUG_ON(ocfs2_inode_is_fast_symlink(inode));
63
64         if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
65                 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
66                      (unsigned long long)iblock);
67                 goto bail;
68         }
69
70         status = ocfs2_read_block(OCFS2_SB(inode->i_sb),
71                                   OCFS2_I(inode)->ip_blkno,
72                                   &bh, OCFS2_BH_CACHED, inode);
73         if (status < 0) {
74                 mlog_errno(status);
75                 goto bail;
76         }
77         fe = (struct ocfs2_dinode *) bh->b_data;
78
79         if (!OCFS2_IS_VALID_DINODE(fe)) {
80                 mlog(ML_ERROR, "Invalid dinode #%llu: signature = %.*s\n",
81                      (unsigned long long)fe->i_blkno, 7, fe->i_signature);
82                 goto bail;
83         }
84
85         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
86                                                     le32_to_cpu(fe->i_clusters))) {
87                 mlog(ML_ERROR, "block offset is outside the allocated size: "
88                      "%llu\n", (unsigned long long)iblock);
89                 goto bail;
90         }
91
92         /* We don't use the page cache to create symlink data, so if
93          * need be, copy it over from the buffer cache. */
94         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
95                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
96                             iblock;
97                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
98                 if (!buffer_cache_bh) {
99                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
100                         goto bail;
101                 }
102
103                 /* we haven't locked out transactions, so a commit
104                  * could've happened. Since we've got a reference on
105                  * the bh, even if it commits while we're doing the
106                  * copy, the data is still good. */
107                 if (buffer_jbd(buffer_cache_bh)
108                     && ocfs2_inode_is_new(inode)) {
109                         kaddr = kmap_atomic(bh_result->b_page, KM_USER0);
110                         if (!kaddr) {
111                                 mlog(ML_ERROR, "couldn't kmap!\n");
112                                 goto bail;
113                         }
114                         memcpy(kaddr + (bh_result->b_size * iblock),
115                                buffer_cache_bh->b_data,
116                                bh_result->b_size);
117                         kunmap_atomic(kaddr, KM_USER0);
118                         set_buffer_uptodate(bh_result);
119                 }
120                 brelse(buffer_cache_bh);
121         }
122
123         map_bh(bh_result, inode->i_sb,
124                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
125
126         err = 0;
127
128 bail:
129         if (bh)
130                 brelse(bh);
131
132         mlog_exit(err);
133         return err;
134 }
135
136 static int ocfs2_get_block(struct inode *inode, sector_t iblock,
137                            struct buffer_head *bh_result, int create)
138 {
139         int err = 0;
140         unsigned int ext_flags;
141         u64 p_blkno, past_eof;
142         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
143
144         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
145                    (unsigned long long)iblock, bh_result, create);
146
147         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
148                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
149                      inode, inode->i_ino);
150
151         if (S_ISLNK(inode->i_mode)) {
152                 /* this always does I/O for some reason. */
153                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
154                 goto bail;
155         }
156
157         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, NULL,
158                                           &ext_flags);
159         if (err) {
160                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
161                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
162                      (unsigned long long)p_blkno);
163                 goto bail;
164         }
165
166         /*
167          * ocfs2 never allocates in this function - the only time we
168          * need to use BH_New is when we're extending i_size on a file
169          * system which doesn't support holes, in which case BH_New
170          * allows block_prepare_write() to zero.
171          */
172         mlog_bug_on_msg(create && p_blkno == 0 && ocfs2_sparse_alloc(osb),
173                         "ino %lu, iblock %llu\n", inode->i_ino,
174                         (unsigned long long)iblock);
175
176         /* Treat the unwritten extent as a hole for zeroing purposes. */
177         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
178                 map_bh(bh_result, inode->i_sb, p_blkno);
179
180         if (!ocfs2_sparse_alloc(osb)) {
181                 if (p_blkno == 0) {
182                         err = -EIO;
183                         mlog(ML_ERROR,
184                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
185                              (unsigned long long)iblock,
186                              (unsigned long long)p_blkno,
187                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
188                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
189                         dump_stack();
190                 }
191
192                 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
193                 mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino,
194                      (unsigned long long)past_eof);
195
196                 if (create && (iblock >= past_eof))
197                         set_buffer_new(bh_result);
198         }
199
200 bail:
201         if (err < 0)
202                 err = -EIO;
203
204         mlog_exit(err);
205         return err;
206 }
207
208 static int ocfs2_readpage(struct file *file, struct page *page)
209 {
210         struct inode *inode = page->mapping->host;
211         loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
212         int ret, unlock = 1;
213
214         mlog_entry("(0x%p, %lu)\n", file, (page ? page->index : 0));
215
216         ret = ocfs2_meta_lock_with_page(inode, NULL, 0, page);
217         if (ret != 0) {
218                 if (ret == AOP_TRUNCATED_PAGE)
219                         unlock = 0;
220                 mlog_errno(ret);
221                 goto out;
222         }
223
224         down_read(&OCFS2_I(inode)->ip_alloc_sem);
225
226         /*
227          * i_size might have just been updated as we grabed the meta lock.  We
228          * might now be discovering a truncate that hit on another node.
229          * block_read_full_page->get_block freaks out if it is asked to read
230          * beyond the end of a file, so we check here.  Callers
231          * (generic_file_read, fault->nopage) are clever enough to check i_size
232          * and notice that the page they just read isn't needed.
233          *
234          * XXX sys_readahead() seems to get that wrong?
235          */
236         if (start >= i_size_read(inode)) {
237                 char *addr = kmap(page);
238                 memset(addr, 0, PAGE_SIZE);
239                 flush_dcache_page(page);
240                 kunmap(page);
241                 SetPageUptodate(page);
242                 ret = 0;
243                 goto out_alloc;
244         }
245
246         ret = ocfs2_data_lock_with_page(inode, 0, page);
247         if (ret != 0) {
248                 if (ret == AOP_TRUNCATED_PAGE)
249                         unlock = 0;
250                 mlog_errno(ret);
251                 goto out_alloc;
252         }
253
254         ret = block_read_full_page(page, ocfs2_get_block);
255         unlock = 0;
256
257         ocfs2_data_unlock(inode, 0);
258 out_alloc:
259         up_read(&OCFS2_I(inode)->ip_alloc_sem);
260         ocfs2_meta_unlock(inode, 0);
261 out:
262         if (unlock)
263                 unlock_page(page);
264         mlog_exit(ret);
265         return ret;
266 }
267
268 /* Note: Because we don't support holes, our allocation has
269  * already happened (allocation writes zeros to the file data)
270  * so we don't have to worry about ordered writes in
271  * ocfs2_writepage.
272  *
273  * ->writepage is called during the process of invalidating the page cache
274  * during blocked lock processing.  It can't block on any cluster locks
275  * to during block mapping.  It's relying on the fact that the block
276  * mapping can't have disappeared under the dirty pages that it is
277  * being asked to write back.
278  */
279 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
280 {
281         int ret;
282
283         mlog_entry("(0x%p)\n", page);
284
285         ret = block_write_full_page(page, ocfs2_get_block, wbc);
286
287         mlog_exit(ret);
288
289         return ret;
290 }
291
292 /*
293  * This is called from ocfs2_write_zero_page() which has handled it's
294  * own cluster locking and has ensured allocation exists for those
295  * blocks to be written.
296  */
297 int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
298                                unsigned from, unsigned to)
299 {
300         int ret;
301
302         down_read(&OCFS2_I(inode)->ip_alloc_sem);
303
304         ret = block_prepare_write(page, from, to, ocfs2_get_block);
305
306         up_read(&OCFS2_I(inode)->ip_alloc_sem);
307
308         return ret;
309 }
310
311 /* Taken from ext3. We don't necessarily need the full blown
312  * functionality yet, but IMHO it's better to cut and paste the whole
313  * thing so we can avoid introducing our own bugs (and easily pick up
314  * their fixes when they happen) --Mark */
315 int walk_page_buffers(  handle_t *handle,
316                         struct buffer_head *head,
317                         unsigned from,
318                         unsigned to,
319                         int *partial,
320                         int (*fn)(      handle_t *handle,
321                                         struct buffer_head *bh))
322 {
323         struct buffer_head *bh;
324         unsigned block_start, block_end;
325         unsigned blocksize = head->b_size;
326         int err, ret = 0;
327         struct buffer_head *next;
328
329         for (   bh = head, block_start = 0;
330                 ret == 0 && (bh != head || !block_start);
331                 block_start = block_end, bh = next)
332         {
333                 next = bh->b_this_page;
334                 block_end = block_start + blocksize;
335                 if (block_end <= from || block_start >= to) {
336                         if (partial && !buffer_uptodate(bh))
337                                 *partial = 1;
338                         continue;
339                 }
340                 err = (*fn)(handle, bh);
341                 if (!ret)
342                         ret = err;
343         }
344         return ret;
345 }
346
347 handle_t *ocfs2_start_walk_page_trans(struct inode *inode,
348                                                          struct page *page,
349                                                          unsigned from,
350                                                          unsigned to)
351 {
352         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
353         handle_t *handle = NULL;
354         int ret = 0;
355
356         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
357         if (!handle) {
358                 ret = -ENOMEM;
359                 mlog_errno(ret);
360                 goto out;
361         }
362
363         if (ocfs2_should_order_data(inode)) {
364                 ret = walk_page_buffers(handle,
365                                         page_buffers(page),
366                                         from, to, NULL,
367                                         ocfs2_journal_dirty_data);
368                 if (ret < 0) 
369                         mlog_errno(ret);
370         }
371 out:
372         if (ret) {
373                 if (handle)
374                         ocfs2_commit_trans(osb, handle);
375                 handle = ERR_PTR(ret);
376         }
377         return handle;
378 }
379
380 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
381 {
382         sector_t status;
383         u64 p_blkno = 0;
384         int err = 0;
385         struct inode *inode = mapping->host;
386
387         mlog_entry("(block = %llu)\n", (unsigned long long)block);
388
389         /* We don't need to lock journal system files, since they aren't
390          * accessed concurrently from multiple nodes.
391          */
392         if (!INODE_JOURNAL(inode)) {
393                 err = ocfs2_meta_lock(inode, NULL, 0);
394                 if (err) {
395                         if (err != -ENOENT)
396                                 mlog_errno(err);
397                         goto bail;
398                 }
399                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
400         }
401
402         err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL, NULL);
403
404         if (!INODE_JOURNAL(inode)) {
405                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
406                 ocfs2_meta_unlock(inode, 0);
407         }
408
409         if (err) {
410                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
411                      (unsigned long long)block);
412                 mlog_errno(err);
413                 goto bail;
414         }
415
416
417 bail:
418         status = err ? 0 : p_blkno;
419
420         mlog_exit((int)status);
421
422         return status;
423 }
424
425 /*
426  * TODO: Make this into a generic get_blocks function.
427  *
428  * From do_direct_io in direct-io.c:
429  *  "So what we do is to permit the ->get_blocks function to populate
430  *   bh.b_size with the size of IO which is permitted at this offset and
431  *   this i_blkbits."
432  *
433  * This function is called directly from get_more_blocks in direct-io.c.
434  *
435  * called like this: dio->get_blocks(dio->inode, fs_startblk,
436  *                                      fs_count, map_bh, dio->rw == WRITE);
437  */
438 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
439                                      struct buffer_head *bh_result, int create)
440 {
441         int ret;
442         u64 p_blkno, inode_blocks, contig_blocks;
443         unsigned int ext_flags;
444         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
445         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
446
447         /* This function won't even be called if the request isn't all
448          * nicely aligned and of the right size, so there's no need
449          * for us to check any of that. */
450
451         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
452
453         /*
454          * Any write past EOF is not allowed because we'd be extending.
455          */
456         if (create && (iblock + max_blocks) > inode_blocks) {
457                 ret = -EIO;
458                 goto bail;
459         }
460
461         /* This figures out the size of the next contiguous block, and
462          * our logical offset */
463         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
464                                           &contig_blocks, &ext_flags);
465         if (ret) {
466                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
467                      (unsigned long long)iblock);
468                 ret = -EIO;
469                 goto bail;
470         }
471
472         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)) && !p_blkno) {
473                 ocfs2_error(inode->i_sb,
474                             "Inode %llu has a hole at block %llu\n",
475                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
476                             (unsigned long long)iblock);
477                 ret = -EROFS;
478                 goto bail;
479         }
480
481         /*
482          * get_more_blocks() expects us to describe a hole by clearing
483          * the mapped bit on bh_result().
484          *
485          * Consider an unwritten extent as a hole.
486          */
487         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
488                 map_bh(bh_result, inode->i_sb, p_blkno);
489         else {
490                 /*
491                  * ocfs2_prepare_inode_for_write() should have caught
492                  * the case where we'd be filling a hole and triggered
493                  * a buffered write instead.
494                  */
495                 if (create) {
496                         ret = -EIO;
497                         mlog_errno(ret);
498                         goto bail;
499                 }
500
501                 clear_buffer_mapped(bh_result);
502         }
503
504         /* make sure we don't map more than max_blocks blocks here as
505            that's all the kernel will handle at this point. */
506         if (max_blocks < contig_blocks)
507                 contig_blocks = max_blocks;
508         bh_result->b_size = contig_blocks << blocksize_bits;
509 bail:
510         return ret;
511 }
512
513 /* 
514  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
515  * particularly interested in the aio/dio case.  Like the core uses
516  * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from
517  * truncation on another.
518  */
519 static void ocfs2_dio_end_io(struct kiocb *iocb,
520                              loff_t offset,
521                              ssize_t bytes,
522                              void *private)
523 {
524         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
525         int level;
526
527         /* this io's submitter should not have unlocked this before we could */
528         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
529
530         ocfs2_iocb_clear_rw_locked(iocb);
531
532         level = ocfs2_iocb_rw_locked_level(iocb);
533         if (!level)
534                 up_read(&inode->i_alloc_sem);
535         ocfs2_rw_unlock(inode, level);
536 }
537
538 /*
539  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
540  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
541  * do journalled data.
542  */
543 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
544 {
545         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
546
547         journal_invalidatepage(journal, page, offset);
548 }
549
550 static int ocfs2_releasepage(struct page *page, gfp_t wait)
551 {
552         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
553
554         if (!page_has_buffers(page))
555                 return 0;
556         return journal_try_to_free_buffers(journal, page, wait);
557 }
558
559 static ssize_t ocfs2_direct_IO(int rw,
560                                struct kiocb *iocb,
561                                const struct iovec *iov,
562                                loff_t offset,
563                                unsigned long nr_segs)
564 {
565         struct file *file = iocb->ki_filp;
566         struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
567         int ret;
568
569         mlog_entry_void();
570
571         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
572                 /*
573                  * We get PR data locks even for O_DIRECT.  This
574                  * allows concurrent O_DIRECT I/O but doesn't let
575                  * O_DIRECT with extending and buffered zeroing writes
576                  * race.  If they did race then the buffered zeroing
577                  * could be written back after the O_DIRECT I/O.  It's
578                  * one thing to tell people not to mix buffered and
579                  * O_DIRECT writes, but expecting them to understand
580                  * that file extension is also an implicit buffered
581                  * write is too much.  By getting the PR we force
582                  * writeback of the buffered zeroing before
583                  * proceeding.
584                  */
585                 ret = ocfs2_data_lock(inode, 0);
586                 if (ret < 0) {
587                         mlog_errno(ret);
588                         goto out;
589                 }
590                 ocfs2_data_unlock(inode, 0);
591         }
592
593         ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
594                                             inode->i_sb->s_bdev, iov, offset,
595                                             nr_segs, 
596                                             ocfs2_direct_IO_get_blocks,
597                                             ocfs2_dio_end_io);
598 out:
599         mlog_exit(ret);
600         return ret;
601 }
602
603 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
604                                             u32 cpos,
605                                             unsigned int *start,
606                                             unsigned int *end)
607 {
608         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
609
610         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
611                 unsigned int cpp;
612
613                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
614
615                 cluster_start = cpos % cpp;
616                 cluster_start = cluster_start << osb->s_clustersize_bits;
617
618                 cluster_end = cluster_start + osb->s_clustersize;
619         }
620
621         BUG_ON(cluster_start > PAGE_SIZE);
622         BUG_ON(cluster_end > PAGE_SIZE);
623
624         if (start)
625                 *start = cluster_start;
626         if (end)
627                 *end = cluster_end;
628 }
629
630 /*
631  * 'from' and 'to' are the region in the page to avoid zeroing.
632  *
633  * If pagesize > clustersize, this function will avoid zeroing outside
634  * of the cluster boundary.
635  *
636  * from == to == 0 is code for "zero the entire cluster region"
637  */
638 static void ocfs2_clear_page_regions(struct page *page,
639                                      struct ocfs2_super *osb, u32 cpos,
640                                      unsigned from, unsigned to)
641 {
642         void *kaddr;
643         unsigned int cluster_start, cluster_end;
644
645         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
646
647         kaddr = kmap_atomic(page, KM_USER0);
648
649         if (from || to) {
650                 if (from > cluster_start)
651                         memset(kaddr + cluster_start, 0, from - cluster_start);
652                 if (to < cluster_end)
653                         memset(kaddr + to, 0, cluster_end - to);
654         } else {
655                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
656         }
657
658         kunmap_atomic(kaddr, KM_USER0);
659 }
660
661 /*
662  * Some of this taken from block_prepare_write(). We already have our
663  * mapping by now though, and the entire write will be allocating or
664  * it won't, so not much need to use BH_New.
665  *
666  * This will also skip zeroing, which is handled externally.
667  */
668 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
669                           struct inode *inode, unsigned int from,
670                           unsigned int to, int new)
671 {
672         int ret = 0;
673         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
674         unsigned int block_end, block_start;
675         unsigned int bsize = 1 << inode->i_blkbits;
676
677         if (!page_has_buffers(page))
678                 create_empty_buffers(page, bsize, 0);
679
680         head = page_buffers(page);
681         for (bh = head, block_start = 0; bh != head || !block_start;
682              bh = bh->b_this_page, block_start += bsize) {
683                 block_end = block_start + bsize;
684
685                 /*
686                  * Ignore blocks outside of our i/o range -
687                  * they may belong to unallocated clusters.
688                  */
689                 if (block_start >= to || block_end <= from) {
690                         if (PageUptodate(page))
691                                 set_buffer_uptodate(bh);
692                         continue;
693                 }
694
695                 /*
696                  * For an allocating write with cluster size >= page
697                  * size, we always write the entire page.
698                  */
699
700                 if (buffer_new(bh))
701                         clear_buffer_new(bh);
702
703                 if (!buffer_mapped(bh)) {
704                         map_bh(bh, inode->i_sb, *p_blkno);
705                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
706                 }
707
708                 if (PageUptodate(page)) {
709                         if (!buffer_uptodate(bh))
710                                 set_buffer_uptodate(bh);
711                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
712                      (block_start < from || block_end > to)) {
713                         ll_rw_block(READ, 1, &bh);
714                         *wait_bh++=bh;
715                 }
716
717                 *p_blkno = *p_blkno + 1;
718         }
719
720         /*
721          * If we issued read requests - let them complete.
722          */
723         while(wait_bh > wait) {
724                 wait_on_buffer(*--wait_bh);
725                 if (!buffer_uptodate(*wait_bh))
726                         ret = -EIO;
727         }
728
729         if (ret == 0 || !new)
730                 return ret;
731
732         /*
733          * If we get -EIO above, zero out any newly allocated blocks
734          * to avoid exposing stale data.
735          */
736         bh = head;
737         block_start = 0;
738         do {
739                 void *kaddr;
740
741                 block_end = block_start + bsize;
742                 if (block_end <= from)
743                         goto next_bh;
744                 if (block_start >= to)
745                         break;
746
747                 kaddr = kmap_atomic(page, KM_USER0);
748                 memset(kaddr+block_start, 0, bh->b_size);
749                 flush_dcache_page(page);
750                 kunmap_atomic(kaddr, KM_USER0);
751                 set_buffer_uptodate(bh);
752                 mark_buffer_dirty(bh);
753
754 next_bh:
755                 block_start = block_end;
756                 bh = bh->b_this_page;
757         } while (bh != head);
758
759         return ret;
760 }
761
762 /*
763  * This will copy user data from the buffer page in the splice
764  * context.
765  *
766  * For now, we ignore SPLICE_F_MOVE as that would require some extra
767  * communication out all the way to ocfs2_write().
768  */
769 int ocfs2_map_and_write_splice_data(struct inode *inode,
770                                   struct ocfs2_write_ctxt *wc, u64 *p_blkno,
771                                   unsigned int *ret_from, unsigned int *ret_to)
772 {
773         int ret;
774         unsigned int to, from, cluster_start, cluster_end;
775         char *src, *dst;
776         struct ocfs2_splice_write_priv *sp = wc->w_private;
777         struct pipe_buffer *buf = sp->s_buf;
778         unsigned long bytes, src_from;
779         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
780
781         ocfs2_figure_cluster_boundaries(osb, wc->w_cpos, &cluster_start,
782                                         &cluster_end);
783
784         from = sp->s_offset;
785         src_from = sp->s_buf_offset;
786         bytes = wc->w_count;
787
788         if (wc->w_large_pages) {
789                 /*
790                  * For cluster size < page size, we have to
791                  * calculate pos within the cluster and obey
792                  * the rightmost boundary.
793                  */
794                 bytes = min(bytes, (unsigned long)(osb->s_clustersize
795                                    - (wc->w_pos & (osb->s_clustersize - 1))));
796         }
797         to = from + bytes;
798
799         if (wc->w_this_page_new)
800                 ret = ocfs2_map_page_blocks(wc->w_this_page, p_blkno, inode,
801                                             cluster_start, cluster_end, 1);
802         else
803                 ret = ocfs2_map_page_blocks(wc->w_this_page, p_blkno, inode,
804                                             from, to, 0);
805         if (ret) {
806                 mlog_errno(ret);
807                 goto out;
808         }
809
810         BUG_ON(from > PAGE_CACHE_SIZE);
811         BUG_ON(to > PAGE_CACHE_SIZE);
812         BUG_ON(from > osb->s_clustersize);
813         BUG_ON(to > osb->s_clustersize);
814
815         src = buf->ops->map(sp->s_pipe, buf, 1);
816         dst = kmap_atomic(wc->w_this_page, KM_USER1);
817         memcpy(dst + from, src + src_from, bytes);
818         kunmap_atomic(wc->w_this_page, KM_USER1);
819         buf->ops->unmap(sp->s_pipe, buf, src);
820
821         wc->w_finished_copy = 1;
822
823         *ret_from = from;
824         *ret_to = to;
825 out:
826
827         return bytes ? (unsigned int)bytes : ret;
828 }
829
830 /*
831  * This will copy user data from the iovec in the buffered write
832  * context.
833  */
834 int ocfs2_map_and_write_user_data(struct inode *inode,
835                                   struct ocfs2_write_ctxt *wc, u64 *p_blkno,
836                                   unsigned int *ret_from, unsigned int *ret_to)
837 {
838         int ret;
839         unsigned int to, from, cluster_start, cluster_end;
840         unsigned long bytes, src_from;
841         char *dst;
842         struct ocfs2_buffered_write_priv *bp = wc->w_private;
843         const struct iovec *cur_iov = bp->b_cur_iov;
844         char __user *buf;
845         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
846
847         ocfs2_figure_cluster_boundaries(osb, wc->w_cpos, &cluster_start,
848                                         &cluster_end);
849
850         buf = cur_iov->iov_base + bp->b_cur_off;
851         src_from = (unsigned long)buf & ~PAGE_CACHE_MASK;
852
853         from = wc->w_pos & (PAGE_CACHE_SIZE - 1);
854
855         /*
856          * This is a lot of comparisons, but it reads quite
857          * easily, which is important here.
858          */
859         /* Stay within the src page */
860         bytes = PAGE_SIZE - src_from;
861         /* Stay within the vector */
862         bytes = min(bytes,
863                     (unsigned long)(cur_iov->iov_len - bp->b_cur_off));
864         /* Stay within count */
865         bytes = min(bytes, (unsigned long)wc->w_count);
866         /*
867          * For clustersize > page size, just stay within
868          * target page, otherwise we have to calculate pos
869          * within the cluster and obey the rightmost
870          * boundary.
871          */
872         if (wc->w_large_pages) {
873                 /*
874                  * For cluster size < page size, we have to
875                  * calculate pos within the cluster and obey
876                  * the rightmost boundary.
877                  */
878                 bytes = min(bytes, (unsigned long)(osb->s_clustersize
879                                    - (wc->w_pos & (osb->s_clustersize - 1))));
880         } else {
881                 /*
882                  * cluster size > page size is the most common
883                  * case - we just stay within the target page
884                  * boundary.
885                  */
886                 bytes = min(bytes, PAGE_CACHE_SIZE - from);
887         }
888
889         to = from + bytes;
890
891         if (wc->w_this_page_new)
892                 ret = ocfs2_map_page_blocks(wc->w_this_page, p_blkno, inode,
893                                             cluster_start, cluster_end, 1);
894         else
895                 ret = ocfs2_map_page_blocks(wc->w_this_page, p_blkno, inode,
896                                             from, to, 0);
897         if (ret) {
898                 mlog_errno(ret);
899                 goto out;
900         }
901
902         BUG_ON(from > PAGE_CACHE_SIZE);
903         BUG_ON(to > PAGE_CACHE_SIZE);
904         BUG_ON(from > osb->s_clustersize);
905         BUG_ON(to > osb->s_clustersize);
906
907         dst = kmap(wc->w_this_page);
908         memcpy(dst + from, bp->b_src_buf + src_from, bytes);
909         kunmap(wc->w_this_page);
910
911         /*
912          * XXX: This is slow, but simple. The caller of
913          * ocfs2_buffered_write_cluster() is responsible for
914          * passing through the iovecs, so it's difficult to
915          * predict what our next step is in here after our
916          * initial write. A future version should be pushing
917          * that iovec manipulation further down.
918          *
919          * By setting this, we indicate that a copy from user
920          * data was done, and subsequent calls for this
921          * cluster will skip copying more data.
922          */
923         wc->w_finished_copy = 1;
924
925         *ret_from = from;
926         *ret_to = to;
927 out:
928
929         return bytes ? (unsigned int)bytes : ret;
930 }
931
932 /*
933  * Map, fill and write a page to disk.
934  *
935  * The work of copying data is done via callback.  Newly allocated
936  * pages which don't take user data will be zero'd (set 'new' to
937  * indicate an allocating write)
938  *
939  * Returns a negative error code or the number of bytes copied into
940  * the page.
941  */
942 int ocfs2_write_data_page(struct inode *inode, handle_t *handle,
943                           u64 *p_blkno, struct page *page,
944                           struct ocfs2_write_ctxt *wc, int new)
945 {
946         int ret, copied = 0;
947         unsigned int from = 0, to = 0;
948         unsigned int cluster_start, cluster_end;
949         unsigned int zero_from = 0, zero_to = 0;
950
951         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), wc->w_cpos,
952                                         &cluster_start, &cluster_end);
953
954         if ((wc->w_pos >> PAGE_CACHE_SHIFT) == page->index
955             && !wc->w_finished_copy) {
956
957                 wc->w_this_page = page;
958                 wc->w_this_page_new = new;
959                 ret = wc->w_write_data_page(inode, wc, p_blkno, &from, &to);
960                 if (ret < 0) {
961                         mlog_errno(ret);
962                         goto out;
963                 }
964
965                 copied = ret;
966
967                 zero_from = from;
968                 zero_to = to;
969                 if (new) {
970                         from = cluster_start;
971                         to = cluster_end;
972                 }
973         } else {
974                 /*
975                  * If we haven't allocated the new page yet, we
976                  * shouldn't be writing it out without copying user
977                  * data. This is likely a math error from the caller.
978                  */
979                 BUG_ON(!new);
980
981                 from = cluster_start;
982                 to = cluster_end;
983
984                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
985                                             cluster_start, cluster_end, 1);
986                 if (ret) {
987                         mlog_errno(ret);
988                         goto out;
989                 }
990         }
991
992         /*
993          * Parts of newly allocated pages need to be zero'd.
994          *
995          * Above, we have also rewritten 'to' and 'from' - as far as
996          * the rest of the function is concerned, the entire cluster
997          * range inside of a page needs to be written.
998          *
999          * We can skip this if the page is up to date - it's already
1000          * been zero'd from being read in as a hole.
1001          */
1002         if (new && !PageUptodate(page))
1003                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1004                                          wc->w_cpos, zero_from, zero_to);
1005
1006         flush_dcache_page(page);
1007
1008         if (ocfs2_should_order_data(inode)) {
1009                 ret = walk_page_buffers(handle,
1010                                         page_buffers(page),
1011                                         from, to, NULL,
1012                                         ocfs2_journal_dirty_data);
1013                 if (ret < 0)
1014                         mlog_errno(ret);
1015         }
1016
1017         /*
1018          * We don't use generic_commit_write() because we need to
1019          * handle our own i_size update.
1020          */
1021         ret = block_commit_write(page, from, to);
1022         if (ret)
1023                 mlog_errno(ret);
1024 out:
1025
1026         return copied ? copied : ret;
1027 }
1028
1029 /*
1030  * Do the actual write of some data into an inode. Optionally allocate
1031  * in order to fulfill the write.
1032  *
1033  * cpos is the logical cluster offset within the file to write at
1034  *
1035  * 'phys' is the physical mapping of that offset. a 'phys' value of
1036  * zero indicates that allocation is required. In this case, data_ac
1037  * and meta_ac should be valid (meta_ac can be null if metadata
1038  * allocation isn't required).
1039  */
1040 static ssize_t ocfs2_write(struct file *file, u32 phys, handle_t *handle,
1041                            struct buffer_head *di_bh,
1042                            struct ocfs2_alloc_context *data_ac,
1043                            struct ocfs2_alloc_context *meta_ac,
1044                            struct ocfs2_write_ctxt *wc)
1045 {
1046         int ret, i, numpages = 1, new;
1047         unsigned int copied = 0;
1048         u32 tmp_pos;
1049         u64 v_blkno, p_blkno;
1050         struct address_space *mapping = file->f_mapping;
1051         struct inode *inode = mapping->host;
1052         unsigned long index, start;
1053         struct page **cpages;
1054
1055         new = phys == 0 ? 1 : 0;
1056
1057         /*
1058          * Figure out how many pages we'll be manipulating here. For
1059          * non allocating write, we just change the one
1060          * page. Otherwise, we'll need a whole clusters worth.
1061          */
1062         if (new)
1063                 numpages = ocfs2_pages_per_cluster(inode->i_sb);
1064
1065         cpages = kzalloc(sizeof(*cpages) * numpages, GFP_NOFS);
1066         if (!cpages) {
1067                 ret = -ENOMEM;
1068                 mlog_errno(ret);
1069                 return ret;
1070         }
1071
1072         /*
1073          * Fill our page array first. That way we've grabbed enough so
1074          * that we can zero and flush if we error after adding the
1075          * extent.
1076          */
1077         if (new) {
1078                 start = ocfs2_align_clusters_to_page_index(inode->i_sb,
1079                                                            wc->w_cpos);
1080                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, wc->w_cpos);
1081         } else {
1082                 start = wc->w_pos >> PAGE_CACHE_SHIFT;
1083                 v_blkno = wc->w_pos >> inode->i_sb->s_blocksize_bits;
1084         }
1085
1086         for(i = 0; i < numpages; i++) {
1087                 index = start + i;
1088
1089                 cpages[i] = grab_cache_page(mapping, index);
1090                 if (!cpages[i]) {
1091                         ret = -ENOMEM;
1092                         mlog_errno(ret);
1093                         goto out;
1094                 }
1095         }
1096
1097         if (new) {
1098                 /*
1099                  * This is safe to call with the page locks - it won't take
1100                  * any additional semaphores or cluster locks.
1101                  */
1102                 tmp_pos = wc->w_cpos;
1103                 ret = ocfs2_do_extend_allocation(OCFS2_SB(inode->i_sb), inode,
1104                                                  &tmp_pos, 1, di_bh, handle,
1105                                                  data_ac, meta_ac, NULL);
1106                 /*
1107                  * This shouldn't happen because we must have already
1108                  * calculated the correct meta data allocation required. The
1109                  * internal tree allocation code should know how to increase
1110                  * transaction credits itself.
1111                  *
1112                  * If need be, we could handle -EAGAIN for a
1113                  * RESTART_TRANS here.
1114                  */
1115                 mlog_bug_on_msg(ret == -EAGAIN,
1116                                 "Inode %llu: EAGAIN return during allocation.\n",
1117                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1118                 if (ret < 0) {
1119                         mlog_errno(ret);
1120                         goto out;
1121                 }
1122         }
1123
1124         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1125                                           NULL);
1126         if (ret < 0) {
1127
1128                 /*
1129                  * XXX: Should we go readonly here?
1130                  */
1131
1132                 mlog_errno(ret);
1133                 goto out;
1134         }
1135
1136         BUG_ON(p_blkno == 0);
1137
1138         for(i = 0; i < numpages; i++) {
1139                 ret = ocfs2_write_data_page(inode, handle, &p_blkno, cpages[i],
1140                                             wc, new);
1141                 if (ret < 0) {
1142                         mlog_errno(ret);
1143                         goto out;
1144                 }
1145
1146                 copied += ret;
1147         }
1148
1149 out:
1150         for(i = 0; i < numpages; i++) {
1151                 unlock_page(cpages[i]);
1152                 mark_page_accessed(cpages[i]);
1153                 page_cache_release(cpages[i]);
1154         }
1155         kfree(cpages);
1156
1157         return copied ? copied : ret;
1158 }
1159
1160 static void ocfs2_write_ctxt_init(struct ocfs2_write_ctxt *wc,
1161                                   struct ocfs2_super *osb, loff_t pos,
1162                                   size_t count, ocfs2_page_writer *cb,
1163                                   void *cb_priv)
1164 {
1165         wc->w_count = count;
1166         wc->w_pos = pos;
1167         wc->w_cpos = wc->w_pos >> osb->s_clustersize_bits;
1168         wc->w_finished_copy = 0;
1169
1170         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
1171                 wc->w_large_pages = 1;
1172         else
1173                 wc->w_large_pages = 0;
1174
1175         wc->w_write_data_page = cb;
1176         wc->w_private = cb_priv;
1177 }
1178
1179 /*
1180  * Write a cluster to an inode. The cluster may not be allocated yet,
1181  * in which case it will be. This only exists for buffered writes -
1182  * O_DIRECT takes a more "traditional" path through the kernel.
1183  *
1184  * The caller is responsible for incrementing pos, written counts, etc
1185  *
1186  * For file systems that don't support sparse files, pre-allocation
1187  * and page zeroing up until cpos should be done prior to this
1188  * function call.
1189  *
1190  * Callers should be holding i_sem, and the rw cluster lock.
1191  *
1192  * Returns the number of user bytes written, or less than zero for
1193  * error.
1194  */
1195 ssize_t ocfs2_buffered_write_cluster(struct file *file, loff_t pos,
1196                                      size_t count, ocfs2_page_writer *actor,
1197                                      void *priv)
1198 {
1199         int ret, credits = OCFS2_INODE_UPDATE_CREDITS;
1200         ssize_t written = 0;
1201         u32 phys;
1202         struct inode *inode = file->f_mapping->host;
1203         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1204         struct buffer_head *di_bh = NULL;
1205         struct ocfs2_dinode *di;
1206         struct ocfs2_alloc_context *data_ac = NULL;
1207         struct ocfs2_alloc_context *meta_ac = NULL;
1208         handle_t *handle;
1209         struct ocfs2_write_ctxt wc;
1210
1211         ocfs2_write_ctxt_init(&wc, osb, pos, count, actor, priv);
1212
1213         ret = ocfs2_meta_lock(inode, &di_bh, 1);
1214         if (ret) {
1215                 mlog_errno(ret);
1216                 goto out;
1217         }
1218         di = (struct ocfs2_dinode *)di_bh->b_data;
1219
1220         /*
1221          * Take alloc sem here to prevent concurrent lookups. That way
1222          * the mapping, zeroing and tree manipulation within
1223          * ocfs2_write() will be safe against ->readpage(). This
1224          * should also serve to lock out allocation from a shared
1225          * writeable region.
1226          */
1227         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1228
1229         ret = ocfs2_get_clusters(inode, wc.w_cpos, &phys, NULL, NULL);
1230         if (ret) {
1231                 mlog_errno(ret);
1232                 goto out_meta;
1233         }
1234
1235         /* phys == 0 means that allocation is required. */
1236         if (phys == 0) {
1237                 ret = ocfs2_lock_allocators(inode, di, 1, &data_ac, &meta_ac);
1238                 if (ret) {
1239                         mlog_errno(ret);
1240                         goto out_meta;
1241                 }
1242
1243                 credits = ocfs2_calc_extend_credits(inode->i_sb, di, 1);
1244         }
1245
1246         ret = ocfs2_data_lock(inode, 1);
1247         if (ret) {
1248                 mlog_errno(ret);
1249                 goto out_meta;
1250         }
1251
1252         handle = ocfs2_start_trans(osb, credits);
1253         if (IS_ERR(handle)) {
1254                 ret = PTR_ERR(handle);
1255                 mlog_errno(ret);
1256                 goto out_data;
1257         }
1258
1259         written = ocfs2_write(file, phys, handle, di_bh, data_ac,
1260                               meta_ac, &wc);
1261         if (written < 0) {
1262                 ret = written;
1263                 mlog_errno(ret);
1264                 goto out_commit;
1265         }
1266
1267         ret = ocfs2_journal_access(handle, inode, di_bh,
1268                                    OCFS2_JOURNAL_ACCESS_WRITE);
1269         if (ret) {
1270                 mlog_errno(ret);
1271                 goto out_commit;
1272         }
1273
1274         pos += written;
1275         if (pos > inode->i_size) {
1276                 i_size_write(inode, pos);
1277                 mark_inode_dirty(inode);
1278         }
1279         inode->i_blocks = ocfs2_inode_sector_count(inode);
1280         di->i_size = cpu_to_le64((u64)i_size_read(inode));
1281         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1282         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
1283         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
1284
1285         ret = ocfs2_journal_dirty(handle, di_bh);
1286         if (ret)
1287                 mlog_errno(ret);
1288
1289 out_commit:
1290         ocfs2_commit_trans(osb, handle);
1291
1292 out_data:
1293         ocfs2_data_unlock(inode, 1);
1294
1295 out_meta:
1296         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1297         ocfs2_meta_unlock(inode, 1);
1298
1299 out:
1300         brelse(di_bh);
1301         if (data_ac)
1302                 ocfs2_free_alloc_context(data_ac);
1303         if (meta_ac)
1304                 ocfs2_free_alloc_context(meta_ac);
1305
1306         return written ? written : ret;
1307 }
1308
1309 const struct address_space_operations ocfs2_aops = {
1310         .readpage       = ocfs2_readpage,
1311         .writepage      = ocfs2_writepage,
1312         .bmap           = ocfs2_bmap,
1313         .sync_page      = block_sync_page,
1314         .direct_IO      = ocfs2_direct_IO,
1315         .invalidatepage = ocfs2_invalidatepage,
1316         .releasepage    = ocfs2_releasepage,
1317         .migratepage    = buffer_migrate_page,
1318 };