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