drm/radeon: add new RS880 pci id
[linux-2.6.git] / 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 #include <linux/mpage.h>
30 #include <linux/quotaops.h>
31
32 #define MLOG_MASK_PREFIX ML_FILE_IO
33 #include <cluster/masklog.h>
34
35 #include "ocfs2.h"
36
37 #include "alloc.h"
38 #include "aops.h"
39 #include "dlmglue.h"
40 #include "extent_map.h"
41 #include "file.h"
42 #include "inode.h"
43 #include "journal.h"
44 #include "suballoc.h"
45 #include "super.h"
46 #include "symlink.h"
47 #include "refcounttree.h"
48
49 #include "buffer_head_io.h"
50
51 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
52                                    struct buffer_head *bh_result, int create)
53 {
54         int err = -EIO;
55         int status;
56         struct ocfs2_dinode *fe = NULL;
57         struct buffer_head *bh = NULL;
58         struct buffer_head *buffer_cache_bh = NULL;
59         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
60         void *kaddr;
61
62         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
63                    (unsigned long long)iblock, bh_result, create);
64
65         BUG_ON(ocfs2_inode_is_fast_symlink(inode));
66
67         if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
68                 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
69                      (unsigned long long)iblock);
70                 goto bail;
71         }
72
73         status = ocfs2_read_inode_block(inode, &bh);
74         if (status < 0) {
75                 mlog_errno(status);
76                 goto bail;
77         }
78         fe = (struct ocfs2_dinode *) bh->b_data;
79
80         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
81                                                     le32_to_cpu(fe->i_clusters))) {
82                 mlog(ML_ERROR, "block offset is outside the allocated size: "
83                      "%llu\n", (unsigned long long)iblock);
84                 goto bail;
85         }
86
87         /* We don't use the page cache to create symlink data, so if
88          * need be, copy it over from the buffer cache. */
89         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
90                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
91                             iblock;
92                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
93                 if (!buffer_cache_bh) {
94                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
95                         goto bail;
96                 }
97
98                 /* we haven't locked out transactions, so a commit
99                  * could've happened. Since we've got a reference on
100                  * the bh, even if it commits while we're doing the
101                  * copy, the data is still good. */
102                 if (buffer_jbd(buffer_cache_bh)
103                     && ocfs2_inode_is_new(inode)) {
104                         kaddr = kmap_atomic(bh_result->b_page, KM_USER0);
105                         if (!kaddr) {
106                                 mlog(ML_ERROR, "couldn't kmap!\n");
107                                 goto bail;
108                         }
109                         memcpy(kaddr + (bh_result->b_size * iblock),
110                                buffer_cache_bh->b_data,
111                                bh_result->b_size);
112                         kunmap_atomic(kaddr, KM_USER0);
113                         set_buffer_uptodate(bh_result);
114                 }
115                 brelse(buffer_cache_bh);
116         }
117
118         map_bh(bh_result, inode->i_sb,
119                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
120
121         err = 0;
122
123 bail:
124         brelse(bh);
125
126         mlog_exit(err);
127         return err;
128 }
129
130 int ocfs2_get_block(struct inode *inode, sector_t iblock,
131                     struct buffer_head *bh_result, int create)
132 {
133         int err = 0;
134         unsigned int ext_flags;
135         u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
136         u64 p_blkno, count, past_eof;
137         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
138
139         mlog_entry("(0x%p, %llu, 0x%p, %d)\n", inode,
140                    (unsigned long long)iblock, bh_result, create);
141
142         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
143                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
144                      inode, inode->i_ino);
145
146         if (S_ISLNK(inode->i_mode)) {
147                 /* this always does I/O for some reason. */
148                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
149                 goto bail;
150         }
151
152         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
153                                           &ext_flags);
154         if (err) {
155                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
156                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
157                      (unsigned long long)p_blkno);
158                 goto bail;
159         }
160
161         if (max_blocks < count)
162                 count = max_blocks;
163
164         /*
165          * ocfs2 never allocates in this function - the only time we
166          * need to use BH_New is when we're extending i_size on a file
167          * system which doesn't support holes, in which case BH_New
168          * allows block_prepare_write() to zero.
169          *
170          * If we see this on a sparse file system, then a truncate has
171          * raced us and removed the cluster. In this case, we clear
172          * the buffers dirty and uptodate bits and let the buffer code
173          * ignore it as a hole.
174          */
175         if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
176                 clear_buffer_dirty(bh_result);
177                 clear_buffer_uptodate(bh_result);
178                 goto bail;
179         }
180
181         /* Treat the unwritten extent as a hole for zeroing purposes. */
182         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
183                 map_bh(bh_result, inode->i_sb, p_blkno);
184
185         bh_result->b_size = count << inode->i_blkbits;
186
187         if (!ocfs2_sparse_alloc(osb)) {
188                 if (p_blkno == 0) {
189                         err = -EIO;
190                         mlog(ML_ERROR,
191                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
192                              (unsigned long long)iblock,
193                              (unsigned long long)p_blkno,
194                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
195                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
196                         dump_stack();
197                         goto bail;
198                 }
199
200                 past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
201                 mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino,
202                      (unsigned long long)past_eof);
203
204                 if (create && (iblock >= past_eof))
205                         set_buffer_new(bh_result);
206         }
207
208 bail:
209         if (err < 0)
210                 err = -EIO;
211
212         mlog_exit(err);
213         return err;
214 }
215
216 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
217                            struct buffer_head *di_bh)
218 {
219         void *kaddr;
220         loff_t size;
221         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
222
223         if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
224                 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag",
225                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
226                 return -EROFS;
227         }
228
229         size = i_size_read(inode);
230
231         if (size > PAGE_CACHE_SIZE ||
232             size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
233                 ocfs2_error(inode->i_sb,
234                             "Inode %llu has with inline data has bad size: %Lu",
235                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
236                             (unsigned long long)size);
237                 return -EROFS;
238         }
239
240         kaddr = kmap_atomic(page, KM_USER0);
241         if (size)
242                 memcpy(kaddr, di->id2.i_data.id_data, size);
243         /* Clear the remaining part of the page */
244         memset(kaddr + size, 0, PAGE_CACHE_SIZE - size);
245         flush_dcache_page(page);
246         kunmap_atomic(kaddr, KM_USER0);
247
248         SetPageUptodate(page);
249
250         return 0;
251 }
252
253 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
254 {
255         int ret;
256         struct buffer_head *di_bh = NULL;
257
258         BUG_ON(!PageLocked(page));
259         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
260
261         ret = ocfs2_read_inode_block(inode, &di_bh);
262         if (ret) {
263                 mlog_errno(ret);
264                 goto out;
265         }
266
267         ret = ocfs2_read_inline_data(inode, page, di_bh);
268 out:
269         unlock_page(page);
270
271         brelse(di_bh);
272         return ret;
273 }
274
275 static int ocfs2_readpage(struct file *file, struct page *page)
276 {
277         struct inode *inode = page->mapping->host;
278         struct ocfs2_inode_info *oi = OCFS2_I(inode);
279         loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
280         int ret, unlock = 1;
281
282         mlog_entry("(0x%p, %lu)\n", file, (page ? page->index : 0));
283
284         ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
285         if (ret != 0) {
286                 if (ret == AOP_TRUNCATED_PAGE)
287                         unlock = 0;
288                 mlog_errno(ret);
289                 goto out;
290         }
291
292         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
293                 ret = AOP_TRUNCATED_PAGE;
294                 goto out_inode_unlock;
295         }
296
297         /*
298          * i_size might have just been updated as we grabed the meta lock.  We
299          * might now be discovering a truncate that hit on another node.
300          * block_read_full_page->get_block freaks out if it is asked to read
301          * beyond the end of a file, so we check here.  Callers
302          * (generic_file_read, vm_ops->fault) are clever enough to check i_size
303          * and notice that the page they just read isn't needed.
304          *
305          * XXX sys_readahead() seems to get that wrong?
306          */
307         if (start >= i_size_read(inode)) {
308                 zero_user(page, 0, PAGE_SIZE);
309                 SetPageUptodate(page);
310                 ret = 0;
311                 goto out_alloc;
312         }
313
314         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
315                 ret = ocfs2_readpage_inline(inode, page);
316         else
317                 ret = block_read_full_page(page, ocfs2_get_block);
318         unlock = 0;
319
320 out_alloc:
321         up_read(&OCFS2_I(inode)->ip_alloc_sem);
322 out_inode_unlock:
323         ocfs2_inode_unlock(inode, 0);
324 out:
325         if (unlock)
326                 unlock_page(page);
327         mlog_exit(ret);
328         return ret;
329 }
330
331 /*
332  * This is used only for read-ahead. Failures or difficult to handle
333  * situations are safe to ignore.
334  *
335  * Right now, we don't bother with BH_Boundary - in-inode extent lists
336  * are quite large (243 extents on 4k blocks), so most inodes don't
337  * grow out to a tree. If need be, detecting boundary extents could
338  * trivially be added in a future version of ocfs2_get_block().
339  */
340 static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
341                            struct list_head *pages, unsigned nr_pages)
342 {
343         int ret, err = -EIO;
344         struct inode *inode = mapping->host;
345         struct ocfs2_inode_info *oi = OCFS2_I(inode);
346         loff_t start;
347         struct page *last;
348
349         /*
350          * Use the nonblocking flag for the dlm code to avoid page
351          * lock inversion, but don't bother with retrying.
352          */
353         ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
354         if (ret)
355                 return err;
356
357         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
358                 ocfs2_inode_unlock(inode, 0);
359                 return err;
360         }
361
362         /*
363          * Don't bother with inline-data. There isn't anything
364          * to read-ahead in that case anyway...
365          */
366         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
367                 goto out_unlock;
368
369         /*
370          * Check whether a remote node truncated this file - we just
371          * drop out in that case as it's not worth handling here.
372          */
373         last = list_entry(pages->prev, struct page, lru);
374         start = (loff_t)last->index << PAGE_CACHE_SHIFT;
375         if (start >= i_size_read(inode))
376                 goto out_unlock;
377
378         err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
379
380 out_unlock:
381         up_read(&oi->ip_alloc_sem);
382         ocfs2_inode_unlock(inode, 0);
383
384         return err;
385 }
386
387 /* Note: Because we don't support holes, our allocation has
388  * already happened (allocation writes zeros to the file data)
389  * so we don't have to worry about ordered writes in
390  * ocfs2_writepage.
391  *
392  * ->writepage is called during the process of invalidating the page cache
393  * during blocked lock processing.  It can't block on any cluster locks
394  * to during block mapping.  It's relying on the fact that the block
395  * mapping can't have disappeared under the dirty pages that it is
396  * being asked to write back.
397  */
398 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
399 {
400         int ret;
401
402         mlog_entry("(0x%p)\n", page);
403
404         ret = block_write_full_page(page, ocfs2_get_block, wbc);
405
406         mlog_exit(ret);
407
408         return ret;
409 }
410
411 /*
412  * This is called from ocfs2_write_zero_page() which has handled it's
413  * own cluster locking and has ensured allocation exists for those
414  * blocks to be written.
415  */
416 int ocfs2_prepare_write_nolock(struct inode *inode, struct page *page,
417                                unsigned from, unsigned to)
418 {
419         int ret;
420
421         ret = block_prepare_write(page, from, to, ocfs2_get_block);
422
423         return ret;
424 }
425
426 /* Taken from ext3. We don't necessarily need the full blown
427  * functionality yet, but IMHO it's better to cut and paste the whole
428  * thing so we can avoid introducing our own bugs (and easily pick up
429  * their fixes when they happen) --Mark */
430 int walk_page_buffers(  handle_t *handle,
431                         struct buffer_head *head,
432                         unsigned from,
433                         unsigned to,
434                         int *partial,
435                         int (*fn)(      handle_t *handle,
436                                         struct buffer_head *bh))
437 {
438         struct buffer_head *bh;
439         unsigned block_start, block_end;
440         unsigned blocksize = head->b_size;
441         int err, ret = 0;
442         struct buffer_head *next;
443
444         for (   bh = head, block_start = 0;
445                 ret == 0 && (bh != head || !block_start);
446                 block_start = block_end, bh = next)
447         {
448                 next = bh->b_this_page;
449                 block_end = block_start + blocksize;
450                 if (block_end <= from || block_start >= to) {
451                         if (partial && !buffer_uptodate(bh))
452                                 *partial = 1;
453                         continue;
454                 }
455                 err = (*fn)(handle, bh);
456                 if (!ret)
457                         ret = err;
458         }
459         return ret;
460 }
461
462 handle_t *ocfs2_start_walk_page_trans(struct inode *inode,
463                                                          struct page *page,
464                                                          unsigned from,
465                                                          unsigned to)
466 {
467         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
468         handle_t *handle;
469         int ret = 0;
470
471         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
472         if (IS_ERR(handle)) {
473                 ret = -ENOMEM;
474                 mlog_errno(ret);
475                 goto out;
476         }
477
478         if (ocfs2_should_order_data(inode)) {
479                 ret = ocfs2_jbd2_file_inode(handle, inode);
480                 if (ret < 0)
481                         mlog_errno(ret);
482         }
483 out:
484         if (ret) {
485                 if (!IS_ERR(handle))
486                         ocfs2_commit_trans(osb, handle);
487                 handle = ERR_PTR(ret);
488         }
489         return handle;
490 }
491
492 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
493 {
494         sector_t status;
495         u64 p_blkno = 0;
496         int err = 0;
497         struct inode *inode = mapping->host;
498
499         mlog_entry("(block = %llu)\n", (unsigned long long)block);
500
501         /* We don't need to lock journal system files, since they aren't
502          * accessed concurrently from multiple nodes.
503          */
504         if (!INODE_JOURNAL(inode)) {
505                 err = ocfs2_inode_lock(inode, NULL, 0);
506                 if (err) {
507                         if (err != -ENOENT)
508                                 mlog_errno(err);
509                         goto bail;
510                 }
511                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
512         }
513
514         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
515                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
516                                                   NULL);
517
518         if (!INODE_JOURNAL(inode)) {
519                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
520                 ocfs2_inode_unlock(inode, 0);
521         }
522
523         if (err) {
524                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
525                      (unsigned long long)block);
526                 mlog_errno(err);
527                 goto bail;
528         }
529
530 bail:
531         status = err ? 0 : p_blkno;
532
533         mlog_exit((int)status);
534
535         return status;
536 }
537
538 /*
539  * TODO: Make this into a generic get_blocks function.
540  *
541  * From do_direct_io in direct-io.c:
542  *  "So what we do is to permit the ->get_blocks function to populate
543  *   bh.b_size with the size of IO which is permitted at this offset and
544  *   this i_blkbits."
545  *
546  * This function is called directly from get_more_blocks in direct-io.c.
547  *
548  * called like this: dio->get_blocks(dio->inode, fs_startblk,
549  *                                      fs_count, map_bh, dio->rw == WRITE);
550  *
551  * Note that we never bother to allocate blocks here, and thus ignore the
552  * create argument.
553  */
554 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
555                                      struct buffer_head *bh_result, int create)
556 {
557         int ret;
558         u64 p_blkno, inode_blocks, contig_blocks;
559         unsigned int ext_flags;
560         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
561         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
562
563         /* This function won't even be called if the request isn't all
564          * nicely aligned and of the right size, so there's no need
565          * for us to check any of that. */
566
567         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
568
569         /* This figures out the size of the next contiguous block, and
570          * our logical offset */
571         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
572                                           &contig_blocks, &ext_flags);
573         if (ret) {
574                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
575                      (unsigned long long)iblock);
576                 ret = -EIO;
577                 goto bail;
578         }
579
580         /* We should already CoW the refcounted extent. */
581         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
582         /*
583          * get_more_blocks() expects us to describe a hole by clearing
584          * the mapped bit on bh_result().
585          *
586          * Consider an unwritten extent as a hole.
587          */
588         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
589                 map_bh(bh_result, inode->i_sb, p_blkno);
590         else
591                 clear_buffer_mapped(bh_result);
592
593         /* make sure we don't map more than max_blocks blocks here as
594            that's all the kernel will handle at this point. */
595         if (max_blocks < contig_blocks)
596                 contig_blocks = max_blocks;
597         bh_result->b_size = contig_blocks << blocksize_bits;
598 bail:
599         return ret;
600 }
601
602 /*
603  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
604  * particularly interested in the aio/dio case.  Like the core uses
605  * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from
606  * truncation on another.
607  */
608 static void ocfs2_dio_end_io(struct kiocb *iocb,
609                              loff_t offset,
610                              ssize_t bytes,
611                              void *private)
612 {
613         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
614         int level;
615
616         /* this io's submitter should not have unlocked this before we could */
617         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
618
619         ocfs2_iocb_clear_rw_locked(iocb);
620
621         level = ocfs2_iocb_rw_locked_level(iocb);
622         if (!level)
623                 up_read(&inode->i_alloc_sem);
624         ocfs2_rw_unlock(inode, level);
625 }
626
627 /*
628  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
629  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
630  * do journalled data.
631  */
632 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
633 {
634         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
635
636         jbd2_journal_invalidatepage(journal, page, offset);
637 }
638
639 static int ocfs2_releasepage(struct page *page, gfp_t wait)
640 {
641         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
642
643         if (!page_has_buffers(page))
644                 return 0;
645         return jbd2_journal_try_to_free_buffers(journal, page, wait);
646 }
647
648 static ssize_t ocfs2_direct_IO(int rw,
649                                struct kiocb *iocb,
650                                const struct iovec *iov,
651                                loff_t offset,
652                                unsigned long nr_segs)
653 {
654         struct file *file = iocb->ki_filp;
655         struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
656         int ret;
657
658         mlog_entry_void();
659
660         /*
661          * Fallback to buffered I/O if we see an inode without
662          * extents.
663          */
664         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
665                 return 0;
666
667         /* Fallback to buffered I/O if we are appending. */
668         if (i_size_read(inode) <= offset)
669                 return 0;
670
671         ret = blockdev_direct_IO_no_locking(rw, iocb, inode,
672                                             inode->i_sb->s_bdev, iov, offset,
673                                             nr_segs,
674                                             ocfs2_direct_IO_get_blocks,
675                                             ocfs2_dio_end_io);
676
677         mlog_exit(ret);
678         return ret;
679 }
680
681 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
682                                             u32 cpos,
683                                             unsigned int *start,
684                                             unsigned int *end)
685 {
686         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
687
688         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
689                 unsigned int cpp;
690
691                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
692
693                 cluster_start = cpos % cpp;
694                 cluster_start = cluster_start << osb->s_clustersize_bits;
695
696                 cluster_end = cluster_start + osb->s_clustersize;
697         }
698
699         BUG_ON(cluster_start > PAGE_SIZE);
700         BUG_ON(cluster_end > PAGE_SIZE);
701
702         if (start)
703                 *start = cluster_start;
704         if (end)
705                 *end = cluster_end;
706 }
707
708 /*
709  * 'from' and 'to' are the region in the page to avoid zeroing.
710  *
711  * If pagesize > clustersize, this function will avoid zeroing outside
712  * of the cluster boundary.
713  *
714  * from == to == 0 is code for "zero the entire cluster region"
715  */
716 static void ocfs2_clear_page_regions(struct page *page,
717                                      struct ocfs2_super *osb, u32 cpos,
718                                      unsigned from, unsigned to)
719 {
720         void *kaddr;
721         unsigned int cluster_start, cluster_end;
722
723         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
724
725         kaddr = kmap_atomic(page, KM_USER0);
726
727         if (from || to) {
728                 if (from > cluster_start)
729                         memset(kaddr + cluster_start, 0, from - cluster_start);
730                 if (to < cluster_end)
731                         memset(kaddr + to, 0, cluster_end - to);
732         } else {
733                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
734         }
735
736         kunmap_atomic(kaddr, KM_USER0);
737 }
738
739 /*
740  * Nonsparse file systems fully allocate before we get to the write
741  * code. This prevents ocfs2_write() from tagging the write as an
742  * allocating one, which means ocfs2_map_page_blocks() might try to
743  * read-in the blocks at the tail of our file. Avoid reading them by
744  * testing i_size against each block offset.
745  */
746 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
747                                  unsigned int block_start)
748 {
749         u64 offset = page_offset(page) + block_start;
750
751         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
752                 return 1;
753
754         if (i_size_read(inode) > offset)
755                 return 1;
756
757         return 0;
758 }
759
760 /*
761  * Some of this taken from block_prepare_write(). We already have our
762  * mapping by now though, and the entire write will be allocating or
763  * it won't, so not much need to use BH_New.
764  *
765  * This will also skip zeroing, which is handled externally.
766  */
767 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
768                           struct inode *inode, unsigned int from,
769                           unsigned int to, int new)
770 {
771         int ret = 0;
772         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
773         unsigned int block_end, block_start;
774         unsigned int bsize = 1 << inode->i_blkbits;
775
776         if (!page_has_buffers(page))
777                 create_empty_buffers(page, bsize, 0);
778
779         head = page_buffers(page);
780         for (bh = head, block_start = 0; bh != head || !block_start;
781              bh = bh->b_this_page, block_start += bsize) {
782                 block_end = block_start + bsize;
783
784                 clear_buffer_new(bh);
785
786                 /*
787                  * Ignore blocks outside of our i/o range -
788                  * they may belong to unallocated clusters.
789                  */
790                 if (block_start >= to || block_end <= from) {
791                         if (PageUptodate(page))
792                                 set_buffer_uptodate(bh);
793                         continue;
794                 }
795
796                 /*
797                  * For an allocating write with cluster size >= page
798                  * size, we always write the entire page.
799                  */
800                 if (new)
801                         set_buffer_new(bh);
802
803                 if (!buffer_mapped(bh)) {
804                         map_bh(bh, inode->i_sb, *p_blkno);
805                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
806                 }
807
808                 if (PageUptodate(page)) {
809                         if (!buffer_uptodate(bh))
810                                 set_buffer_uptodate(bh);
811                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
812                            !buffer_new(bh) &&
813                            ocfs2_should_read_blk(inode, page, block_start) &&
814                            (block_start < from || block_end > to)) {
815                         ll_rw_block(READ, 1, &bh);
816                         *wait_bh++=bh;
817                 }
818
819                 *p_blkno = *p_blkno + 1;
820         }
821
822         /*
823          * If we issued read requests - let them complete.
824          */
825         while(wait_bh > wait) {
826                 wait_on_buffer(*--wait_bh);
827                 if (!buffer_uptodate(*wait_bh))
828                         ret = -EIO;
829         }
830
831         if (ret == 0 || !new)
832                 return ret;
833
834         /*
835          * If we get -EIO above, zero out any newly allocated blocks
836          * to avoid exposing stale data.
837          */
838         bh = head;
839         block_start = 0;
840         do {
841                 block_end = block_start + bsize;
842                 if (block_end <= from)
843                         goto next_bh;
844                 if (block_start >= to)
845                         break;
846
847                 zero_user(page, block_start, bh->b_size);
848                 set_buffer_uptodate(bh);
849                 mark_buffer_dirty(bh);
850
851 next_bh:
852                 block_start = block_end;
853                 bh = bh->b_this_page;
854         } while (bh != head);
855
856         return ret;
857 }
858
859 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
860 #define OCFS2_MAX_CTXT_PAGES    1
861 #else
862 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
863 #endif
864
865 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
866
867 /*
868  * Describe the state of a single cluster to be written to.
869  */
870 struct ocfs2_write_cluster_desc {
871         u32             c_cpos;
872         u32             c_phys;
873         /*
874          * Give this a unique field because c_phys eventually gets
875          * filled.
876          */
877         unsigned        c_new;
878         unsigned        c_unwritten;
879         unsigned        c_needs_zero;
880 };
881
882 struct ocfs2_write_ctxt {
883         /* Logical cluster position / len of write */
884         u32                             w_cpos;
885         u32                             w_clen;
886
887         /* First cluster allocated in a nonsparse extend */
888         u32                             w_first_new_cpos;
889
890         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
891
892         /*
893          * This is true if page_size > cluster_size.
894          *
895          * It triggers a set of special cases during write which might
896          * have to deal with allocating writes to partial pages.
897          */
898         unsigned int                    w_large_pages;
899
900         /*
901          * Pages involved in this write.
902          *
903          * w_target_page is the page being written to by the user.
904          *
905          * w_pages is an array of pages which always contains
906          * w_target_page, and in the case of an allocating write with
907          * page_size < cluster size, it will contain zero'd and mapped
908          * pages adjacent to w_target_page which need to be written
909          * out in so that future reads from that region will get
910          * zero's.
911          */
912         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
913         unsigned int                    w_num_pages;
914         struct page                     *w_target_page;
915
916         /*
917          * ocfs2_write_end() uses this to know what the real range to
918          * write in the target should be.
919          */
920         unsigned int                    w_target_from;
921         unsigned int                    w_target_to;
922
923         /*
924          * We could use journal_current_handle() but this is cleaner,
925          * IMHO -Mark
926          */
927         handle_t                        *w_handle;
928
929         struct buffer_head              *w_di_bh;
930
931         struct ocfs2_cached_dealloc_ctxt w_dealloc;
932 };
933
934 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
935 {
936         int i;
937
938         for(i = 0; i < num_pages; i++) {
939                 if (pages[i]) {
940                         unlock_page(pages[i]);
941                         mark_page_accessed(pages[i]);
942                         page_cache_release(pages[i]);
943                 }
944         }
945 }
946
947 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
948 {
949         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
950
951         brelse(wc->w_di_bh);
952         kfree(wc);
953 }
954
955 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
956                                   struct ocfs2_super *osb, loff_t pos,
957                                   unsigned len, struct buffer_head *di_bh)
958 {
959         u32 cend;
960         struct ocfs2_write_ctxt *wc;
961
962         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
963         if (!wc)
964                 return -ENOMEM;
965
966         wc->w_cpos = pos >> osb->s_clustersize_bits;
967         wc->w_first_new_cpos = UINT_MAX;
968         cend = (pos + len - 1) >> osb->s_clustersize_bits;
969         wc->w_clen = cend - wc->w_cpos + 1;
970         get_bh(di_bh);
971         wc->w_di_bh = di_bh;
972
973         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
974                 wc->w_large_pages = 1;
975         else
976                 wc->w_large_pages = 0;
977
978         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
979
980         *wcp = wc;
981
982         return 0;
983 }
984
985 /*
986  * If a page has any new buffers, zero them out here, and mark them uptodate
987  * and dirty so they'll be written out (in order to prevent uninitialised
988  * block data from leaking). And clear the new bit.
989  */
990 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
991 {
992         unsigned int block_start, block_end;
993         struct buffer_head *head, *bh;
994
995         BUG_ON(!PageLocked(page));
996         if (!page_has_buffers(page))
997                 return;
998
999         bh = head = page_buffers(page);
1000         block_start = 0;
1001         do {
1002                 block_end = block_start + bh->b_size;
1003
1004                 if (buffer_new(bh)) {
1005                         if (block_end > from && block_start < to) {
1006                                 if (!PageUptodate(page)) {
1007                                         unsigned start, end;
1008
1009                                         start = max(from, block_start);
1010                                         end = min(to, block_end);
1011
1012                                         zero_user_segment(page, start, end);
1013                                         set_buffer_uptodate(bh);
1014                                 }
1015
1016                                 clear_buffer_new(bh);
1017                                 mark_buffer_dirty(bh);
1018                         }
1019                 }
1020
1021                 block_start = block_end;
1022                 bh = bh->b_this_page;
1023         } while (bh != head);
1024 }
1025
1026 /*
1027  * Only called when we have a failure during allocating write to write
1028  * zero's to the newly allocated region.
1029  */
1030 static void ocfs2_write_failure(struct inode *inode,
1031                                 struct ocfs2_write_ctxt *wc,
1032                                 loff_t user_pos, unsigned user_len)
1033 {
1034         int i;
1035         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
1036                 to = user_pos + user_len;
1037         struct page *tmppage;
1038
1039         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
1040
1041         for(i = 0; i < wc->w_num_pages; i++) {
1042                 tmppage = wc->w_pages[i];
1043
1044                 if (page_has_buffers(tmppage)) {
1045                         if (ocfs2_should_order_data(inode))
1046                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1047
1048                         block_commit_write(tmppage, from, to);
1049                 }
1050         }
1051 }
1052
1053 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1054                                         struct ocfs2_write_ctxt *wc,
1055                                         struct page *page, u32 cpos,
1056                                         loff_t user_pos, unsigned user_len,
1057                                         int new)
1058 {
1059         int ret;
1060         unsigned int map_from = 0, map_to = 0;
1061         unsigned int cluster_start, cluster_end;
1062         unsigned int user_data_from = 0, user_data_to = 0;
1063
1064         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1065                                         &cluster_start, &cluster_end);
1066
1067         if (page == wc->w_target_page) {
1068                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1069                 map_to = map_from + user_len;
1070
1071                 if (new)
1072                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1073                                                     cluster_start, cluster_end,
1074                                                     new);
1075                 else
1076                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1077                                                     map_from, map_to, new);
1078                 if (ret) {
1079                         mlog_errno(ret);
1080                         goto out;
1081                 }
1082
1083                 user_data_from = map_from;
1084                 user_data_to = map_to;
1085                 if (new) {
1086                         map_from = cluster_start;
1087                         map_to = cluster_end;
1088                 }
1089         } else {
1090                 /*
1091                  * If we haven't allocated the new page yet, we
1092                  * shouldn't be writing it out without copying user
1093                  * data. This is likely a math error from the caller.
1094                  */
1095                 BUG_ON(!new);
1096
1097                 map_from = cluster_start;
1098                 map_to = cluster_end;
1099
1100                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1101                                             cluster_start, cluster_end, new);
1102                 if (ret) {
1103                         mlog_errno(ret);
1104                         goto out;
1105                 }
1106         }
1107
1108         /*
1109          * Parts of newly allocated pages need to be zero'd.
1110          *
1111          * Above, we have also rewritten 'to' and 'from' - as far as
1112          * the rest of the function is concerned, the entire cluster
1113          * range inside of a page needs to be written.
1114          *
1115          * We can skip this if the page is up to date - it's already
1116          * been zero'd from being read in as a hole.
1117          */
1118         if (new && !PageUptodate(page))
1119                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1120                                          cpos, user_data_from, user_data_to);
1121
1122         flush_dcache_page(page);
1123
1124 out:
1125         return ret;
1126 }
1127
1128 /*
1129  * This function will only grab one clusters worth of pages.
1130  */
1131 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1132                                       struct ocfs2_write_ctxt *wc,
1133                                       u32 cpos, loff_t user_pos, int new,
1134                                       struct page *mmap_page)
1135 {
1136         int ret = 0, i;
1137         unsigned long start, target_index, index;
1138         struct inode *inode = mapping->host;
1139
1140         target_index = user_pos >> PAGE_CACHE_SHIFT;
1141
1142         /*
1143          * Figure out how many pages we'll be manipulating here. For
1144          * non allocating write, we just change the one
1145          * page. Otherwise, we'll need a whole clusters worth.
1146          */
1147         if (new) {
1148                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1149                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1150         } else {
1151                 wc->w_num_pages = 1;
1152                 start = target_index;
1153         }
1154
1155         for(i = 0; i < wc->w_num_pages; i++) {
1156                 index = start + i;
1157
1158                 if (index == target_index && mmap_page) {
1159                         /*
1160                          * ocfs2_pagemkwrite() is a little different
1161                          * and wants us to directly use the page
1162                          * passed in.
1163                          */
1164                         lock_page(mmap_page);
1165
1166                         if (mmap_page->mapping != mapping) {
1167                                 unlock_page(mmap_page);
1168                                 /*
1169                                  * Sanity check - the locking in
1170                                  * ocfs2_pagemkwrite() should ensure
1171                                  * that this code doesn't trigger.
1172                                  */
1173                                 ret = -EINVAL;
1174                                 mlog_errno(ret);
1175                                 goto out;
1176                         }
1177
1178                         page_cache_get(mmap_page);
1179                         wc->w_pages[i] = mmap_page;
1180                 } else {
1181                         wc->w_pages[i] = find_or_create_page(mapping, index,
1182                                                              GFP_NOFS);
1183                         if (!wc->w_pages[i]) {
1184                                 ret = -ENOMEM;
1185                                 mlog_errno(ret);
1186                                 goto out;
1187                         }
1188                 }
1189
1190                 if (index == target_index)
1191                         wc->w_target_page = wc->w_pages[i];
1192         }
1193 out:
1194         return ret;
1195 }
1196
1197 /*
1198  * Prepare a single cluster for write one cluster into the file.
1199  */
1200 static int ocfs2_write_cluster(struct address_space *mapping,
1201                                u32 phys, unsigned int unwritten,
1202                                unsigned int should_zero,
1203                                struct ocfs2_alloc_context *data_ac,
1204                                struct ocfs2_alloc_context *meta_ac,
1205                                struct ocfs2_write_ctxt *wc, u32 cpos,
1206                                loff_t user_pos, unsigned user_len)
1207 {
1208         int ret, i, new;
1209         u64 v_blkno, p_blkno;
1210         struct inode *inode = mapping->host;
1211         struct ocfs2_extent_tree et;
1212
1213         new = phys == 0 ? 1 : 0;
1214         if (new) {
1215                 u32 tmp_pos;
1216
1217                 /*
1218                  * This is safe to call with the page locks - it won't take
1219                  * any additional semaphores or cluster locks.
1220                  */
1221                 tmp_pos = cpos;
1222                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1223                                            &tmp_pos, 1, 0, wc->w_di_bh,
1224                                            wc->w_handle, data_ac,
1225                                            meta_ac, NULL);
1226                 /*
1227                  * This shouldn't happen because we must have already
1228                  * calculated the correct meta data allocation required. The
1229                  * internal tree allocation code should know how to increase
1230                  * transaction credits itself.
1231                  *
1232                  * If need be, we could handle -EAGAIN for a
1233                  * RESTART_TRANS here.
1234                  */
1235                 mlog_bug_on_msg(ret == -EAGAIN,
1236                                 "Inode %llu: EAGAIN return during allocation.\n",
1237                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1238                 if (ret < 0) {
1239                         mlog_errno(ret);
1240                         goto out;
1241                 }
1242         } else if (unwritten) {
1243                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1244                                               wc->w_di_bh);
1245                 ret = ocfs2_mark_extent_written(inode, &et,
1246                                                 wc->w_handle, cpos, 1, phys,
1247                                                 meta_ac, &wc->w_dealloc);
1248                 if (ret < 0) {
1249                         mlog_errno(ret);
1250                         goto out;
1251                 }
1252         }
1253
1254         if (should_zero)
1255                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1256         else
1257                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1258
1259         /*
1260          * The only reason this should fail is due to an inability to
1261          * find the extent added.
1262          */
1263         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1264                                           NULL);
1265         if (ret < 0) {
1266                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1267                             "at logical block %llu",
1268                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1269                             (unsigned long long)v_blkno);
1270                 goto out;
1271         }
1272
1273         BUG_ON(p_blkno == 0);
1274
1275         for(i = 0; i < wc->w_num_pages; i++) {
1276                 int tmpret;
1277
1278                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1279                                                       wc->w_pages[i], cpos,
1280                                                       user_pos, user_len,
1281                                                       should_zero);
1282                 if (tmpret) {
1283                         mlog_errno(tmpret);
1284                         if (ret == 0)
1285                                 ret = tmpret;
1286                 }
1287         }
1288
1289         /*
1290          * We only have cleanup to do in case of allocating write.
1291          */
1292         if (ret && new)
1293                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1294
1295 out:
1296
1297         return ret;
1298 }
1299
1300 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1301                                        struct ocfs2_alloc_context *data_ac,
1302                                        struct ocfs2_alloc_context *meta_ac,
1303                                        struct ocfs2_write_ctxt *wc,
1304                                        loff_t pos, unsigned len)
1305 {
1306         int ret, i;
1307         loff_t cluster_off;
1308         unsigned int local_len = len;
1309         struct ocfs2_write_cluster_desc *desc;
1310         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1311
1312         for (i = 0; i < wc->w_clen; i++) {
1313                 desc = &wc->w_desc[i];
1314
1315                 /*
1316                  * We have to make sure that the total write passed in
1317                  * doesn't extend past a single cluster.
1318                  */
1319                 local_len = len;
1320                 cluster_off = pos & (osb->s_clustersize - 1);
1321                 if ((cluster_off + local_len) > osb->s_clustersize)
1322                         local_len = osb->s_clustersize - cluster_off;
1323
1324                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1325                                           desc->c_unwritten,
1326                                           desc->c_needs_zero,
1327                                           data_ac, meta_ac,
1328                                           wc, desc->c_cpos, pos, local_len);
1329                 if (ret) {
1330                         mlog_errno(ret);
1331                         goto out;
1332                 }
1333
1334                 len -= local_len;
1335                 pos += local_len;
1336         }
1337
1338         ret = 0;
1339 out:
1340         return ret;
1341 }
1342
1343 /*
1344  * ocfs2_write_end() wants to know which parts of the target page it
1345  * should complete the write on. It's easiest to compute them ahead of
1346  * time when a more complete view of the write is available.
1347  */
1348 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1349                                         struct ocfs2_write_ctxt *wc,
1350                                         loff_t pos, unsigned len, int alloc)
1351 {
1352         struct ocfs2_write_cluster_desc *desc;
1353
1354         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1355         wc->w_target_to = wc->w_target_from + len;
1356
1357         if (alloc == 0)
1358                 return;
1359
1360         /*
1361          * Allocating write - we may have different boundaries based
1362          * on page size and cluster size.
1363          *
1364          * NOTE: We can no longer compute one value from the other as
1365          * the actual write length and user provided length may be
1366          * different.
1367          */
1368
1369         if (wc->w_large_pages) {
1370                 /*
1371                  * We only care about the 1st and last cluster within
1372                  * our range and whether they should be zero'd or not. Either
1373                  * value may be extended out to the start/end of a
1374                  * newly allocated cluster.
1375                  */
1376                 desc = &wc->w_desc[0];
1377                 if (desc->c_needs_zero)
1378                         ocfs2_figure_cluster_boundaries(osb,
1379                                                         desc->c_cpos,
1380                                                         &wc->w_target_from,
1381                                                         NULL);
1382
1383                 desc = &wc->w_desc[wc->w_clen - 1];
1384                 if (desc->c_needs_zero)
1385                         ocfs2_figure_cluster_boundaries(osb,
1386                                                         desc->c_cpos,
1387                                                         NULL,
1388                                                         &wc->w_target_to);
1389         } else {
1390                 wc->w_target_from = 0;
1391                 wc->w_target_to = PAGE_CACHE_SIZE;
1392         }
1393 }
1394
1395 /*
1396  * Populate each single-cluster write descriptor in the write context
1397  * with information about the i/o to be done.
1398  *
1399  * Returns the number of clusters that will have to be allocated, as
1400  * well as a worst case estimate of the number of extent records that
1401  * would have to be created during a write to an unwritten region.
1402  */
1403 static int ocfs2_populate_write_desc(struct inode *inode,
1404                                      struct ocfs2_write_ctxt *wc,
1405                                      unsigned int *clusters_to_alloc,
1406                                      unsigned int *extents_to_split)
1407 {
1408         int ret;
1409         struct ocfs2_write_cluster_desc *desc;
1410         unsigned int num_clusters = 0;
1411         unsigned int ext_flags = 0;
1412         u32 phys = 0;
1413         int i;
1414
1415         *clusters_to_alloc = 0;
1416         *extents_to_split = 0;
1417
1418         for (i = 0; i < wc->w_clen; i++) {
1419                 desc = &wc->w_desc[i];
1420                 desc->c_cpos = wc->w_cpos + i;
1421
1422                 if (num_clusters == 0) {
1423                         /*
1424                          * Need to look up the next extent record.
1425                          */
1426                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1427                                                  &num_clusters, &ext_flags);
1428                         if (ret) {
1429                                 mlog_errno(ret);
1430                                 goto out;
1431                         }
1432
1433                         /* We should already CoW the refcountd extent. */
1434                         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1435
1436                         /*
1437                          * Assume worst case - that we're writing in
1438                          * the middle of the extent.
1439                          *
1440                          * We can assume that the write proceeds from
1441                          * left to right, in which case the extent
1442                          * insert code is smart enough to coalesce the
1443                          * next splits into the previous records created.
1444                          */
1445                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1446                                 *extents_to_split = *extents_to_split + 2;
1447                 } else if (phys) {
1448                         /*
1449                          * Only increment phys if it doesn't describe
1450                          * a hole.
1451                          */
1452                         phys++;
1453                 }
1454
1455                 /*
1456                  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1457                  * file that got extended.  w_first_new_cpos tells us
1458                  * where the newly allocated clusters are so we can
1459                  * zero them.
1460                  */
1461                 if (desc->c_cpos >= wc->w_first_new_cpos) {
1462                         BUG_ON(phys == 0);
1463                         desc->c_needs_zero = 1;
1464                 }
1465
1466                 desc->c_phys = phys;
1467                 if (phys == 0) {
1468                         desc->c_new = 1;
1469                         desc->c_needs_zero = 1;
1470                         *clusters_to_alloc = *clusters_to_alloc + 1;
1471                 }
1472
1473                 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1474                         desc->c_unwritten = 1;
1475                         desc->c_needs_zero = 1;
1476                 }
1477
1478                 num_clusters--;
1479         }
1480
1481         ret = 0;
1482 out:
1483         return ret;
1484 }
1485
1486 static int ocfs2_write_begin_inline(struct address_space *mapping,
1487                                     struct inode *inode,
1488                                     struct ocfs2_write_ctxt *wc)
1489 {
1490         int ret;
1491         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1492         struct page *page;
1493         handle_t *handle;
1494         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1495
1496         page = find_or_create_page(mapping, 0, GFP_NOFS);
1497         if (!page) {
1498                 ret = -ENOMEM;
1499                 mlog_errno(ret);
1500                 goto out;
1501         }
1502         /*
1503          * If we don't set w_num_pages then this page won't get unlocked
1504          * and freed on cleanup of the write context.
1505          */
1506         wc->w_pages[0] = wc->w_target_page = page;
1507         wc->w_num_pages = 1;
1508
1509         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1510         if (IS_ERR(handle)) {
1511                 ret = PTR_ERR(handle);
1512                 mlog_errno(ret);
1513                 goto out;
1514         }
1515
1516         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1517                                       OCFS2_JOURNAL_ACCESS_WRITE);
1518         if (ret) {
1519                 ocfs2_commit_trans(osb, handle);
1520
1521                 mlog_errno(ret);
1522                 goto out;
1523         }
1524
1525         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1526                 ocfs2_set_inode_data_inline(inode, di);
1527
1528         if (!PageUptodate(page)) {
1529                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1530                 if (ret) {
1531                         ocfs2_commit_trans(osb, handle);
1532
1533                         goto out;
1534                 }
1535         }
1536
1537         wc->w_handle = handle;
1538 out:
1539         return ret;
1540 }
1541
1542 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1543 {
1544         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1545
1546         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1547                 return 1;
1548         return 0;
1549 }
1550
1551 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1552                                           struct inode *inode, loff_t pos,
1553                                           unsigned len, struct page *mmap_page,
1554                                           struct ocfs2_write_ctxt *wc)
1555 {
1556         int ret, written = 0;
1557         loff_t end = pos + len;
1558         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1559         struct ocfs2_dinode *di = NULL;
1560
1561         mlog(0, "Inode %llu, write of %u bytes at off %llu. features: 0x%x\n",
1562              (unsigned long long)oi->ip_blkno, len, (unsigned long long)pos,
1563              oi->ip_dyn_features);
1564
1565         /*
1566          * Handle inodes which already have inline data 1st.
1567          */
1568         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1569                 if (mmap_page == NULL &&
1570                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1571                         goto do_inline_write;
1572
1573                 /*
1574                  * The write won't fit - we have to give this inode an
1575                  * inline extent list now.
1576                  */
1577                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1578                 if (ret)
1579                         mlog_errno(ret);
1580                 goto out;
1581         }
1582
1583         /*
1584          * Check whether the inode can accept inline data.
1585          */
1586         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1587                 return 0;
1588
1589         /*
1590          * Check whether the write can fit.
1591          */
1592         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1593         if (mmap_page ||
1594             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1595                 return 0;
1596
1597 do_inline_write:
1598         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1599         if (ret) {
1600                 mlog_errno(ret);
1601                 goto out;
1602         }
1603
1604         /*
1605          * This signals to the caller that the data can be written
1606          * inline.
1607          */
1608         written = 1;
1609 out:
1610         return written ? written : ret;
1611 }
1612
1613 /*
1614  * This function only does anything for file systems which can't
1615  * handle sparse files.
1616  *
1617  * What we want to do here is fill in any hole between the current end
1618  * of allocation and the end of our write. That way the rest of the
1619  * write path can treat it as an non-allocating write, which has no
1620  * special case code for sparse/nonsparse files.
1621  */
1622 static int ocfs2_expand_nonsparse_inode(struct inode *inode, loff_t pos,
1623                                         unsigned len,
1624                                         struct ocfs2_write_ctxt *wc)
1625 {
1626         int ret;
1627         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1628         loff_t newsize = pos + len;
1629
1630         if (ocfs2_sparse_alloc(osb))
1631                 return 0;
1632
1633         if (newsize <= i_size_read(inode))
1634                 return 0;
1635
1636         ret = ocfs2_extend_no_holes(inode, newsize, pos);
1637         if (ret)
1638                 mlog_errno(ret);
1639
1640         wc->w_first_new_cpos =
1641                 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1642
1643         return ret;
1644 }
1645
1646 int ocfs2_write_begin_nolock(struct address_space *mapping,
1647                              loff_t pos, unsigned len, unsigned flags,
1648                              struct page **pagep, void **fsdata,
1649                              struct buffer_head *di_bh, struct page *mmap_page)
1650 {
1651         int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1652         unsigned int clusters_to_alloc, extents_to_split;
1653         struct ocfs2_write_ctxt *wc;
1654         struct inode *inode = mapping->host;
1655         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1656         struct ocfs2_dinode *di;
1657         struct ocfs2_alloc_context *data_ac = NULL;
1658         struct ocfs2_alloc_context *meta_ac = NULL;
1659         handle_t *handle;
1660         struct ocfs2_extent_tree et;
1661
1662         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1663         if (ret) {
1664                 mlog_errno(ret);
1665                 return ret;
1666         }
1667
1668         if (ocfs2_supports_inline_data(osb)) {
1669                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1670                                                      mmap_page, wc);
1671                 if (ret == 1) {
1672                         ret = 0;
1673                         goto success;
1674                 }
1675                 if (ret < 0) {
1676                         mlog_errno(ret);
1677                         goto out;
1678                 }
1679         }
1680
1681         ret = ocfs2_expand_nonsparse_inode(inode, pos, len, wc);
1682         if (ret) {
1683                 mlog_errno(ret);
1684                 goto out;
1685         }
1686
1687         ret = ocfs2_check_range_for_refcount(inode, pos, len);
1688         if (ret < 0) {
1689                 mlog_errno(ret);
1690                 goto out;
1691         } else if (ret == 1) {
1692                 ret = ocfs2_refcount_cow(inode, di_bh,
1693                                          wc->w_cpos, wc->w_clen, UINT_MAX);
1694                 if (ret) {
1695                         mlog_errno(ret);
1696                         goto out;
1697                 }
1698         }
1699
1700         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1701                                         &extents_to_split);
1702         if (ret) {
1703                 mlog_errno(ret);
1704                 goto out;
1705         }
1706
1707         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1708
1709         /*
1710          * We set w_target_from, w_target_to here so that
1711          * ocfs2_write_end() knows which range in the target page to
1712          * write out. An allocation requires that we write the entire
1713          * cluster range.
1714          */
1715         if (clusters_to_alloc || extents_to_split) {
1716                 /*
1717                  * XXX: We are stretching the limits of
1718                  * ocfs2_lock_allocators(). It greatly over-estimates
1719                  * the work to be done.
1720                  */
1721                 mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u,"
1722                      " clusters_to_add = %u, extents_to_split = %u\n",
1723                      (unsigned long long)OCFS2_I(inode)->ip_blkno,
1724                      (long long)i_size_read(inode), le32_to_cpu(di->i_clusters),
1725                      clusters_to_alloc, extents_to_split);
1726
1727                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1728                                               wc->w_di_bh);
1729                 ret = ocfs2_lock_allocators(inode, &et,
1730                                             clusters_to_alloc, extents_to_split,
1731                                             &data_ac, &meta_ac);
1732                 if (ret) {
1733                         mlog_errno(ret);
1734                         goto out;
1735                 }
1736
1737                 credits = ocfs2_calc_extend_credits(inode->i_sb,
1738                                                     &di->id2.i_list,
1739                                                     clusters_to_alloc);
1740
1741         }
1742
1743         /*
1744          * We have to zero sparse allocated clusters, unwritten extent clusters,
1745          * and non-sparse clusters we just extended.  For non-sparse writes,
1746          * we know zeros will only be needed in the first and/or last cluster.
1747          */
1748         if (clusters_to_alloc || extents_to_split ||
1749             (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1750                             wc->w_desc[wc->w_clen - 1].c_needs_zero)))
1751                 cluster_of_pages = 1;
1752         else
1753                 cluster_of_pages = 0;
1754
1755         ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1756
1757         handle = ocfs2_start_trans(osb, credits);
1758         if (IS_ERR(handle)) {
1759                 ret = PTR_ERR(handle);
1760                 mlog_errno(ret);
1761                 goto out;
1762         }
1763
1764         wc->w_handle = handle;
1765
1766         if (clusters_to_alloc && vfs_dq_alloc_space_nodirty(inode,
1767                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc))) {
1768                 ret = -EDQUOT;
1769                 goto out_commit;
1770         }
1771         /*
1772          * We don't want this to fail in ocfs2_write_end(), so do it
1773          * here.
1774          */
1775         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1776                                       OCFS2_JOURNAL_ACCESS_WRITE);
1777         if (ret) {
1778                 mlog_errno(ret);
1779                 goto out_quota;
1780         }
1781
1782         /*
1783          * Fill our page array first. That way we've grabbed enough so
1784          * that we can zero and flush if we error after adding the
1785          * extent.
1786          */
1787         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos,
1788                                          cluster_of_pages, mmap_page);
1789         if (ret) {
1790                 mlog_errno(ret);
1791                 goto out_quota;
1792         }
1793
1794         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1795                                           len);
1796         if (ret) {
1797                 mlog_errno(ret);
1798                 goto out_quota;
1799         }
1800
1801         if (data_ac)
1802                 ocfs2_free_alloc_context(data_ac);
1803         if (meta_ac)
1804                 ocfs2_free_alloc_context(meta_ac);
1805
1806 success:
1807         *pagep = wc->w_target_page;
1808         *fsdata = wc;
1809         return 0;
1810 out_quota:
1811         if (clusters_to_alloc)
1812                 vfs_dq_free_space(inode,
1813                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1814 out_commit:
1815         ocfs2_commit_trans(osb, handle);
1816
1817 out:
1818         ocfs2_free_write_ctxt(wc);
1819
1820         if (data_ac)
1821                 ocfs2_free_alloc_context(data_ac);
1822         if (meta_ac)
1823                 ocfs2_free_alloc_context(meta_ac);
1824         return ret;
1825 }
1826
1827 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1828                              loff_t pos, unsigned len, unsigned flags,
1829                              struct page **pagep, void **fsdata)
1830 {
1831         int ret;
1832         struct buffer_head *di_bh = NULL;
1833         struct inode *inode = mapping->host;
1834
1835         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1836         if (ret) {
1837                 mlog_errno(ret);
1838                 return ret;
1839         }
1840
1841         /*
1842          * Take alloc sem here to prevent concurrent lookups. That way
1843          * the mapping, zeroing and tree manipulation within
1844          * ocfs2_write() will be safe against ->readpage(). This
1845          * should also serve to lock out allocation from a shared
1846          * writeable region.
1847          */
1848         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1849
1850         ret = ocfs2_write_begin_nolock(mapping, pos, len, flags, pagep,
1851                                        fsdata, di_bh, NULL);
1852         if (ret) {
1853                 mlog_errno(ret);
1854                 goto out_fail;
1855         }
1856
1857         brelse(di_bh);
1858
1859         return 0;
1860
1861 out_fail:
1862         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1863
1864         brelse(di_bh);
1865         ocfs2_inode_unlock(inode, 1);
1866
1867         return ret;
1868 }
1869
1870 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1871                                    unsigned len, unsigned *copied,
1872                                    struct ocfs2_dinode *di,
1873                                    struct ocfs2_write_ctxt *wc)
1874 {
1875         void *kaddr;
1876
1877         if (unlikely(*copied < len)) {
1878                 if (!PageUptodate(wc->w_target_page)) {
1879                         *copied = 0;
1880                         return;
1881                 }
1882         }
1883
1884         kaddr = kmap_atomic(wc->w_target_page, KM_USER0);
1885         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1886         kunmap_atomic(kaddr, KM_USER0);
1887
1888         mlog(0, "Data written to inode at offset %llu. "
1889              "id_count = %u, copied = %u, i_dyn_features = 0x%x\n",
1890              (unsigned long long)pos, *copied,
1891              le16_to_cpu(di->id2.i_data.id_count),
1892              le16_to_cpu(di->i_dyn_features));
1893 }
1894
1895 int ocfs2_write_end_nolock(struct address_space *mapping,
1896                            loff_t pos, unsigned len, unsigned copied,
1897                            struct page *page, void *fsdata)
1898 {
1899         int i;
1900         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1901         struct inode *inode = mapping->host;
1902         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1903         struct ocfs2_write_ctxt *wc = fsdata;
1904         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1905         handle_t *handle = wc->w_handle;
1906         struct page *tmppage;
1907
1908         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1909                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1910                 goto out_write_size;
1911         }
1912
1913         if (unlikely(copied < len)) {
1914                 if (!PageUptodate(wc->w_target_page))
1915                         copied = 0;
1916
1917                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1918                                        start+len);
1919         }
1920         flush_dcache_page(wc->w_target_page);
1921
1922         for(i = 0; i < wc->w_num_pages; i++) {
1923                 tmppage = wc->w_pages[i];
1924
1925                 if (tmppage == wc->w_target_page) {
1926                         from = wc->w_target_from;
1927                         to = wc->w_target_to;
1928
1929                         BUG_ON(from > PAGE_CACHE_SIZE ||
1930                                to > PAGE_CACHE_SIZE ||
1931                                to < from);
1932                 } else {
1933                         /*
1934                          * Pages adjacent to the target (if any) imply
1935                          * a hole-filling write in which case we want
1936                          * to flush their entire range.
1937                          */
1938                         from = 0;
1939                         to = PAGE_CACHE_SIZE;
1940                 }
1941
1942                 if (page_has_buffers(tmppage)) {
1943                         if (ocfs2_should_order_data(inode))
1944                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1945                         block_commit_write(tmppage, from, to);
1946                 }
1947         }
1948
1949 out_write_size:
1950         pos += copied;
1951         if (pos > inode->i_size) {
1952                 i_size_write(inode, pos);
1953                 mark_inode_dirty(inode);
1954         }
1955         inode->i_blocks = ocfs2_inode_sector_count(inode);
1956         di->i_size = cpu_to_le64((u64)i_size_read(inode));
1957         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1958         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
1959         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
1960         ocfs2_journal_dirty(handle, wc->w_di_bh);
1961
1962         ocfs2_commit_trans(osb, handle);
1963
1964         ocfs2_run_deallocs(osb, &wc->w_dealloc);
1965
1966         ocfs2_free_write_ctxt(wc);
1967
1968         return copied;
1969 }
1970
1971 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
1972                            loff_t pos, unsigned len, unsigned copied,
1973                            struct page *page, void *fsdata)
1974 {
1975         int ret;
1976         struct inode *inode = mapping->host;
1977
1978         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
1979
1980         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1981         ocfs2_inode_unlock(inode, 1);
1982
1983         return ret;
1984 }
1985
1986 const struct address_space_operations ocfs2_aops = {
1987         .readpage               = ocfs2_readpage,
1988         .readpages              = ocfs2_readpages,
1989         .writepage              = ocfs2_writepage,
1990         .write_begin            = ocfs2_write_begin,
1991         .write_end              = ocfs2_write_end,
1992         .bmap                   = ocfs2_bmap,
1993         .sync_page              = block_sync_page,
1994         .direct_IO              = ocfs2_direct_IO,
1995         .invalidatepage         = ocfs2_invalidatepage,
1996         .releasepage            = ocfs2_releasepage,
1997         .migratepage            = buffer_migrate_page,
1998         .is_partially_uptodate  = block_is_partially_uptodate,
1999         .error_remove_page      = generic_error_remove_page,
2000 };