Merge git://git.kernel.org/pub/scm/linux/kernel/git/sfrench/cifs-2.6
[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)le64_to_cpu(fe->i_blkno), 7,
82                      fe->i_signature);
83                 goto bail;
84         }
85
86         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
87                                                     le32_to_cpu(fe->i_clusters))) {
88                 mlog(ML_ERROR, "block offset is outside the allocated size: "
89                      "%llu\n", (unsigned long long)iblock);
90                 goto bail;
91         }
92
93         /* We don't use the page cache to create symlink data, so if
94          * need be, copy it over from the buffer cache. */
95         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
96                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
97                             iblock;
98                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
99                 if (!buffer_cache_bh) {
100                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
101                         goto bail;
102                 }
103
104                 /* we haven't locked out transactions, so a commit
105                  * could've happened. Since we've got a reference on
106                  * the bh, even if it commits while we're doing the
107                  * copy, the data is still good. */
108                 if (buffer_jbd(buffer_cache_bh)
109                     && ocfs2_inode_is_new(inode)) {
110                         kaddr = kmap_atomic(bh_result->b_page, KM_USER0);
111                         if (!kaddr) {
112                                 mlog(ML_ERROR, "couldn't kmap!\n");
113                                 goto bail;
114                         }
115                         memcpy(kaddr + (bh_result->b_size * iblock),
116                                buffer_cache_bh->b_data,
117                                bh_result->b_size);
118                         kunmap_atomic(kaddr, KM_USER0);
119                         set_buffer_uptodate(bh_result);
120                 }
121                 brelse(buffer_cache_bh);
122         }
123
124         map_bh(bh_result, inode->i_sb,
125                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
126
127         err = 0;
128
129 bail:
130         if (bh)
131                 brelse(bh);
132
133         mlog_exit(err);
134         return err;
135 }
136
137 static int ocfs2_get_block(struct inode *inode, sector_t iblock,
138                            struct buffer_head *bh_result, int create)
139 {
140         int err = 0;
141         unsigned int ext_flags;
142         u64 p_blkno, past_eof;
143         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
144
145         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
146                    (unsigned long long)iblock, bh_result, create);
147
148         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
149                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
150                      inode, inode->i_ino);
151
152         if (S_ISLNK(inode->i_mode)) {
153                 /* this always does I/O for some reason. */
154                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
155                 goto bail;
156         }
157
158         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, NULL,
159                                           &ext_flags);
160         if (err) {
161                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
162                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
163                      (unsigned long long)p_blkno);
164                 goto bail;
165         }
166
167         /*
168          * ocfs2 never allocates in this function - the only time we
169          * need to use BH_New is when we're extending i_size on a file
170          * system which doesn't support holes, in which case BH_New
171          * allows block_prepare_write() to zero.
172          */
173         mlog_bug_on_msg(create && p_blkno == 0 && ocfs2_sparse_alloc(osb),
174                         "ino %lu, iblock %llu\n", inode->i_ino,
175                         (unsigned long long)iblock);
176
177         /* Treat the unwritten extent as a hole for zeroing purposes. */
178         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
179                 map_bh(bh_result, inode->i_sb, p_blkno);
180
181         if (!ocfs2_sparse_alloc(osb)) {
182                 if (p_blkno == 0) {
183                         err = -EIO;
184                         mlog(ML_ERROR,
185                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
186                              (unsigned long long)iblock,
187                              (unsigned long long)p_blkno,
188                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
189                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
190                         dump_stack();
191                 }
192
193                 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
194                 mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino,
195                      (unsigned long long)past_eof);
196
197                 if (create && (iblock >= past_eof))
198                         set_buffer_new(bh_result);
199         }
200
201 bail:
202         if (err < 0)
203                 err = -EIO;
204
205         mlog_exit(err);
206         return err;
207 }
208
209 static int ocfs2_readpage(struct file *file, struct page *page)
210 {
211         struct inode *inode = page->mapping->host;
212         loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
213         int ret, unlock = 1;
214
215         mlog_entry("(0x%p, %lu)\n", file, (page ? page->index : 0));
216
217         ret = ocfs2_meta_lock_with_page(inode, NULL, 0, page);
218         if (ret != 0) {
219                 if (ret == AOP_TRUNCATED_PAGE)
220                         unlock = 0;
221                 mlog_errno(ret);
222                 goto out;
223         }
224
225         if (down_read_trylock(&OCFS2_I(inode)->ip_alloc_sem) == 0) {
226                 ret = AOP_TRUNCATED_PAGE;
227                 goto out_meta_unlock;
228         }
229
230         /*
231          * i_size might have just been updated as we grabed the meta lock.  We
232          * might now be discovering a truncate that hit on another node.
233          * block_read_full_page->get_block freaks out if it is asked to read
234          * beyond the end of a file, so we check here.  Callers
235          * (generic_file_read, fault->nopage) are clever enough to check i_size
236          * and notice that the page they just read isn't needed.
237          *
238          * XXX sys_readahead() seems to get that wrong?
239          */
240         if (start >= i_size_read(inode)) {
241                 zero_user_page(page, 0, PAGE_SIZE, KM_USER0);
242                 SetPageUptodate(page);
243                 ret = 0;
244                 goto out_alloc;
245         }
246
247         ret = ocfs2_data_lock_with_page(inode, 0, page);
248         if (ret != 0) {
249                 if (ret == AOP_TRUNCATED_PAGE)
250                         unlock = 0;
251                 mlog_errno(ret);
252                 goto out_alloc;
253         }
254
255         ret = block_read_full_page(page, ocfs2_get_block);
256         unlock = 0;
257
258         ocfs2_data_unlock(inode, 0);
259 out_alloc:
260         up_read(&OCFS2_I(inode)->ip_alloc_sem);
261 out_meta_unlock:
262         ocfs2_meta_unlock(inode, 0);
263 out:
264         if (unlock)
265                 unlock_page(page);
266         mlog_exit(ret);
267         return ret;
268 }
269
270 /* Note: Because we don't support holes, our allocation has
271  * already happened (allocation writes zeros to the file data)
272  * so we don't have to worry about ordered writes in
273  * ocfs2_writepage.
274  *
275  * ->writepage is called during the process of invalidating the page cache
276  * during blocked lock processing.  It can't block on any cluster locks
277  * to during block mapping.  It's relying on the fact that the block
278  * mapping can't have disappeared under the dirty pages that it is
279  * being asked to write back.
280  */
281 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
282 {
283         int ret;
284
285         mlog_entry("(0x%p)\n", page);
286
287         ret = block_write_full_page(page, ocfs2_get_block, wbc);
288
289         mlog_exit(ret);
290
291         return ret;
292 }
293
294 /*
295  * This is called from ocfs2_write_zero_page() which has handled it's
296  * own cluster locking and has ensured allocation exists for those
297  * blocks to be written.
298  */
299 int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
300                                unsigned from, unsigned to)
301 {
302         int ret;
303
304         down_read(&OCFS2_I(inode)->ip_alloc_sem);
305
306         ret = block_prepare_write(page, from, to, ocfs2_get_block);
307
308         up_read(&OCFS2_I(inode)->ip_alloc_sem);
309
310         return ret;
311 }
312
313 /* Taken from ext3. We don't necessarily need the full blown
314  * functionality yet, but IMHO it's better to cut and paste the whole
315  * thing so we can avoid introducing our own bugs (and easily pick up
316  * their fixes when they happen) --Mark */
317 int walk_page_buffers(  handle_t *handle,
318                         struct buffer_head *head,
319                         unsigned from,
320                         unsigned to,
321                         int *partial,
322                         int (*fn)(      handle_t *handle,
323                                         struct buffer_head *bh))
324 {
325         struct buffer_head *bh;
326         unsigned block_start, block_end;
327         unsigned blocksize = head->b_size;
328         int err, ret = 0;
329         struct buffer_head *next;
330
331         for (   bh = head, block_start = 0;
332                 ret == 0 && (bh != head || !block_start);
333                 block_start = block_end, bh = next)
334         {
335                 next = bh->b_this_page;
336                 block_end = block_start + blocksize;
337                 if (block_end <= from || block_start >= to) {
338                         if (partial && !buffer_uptodate(bh))
339                                 *partial = 1;
340                         continue;
341                 }
342                 err = (*fn)(handle, bh);
343                 if (!ret)
344                         ret = err;
345         }
346         return ret;
347 }
348
349 handle_t *ocfs2_start_walk_page_trans(struct inode *inode,
350                                                          struct page *page,
351                                                          unsigned from,
352                                                          unsigned to)
353 {
354         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
355         handle_t *handle = NULL;
356         int ret = 0;
357
358         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
359         if (!handle) {
360                 ret = -ENOMEM;
361                 mlog_errno(ret);
362                 goto out;
363         }
364
365         if (ocfs2_should_order_data(inode)) {
366                 ret = walk_page_buffers(handle,
367                                         page_buffers(page),
368                                         from, to, NULL,
369                                         ocfs2_journal_dirty_data);
370                 if (ret < 0) 
371                         mlog_errno(ret);
372         }
373 out:
374         if (ret) {
375                 if (handle)
376                         ocfs2_commit_trans(osb, handle);
377                 handle = ERR_PTR(ret);
378         }
379         return handle;
380 }
381
382 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
383 {
384         sector_t status;
385         u64 p_blkno = 0;
386         int err = 0;
387         struct inode *inode = mapping->host;
388
389         mlog_entry("(block = %llu)\n", (unsigned long long)block);
390
391         /* We don't need to lock journal system files, since they aren't
392          * accessed concurrently from multiple nodes.
393          */
394         if (!INODE_JOURNAL(inode)) {
395                 err = ocfs2_meta_lock(inode, NULL, 0);
396                 if (err) {
397                         if (err != -ENOENT)
398                                 mlog_errno(err);
399                         goto bail;
400                 }
401                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
402         }
403
404         err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL, NULL);
405
406         if (!INODE_JOURNAL(inode)) {
407                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
408                 ocfs2_meta_unlock(inode, 0);
409         }
410
411         if (err) {
412                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
413                      (unsigned long long)block);
414                 mlog_errno(err);
415                 goto bail;
416         }
417
418
419 bail:
420         status = err ? 0 : p_blkno;
421
422         mlog_exit((int)status);
423
424         return status;
425 }
426
427 /*
428  * TODO: Make this into a generic get_blocks function.
429  *
430  * From do_direct_io in direct-io.c:
431  *  "So what we do is to permit the ->get_blocks function to populate
432  *   bh.b_size with the size of IO which is permitted at this offset and
433  *   this i_blkbits."
434  *
435  * This function is called directly from get_more_blocks in direct-io.c.
436  *
437  * called like this: dio->get_blocks(dio->inode, fs_startblk,
438  *                                      fs_count, map_bh, dio->rw == WRITE);
439  */
440 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
441                                      struct buffer_head *bh_result, int create)
442 {
443         int ret;
444         u64 p_blkno, inode_blocks, contig_blocks;
445         unsigned int ext_flags;
446         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
447         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
448
449         /* This function won't even be called if the request isn't all
450          * nicely aligned and of the right size, so there's no need
451          * for us to check any of that. */
452
453         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
454
455         /*
456          * Any write past EOF is not allowed because we'd be extending.
457          */
458         if (create && (iblock + max_blocks) > inode_blocks) {
459                 ret = -EIO;
460                 goto bail;
461         }
462
463         /* This figures out the size of the next contiguous block, and
464          * our logical offset */
465         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
466                                           &contig_blocks, &ext_flags);
467         if (ret) {
468                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
469                      (unsigned long long)iblock);
470                 ret = -EIO;
471                 goto bail;
472         }
473
474         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)) && !p_blkno) {
475                 ocfs2_error(inode->i_sb,
476                             "Inode %llu has a hole at block %llu\n",
477                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
478                             (unsigned long long)iblock);
479                 ret = -EROFS;
480                 goto bail;
481         }
482
483         /*
484          * get_more_blocks() expects us to describe a hole by clearing
485          * the mapped bit on bh_result().
486          *
487          * Consider an unwritten extent as a hole.
488          */
489         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
490                 map_bh(bh_result, inode->i_sb, p_blkno);
491         else {
492                 /*
493                  * ocfs2_prepare_inode_for_write() should have caught
494                  * the case where we'd be filling a hole and triggered
495                  * a buffered write instead.
496                  */
497                 if (create) {
498                         ret = -EIO;
499                         mlog_errno(ret);
500                         goto bail;
501                 }
502
503                 clear_buffer_mapped(bh_result);
504         }
505
506         /* make sure we don't map more than max_blocks blocks here as
507            that's all the kernel will handle at this point. */
508         if (max_blocks < contig_blocks)
509                 contig_blocks = max_blocks;
510         bh_result->b_size = contig_blocks << blocksize_bits;
511 bail:
512         return ret;
513 }
514
515 /* 
516  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
517  * particularly interested in the aio/dio case.  Like the core uses
518  * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from
519  * truncation on another.
520  */
521 static void ocfs2_dio_end_io(struct kiocb *iocb,
522                              loff_t offset,
523                              ssize_t bytes,
524                              void *private)
525 {
526         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
527         int level;
528
529         /* this io's submitter should not have unlocked this before we could */
530         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
531
532         ocfs2_iocb_clear_rw_locked(iocb);
533
534         level = ocfs2_iocb_rw_locked_level(iocb);
535         if (!level)
536                 up_read(&inode->i_alloc_sem);
537         ocfs2_rw_unlock(inode, level);
538 }
539
540 /*
541  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
542  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
543  * do journalled data.
544  */
545 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
546 {
547         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
548
549         journal_invalidatepage(journal, page, offset);
550 }
551
552 static int ocfs2_releasepage(struct page *page, gfp_t wait)
553 {
554         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
555
556         if (!page_has_buffers(page))
557                 return 0;
558         return journal_try_to_free_buffers(journal, page, wait);
559 }
560
561 static ssize_t ocfs2_direct_IO(int rw,
562                                struct kiocb *iocb,
563                                const struct iovec *iov,
564                                loff_t offset,
565                                unsigned long nr_segs)
566 {
567         struct file *file = iocb->ki_filp;
568         struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
569         int ret;
570
571         mlog_entry_void();
572
573         if (!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb))) {
574                 /*
575                  * We get PR data locks even for O_DIRECT.  This
576                  * allows concurrent O_DIRECT I/O but doesn't let
577                  * O_DIRECT with extending and buffered zeroing writes
578                  * race.  If they did race then the buffered zeroing
579                  * could be written back after the O_DIRECT I/O.  It's
580                  * one thing to tell people not to mix buffered and
581                  * O_DIRECT writes, but expecting them to understand
582                  * that file extension is also an implicit buffered
583                  * write is too much.  By getting the PR we force
584                  * writeback of the buffered zeroing before
585                  * proceeding.
586                  */
587                 ret = ocfs2_data_lock(inode, 0);
588                 if (ret < 0) {
589                         mlog_errno(ret);
590                         goto out;
591                 }
592                 ocfs2_data_unlock(inode, 0);
593         }
594
595         ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
596                                             inode->i_sb->s_bdev, iov, offset,
597                                             nr_segs, 
598                                             ocfs2_direct_IO_get_blocks,
599                                             ocfs2_dio_end_io);
600 out:
601         mlog_exit(ret);
602         return ret;
603 }
604
605 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
606                                             u32 cpos,
607                                             unsigned int *start,
608                                             unsigned int *end)
609 {
610         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
611
612         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
613                 unsigned int cpp;
614
615                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
616
617                 cluster_start = cpos % cpp;
618                 cluster_start = cluster_start << osb->s_clustersize_bits;
619
620                 cluster_end = cluster_start + osb->s_clustersize;
621         }
622
623         BUG_ON(cluster_start > PAGE_SIZE);
624         BUG_ON(cluster_end > PAGE_SIZE);
625
626         if (start)
627                 *start = cluster_start;
628         if (end)
629                 *end = cluster_end;
630 }
631
632 /*
633  * 'from' and 'to' are the region in the page to avoid zeroing.
634  *
635  * If pagesize > clustersize, this function will avoid zeroing outside
636  * of the cluster boundary.
637  *
638  * from == to == 0 is code for "zero the entire cluster region"
639  */
640 static void ocfs2_clear_page_regions(struct page *page,
641                                      struct ocfs2_super *osb, u32 cpos,
642                                      unsigned from, unsigned to)
643 {
644         void *kaddr;
645         unsigned int cluster_start, cluster_end;
646
647         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
648
649         kaddr = kmap_atomic(page, KM_USER0);
650
651         if (from || to) {
652                 if (from > cluster_start)
653                         memset(kaddr + cluster_start, 0, from - cluster_start);
654                 if (to < cluster_end)
655                         memset(kaddr + to, 0, cluster_end - to);
656         } else {
657                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
658         }
659
660         kunmap_atomic(kaddr, KM_USER0);
661 }
662
663 /*
664  * Some of this taken from block_prepare_write(). We already have our
665  * mapping by now though, and the entire write will be allocating or
666  * it won't, so not much need to use BH_New.
667  *
668  * This will also skip zeroing, which is handled externally.
669  */
670 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
671                           struct inode *inode, unsigned int from,
672                           unsigned int to, int new)
673 {
674         int ret = 0;
675         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
676         unsigned int block_end, block_start;
677         unsigned int bsize = 1 << inode->i_blkbits;
678
679         if (!page_has_buffers(page))
680                 create_empty_buffers(page, bsize, 0);
681
682         head = page_buffers(page);
683         for (bh = head, block_start = 0; bh != head || !block_start;
684              bh = bh->b_this_page, block_start += bsize) {
685                 block_end = block_start + bsize;
686
687                 /*
688                  * Ignore blocks outside of our i/o range -
689                  * they may belong to unallocated clusters.
690                  */
691                 if (block_start >= to || block_end <= from) {
692                         if (PageUptodate(page))
693                                 set_buffer_uptodate(bh);
694                         continue;
695                 }
696
697                 /*
698                  * For an allocating write with cluster size >= page
699                  * size, we always write the entire page.
700                  */
701
702                 if (buffer_new(bh))
703                         clear_buffer_new(bh);
704
705                 if (!buffer_mapped(bh)) {
706                         map_bh(bh, inode->i_sb, *p_blkno);
707                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
708                 }
709
710                 if (PageUptodate(page)) {
711                         if (!buffer_uptodate(bh))
712                                 set_buffer_uptodate(bh);
713                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
714                      (block_start < from || block_end > to)) {
715                         ll_rw_block(READ, 1, &bh);
716                         *wait_bh++=bh;
717                 }
718
719                 *p_blkno = *p_blkno + 1;
720         }
721
722         /*
723          * If we issued read requests - let them complete.
724          */
725         while(wait_bh > wait) {
726                 wait_on_buffer(*--wait_bh);
727                 if (!buffer_uptodate(*wait_bh))
728                         ret = -EIO;
729         }
730
731         if (ret == 0 || !new)
732                 return ret;
733
734         /*
735          * If we get -EIO above, zero out any newly allocated blocks
736          * to avoid exposing stale data.
737          */
738         bh = head;
739         block_start = 0;
740         do {
741                 void *kaddr;
742
743                 block_end = block_start + bsize;
744                 if (block_end <= from)
745                         goto next_bh;
746                 if (block_start >= to)
747                         break;
748
749                 kaddr = kmap_atomic(page, KM_USER0);
750                 memset(kaddr+block_start, 0, bh->b_size);
751                 flush_dcache_page(page);
752                 kunmap_atomic(kaddr, KM_USER0);
753                 set_buffer_uptodate(bh);
754                 mark_buffer_dirty(bh);
755
756 next_bh:
757                 block_start = block_end;
758                 bh = bh->b_this_page;
759         } while (bh != head);
760
761         return ret;
762 }
763
764 /*
765  * This will copy user data from the buffer page in the splice
766  * context.
767  *
768  * For now, we ignore SPLICE_F_MOVE as that would require some extra
769  * communication out all the way to ocfs2_write().
770  */
771 int ocfs2_map_and_write_splice_data(struct inode *inode,
772                                   struct ocfs2_write_ctxt *wc, u64 *p_blkno,
773                                   unsigned int *ret_from, unsigned int *ret_to)
774 {
775         int ret;
776         unsigned int to, from, cluster_start, cluster_end;
777         char *src, *dst;
778         struct ocfs2_splice_write_priv *sp = wc->w_private;
779         struct pipe_buffer *buf = sp->s_buf;
780         unsigned long bytes, src_from;
781         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
782
783         ocfs2_figure_cluster_boundaries(osb, wc->w_cpos, &cluster_start,
784                                         &cluster_end);
785
786         from = sp->s_offset;
787         src_from = sp->s_buf_offset;
788         bytes = wc->w_count;
789
790         if (wc->w_large_pages) {
791                 /*
792                  * For cluster size < page size, we have to
793                  * calculate pos within the cluster and obey
794                  * the rightmost boundary.
795                  */
796                 bytes = min(bytes, (unsigned long)(osb->s_clustersize
797                                    - (wc->w_pos & (osb->s_clustersize - 1))));
798         }
799         to = from + bytes;
800
801         BUG_ON(from > PAGE_CACHE_SIZE);
802         BUG_ON(to > PAGE_CACHE_SIZE);
803         BUG_ON(from < cluster_start);
804         BUG_ON(to > cluster_end);
805
806         if (wc->w_this_page_new)
807                 ret = ocfs2_map_page_blocks(wc->w_this_page, p_blkno, inode,
808                                             cluster_start, cluster_end, 1);
809         else
810                 ret = ocfs2_map_page_blocks(wc->w_this_page, p_blkno, inode,
811                                             from, to, 0);
812         if (ret) {
813                 mlog_errno(ret);
814                 goto out;
815         }
816
817         src = buf->ops->map(sp->s_pipe, buf, 1);
818         dst = kmap_atomic(wc->w_this_page, KM_USER1);
819         memcpy(dst + from, src + src_from, bytes);
820         kunmap_atomic(wc->w_this_page, KM_USER1);
821         buf->ops->unmap(sp->s_pipe, buf, src);
822
823         wc->w_finished_copy = 1;
824
825         *ret_from = from;
826         *ret_to = to;
827 out:
828
829         return bytes ? (unsigned int)bytes : ret;
830 }
831
832 /*
833  * This will copy user data from the iovec in the buffered write
834  * context.
835  */
836 int ocfs2_map_and_write_user_data(struct inode *inode,
837                                   struct ocfs2_write_ctxt *wc, u64 *p_blkno,
838                                   unsigned int *ret_from, unsigned int *ret_to)
839 {
840         int ret;
841         unsigned int to, from, cluster_start, cluster_end;
842         unsigned long bytes, src_from;
843         char *dst;
844         struct ocfs2_buffered_write_priv *bp = wc->w_private;
845         const struct iovec *cur_iov = bp->b_cur_iov;
846         char __user *buf;
847         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
848
849         ocfs2_figure_cluster_boundaries(osb, wc->w_cpos, &cluster_start,
850                                         &cluster_end);
851
852         buf = cur_iov->iov_base + bp->b_cur_off;
853         src_from = (unsigned long)buf & ~PAGE_CACHE_MASK;
854
855         from = wc->w_pos & (PAGE_CACHE_SIZE - 1);
856
857         /*
858          * This is a lot of comparisons, but it reads quite
859          * easily, which is important here.
860          */
861         /* Stay within the src page */
862         bytes = PAGE_SIZE - src_from;
863         /* Stay within the vector */
864         bytes = min(bytes,
865                     (unsigned long)(cur_iov->iov_len - bp->b_cur_off));
866         /* Stay within count */
867         bytes = min(bytes, (unsigned long)wc->w_count);
868         /*
869          * For clustersize > page size, just stay within
870          * target page, otherwise we have to calculate pos
871          * within the cluster and obey the rightmost
872          * boundary.
873          */
874         if (wc->w_large_pages) {
875                 /*
876                  * For cluster size < page size, we have to
877                  * calculate pos within the cluster and obey
878                  * the rightmost boundary.
879                  */
880                 bytes = min(bytes, (unsigned long)(osb->s_clustersize
881                                    - (wc->w_pos & (osb->s_clustersize - 1))));
882         } else {
883                 /*
884                  * cluster size > page size is the most common
885                  * case - we just stay within the target page
886                  * boundary.
887                  */
888                 bytes = min(bytes, PAGE_CACHE_SIZE - from);
889         }
890
891         to = from + bytes;
892
893         BUG_ON(from > PAGE_CACHE_SIZE);
894         BUG_ON(to > PAGE_CACHE_SIZE);
895         BUG_ON(from < cluster_start);
896         BUG_ON(to > cluster_end);
897
898         if (wc->w_this_page_new)
899                 ret = ocfs2_map_page_blocks(wc->w_this_page, p_blkno, inode,
900                                             cluster_start, cluster_end, 1);
901         else
902                 ret = ocfs2_map_page_blocks(wc->w_this_page, p_blkno, inode,
903                                             from, to, 0);
904         if (ret) {
905                 mlog_errno(ret);
906                 goto out;
907         }
908
909         dst = kmap(wc->w_this_page);
910         memcpy(dst + from, bp->b_src_buf + src_from, bytes);
911         kunmap(wc->w_this_page);
912
913         /*
914          * XXX: This is slow, but simple. The caller of
915          * ocfs2_buffered_write_cluster() is responsible for
916          * passing through the iovecs, so it's difficult to
917          * predict what our next step is in here after our
918          * initial write. A future version should be pushing
919          * that iovec manipulation further down.
920          *
921          * By setting this, we indicate that a copy from user
922          * data was done, and subsequent calls for this
923          * cluster will skip copying more data.
924          */
925         wc->w_finished_copy = 1;
926
927         *ret_from = from;
928         *ret_to = to;
929 out:
930
931         return bytes ? (unsigned int)bytes : ret;
932 }
933
934 /*
935  * Map, fill and write a page to disk.
936  *
937  * The work of copying data is done via callback.  Newly allocated
938  * pages which don't take user data will be zero'd (set 'new' to
939  * indicate an allocating write)
940  *
941  * Returns a negative error code or the number of bytes copied into
942  * the page.
943  */
944 static int ocfs2_write_data_page(struct inode *inode, handle_t *handle,
945                                  u64 *p_blkno, struct page *page,
946                                  struct ocfs2_write_ctxt *wc, int new)
947 {
948         int ret, copied = 0;
949         unsigned int from = 0, to = 0;
950         unsigned int cluster_start, cluster_end;
951         unsigned int zero_from = 0, zero_to = 0;
952
953         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), wc->w_cpos,
954                                         &cluster_start, &cluster_end);
955
956         if ((wc->w_pos >> PAGE_CACHE_SHIFT) == page->index
957             && !wc->w_finished_copy) {
958
959                 wc->w_this_page = page;
960                 wc->w_this_page_new = new;
961                 ret = wc->w_write_data_page(inode, wc, p_blkno, &from, &to);
962                 if (ret < 0) {
963                         mlog_errno(ret);
964                         goto out;
965                 }
966
967                 copied = ret;
968
969                 zero_from = from;
970                 zero_to = to;
971                 if (new) {
972                         from = cluster_start;
973                         to = cluster_end;
974                 }
975         } else {
976                 /*
977                  * If we haven't allocated the new page yet, we
978                  * shouldn't be writing it out without copying user
979                  * data. This is likely a math error from the caller.
980                  */
981                 BUG_ON(!new);
982
983                 from = cluster_start;
984                 to = cluster_end;
985
986                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
987                                             cluster_start, cluster_end, 1);
988                 if (ret) {
989                         mlog_errno(ret);
990                         goto out;
991                 }
992         }
993
994         /*
995          * Parts of newly allocated pages need to be zero'd.
996          *
997          * Above, we have also rewritten 'to' and 'from' - as far as
998          * the rest of the function is concerned, the entire cluster
999          * range inside of a page needs to be written.
1000          *
1001          * We can skip this if the page is up to date - it's already
1002          * been zero'd from being read in as a hole.
1003          */
1004         if (new && !PageUptodate(page))
1005                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1006                                          wc->w_cpos, zero_from, zero_to);
1007
1008         flush_dcache_page(page);
1009
1010         if (ocfs2_should_order_data(inode)) {
1011                 ret = walk_page_buffers(handle,
1012                                         page_buffers(page),
1013                                         from, to, NULL,
1014                                         ocfs2_journal_dirty_data);
1015                 if (ret < 0)
1016                         mlog_errno(ret);
1017         }
1018
1019         /*
1020          * We don't use generic_commit_write() because we need to
1021          * handle our own i_size update.
1022          */
1023         ret = block_commit_write(page, from, to);
1024         if (ret)
1025                 mlog_errno(ret);
1026 out:
1027
1028         return copied ? copied : ret;
1029 }
1030
1031 /*
1032  * Do the actual write of some data into an inode. Optionally allocate
1033  * in order to fulfill the write.
1034  *
1035  * cpos is the logical cluster offset within the file to write at
1036  *
1037  * 'phys' is the physical mapping of that offset. a 'phys' value of
1038  * zero indicates that allocation is required. In this case, data_ac
1039  * and meta_ac should be valid (meta_ac can be null if metadata
1040  * allocation isn't required).
1041  */
1042 static ssize_t ocfs2_write(struct file *file, u32 phys, handle_t *handle,
1043                            struct buffer_head *di_bh,
1044                            struct ocfs2_alloc_context *data_ac,
1045                            struct ocfs2_alloc_context *meta_ac,
1046                            struct ocfs2_write_ctxt *wc)
1047 {
1048         int ret, i, numpages = 1, new;
1049         unsigned int copied = 0;
1050         u32 tmp_pos;
1051         u64 v_blkno, p_blkno;
1052         struct address_space *mapping = file->f_mapping;
1053         struct inode *inode = mapping->host;
1054         unsigned long index, start;
1055         struct page **cpages;
1056
1057         new = phys == 0 ? 1 : 0;
1058
1059         /*
1060          * Figure out how many pages we'll be manipulating here. For
1061          * non allocating write, we just change the one
1062          * page. Otherwise, we'll need a whole clusters worth.
1063          */
1064         if (new)
1065                 numpages = ocfs2_pages_per_cluster(inode->i_sb);
1066
1067         cpages = kzalloc(sizeof(*cpages) * numpages, GFP_NOFS);
1068         if (!cpages) {
1069                 ret = -ENOMEM;
1070                 mlog_errno(ret);
1071                 return ret;
1072         }
1073
1074         /*
1075          * Fill our page array first. That way we've grabbed enough so
1076          * that we can zero and flush if we error after adding the
1077          * extent.
1078          */
1079         if (new) {
1080                 start = ocfs2_align_clusters_to_page_index(inode->i_sb,
1081                                                            wc->w_cpos);
1082                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, wc->w_cpos);
1083         } else {
1084                 start = wc->w_pos >> PAGE_CACHE_SHIFT;
1085                 v_blkno = wc->w_pos >> inode->i_sb->s_blocksize_bits;
1086         }
1087
1088         for(i = 0; i < numpages; i++) {
1089                 index = start + i;
1090
1091                 cpages[i] = find_or_create_page(mapping, index, GFP_NOFS);
1092                 if (!cpages[i]) {
1093                         ret = -ENOMEM;
1094                         mlog_errno(ret);
1095                         goto out;
1096                 }
1097         }
1098
1099         if (new) {
1100                 /*
1101                  * This is safe to call with the page locks - it won't take
1102                  * any additional semaphores or cluster locks.
1103                  */
1104                 tmp_pos = wc->w_cpos;
1105                 ret = ocfs2_do_extend_allocation(OCFS2_SB(inode->i_sb), inode,
1106                                                  &tmp_pos, 1, di_bh, handle,
1107                                                  data_ac, meta_ac, NULL);
1108                 /*
1109                  * This shouldn't happen because we must have already
1110                  * calculated the correct meta data allocation required. The
1111                  * internal tree allocation code should know how to increase
1112                  * transaction credits itself.
1113                  *
1114                  * If need be, we could handle -EAGAIN for a
1115                  * RESTART_TRANS here.
1116                  */
1117                 mlog_bug_on_msg(ret == -EAGAIN,
1118                                 "Inode %llu: EAGAIN return during allocation.\n",
1119                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1120                 if (ret < 0) {
1121                         mlog_errno(ret);
1122                         goto out;
1123                 }
1124         }
1125
1126         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1127                                           NULL);
1128         if (ret < 0) {
1129
1130                 /*
1131                  * XXX: Should we go readonly here?
1132                  */
1133
1134                 mlog_errno(ret);
1135                 goto out;
1136         }
1137
1138         BUG_ON(p_blkno == 0);
1139
1140         for(i = 0; i < numpages; i++) {
1141                 ret = ocfs2_write_data_page(inode, handle, &p_blkno, cpages[i],
1142                                             wc, new);
1143                 if (ret < 0) {
1144                         mlog_errno(ret);
1145                         goto out;
1146                 }
1147
1148                 copied += ret;
1149         }
1150
1151 out:
1152         for(i = 0; i < numpages; i++) {
1153                 unlock_page(cpages[i]);
1154                 mark_page_accessed(cpages[i]);
1155                 page_cache_release(cpages[i]);
1156         }
1157         kfree(cpages);
1158
1159         return copied ? copied : ret;
1160 }
1161
1162 static void ocfs2_write_ctxt_init(struct ocfs2_write_ctxt *wc,
1163                                   struct ocfs2_super *osb, loff_t pos,
1164                                   size_t count, ocfs2_page_writer *cb,
1165                                   void *cb_priv)
1166 {
1167         wc->w_count = count;
1168         wc->w_pos = pos;
1169         wc->w_cpos = wc->w_pos >> osb->s_clustersize_bits;
1170         wc->w_finished_copy = 0;
1171
1172         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
1173                 wc->w_large_pages = 1;
1174         else
1175                 wc->w_large_pages = 0;
1176
1177         wc->w_write_data_page = cb;
1178         wc->w_private = cb_priv;
1179 }
1180
1181 /*
1182  * Write a cluster to an inode. The cluster may not be allocated yet,
1183  * in which case it will be. This only exists for buffered writes -
1184  * O_DIRECT takes a more "traditional" path through the kernel.
1185  *
1186  * The caller is responsible for incrementing pos, written counts, etc
1187  *
1188  * For file systems that don't support sparse files, pre-allocation
1189  * and page zeroing up until cpos should be done prior to this
1190  * function call.
1191  *
1192  * Callers should be holding i_sem, and the rw cluster lock.
1193  *
1194  * Returns the number of user bytes written, or less than zero for
1195  * error.
1196  */
1197 ssize_t ocfs2_buffered_write_cluster(struct file *file, loff_t pos,
1198                                      size_t count, ocfs2_page_writer *actor,
1199                                      void *priv)
1200 {
1201         int ret, credits = OCFS2_INODE_UPDATE_CREDITS;
1202         ssize_t written = 0;
1203         u32 phys;
1204         struct inode *inode = file->f_mapping->host;
1205         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1206         struct buffer_head *di_bh = NULL;
1207         struct ocfs2_dinode *di;
1208         struct ocfs2_alloc_context *data_ac = NULL;
1209         struct ocfs2_alloc_context *meta_ac = NULL;
1210         handle_t *handle;
1211         struct ocfs2_write_ctxt wc;
1212
1213         ocfs2_write_ctxt_init(&wc, osb, pos, count, actor, priv);
1214
1215         ret = ocfs2_meta_lock(inode, &di_bh, 1);
1216         if (ret) {
1217                 mlog_errno(ret);
1218                 goto out;
1219         }
1220         di = (struct ocfs2_dinode *)di_bh->b_data;
1221
1222         /*
1223          * Take alloc sem here to prevent concurrent lookups. That way
1224          * the mapping, zeroing and tree manipulation within
1225          * ocfs2_write() will be safe against ->readpage(). This
1226          * should also serve to lock out allocation from a shared
1227          * writeable region.
1228          */
1229         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1230
1231         ret = ocfs2_get_clusters(inode, wc.w_cpos, &phys, NULL, NULL);
1232         if (ret) {
1233                 mlog_errno(ret);
1234                 goto out_meta;
1235         }
1236
1237         /* phys == 0 means that allocation is required. */
1238         if (phys == 0) {
1239                 ret = ocfs2_lock_allocators(inode, di, 1, &data_ac, &meta_ac);
1240                 if (ret) {
1241                         mlog_errno(ret);
1242                         goto out_meta;
1243                 }
1244
1245                 credits = ocfs2_calc_extend_credits(inode->i_sb, di, 1);
1246         }
1247
1248         ret = ocfs2_data_lock(inode, 1);
1249         if (ret) {
1250                 mlog_errno(ret);
1251                 goto out_meta;
1252         }
1253
1254         handle = ocfs2_start_trans(osb, credits);
1255         if (IS_ERR(handle)) {
1256                 ret = PTR_ERR(handle);
1257                 mlog_errno(ret);
1258                 goto out_data;
1259         }
1260
1261         written = ocfs2_write(file, phys, handle, di_bh, data_ac,
1262                               meta_ac, &wc);
1263         if (written < 0) {
1264                 ret = written;
1265                 mlog_errno(ret);
1266                 goto out_commit;
1267         }
1268
1269         ret = ocfs2_journal_access(handle, inode, di_bh,
1270                                    OCFS2_JOURNAL_ACCESS_WRITE);
1271         if (ret) {
1272                 mlog_errno(ret);
1273                 goto out_commit;
1274         }
1275
1276         pos += written;
1277         if (pos > inode->i_size) {
1278                 i_size_write(inode, pos);
1279                 mark_inode_dirty(inode);
1280         }
1281         inode->i_blocks = ocfs2_inode_sector_count(inode);
1282         di->i_size = cpu_to_le64((u64)i_size_read(inode));
1283         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1284         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
1285         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
1286
1287         ret = ocfs2_journal_dirty(handle, di_bh);
1288         if (ret)
1289                 mlog_errno(ret);
1290
1291 out_commit:
1292         ocfs2_commit_trans(osb, handle);
1293
1294 out_data:
1295         ocfs2_data_unlock(inode, 1);
1296
1297 out_meta:
1298         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1299         ocfs2_meta_unlock(inode, 1);
1300
1301 out:
1302         brelse(di_bh);
1303         if (data_ac)
1304                 ocfs2_free_alloc_context(data_ac);
1305         if (meta_ac)
1306                 ocfs2_free_alloc_context(meta_ac);
1307
1308         return written ? written : ret;
1309 }
1310
1311 const struct address_space_operations ocfs2_aops = {
1312         .readpage       = ocfs2_readpage,
1313         .writepage      = ocfs2_writepage,
1314         .bmap           = ocfs2_bmap,
1315         .sync_page      = block_sync_page,
1316         .direct_IO      = ocfs2_direct_IO,
1317         .invalidatepage = ocfs2_invalidatepage,
1318         .releasepage    = ocfs2_releasepage,
1319         .migratepage    = buffer_migrate_page,
1320 };