fs: kill block_prepare_write
[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_write_begin() 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
201         past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
202         mlog(0, "Inode %lu, past_eof = %llu\n", inode->i_ino,
203              (unsigned long long)past_eof);
204         if (create && (iblock >= past_eof))
205                 set_buffer_new(bh_result);
206
207 bail:
208         if (err < 0)
209                 err = -EIO;
210
211         mlog_exit(err);
212         return err;
213 }
214
215 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
216                            struct buffer_head *di_bh)
217 {
218         void *kaddr;
219         loff_t size;
220         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
221
222         if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
223                 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag",
224                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
225                 return -EROFS;
226         }
227
228         size = i_size_read(inode);
229
230         if (size > PAGE_CACHE_SIZE ||
231             size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
232                 ocfs2_error(inode->i_sb,
233                             "Inode %llu has with inline data has bad size: %Lu",
234                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
235                             (unsigned long long)size);
236                 return -EROFS;
237         }
238
239         kaddr = kmap_atomic(page, KM_USER0);
240         if (size)
241                 memcpy(kaddr, di->id2.i_data.id_data, size);
242         /* Clear the remaining part of the page */
243         memset(kaddr + size, 0, PAGE_CACHE_SIZE - size);
244         flush_dcache_page(page);
245         kunmap_atomic(kaddr, KM_USER0);
246
247         SetPageUptodate(page);
248
249         return 0;
250 }
251
252 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
253 {
254         int ret;
255         struct buffer_head *di_bh = NULL;
256
257         BUG_ON(!PageLocked(page));
258         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
259
260         ret = ocfs2_read_inode_block(inode, &di_bh);
261         if (ret) {
262                 mlog_errno(ret);
263                 goto out;
264         }
265
266         ret = ocfs2_read_inline_data(inode, page, di_bh);
267 out:
268         unlock_page(page);
269
270         brelse(di_bh);
271         return ret;
272 }
273
274 static int ocfs2_readpage(struct file *file, struct page *page)
275 {
276         struct inode *inode = page->mapping->host;
277         struct ocfs2_inode_info *oi = OCFS2_I(inode);
278         loff_t start = (loff_t)page->index << PAGE_CACHE_SHIFT;
279         int ret, unlock = 1;
280
281         mlog_entry("(0x%p, %lu)\n", file, (page ? page->index : 0));
282
283         ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
284         if (ret != 0) {
285                 if (ret == AOP_TRUNCATED_PAGE)
286                         unlock = 0;
287                 mlog_errno(ret);
288                 goto out;
289         }
290
291         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
292                 ret = AOP_TRUNCATED_PAGE;
293                 goto out_inode_unlock;
294         }
295
296         /*
297          * i_size might have just been updated as we grabed the meta lock.  We
298          * might now be discovering a truncate that hit on another node.
299          * block_read_full_page->get_block freaks out if it is asked to read
300          * beyond the end of a file, so we check here.  Callers
301          * (generic_file_read, vm_ops->fault) are clever enough to check i_size
302          * and notice that the page they just read isn't needed.
303          *
304          * XXX sys_readahead() seems to get that wrong?
305          */
306         if (start >= i_size_read(inode)) {
307                 zero_user(page, 0, PAGE_SIZE);
308                 SetPageUptodate(page);
309                 ret = 0;
310                 goto out_alloc;
311         }
312
313         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
314                 ret = ocfs2_readpage_inline(inode, page);
315         else
316                 ret = block_read_full_page(page, ocfs2_get_block);
317         unlock = 0;
318
319 out_alloc:
320         up_read(&OCFS2_I(inode)->ip_alloc_sem);
321 out_inode_unlock:
322         ocfs2_inode_unlock(inode, 0);
323 out:
324         if (unlock)
325                 unlock_page(page);
326         mlog_exit(ret);
327         return ret;
328 }
329
330 /*
331  * This is used only for read-ahead. Failures or difficult to handle
332  * situations are safe to ignore.
333  *
334  * Right now, we don't bother with BH_Boundary - in-inode extent lists
335  * are quite large (243 extents on 4k blocks), so most inodes don't
336  * grow out to a tree. If need be, detecting boundary extents could
337  * trivially be added in a future version of ocfs2_get_block().
338  */
339 static int ocfs2_readpages(struct file *filp, struct address_space *mapping,
340                            struct list_head *pages, unsigned nr_pages)
341 {
342         int ret, err = -EIO;
343         struct inode *inode = mapping->host;
344         struct ocfs2_inode_info *oi = OCFS2_I(inode);
345         loff_t start;
346         struct page *last;
347
348         /*
349          * Use the nonblocking flag for the dlm code to avoid page
350          * lock inversion, but don't bother with retrying.
351          */
352         ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
353         if (ret)
354                 return err;
355
356         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
357                 ocfs2_inode_unlock(inode, 0);
358                 return err;
359         }
360
361         /*
362          * Don't bother with inline-data. There isn't anything
363          * to read-ahead in that case anyway...
364          */
365         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
366                 goto out_unlock;
367
368         /*
369          * Check whether a remote node truncated this file - we just
370          * drop out in that case as it's not worth handling here.
371          */
372         last = list_entry(pages->prev, struct page, lru);
373         start = (loff_t)last->index << PAGE_CACHE_SHIFT;
374         if (start >= i_size_read(inode))
375                 goto out_unlock;
376
377         err = mpage_readpages(mapping, pages, nr_pages, ocfs2_get_block);
378
379 out_unlock:
380         up_read(&oi->ip_alloc_sem);
381         ocfs2_inode_unlock(inode, 0);
382
383         return err;
384 }
385
386 /* Note: Because we don't support holes, our allocation has
387  * already happened (allocation writes zeros to the file data)
388  * so we don't have to worry about ordered writes in
389  * ocfs2_writepage.
390  *
391  * ->writepage is called during the process of invalidating the page cache
392  * during blocked lock processing.  It can't block on any cluster locks
393  * to during block mapping.  It's relying on the fact that the block
394  * mapping can't have disappeared under the dirty pages that it is
395  * being asked to write back.
396  */
397 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
398 {
399         int ret;
400
401         mlog_entry("(0x%p)\n", page);
402
403         ret = block_write_full_page(page, ocfs2_get_block, wbc);
404
405         mlog_exit(ret);
406
407         return ret;
408 }
409
410 /* Taken from ext3. We don't necessarily need the full blown
411  * functionality yet, but IMHO it's better to cut and paste the whole
412  * thing so we can avoid introducing our own bugs (and easily pick up
413  * their fixes when they happen) --Mark */
414 int walk_page_buffers(  handle_t *handle,
415                         struct buffer_head *head,
416                         unsigned from,
417                         unsigned to,
418                         int *partial,
419                         int (*fn)(      handle_t *handle,
420                                         struct buffer_head *bh))
421 {
422         struct buffer_head *bh;
423         unsigned block_start, block_end;
424         unsigned blocksize = head->b_size;
425         int err, ret = 0;
426         struct buffer_head *next;
427
428         for (   bh = head, block_start = 0;
429                 ret == 0 && (bh != head || !block_start);
430                 block_start = block_end, bh = next)
431         {
432                 next = bh->b_this_page;
433                 block_end = block_start + blocksize;
434                 if (block_end <= from || block_start >= to) {
435                         if (partial && !buffer_uptodate(bh))
436                                 *partial = 1;
437                         continue;
438                 }
439                 err = (*fn)(handle, bh);
440                 if (!ret)
441                         ret = err;
442         }
443         return ret;
444 }
445
446 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
447 {
448         sector_t status;
449         u64 p_blkno = 0;
450         int err = 0;
451         struct inode *inode = mapping->host;
452
453         mlog_entry("(block = %llu)\n", (unsigned long long)block);
454
455         /* We don't need to lock journal system files, since they aren't
456          * accessed concurrently from multiple nodes.
457          */
458         if (!INODE_JOURNAL(inode)) {
459                 err = ocfs2_inode_lock(inode, NULL, 0);
460                 if (err) {
461                         if (err != -ENOENT)
462                                 mlog_errno(err);
463                         goto bail;
464                 }
465                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
466         }
467
468         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
469                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
470                                                   NULL);
471
472         if (!INODE_JOURNAL(inode)) {
473                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
474                 ocfs2_inode_unlock(inode, 0);
475         }
476
477         if (err) {
478                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
479                      (unsigned long long)block);
480                 mlog_errno(err);
481                 goto bail;
482         }
483
484 bail:
485         status = err ? 0 : p_blkno;
486
487         mlog_exit((int)status);
488
489         return status;
490 }
491
492 /*
493  * TODO: Make this into a generic get_blocks function.
494  *
495  * From do_direct_io in direct-io.c:
496  *  "So what we do is to permit the ->get_blocks function to populate
497  *   bh.b_size with the size of IO which is permitted at this offset and
498  *   this i_blkbits."
499  *
500  * This function is called directly from get_more_blocks in direct-io.c.
501  *
502  * called like this: dio->get_blocks(dio->inode, fs_startblk,
503  *                                      fs_count, map_bh, dio->rw == WRITE);
504  *
505  * Note that we never bother to allocate blocks here, and thus ignore the
506  * create argument.
507  */
508 static int ocfs2_direct_IO_get_blocks(struct inode *inode, sector_t iblock,
509                                      struct buffer_head *bh_result, int create)
510 {
511         int ret;
512         u64 p_blkno, inode_blocks, contig_blocks;
513         unsigned int ext_flags;
514         unsigned char blocksize_bits = inode->i_sb->s_blocksize_bits;
515         unsigned long max_blocks = bh_result->b_size >> inode->i_blkbits;
516
517         /* This function won't even be called if the request isn't all
518          * nicely aligned and of the right size, so there's no need
519          * for us to check any of that. */
520
521         inode_blocks = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
522
523         /* This figures out the size of the next contiguous block, and
524          * our logical offset */
525         ret = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno,
526                                           &contig_blocks, &ext_flags);
527         if (ret) {
528                 mlog(ML_ERROR, "get_blocks() failed iblock=%llu\n",
529                      (unsigned long long)iblock);
530                 ret = -EIO;
531                 goto bail;
532         }
533
534         /* We should already CoW the refcounted extent in case of create. */
535         BUG_ON(create && (ext_flags & OCFS2_EXT_REFCOUNTED));
536
537         /*
538          * get_more_blocks() expects us to describe a hole by clearing
539          * the mapped bit on bh_result().
540          *
541          * Consider an unwritten extent as a hole.
542          */
543         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
544                 map_bh(bh_result, inode->i_sb, p_blkno);
545         else
546                 clear_buffer_mapped(bh_result);
547
548         /* make sure we don't map more than max_blocks blocks here as
549            that's all the kernel will handle at this point. */
550         if (max_blocks < contig_blocks)
551                 contig_blocks = max_blocks;
552         bh_result->b_size = contig_blocks << blocksize_bits;
553 bail:
554         return ret;
555 }
556
557 /*
558  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
559  * particularly interested in the aio/dio case.  Like the core uses
560  * i_alloc_sem, we use the rw_lock DLM lock to protect io on one node from
561  * truncation on another.
562  */
563 static void ocfs2_dio_end_io(struct kiocb *iocb,
564                              loff_t offset,
565                              ssize_t bytes,
566                              void *private,
567                              int ret,
568                              bool is_async)
569 {
570         struct inode *inode = iocb->ki_filp->f_path.dentry->d_inode;
571         int level;
572
573         /* this io's submitter should not have unlocked this before we could */
574         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
575
576         ocfs2_iocb_clear_rw_locked(iocb);
577
578         level = ocfs2_iocb_rw_locked_level(iocb);
579         if (!level)
580                 up_read(&inode->i_alloc_sem);
581         ocfs2_rw_unlock(inode, level);
582
583         if (is_async)
584                 aio_complete(iocb, ret, 0);
585 }
586
587 /*
588  * ocfs2_invalidatepage() and ocfs2_releasepage() are shamelessly stolen
589  * from ext3.  PageChecked() bits have been removed as OCFS2 does not
590  * do journalled data.
591  */
592 static void ocfs2_invalidatepage(struct page *page, unsigned long offset)
593 {
594         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
595
596         jbd2_journal_invalidatepage(journal, page, offset);
597 }
598
599 static int ocfs2_releasepage(struct page *page, gfp_t wait)
600 {
601         journal_t *journal = OCFS2_SB(page->mapping->host->i_sb)->journal->j_journal;
602
603         if (!page_has_buffers(page))
604                 return 0;
605         return jbd2_journal_try_to_free_buffers(journal, page, wait);
606 }
607
608 static ssize_t ocfs2_direct_IO(int rw,
609                                struct kiocb *iocb,
610                                const struct iovec *iov,
611                                loff_t offset,
612                                unsigned long nr_segs)
613 {
614         struct file *file = iocb->ki_filp;
615         struct inode *inode = file->f_path.dentry->d_inode->i_mapping->host;
616         int ret;
617
618         mlog_entry_void();
619
620         /*
621          * Fallback to buffered I/O if we see an inode without
622          * extents.
623          */
624         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
625                 return 0;
626
627         /* Fallback to buffered I/O if we are appending. */
628         if (i_size_read(inode) <= offset)
629                 return 0;
630
631         ret = __blockdev_direct_IO(rw, iocb, inode, inode->i_sb->s_bdev,
632                                    iov, offset, nr_segs,
633                                    ocfs2_direct_IO_get_blocks,
634                                    ocfs2_dio_end_io, NULL, 0);
635
636         mlog_exit(ret);
637         return ret;
638 }
639
640 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
641                                             u32 cpos,
642                                             unsigned int *start,
643                                             unsigned int *end)
644 {
645         unsigned int cluster_start = 0, cluster_end = PAGE_CACHE_SIZE;
646
647         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits)) {
648                 unsigned int cpp;
649
650                 cpp = 1 << (PAGE_CACHE_SHIFT - osb->s_clustersize_bits);
651
652                 cluster_start = cpos % cpp;
653                 cluster_start = cluster_start << osb->s_clustersize_bits;
654
655                 cluster_end = cluster_start + osb->s_clustersize;
656         }
657
658         BUG_ON(cluster_start > PAGE_SIZE);
659         BUG_ON(cluster_end > PAGE_SIZE);
660
661         if (start)
662                 *start = cluster_start;
663         if (end)
664                 *end = cluster_end;
665 }
666
667 /*
668  * 'from' and 'to' are the region in the page to avoid zeroing.
669  *
670  * If pagesize > clustersize, this function will avoid zeroing outside
671  * of the cluster boundary.
672  *
673  * from == to == 0 is code for "zero the entire cluster region"
674  */
675 static void ocfs2_clear_page_regions(struct page *page,
676                                      struct ocfs2_super *osb, u32 cpos,
677                                      unsigned from, unsigned to)
678 {
679         void *kaddr;
680         unsigned int cluster_start, cluster_end;
681
682         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
683
684         kaddr = kmap_atomic(page, KM_USER0);
685
686         if (from || to) {
687                 if (from > cluster_start)
688                         memset(kaddr + cluster_start, 0, from - cluster_start);
689                 if (to < cluster_end)
690                         memset(kaddr + to, 0, cluster_end - to);
691         } else {
692                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
693         }
694
695         kunmap_atomic(kaddr, KM_USER0);
696 }
697
698 /*
699  * Nonsparse file systems fully allocate before we get to the write
700  * code. This prevents ocfs2_write() from tagging the write as an
701  * allocating one, which means ocfs2_map_page_blocks() might try to
702  * read-in the blocks at the tail of our file. Avoid reading them by
703  * testing i_size against each block offset.
704  */
705 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
706                                  unsigned int block_start)
707 {
708         u64 offset = page_offset(page) + block_start;
709
710         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
711                 return 1;
712
713         if (i_size_read(inode) > offset)
714                 return 1;
715
716         return 0;
717 }
718
719 /*
720  * Some of this taken from __block_write_begin(). We already have our
721  * mapping by now though, and the entire write will be allocating or
722  * it won't, so not much need to use BH_New.
723  *
724  * This will also skip zeroing, which is handled externally.
725  */
726 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
727                           struct inode *inode, unsigned int from,
728                           unsigned int to, int new)
729 {
730         int ret = 0;
731         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
732         unsigned int block_end, block_start;
733         unsigned int bsize = 1 << inode->i_blkbits;
734
735         if (!page_has_buffers(page))
736                 create_empty_buffers(page, bsize, 0);
737
738         head = page_buffers(page);
739         for (bh = head, block_start = 0; bh != head || !block_start;
740              bh = bh->b_this_page, block_start += bsize) {
741                 block_end = block_start + bsize;
742
743                 clear_buffer_new(bh);
744
745                 /*
746                  * Ignore blocks outside of our i/o range -
747                  * they may belong to unallocated clusters.
748                  */
749                 if (block_start >= to || block_end <= from) {
750                         if (PageUptodate(page))
751                                 set_buffer_uptodate(bh);
752                         continue;
753                 }
754
755                 /*
756                  * For an allocating write with cluster size >= page
757                  * size, we always write the entire page.
758                  */
759                 if (new)
760                         set_buffer_new(bh);
761
762                 if (!buffer_mapped(bh)) {
763                         map_bh(bh, inode->i_sb, *p_blkno);
764                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
765                 }
766
767                 if (PageUptodate(page)) {
768                         if (!buffer_uptodate(bh))
769                                 set_buffer_uptodate(bh);
770                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
771                            !buffer_new(bh) &&
772                            ocfs2_should_read_blk(inode, page, block_start) &&
773                            (block_start < from || block_end > to)) {
774                         ll_rw_block(READ, 1, &bh);
775                         *wait_bh++=bh;
776                 }
777
778                 *p_blkno = *p_blkno + 1;
779         }
780
781         /*
782          * If we issued read requests - let them complete.
783          */
784         while(wait_bh > wait) {
785                 wait_on_buffer(*--wait_bh);
786                 if (!buffer_uptodate(*wait_bh))
787                         ret = -EIO;
788         }
789
790         if (ret == 0 || !new)
791                 return ret;
792
793         /*
794          * If we get -EIO above, zero out any newly allocated blocks
795          * to avoid exposing stale data.
796          */
797         bh = head;
798         block_start = 0;
799         do {
800                 block_end = block_start + bsize;
801                 if (block_end <= from)
802                         goto next_bh;
803                 if (block_start >= to)
804                         break;
805
806                 zero_user(page, block_start, bh->b_size);
807                 set_buffer_uptodate(bh);
808                 mark_buffer_dirty(bh);
809
810 next_bh:
811                 block_start = block_end;
812                 bh = bh->b_this_page;
813         } while (bh != head);
814
815         return ret;
816 }
817
818 #if (PAGE_CACHE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
819 #define OCFS2_MAX_CTXT_PAGES    1
820 #else
821 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_CACHE_SIZE)
822 #endif
823
824 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_CACHE_SIZE / OCFS2_MIN_CLUSTERSIZE)
825
826 /*
827  * Describe the state of a single cluster to be written to.
828  */
829 struct ocfs2_write_cluster_desc {
830         u32             c_cpos;
831         u32             c_phys;
832         /*
833          * Give this a unique field because c_phys eventually gets
834          * filled.
835          */
836         unsigned        c_new;
837         unsigned        c_unwritten;
838         unsigned        c_needs_zero;
839 };
840
841 struct ocfs2_write_ctxt {
842         /* Logical cluster position / len of write */
843         u32                             w_cpos;
844         u32                             w_clen;
845
846         /* First cluster allocated in a nonsparse extend */
847         u32                             w_first_new_cpos;
848
849         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
850
851         /*
852          * This is true if page_size > cluster_size.
853          *
854          * It triggers a set of special cases during write which might
855          * have to deal with allocating writes to partial pages.
856          */
857         unsigned int                    w_large_pages;
858
859         /*
860          * Pages involved in this write.
861          *
862          * w_target_page is the page being written to by the user.
863          *
864          * w_pages is an array of pages which always contains
865          * w_target_page, and in the case of an allocating write with
866          * page_size < cluster size, it will contain zero'd and mapped
867          * pages adjacent to w_target_page which need to be written
868          * out in so that future reads from that region will get
869          * zero's.
870          */
871         unsigned int                    w_num_pages;
872         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
873         struct page                     *w_target_page;
874
875         /*
876          * ocfs2_write_end() uses this to know what the real range to
877          * write in the target should be.
878          */
879         unsigned int                    w_target_from;
880         unsigned int                    w_target_to;
881
882         /*
883          * We could use journal_current_handle() but this is cleaner,
884          * IMHO -Mark
885          */
886         handle_t                        *w_handle;
887
888         struct buffer_head              *w_di_bh;
889
890         struct ocfs2_cached_dealloc_ctxt w_dealloc;
891 };
892
893 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
894 {
895         int i;
896
897         for(i = 0; i < num_pages; i++) {
898                 if (pages[i]) {
899                         unlock_page(pages[i]);
900                         mark_page_accessed(pages[i]);
901                         page_cache_release(pages[i]);
902                 }
903         }
904 }
905
906 static void ocfs2_free_write_ctxt(struct ocfs2_write_ctxt *wc)
907 {
908         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
909
910         brelse(wc->w_di_bh);
911         kfree(wc);
912 }
913
914 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
915                                   struct ocfs2_super *osb, loff_t pos,
916                                   unsigned len, struct buffer_head *di_bh)
917 {
918         u32 cend;
919         struct ocfs2_write_ctxt *wc;
920
921         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
922         if (!wc)
923                 return -ENOMEM;
924
925         wc->w_cpos = pos >> osb->s_clustersize_bits;
926         wc->w_first_new_cpos = UINT_MAX;
927         cend = (pos + len - 1) >> osb->s_clustersize_bits;
928         wc->w_clen = cend - wc->w_cpos + 1;
929         get_bh(di_bh);
930         wc->w_di_bh = di_bh;
931
932         if (unlikely(PAGE_CACHE_SHIFT > osb->s_clustersize_bits))
933                 wc->w_large_pages = 1;
934         else
935                 wc->w_large_pages = 0;
936
937         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
938
939         *wcp = wc;
940
941         return 0;
942 }
943
944 /*
945  * If a page has any new buffers, zero them out here, and mark them uptodate
946  * and dirty so they'll be written out (in order to prevent uninitialised
947  * block data from leaking). And clear the new bit.
948  */
949 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
950 {
951         unsigned int block_start, block_end;
952         struct buffer_head *head, *bh;
953
954         BUG_ON(!PageLocked(page));
955         if (!page_has_buffers(page))
956                 return;
957
958         bh = head = page_buffers(page);
959         block_start = 0;
960         do {
961                 block_end = block_start + bh->b_size;
962
963                 if (buffer_new(bh)) {
964                         if (block_end > from && block_start < to) {
965                                 if (!PageUptodate(page)) {
966                                         unsigned start, end;
967
968                                         start = max(from, block_start);
969                                         end = min(to, block_end);
970
971                                         zero_user_segment(page, start, end);
972                                         set_buffer_uptodate(bh);
973                                 }
974
975                                 clear_buffer_new(bh);
976                                 mark_buffer_dirty(bh);
977                         }
978                 }
979
980                 block_start = block_end;
981                 bh = bh->b_this_page;
982         } while (bh != head);
983 }
984
985 /*
986  * Only called when we have a failure during allocating write to write
987  * zero's to the newly allocated region.
988  */
989 static void ocfs2_write_failure(struct inode *inode,
990                                 struct ocfs2_write_ctxt *wc,
991                                 loff_t user_pos, unsigned user_len)
992 {
993         int i;
994         unsigned from = user_pos & (PAGE_CACHE_SIZE - 1),
995                 to = user_pos + user_len;
996         struct page *tmppage;
997
998         ocfs2_zero_new_buffers(wc->w_target_page, from, to);
999
1000         for(i = 0; i < wc->w_num_pages; i++) {
1001                 tmppage = wc->w_pages[i];
1002
1003                 if (page_has_buffers(tmppage)) {
1004                         if (ocfs2_should_order_data(inode))
1005                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1006
1007                         block_commit_write(tmppage, from, to);
1008                 }
1009         }
1010 }
1011
1012 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
1013                                         struct ocfs2_write_ctxt *wc,
1014                                         struct page *page, u32 cpos,
1015                                         loff_t user_pos, unsigned user_len,
1016                                         int new)
1017 {
1018         int ret;
1019         unsigned int map_from = 0, map_to = 0;
1020         unsigned int cluster_start, cluster_end;
1021         unsigned int user_data_from = 0, user_data_to = 0;
1022
1023         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
1024                                         &cluster_start, &cluster_end);
1025
1026         if (page == wc->w_target_page) {
1027                 map_from = user_pos & (PAGE_CACHE_SIZE - 1);
1028                 map_to = map_from + user_len;
1029
1030                 if (new)
1031                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1032                                                     cluster_start, cluster_end,
1033                                                     new);
1034                 else
1035                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1036                                                     map_from, map_to, new);
1037                 if (ret) {
1038                         mlog_errno(ret);
1039                         goto out;
1040                 }
1041
1042                 user_data_from = map_from;
1043                 user_data_to = map_to;
1044                 if (new) {
1045                         map_from = cluster_start;
1046                         map_to = cluster_end;
1047                 }
1048         } else {
1049                 /*
1050                  * If we haven't allocated the new page yet, we
1051                  * shouldn't be writing it out without copying user
1052                  * data. This is likely a math error from the caller.
1053                  */
1054                 BUG_ON(!new);
1055
1056                 map_from = cluster_start;
1057                 map_to = cluster_end;
1058
1059                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
1060                                             cluster_start, cluster_end, new);
1061                 if (ret) {
1062                         mlog_errno(ret);
1063                         goto out;
1064                 }
1065         }
1066
1067         /*
1068          * Parts of newly allocated pages need to be zero'd.
1069          *
1070          * Above, we have also rewritten 'to' and 'from' - as far as
1071          * the rest of the function is concerned, the entire cluster
1072          * range inside of a page needs to be written.
1073          *
1074          * We can skip this if the page is up to date - it's already
1075          * been zero'd from being read in as a hole.
1076          */
1077         if (new && !PageUptodate(page))
1078                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1079                                          cpos, user_data_from, user_data_to);
1080
1081         flush_dcache_page(page);
1082
1083 out:
1084         return ret;
1085 }
1086
1087 /*
1088  * This function will only grab one clusters worth of pages.
1089  */
1090 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1091                                       struct ocfs2_write_ctxt *wc,
1092                                       u32 cpos, loff_t user_pos,
1093                                       unsigned user_len, int new,
1094                                       struct page *mmap_page)
1095 {
1096         int ret = 0, i;
1097         unsigned long start, target_index, end_index, index;
1098         struct inode *inode = mapping->host;
1099         loff_t last_byte;
1100
1101         target_index = user_pos >> PAGE_CACHE_SHIFT;
1102
1103         /*
1104          * Figure out how many pages we'll be manipulating here. For
1105          * non allocating write, we just change the one
1106          * page. Otherwise, we'll need a whole clusters worth.  If we're
1107          * writing past i_size, we only need enough pages to cover the
1108          * last page of the write.
1109          */
1110         if (new) {
1111                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1112                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1113                 /*
1114                  * We need the index *past* the last page we could possibly
1115                  * touch.  This is the page past the end of the write or
1116                  * i_size, whichever is greater.
1117                  */
1118                 last_byte = max(user_pos + user_len, i_size_read(inode));
1119                 BUG_ON(last_byte < 1);
1120                 end_index = ((last_byte - 1) >> PAGE_CACHE_SHIFT) + 1;
1121                 if ((start + wc->w_num_pages) > end_index)
1122                         wc->w_num_pages = end_index - start;
1123         } else {
1124                 wc->w_num_pages = 1;
1125                 start = target_index;
1126         }
1127
1128         for(i = 0; i < wc->w_num_pages; i++) {
1129                 index = start + i;
1130
1131                 if (index == target_index && mmap_page) {
1132                         /*
1133                          * ocfs2_pagemkwrite() is a little different
1134                          * and wants us to directly use the page
1135                          * passed in.
1136                          */
1137                         lock_page(mmap_page);
1138
1139                         if (mmap_page->mapping != mapping) {
1140                                 unlock_page(mmap_page);
1141                                 /*
1142                                  * Sanity check - the locking in
1143                                  * ocfs2_pagemkwrite() should ensure
1144                                  * that this code doesn't trigger.
1145                                  */
1146                                 ret = -EINVAL;
1147                                 mlog_errno(ret);
1148                                 goto out;
1149                         }
1150
1151                         page_cache_get(mmap_page);
1152                         wc->w_pages[i] = mmap_page;
1153                 } else {
1154                         wc->w_pages[i] = find_or_create_page(mapping, index,
1155                                                              GFP_NOFS);
1156                         if (!wc->w_pages[i]) {
1157                                 ret = -ENOMEM;
1158                                 mlog_errno(ret);
1159                                 goto out;
1160                         }
1161                 }
1162
1163                 if (index == target_index)
1164                         wc->w_target_page = wc->w_pages[i];
1165         }
1166 out:
1167         return ret;
1168 }
1169
1170 /*
1171  * Prepare a single cluster for write one cluster into the file.
1172  */
1173 static int ocfs2_write_cluster(struct address_space *mapping,
1174                                u32 phys, unsigned int unwritten,
1175                                unsigned int should_zero,
1176                                struct ocfs2_alloc_context *data_ac,
1177                                struct ocfs2_alloc_context *meta_ac,
1178                                struct ocfs2_write_ctxt *wc, u32 cpos,
1179                                loff_t user_pos, unsigned user_len)
1180 {
1181         int ret, i, new;
1182         u64 v_blkno, p_blkno;
1183         struct inode *inode = mapping->host;
1184         struct ocfs2_extent_tree et;
1185
1186         new = phys == 0 ? 1 : 0;
1187         if (new) {
1188                 u32 tmp_pos;
1189
1190                 /*
1191                  * This is safe to call with the page locks - it won't take
1192                  * any additional semaphores or cluster locks.
1193                  */
1194                 tmp_pos = cpos;
1195                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1196                                            &tmp_pos, 1, 0, wc->w_di_bh,
1197                                            wc->w_handle, data_ac,
1198                                            meta_ac, NULL);
1199                 /*
1200                  * This shouldn't happen because we must have already
1201                  * calculated the correct meta data allocation required. The
1202                  * internal tree allocation code should know how to increase
1203                  * transaction credits itself.
1204                  *
1205                  * If need be, we could handle -EAGAIN for a
1206                  * RESTART_TRANS here.
1207                  */
1208                 mlog_bug_on_msg(ret == -EAGAIN,
1209                                 "Inode %llu: EAGAIN return during allocation.\n",
1210                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1211                 if (ret < 0) {
1212                         mlog_errno(ret);
1213                         goto out;
1214                 }
1215         } else if (unwritten) {
1216                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1217                                               wc->w_di_bh);
1218                 ret = ocfs2_mark_extent_written(inode, &et,
1219                                                 wc->w_handle, cpos, 1, phys,
1220                                                 meta_ac, &wc->w_dealloc);
1221                 if (ret < 0) {
1222                         mlog_errno(ret);
1223                         goto out;
1224                 }
1225         }
1226
1227         if (should_zero)
1228                 v_blkno = ocfs2_clusters_to_blocks(inode->i_sb, cpos);
1229         else
1230                 v_blkno = user_pos >> inode->i_sb->s_blocksize_bits;
1231
1232         /*
1233          * The only reason this should fail is due to an inability to
1234          * find the extent added.
1235          */
1236         ret = ocfs2_extent_map_get_blocks(inode, v_blkno, &p_blkno, NULL,
1237                                           NULL);
1238         if (ret < 0) {
1239                 ocfs2_error(inode->i_sb, "Corrupting extend for inode %llu, "
1240                             "at logical block %llu",
1241                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
1242                             (unsigned long long)v_blkno);
1243                 goto out;
1244         }
1245
1246         BUG_ON(p_blkno == 0);
1247
1248         for(i = 0; i < wc->w_num_pages; i++) {
1249                 int tmpret;
1250
1251                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1252                                                       wc->w_pages[i], cpos,
1253                                                       user_pos, user_len,
1254                                                       should_zero);
1255                 if (tmpret) {
1256                         mlog_errno(tmpret);
1257                         if (ret == 0)
1258                                 ret = tmpret;
1259                 }
1260         }
1261
1262         /*
1263          * We only have cleanup to do in case of allocating write.
1264          */
1265         if (ret && new)
1266                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1267
1268 out:
1269
1270         return ret;
1271 }
1272
1273 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1274                                        struct ocfs2_alloc_context *data_ac,
1275                                        struct ocfs2_alloc_context *meta_ac,
1276                                        struct ocfs2_write_ctxt *wc,
1277                                        loff_t pos, unsigned len)
1278 {
1279         int ret, i;
1280         loff_t cluster_off;
1281         unsigned int local_len = len;
1282         struct ocfs2_write_cluster_desc *desc;
1283         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1284
1285         for (i = 0; i < wc->w_clen; i++) {
1286                 desc = &wc->w_desc[i];
1287
1288                 /*
1289                  * We have to make sure that the total write passed in
1290                  * doesn't extend past a single cluster.
1291                  */
1292                 local_len = len;
1293                 cluster_off = pos & (osb->s_clustersize - 1);
1294                 if ((cluster_off + local_len) > osb->s_clustersize)
1295                         local_len = osb->s_clustersize - cluster_off;
1296
1297                 ret = ocfs2_write_cluster(mapping, desc->c_phys,
1298                                           desc->c_unwritten,
1299                                           desc->c_needs_zero,
1300                                           data_ac, meta_ac,
1301                                           wc, desc->c_cpos, pos, local_len);
1302                 if (ret) {
1303                         mlog_errno(ret);
1304                         goto out;
1305                 }
1306
1307                 len -= local_len;
1308                 pos += local_len;
1309         }
1310
1311         ret = 0;
1312 out:
1313         return ret;
1314 }
1315
1316 /*
1317  * ocfs2_write_end() wants to know which parts of the target page it
1318  * should complete the write on. It's easiest to compute them ahead of
1319  * time when a more complete view of the write is available.
1320  */
1321 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1322                                         struct ocfs2_write_ctxt *wc,
1323                                         loff_t pos, unsigned len, int alloc)
1324 {
1325         struct ocfs2_write_cluster_desc *desc;
1326
1327         wc->w_target_from = pos & (PAGE_CACHE_SIZE - 1);
1328         wc->w_target_to = wc->w_target_from + len;
1329
1330         if (alloc == 0)
1331                 return;
1332
1333         /*
1334          * Allocating write - we may have different boundaries based
1335          * on page size and cluster size.
1336          *
1337          * NOTE: We can no longer compute one value from the other as
1338          * the actual write length and user provided length may be
1339          * different.
1340          */
1341
1342         if (wc->w_large_pages) {
1343                 /*
1344                  * We only care about the 1st and last cluster within
1345                  * our range and whether they should be zero'd or not. Either
1346                  * value may be extended out to the start/end of a
1347                  * newly allocated cluster.
1348                  */
1349                 desc = &wc->w_desc[0];
1350                 if (desc->c_needs_zero)
1351                         ocfs2_figure_cluster_boundaries(osb,
1352                                                         desc->c_cpos,
1353                                                         &wc->w_target_from,
1354                                                         NULL);
1355
1356                 desc = &wc->w_desc[wc->w_clen - 1];
1357                 if (desc->c_needs_zero)
1358                         ocfs2_figure_cluster_boundaries(osb,
1359                                                         desc->c_cpos,
1360                                                         NULL,
1361                                                         &wc->w_target_to);
1362         } else {
1363                 wc->w_target_from = 0;
1364                 wc->w_target_to = PAGE_CACHE_SIZE;
1365         }
1366 }
1367
1368 /*
1369  * Populate each single-cluster write descriptor in the write context
1370  * with information about the i/o to be done.
1371  *
1372  * Returns the number of clusters that will have to be allocated, as
1373  * well as a worst case estimate of the number of extent records that
1374  * would have to be created during a write to an unwritten region.
1375  */
1376 static int ocfs2_populate_write_desc(struct inode *inode,
1377                                      struct ocfs2_write_ctxt *wc,
1378                                      unsigned int *clusters_to_alloc,
1379                                      unsigned int *extents_to_split)
1380 {
1381         int ret;
1382         struct ocfs2_write_cluster_desc *desc;
1383         unsigned int num_clusters = 0;
1384         unsigned int ext_flags = 0;
1385         u32 phys = 0;
1386         int i;
1387
1388         *clusters_to_alloc = 0;
1389         *extents_to_split = 0;
1390
1391         for (i = 0; i < wc->w_clen; i++) {
1392                 desc = &wc->w_desc[i];
1393                 desc->c_cpos = wc->w_cpos + i;
1394
1395                 if (num_clusters == 0) {
1396                         /*
1397                          * Need to look up the next extent record.
1398                          */
1399                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1400                                                  &num_clusters, &ext_flags);
1401                         if (ret) {
1402                                 mlog_errno(ret);
1403                                 goto out;
1404                         }
1405
1406                         /* We should already CoW the refcountd extent. */
1407                         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1408
1409                         /*
1410                          * Assume worst case - that we're writing in
1411                          * the middle of the extent.
1412                          *
1413                          * We can assume that the write proceeds from
1414                          * left to right, in which case the extent
1415                          * insert code is smart enough to coalesce the
1416                          * next splits into the previous records created.
1417                          */
1418                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1419                                 *extents_to_split = *extents_to_split + 2;
1420                 } else if (phys) {
1421                         /*
1422                          * Only increment phys if it doesn't describe
1423                          * a hole.
1424                          */
1425                         phys++;
1426                 }
1427
1428                 /*
1429                  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1430                  * file that got extended.  w_first_new_cpos tells us
1431                  * where the newly allocated clusters are so we can
1432                  * zero them.
1433                  */
1434                 if (desc->c_cpos >= wc->w_first_new_cpos) {
1435                         BUG_ON(phys == 0);
1436                         desc->c_needs_zero = 1;
1437                 }
1438
1439                 desc->c_phys = phys;
1440                 if (phys == 0) {
1441                         desc->c_new = 1;
1442                         desc->c_needs_zero = 1;
1443                         *clusters_to_alloc = *clusters_to_alloc + 1;
1444                 }
1445
1446                 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1447                         desc->c_unwritten = 1;
1448                         desc->c_needs_zero = 1;
1449                 }
1450
1451                 num_clusters--;
1452         }
1453
1454         ret = 0;
1455 out:
1456         return ret;
1457 }
1458
1459 static int ocfs2_write_begin_inline(struct address_space *mapping,
1460                                     struct inode *inode,
1461                                     struct ocfs2_write_ctxt *wc)
1462 {
1463         int ret;
1464         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1465         struct page *page;
1466         handle_t *handle;
1467         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1468
1469         page = find_or_create_page(mapping, 0, GFP_NOFS);
1470         if (!page) {
1471                 ret = -ENOMEM;
1472                 mlog_errno(ret);
1473                 goto out;
1474         }
1475         /*
1476          * If we don't set w_num_pages then this page won't get unlocked
1477          * and freed on cleanup of the write context.
1478          */
1479         wc->w_pages[0] = wc->w_target_page = page;
1480         wc->w_num_pages = 1;
1481
1482         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1483         if (IS_ERR(handle)) {
1484                 ret = PTR_ERR(handle);
1485                 mlog_errno(ret);
1486                 goto out;
1487         }
1488
1489         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1490                                       OCFS2_JOURNAL_ACCESS_WRITE);
1491         if (ret) {
1492                 ocfs2_commit_trans(osb, handle);
1493
1494                 mlog_errno(ret);
1495                 goto out;
1496         }
1497
1498         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1499                 ocfs2_set_inode_data_inline(inode, di);
1500
1501         if (!PageUptodate(page)) {
1502                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1503                 if (ret) {
1504                         ocfs2_commit_trans(osb, handle);
1505
1506                         goto out;
1507                 }
1508         }
1509
1510         wc->w_handle = handle;
1511 out:
1512         return ret;
1513 }
1514
1515 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1516 {
1517         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1518
1519         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1520                 return 1;
1521         return 0;
1522 }
1523
1524 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1525                                           struct inode *inode, loff_t pos,
1526                                           unsigned len, struct page *mmap_page,
1527                                           struct ocfs2_write_ctxt *wc)
1528 {
1529         int ret, written = 0;
1530         loff_t end = pos + len;
1531         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1532         struct ocfs2_dinode *di = NULL;
1533
1534         mlog(0, "Inode %llu, write of %u bytes at off %llu. features: 0x%x\n",
1535              (unsigned long long)oi->ip_blkno, len, (unsigned long long)pos,
1536              oi->ip_dyn_features);
1537
1538         /*
1539          * Handle inodes which already have inline data 1st.
1540          */
1541         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1542                 if (mmap_page == NULL &&
1543                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1544                         goto do_inline_write;
1545
1546                 /*
1547                  * The write won't fit - we have to give this inode an
1548                  * inline extent list now.
1549                  */
1550                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1551                 if (ret)
1552                         mlog_errno(ret);
1553                 goto out;
1554         }
1555
1556         /*
1557          * Check whether the inode can accept inline data.
1558          */
1559         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1560                 return 0;
1561
1562         /*
1563          * Check whether the write can fit.
1564          */
1565         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1566         if (mmap_page ||
1567             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1568                 return 0;
1569
1570 do_inline_write:
1571         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1572         if (ret) {
1573                 mlog_errno(ret);
1574                 goto out;
1575         }
1576
1577         /*
1578          * This signals to the caller that the data can be written
1579          * inline.
1580          */
1581         written = 1;
1582 out:
1583         return written ? written : ret;
1584 }
1585
1586 /*
1587  * This function only does anything for file systems which can't
1588  * handle sparse files.
1589  *
1590  * What we want to do here is fill in any hole between the current end
1591  * of allocation and the end of our write. That way the rest of the
1592  * write path can treat it as an non-allocating write, which has no
1593  * special case code for sparse/nonsparse files.
1594  */
1595 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1596                                         struct buffer_head *di_bh,
1597                                         loff_t pos, unsigned len,
1598                                         struct ocfs2_write_ctxt *wc)
1599 {
1600         int ret;
1601         loff_t newsize = pos + len;
1602
1603         BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1604
1605         if (newsize <= i_size_read(inode))
1606                 return 0;
1607
1608         ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1609         if (ret)
1610                 mlog_errno(ret);
1611
1612         wc->w_first_new_cpos =
1613                 ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1614
1615         return ret;
1616 }
1617
1618 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1619                            loff_t pos)
1620 {
1621         int ret = 0;
1622
1623         BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1624         if (pos > i_size_read(inode))
1625                 ret = ocfs2_zero_extend(inode, di_bh, pos);
1626
1627         return ret;
1628 }
1629
1630 int ocfs2_write_begin_nolock(struct file *filp,
1631                              struct address_space *mapping,
1632                              loff_t pos, unsigned len, unsigned flags,
1633                              struct page **pagep, void **fsdata,
1634                              struct buffer_head *di_bh, struct page *mmap_page)
1635 {
1636         int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1637         unsigned int clusters_to_alloc, extents_to_split;
1638         struct ocfs2_write_ctxt *wc;
1639         struct inode *inode = mapping->host;
1640         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1641         struct ocfs2_dinode *di;
1642         struct ocfs2_alloc_context *data_ac = NULL;
1643         struct ocfs2_alloc_context *meta_ac = NULL;
1644         handle_t *handle;
1645         struct ocfs2_extent_tree et;
1646
1647         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, di_bh);
1648         if (ret) {
1649                 mlog_errno(ret);
1650                 return ret;
1651         }
1652
1653         if (ocfs2_supports_inline_data(osb)) {
1654                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1655                                                      mmap_page, wc);
1656                 if (ret == 1) {
1657                         ret = 0;
1658                         goto success;
1659                 }
1660                 if (ret < 0) {
1661                         mlog_errno(ret);
1662                         goto out;
1663                 }
1664         }
1665
1666         if (ocfs2_sparse_alloc(osb))
1667                 ret = ocfs2_zero_tail(inode, di_bh, pos);
1668         else
1669                 ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos, len,
1670                                                    wc);
1671         if (ret) {
1672                 mlog_errno(ret);
1673                 goto out;
1674         }
1675
1676         ret = ocfs2_check_range_for_refcount(inode, pos, len);
1677         if (ret < 0) {
1678                 mlog_errno(ret);
1679                 goto out;
1680         } else if (ret == 1) {
1681                 ret = ocfs2_refcount_cow(inode, filp, di_bh,
1682                                          wc->w_cpos, wc->w_clen, UINT_MAX);
1683                 if (ret) {
1684                         mlog_errno(ret);
1685                         goto out;
1686                 }
1687         }
1688
1689         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1690                                         &extents_to_split);
1691         if (ret) {
1692                 mlog_errno(ret);
1693                 goto out;
1694         }
1695
1696         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1697
1698         /*
1699          * We set w_target_from, w_target_to here so that
1700          * ocfs2_write_end() knows which range in the target page to
1701          * write out. An allocation requires that we write the entire
1702          * cluster range.
1703          */
1704         if (clusters_to_alloc || extents_to_split) {
1705                 /*
1706                  * XXX: We are stretching the limits of
1707                  * ocfs2_lock_allocators(). It greatly over-estimates
1708                  * the work to be done.
1709                  */
1710                 mlog(0, "extend inode %llu, i_size = %lld, di->i_clusters = %u,"
1711                      " clusters_to_add = %u, extents_to_split = %u\n",
1712                      (unsigned long long)OCFS2_I(inode)->ip_blkno,
1713                      (long long)i_size_read(inode), le32_to_cpu(di->i_clusters),
1714                      clusters_to_alloc, extents_to_split);
1715
1716                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1717                                               wc->w_di_bh);
1718                 ret = ocfs2_lock_allocators(inode, &et,
1719                                             clusters_to_alloc, extents_to_split,
1720                                             &data_ac, &meta_ac);
1721                 if (ret) {
1722                         mlog_errno(ret);
1723                         goto out;
1724                 }
1725
1726                 if (data_ac)
1727                         data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1728
1729                 credits = ocfs2_calc_extend_credits(inode->i_sb,
1730                                                     &di->id2.i_list,
1731                                                     clusters_to_alloc);
1732
1733         }
1734
1735         /*
1736          * We have to zero sparse allocated clusters, unwritten extent clusters,
1737          * and non-sparse clusters we just extended.  For non-sparse writes,
1738          * we know zeros will only be needed in the first and/or last cluster.
1739          */
1740         if (clusters_to_alloc || extents_to_split ||
1741             (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1742                             wc->w_desc[wc->w_clen - 1].c_needs_zero)))
1743                 cluster_of_pages = 1;
1744         else
1745                 cluster_of_pages = 0;
1746
1747         ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1748
1749         handle = ocfs2_start_trans(osb, credits);
1750         if (IS_ERR(handle)) {
1751                 ret = PTR_ERR(handle);
1752                 mlog_errno(ret);
1753                 goto out;
1754         }
1755
1756         wc->w_handle = handle;
1757
1758         if (clusters_to_alloc) {
1759                 ret = dquot_alloc_space_nodirty(inode,
1760                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1761                 if (ret)
1762                         goto out_commit;
1763         }
1764         /*
1765          * We don't want this to fail in ocfs2_write_end(), so do it
1766          * here.
1767          */
1768         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1769                                       OCFS2_JOURNAL_ACCESS_WRITE);
1770         if (ret) {
1771                 mlog_errno(ret);
1772                 goto out_quota;
1773         }
1774
1775         /*
1776          * Fill our page array first. That way we've grabbed enough so
1777          * that we can zero and flush if we error after adding the
1778          * extent.
1779          */
1780         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1781                                          cluster_of_pages, mmap_page);
1782         if (ret) {
1783                 mlog_errno(ret);
1784                 goto out_quota;
1785         }
1786
1787         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1788                                           len);
1789         if (ret) {
1790                 mlog_errno(ret);
1791                 goto out_quota;
1792         }
1793
1794         if (data_ac)
1795                 ocfs2_free_alloc_context(data_ac);
1796         if (meta_ac)
1797                 ocfs2_free_alloc_context(meta_ac);
1798
1799 success:
1800         *pagep = wc->w_target_page;
1801         *fsdata = wc;
1802         return 0;
1803 out_quota:
1804         if (clusters_to_alloc)
1805                 dquot_free_space(inode,
1806                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1807 out_commit:
1808         ocfs2_commit_trans(osb, handle);
1809
1810 out:
1811         ocfs2_free_write_ctxt(wc);
1812
1813         if (data_ac)
1814                 ocfs2_free_alloc_context(data_ac);
1815         if (meta_ac)
1816                 ocfs2_free_alloc_context(meta_ac);
1817         return ret;
1818 }
1819
1820 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1821                              loff_t pos, unsigned len, unsigned flags,
1822                              struct page **pagep, void **fsdata)
1823 {
1824         int ret;
1825         struct buffer_head *di_bh = NULL;
1826         struct inode *inode = mapping->host;
1827
1828         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1829         if (ret) {
1830                 mlog_errno(ret);
1831                 return ret;
1832         }
1833
1834         /*
1835          * Take alloc sem here to prevent concurrent lookups. That way
1836          * the mapping, zeroing and tree manipulation within
1837          * ocfs2_write() will be safe against ->readpage(). This
1838          * should also serve to lock out allocation from a shared
1839          * writeable region.
1840          */
1841         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1842
1843         ret = ocfs2_write_begin_nolock(file, mapping, pos, len, flags, pagep,
1844                                        fsdata, di_bh, NULL);
1845         if (ret) {
1846                 mlog_errno(ret);
1847                 goto out_fail;
1848         }
1849
1850         brelse(di_bh);
1851
1852         return 0;
1853
1854 out_fail:
1855         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1856
1857         brelse(di_bh);
1858         ocfs2_inode_unlock(inode, 1);
1859
1860         return ret;
1861 }
1862
1863 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1864                                    unsigned len, unsigned *copied,
1865                                    struct ocfs2_dinode *di,
1866                                    struct ocfs2_write_ctxt *wc)
1867 {
1868         void *kaddr;
1869
1870         if (unlikely(*copied < len)) {
1871                 if (!PageUptodate(wc->w_target_page)) {
1872                         *copied = 0;
1873                         return;
1874                 }
1875         }
1876
1877         kaddr = kmap_atomic(wc->w_target_page, KM_USER0);
1878         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1879         kunmap_atomic(kaddr, KM_USER0);
1880
1881         mlog(0, "Data written to inode at offset %llu. "
1882              "id_count = %u, copied = %u, i_dyn_features = 0x%x\n",
1883              (unsigned long long)pos, *copied,
1884              le16_to_cpu(di->id2.i_data.id_count),
1885              le16_to_cpu(di->i_dyn_features));
1886 }
1887
1888 int ocfs2_write_end_nolock(struct address_space *mapping,
1889                            loff_t pos, unsigned len, unsigned copied,
1890                            struct page *page, void *fsdata)
1891 {
1892         int i;
1893         unsigned from, to, start = pos & (PAGE_CACHE_SIZE - 1);
1894         struct inode *inode = mapping->host;
1895         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1896         struct ocfs2_write_ctxt *wc = fsdata;
1897         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1898         handle_t *handle = wc->w_handle;
1899         struct page *tmppage;
1900
1901         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1902                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1903                 goto out_write_size;
1904         }
1905
1906         if (unlikely(copied < len)) {
1907                 if (!PageUptodate(wc->w_target_page))
1908                         copied = 0;
1909
1910                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1911                                        start+len);
1912         }
1913         flush_dcache_page(wc->w_target_page);
1914
1915         for(i = 0; i < wc->w_num_pages; i++) {
1916                 tmppage = wc->w_pages[i];
1917
1918                 if (tmppage == wc->w_target_page) {
1919                         from = wc->w_target_from;
1920                         to = wc->w_target_to;
1921
1922                         BUG_ON(from > PAGE_CACHE_SIZE ||
1923                                to > PAGE_CACHE_SIZE ||
1924                                to < from);
1925                 } else {
1926                         /*
1927                          * Pages adjacent to the target (if any) imply
1928                          * a hole-filling write in which case we want
1929                          * to flush their entire range.
1930                          */
1931                         from = 0;
1932                         to = PAGE_CACHE_SIZE;
1933                 }
1934
1935                 if (page_has_buffers(tmppage)) {
1936                         if (ocfs2_should_order_data(inode))
1937                                 ocfs2_jbd2_file_inode(wc->w_handle, inode);
1938                         block_commit_write(tmppage, from, to);
1939                 }
1940         }
1941
1942 out_write_size:
1943         pos += copied;
1944         if (pos > inode->i_size) {
1945                 i_size_write(inode, pos);
1946                 mark_inode_dirty(inode);
1947         }
1948         inode->i_blocks = ocfs2_inode_sector_count(inode);
1949         di->i_size = cpu_to_le64((u64)i_size_read(inode));
1950         inode->i_mtime = inode->i_ctime = CURRENT_TIME;
1951         di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
1952         di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
1953         ocfs2_journal_dirty(handle, wc->w_di_bh);
1954
1955         ocfs2_commit_trans(osb, handle);
1956
1957         ocfs2_run_deallocs(osb, &wc->w_dealloc);
1958
1959         ocfs2_free_write_ctxt(wc);
1960
1961         return copied;
1962 }
1963
1964 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
1965                            loff_t pos, unsigned len, unsigned copied,
1966                            struct page *page, void *fsdata)
1967 {
1968         int ret;
1969         struct inode *inode = mapping->host;
1970
1971         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, page, fsdata);
1972
1973         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1974         ocfs2_inode_unlock(inode, 1);
1975
1976         return ret;
1977 }
1978
1979 const struct address_space_operations ocfs2_aops = {
1980         .readpage               = ocfs2_readpage,
1981         .readpages              = ocfs2_readpages,
1982         .writepage              = ocfs2_writepage,
1983         .write_begin            = ocfs2_write_begin,
1984         .write_end              = ocfs2_write_end,
1985         .bmap                   = ocfs2_bmap,
1986         .sync_page              = block_sync_page,
1987         .direct_IO              = ocfs2_direct_IO,
1988         .invalidatepage         = ocfs2_invalidatepage,
1989         .releasepage            = ocfs2_releasepage,
1990         .migratepage            = buffer_migrate_page,
1991         .is_partially_uptodate  = block_is_partially_uptodate,
1992         .error_remove_page      = generic_error_remove_page,
1993 };