~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/fs/ocfs2/aops.c

Version: ~ [ linux-5.15-rc5 ] ~ [ linux-5.14.11 ] ~ [ linux-5.13.19 ] ~ [ linux-5.12.19 ] ~ [ linux-5.11.22 ] ~ [ linux-5.10.72 ] ~ [ linux-5.9.16 ] ~ [ linux-5.8.18 ] ~ [ linux-5.7.19 ] ~ [ linux-5.6.19 ] ~ [ linux-5.5.19 ] ~ [ linux-5.4.152 ] ~ [ linux-5.3.18 ] ~ [ linux-5.2.21 ] ~ [ linux-5.1.21 ] ~ [ linux-5.0.21 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.210 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.250 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.286 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.288 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.18.140 ] ~ [ linux-3.16.85 ] ~ [ linux-3.14.79 ] ~ [ linux-3.12.74 ] ~ [ linux-3.10.108 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 // SPDX-License-Identifier: GPL-2.0-or-later
  2 /* -*- mode: c; c-basic-offset: 8; -*-
  3  * vim: noexpandtab sw=8 ts=8 sts=0:
  4  *
  5  * Copyright (C) 2002, 2004 Oracle.  All rights reserved.
  6  */
  7 
  8 #include <linux/fs.h>
  9 #include <linux/slab.h>
 10 #include <linux/highmem.h>
 11 #include <linux/pagemap.h>
 12 #include <asm/byteorder.h>
 13 #include <linux/swap.h>
 14 #include <linux/mpage.h>
 15 #include <linux/quotaops.h>
 16 #include <linux/blkdev.h>
 17 #include <linux/uio.h>
 18 #include <linux/mm.h>
 19 
 20 #include <cluster/masklog.h>
 21 
 22 #include "ocfs2.h"
 23 
 24 #include "alloc.h"
 25 #include "aops.h"
 26 #include "dlmglue.h"
 27 #include "extent_map.h"
 28 #include "file.h"
 29 #include "inode.h"
 30 #include "journal.h"
 31 #include "suballoc.h"
 32 #include "super.h"
 33 #include "symlink.h"
 34 #include "refcounttree.h"
 35 #include "ocfs2_trace.h"
 36 
 37 #include "buffer_head_io.h"
 38 #include "dir.h"
 39 #include "namei.h"
 40 #include "sysfile.h"
 41 
 42 static int ocfs2_symlink_get_block(struct inode *inode, sector_t iblock,
 43                                    struct buffer_head *bh_result, int create)
 44 {
 45         int err = -EIO;
 46         int status;
 47         struct ocfs2_dinode *fe = NULL;
 48         struct buffer_head *bh = NULL;
 49         struct buffer_head *buffer_cache_bh = NULL;
 50         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
 51         void *kaddr;
 52 
 53         trace_ocfs2_symlink_get_block(
 54                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
 55                         (unsigned long long)iblock, bh_result, create);
 56 
 57         BUG_ON(ocfs2_inode_is_fast_symlink(inode));
 58 
 59         if ((iblock << inode->i_sb->s_blocksize_bits) > PATH_MAX + 1) {
 60                 mlog(ML_ERROR, "block offset > PATH_MAX: %llu",
 61                      (unsigned long long)iblock);
 62                 goto bail;
 63         }
 64 
 65         status = ocfs2_read_inode_block(inode, &bh);
 66         if (status < 0) {
 67                 mlog_errno(status);
 68                 goto bail;
 69         }
 70         fe = (struct ocfs2_dinode *) bh->b_data;
 71 
 72         if ((u64)iblock >= ocfs2_clusters_to_blocks(inode->i_sb,
 73                                                     le32_to_cpu(fe->i_clusters))) {
 74                 err = -ENOMEM;
 75                 mlog(ML_ERROR, "block offset is outside the allocated size: "
 76                      "%llu\n", (unsigned long long)iblock);
 77                 goto bail;
 78         }
 79 
 80         /* We don't use the page cache to create symlink data, so if
 81          * need be, copy it over from the buffer cache. */
 82         if (!buffer_uptodate(bh_result) && ocfs2_inode_is_new(inode)) {
 83                 u64 blkno = le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) +
 84                             iblock;
 85                 buffer_cache_bh = sb_getblk(osb->sb, blkno);
 86                 if (!buffer_cache_bh) {
 87                         err = -ENOMEM;
 88                         mlog(ML_ERROR, "couldn't getblock for symlink!\n");
 89                         goto bail;
 90                 }
 91 
 92                 /* we haven't locked out transactions, so a commit
 93                  * could've happened. Since we've got a reference on
 94                  * the bh, even if it commits while we're doing the
 95                  * copy, the data is still good. */
 96                 if (buffer_jbd(buffer_cache_bh)
 97                     && ocfs2_inode_is_new(inode)) {
 98                         kaddr = kmap_atomic(bh_result->b_page);
 99                         if (!kaddr) {
100                                 mlog(ML_ERROR, "couldn't kmap!\n");
101                                 goto bail;
102                         }
103                         memcpy(kaddr + (bh_result->b_size * iblock),
104                                buffer_cache_bh->b_data,
105                                bh_result->b_size);
106                         kunmap_atomic(kaddr);
107                         set_buffer_uptodate(bh_result);
108                 }
109                 brelse(buffer_cache_bh);
110         }
111 
112         map_bh(bh_result, inode->i_sb,
113                le64_to_cpu(fe->id2.i_list.l_recs[0].e_blkno) + iblock);
114 
115         err = 0;
116 
117 bail:
118         brelse(bh);
119 
120         return err;
121 }
122 
123 static int ocfs2_lock_get_block(struct inode *inode, sector_t iblock,
124                     struct buffer_head *bh_result, int create)
125 {
126         int ret = 0;
127         struct ocfs2_inode_info *oi = OCFS2_I(inode);
128 
129         down_read(&oi->ip_alloc_sem);
130         ret = ocfs2_get_block(inode, iblock, bh_result, create);
131         up_read(&oi->ip_alloc_sem);
132 
133         return ret;
134 }
135 
136 int ocfs2_get_block(struct inode *inode, sector_t iblock,
137                     struct buffer_head *bh_result, int create)
138 {
139         int err = 0;
140         unsigned int ext_flags;
141         u64 max_blocks = bh_result->b_size >> inode->i_blkbits;
142         u64 p_blkno, count, past_eof;
143         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
144 
145         trace_ocfs2_get_block((unsigned long long)OCFS2_I(inode)->ip_blkno,
146                               (unsigned long long)iblock, bh_result, create);
147 
148         if (OCFS2_I(inode)->ip_flags & OCFS2_INODE_SYSTEM_FILE)
149                 mlog(ML_NOTICE, "get_block on system inode 0x%p (%lu)\n",
150                      inode, inode->i_ino);
151 
152         if (S_ISLNK(inode->i_mode)) {
153                 /* this always does I/O for some reason. */
154                 err = ocfs2_symlink_get_block(inode, iblock, bh_result, create);
155                 goto bail;
156         }
157 
158         err = ocfs2_extent_map_get_blocks(inode, iblock, &p_blkno, &count,
159                                           &ext_flags);
160         if (err) {
161                 mlog(ML_ERROR, "Error %d from get_blocks(0x%p, %llu, 1, "
162                      "%llu, NULL)\n", err, inode, (unsigned long long)iblock,
163                      (unsigned long long)p_blkno);
164                 goto bail;
165         }
166 
167         if (max_blocks < count)
168                 count = max_blocks;
169 
170         /*
171          * ocfs2 never allocates in this function - the only time we
172          * need to use BH_New is when we're extending i_size on a file
173          * system which doesn't support holes, in which case BH_New
174          * allows __block_write_begin() to zero.
175          *
176          * If we see this on a sparse file system, then a truncate has
177          * raced us and removed the cluster. In this case, we clear
178          * the buffers dirty and uptodate bits and let the buffer code
179          * ignore it as a hole.
180          */
181         if (create && p_blkno == 0 && ocfs2_sparse_alloc(osb)) {
182                 clear_buffer_dirty(bh_result);
183                 clear_buffer_uptodate(bh_result);
184                 goto bail;
185         }
186 
187         /* Treat the unwritten extent as a hole for zeroing purposes. */
188         if (p_blkno && !(ext_flags & OCFS2_EXT_UNWRITTEN))
189                 map_bh(bh_result, inode->i_sb, p_blkno);
190 
191         bh_result->b_size = count << inode->i_blkbits;
192 
193         if (!ocfs2_sparse_alloc(osb)) {
194                 if (p_blkno == 0) {
195                         err = -EIO;
196                         mlog(ML_ERROR,
197                              "iblock = %llu p_blkno = %llu blkno=(%llu)\n",
198                              (unsigned long long)iblock,
199                              (unsigned long long)p_blkno,
200                              (unsigned long long)OCFS2_I(inode)->ip_blkno);
201                         mlog(ML_ERROR, "Size %llu, clusters %u\n", (unsigned long long)i_size_read(inode), OCFS2_I(inode)->ip_clusters);
202                         dump_stack();
203                         goto bail;
204                 }
205         }
206 
207         past_eof = ocfs2_blocks_for_bytes(inode->i_sb, i_size_read(inode));
208 
209         trace_ocfs2_get_block_end((unsigned long long)OCFS2_I(inode)->ip_blkno,
210                                   (unsigned long long)past_eof);
211         if (create && (iblock >= past_eof))
212                 set_buffer_new(bh_result);
213 
214 bail:
215         if (err < 0)
216                 err = -EIO;
217 
218         return err;
219 }
220 
221 int ocfs2_read_inline_data(struct inode *inode, struct page *page,
222                            struct buffer_head *di_bh)
223 {
224         void *kaddr;
225         loff_t size;
226         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
227 
228         if (!(le16_to_cpu(di->i_dyn_features) & OCFS2_INLINE_DATA_FL)) {
229                 ocfs2_error(inode->i_sb, "Inode %llu lost inline data flag\n",
230                             (unsigned long long)OCFS2_I(inode)->ip_blkno);
231                 return -EROFS;
232         }
233 
234         size = i_size_read(inode);
235 
236         if (size > PAGE_SIZE ||
237             size > ocfs2_max_inline_data_with_xattr(inode->i_sb, di)) {
238                 ocfs2_error(inode->i_sb,
239                             "Inode %llu has with inline data has bad size: %Lu\n",
240                             (unsigned long long)OCFS2_I(inode)->ip_blkno,
241                             (unsigned long long)size);
242                 return -EROFS;
243         }
244 
245         kaddr = kmap_atomic(page);
246         if (size)
247                 memcpy(kaddr, di->id2.i_data.id_data, size);
248         /* Clear the remaining part of the page */
249         memset(kaddr + size, 0, PAGE_SIZE - size);
250         flush_dcache_page(page);
251         kunmap_atomic(kaddr);
252 
253         SetPageUptodate(page);
254 
255         return 0;
256 }
257 
258 static int ocfs2_readpage_inline(struct inode *inode, struct page *page)
259 {
260         int ret;
261         struct buffer_head *di_bh = NULL;
262 
263         BUG_ON(!PageLocked(page));
264         BUG_ON(!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL));
265 
266         ret = ocfs2_read_inode_block(inode, &di_bh);
267         if (ret) {
268                 mlog_errno(ret);
269                 goto out;
270         }
271 
272         ret = ocfs2_read_inline_data(inode, page, di_bh);
273 out:
274         unlock_page(page);
275 
276         brelse(di_bh);
277         return ret;
278 }
279 
280 static int ocfs2_readpage(struct file *file, struct page *page)
281 {
282         struct inode *inode = page->mapping->host;
283         struct ocfs2_inode_info *oi = OCFS2_I(inode);
284         loff_t start = (loff_t)page->index << PAGE_SHIFT;
285         int ret, unlock = 1;
286 
287         trace_ocfs2_readpage((unsigned long long)oi->ip_blkno,
288                              (page ? page->index : 0));
289 
290         ret = ocfs2_inode_lock_with_page(inode, NULL, 0, page);
291         if (ret != 0) {
292                 if (ret == AOP_TRUNCATED_PAGE)
293                         unlock = 0;
294                 mlog_errno(ret);
295                 goto out;
296         }
297 
298         if (down_read_trylock(&oi->ip_alloc_sem) == 0) {
299                 /*
300                  * Unlock the page and cycle ip_alloc_sem so that we don't
301                  * busyloop waiting for ip_alloc_sem to unlock
302                  */
303                 ret = AOP_TRUNCATED_PAGE;
304                 unlock_page(page);
305                 unlock = 0;
306                 down_read(&oi->ip_alloc_sem);
307                 up_read(&oi->ip_alloc_sem);
308                 goto out_inode_unlock;
309         }
310 
311         /*
312          * i_size might have just been updated as we grabed the meta lock.  We
313          * might now be discovering a truncate that hit on another node.
314          * block_read_full_page->get_block freaks out if it is asked to read
315          * beyond the end of a file, so we check here.  Callers
316          * (generic_file_read, vm_ops->fault) are clever enough to check i_size
317          * and notice that the page they just read isn't needed.
318          *
319          * XXX sys_readahead() seems to get that wrong?
320          */
321         if (start >= i_size_read(inode)) {
322                 zero_user(page, 0, PAGE_SIZE);
323                 SetPageUptodate(page);
324                 ret = 0;
325                 goto out_alloc;
326         }
327 
328         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
329                 ret = ocfs2_readpage_inline(inode, page);
330         else
331                 ret = block_read_full_page(page, ocfs2_get_block);
332         unlock = 0;
333 
334 out_alloc:
335         up_read(&oi->ip_alloc_sem);
336 out_inode_unlock:
337         ocfs2_inode_unlock(inode, 0);
338 out:
339         if (unlock)
340                 unlock_page(page);
341         return ret;
342 }
343 
344 /*
345  * This is used only for read-ahead. Failures or difficult to handle
346  * situations are safe to ignore.
347  *
348  * Right now, we don't bother with BH_Boundary - in-inode extent lists
349  * are quite large (243 extents on 4k blocks), so most inodes don't
350  * grow out to a tree. If need be, detecting boundary extents could
351  * trivially be added in a future version of ocfs2_get_block().
352  */
353 static void ocfs2_readahead(struct readahead_control *rac)
354 {
355         int ret;
356         struct inode *inode = rac->mapping->host;
357         struct ocfs2_inode_info *oi = OCFS2_I(inode);
358 
359         /*
360          * Use the nonblocking flag for the dlm code to avoid page
361          * lock inversion, but don't bother with retrying.
362          */
363         ret = ocfs2_inode_lock_full(inode, NULL, 0, OCFS2_LOCK_NONBLOCK);
364         if (ret)
365                 return;
366 
367         if (down_read_trylock(&oi->ip_alloc_sem) == 0)
368                 goto out_unlock;
369 
370         /*
371          * Don't bother with inline-data. There isn't anything
372          * to read-ahead in that case anyway...
373          */
374         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL)
375                 goto out_up;
376 
377         /*
378          * Check whether a remote node truncated this file - we just
379          * drop out in that case as it's not worth handling here.
380          */
381         if (readahead_pos(rac) >= i_size_read(inode))
382                 goto out_up;
383 
384         mpage_readahead(rac, ocfs2_get_block);
385 
386 out_up:
387         up_read(&oi->ip_alloc_sem);
388 out_unlock:
389         ocfs2_inode_unlock(inode, 0);
390 }
391 
392 /* Note: Because we don't support holes, our allocation has
393  * already happened (allocation writes zeros to the file data)
394  * so we don't have to worry about ordered writes in
395  * ocfs2_writepage.
396  *
397  * ->writepage is called during the process of invalidating the page cache
398  * during blocked lock processing.  It can't block on any cluster locks
399  * to during block mapping.  It's relying on the fact that the block
400  * mapping can't have disappeared under the dirty pages that it is
401  * being asked to write back.
402  */
403 static int ocfs2_writepage(struct page *page, struct writeback_control *wbc)
404 {
405         trace_ocfs2_writepage(
406                 (unsigned long long)OCFS2_I(page->mapping->host)->ip_blkno,
407                 page->index);
408 
409         return block_write_full_page(page, ocfs2_get_block, wbc);
410 }
411 
412 /* Taken from ext3. We don't necessarily need the full blown
413  * functionality yet, but IMHO it's better to cut and paste the whole
414  * thing so we can avoid introducing our own bugs (and easily pick up
415  * their fixes when they happen) --Mark */
416 int walk_page_buffers(  handle_t *handle,
417                         struct buffer_head *head,
418                         unsigned from,
419                         unsigned to,
420                         int *partial,
421                         int (*fn)(      handle_t *handle,
422                                         struct buffer_head *bh))
423 {
424         struct buffer_head *bh;
425         unsigned block_start, block_end;
426         unsigned blocksize = head->b_size;
427         int err, ret = 0;
428         struct buffer_head *next;
429 
430         for (   bh = head, block_start = 0;
431                 ret == 0 && (bh != head || !block_start);
432                 block_start = block_end, bh = next)
433         {
434                 next = bh->b_this_page;
435                 block_end = block_start + blocksize;
436                 if (block_end <= from || block_start >= to) {
437                         if (partial && !buffer_uptodate(bh))
438                                 *partial = 1;
439                         continue;
440                 }
441                 err = (*fn)(handle, bh);
442                 if (!ret)
443                         ret = err;
444         }
445         return ret;
446 }
447 
448 static sector_t ocfs2_bmap(struct address_space *mapping, sector_t block)
449 {
450         sector_t status;
451         u64 p_blkno = 0;
452         int err = 0;
453         struct inode *inode = mapping->host;
454 
455         trace_ocfs2_bmap((unsigned long long)OCFS2_I(inode)->ip_blkno,
456                          (unsigned long long)block);
457 
458         /*
459          * The swap code (ab-)uses ->bmap to get a block mapping and then
460          * bypasseѕ the file system for actual I/O.  We really can't allow
461          * that on refcounted inodes, so we have to skip out here.  And yes,
462          * 0 is the magic code for a bmap error..
463          */
464         if (ocfs2_is_refcount_inode(inode))
465                 return 0;
466 
467         /* We don't need to lock journal system files, since they aren't
468          * accessed concurrently from multiple nodes.
469          */
470         if (!INODE_JOURNAL(inode)) {
471                 err = ocfs2_inode_lock(inode, NULL, 0);
472                 if (err) {
473                         if (err != -ENOENT)
474                                 mlog_errno(err);
475                         goto bail;
476                 }
477                 down_read(&OCFS2_I(inode)->ip_alloc_sem);
478         }
479 
480         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
481                 err = ocfs2_extent_map_get_blocks(inode, block, &p_blkno, NULL,
482                                                   NULL);
483 
484         if (!INODE_JOURNAL(inode)) {
485                 up_read(&OCFS2_I(inode)->ip_alloc_sem);
486                 ocfs2_inode_unlock(inode, 0);
487         }
488 
489         if (err) {
490                 mlog(ML_ERROR, "get_blocks() failed, block = %llu\n",
491                      (unsigned long long)block);
492                 mlog_errno(err);
493                 goto bail;
494         }
495 
496 bail:
497         status = err ? 0 : p_blkno;
498 
499         return status;
500 }
501 
502 static int ocfs2_releasepage(struct page *page, gfp_t wait)
503 {
504         if (!page_has_buffers(page))
505                 return 0;
506         return try_to_free_buffers(page);
507 }
508 
509 static void ocfs2_figure_cluster_boundaries(struct ocfs2_super *osb,
510                                             u32 cpos,
511                                             unsigned int *start,
512                                             unsigned int *end)
513 {
514         unsigned int cluster_start = 0, cluster_end = PAGE_SIZE;
515 
516         if (unlikely(PAGE_SHIFT > osb->s_clustersize_bits)) {
517                 unsigned int cpp;
518 
519                 cpp = 1 << (PAGE_SHIFT - osb->s_clustersize_bits);
520 
521                 cluster_start = cpos % cpp;
522                 cluster_start = cluster_start << osb->s_clustersize_bits;
523 
524                 cluster_end = cluster_start + osb->s_clustersize;
525         }
526 
527         BUG_ON(cluster_start > PAGE_SIZE);
528         BUG_ON(cluster_end > PAGE_SIZE);
529 
530         if (start)
531                 *start = cluster_start;
532         if (end)
533                 *end = cluster_end;
534 }
535 
536 /*
537  * 'from' and 'to' are the region in the page to avoid zeroing.
538  *
539  * If pagesize > clustersize, this function will avoid zeroing outside
540  * of the cluster boundary.
541  *
542  * from == to == 0 is code for "zero the entire cluster region"
543  */
544 static void ocfs2_clear_page_regions(struct page *page,
545                                      struct ocfs2_super *osb, u32 cpos,
546                                      unsigned from, unsigned to)
547 {
548         void *kaddr;
549         unsigned int cluster_start, cluster_end;
550 
551         ocfs2_figure_cluster_boundaries(osb, cpos, &cluster_start, &cluster_end);
552 
553         kaddr = kmap_atomic(page);
554 
555         if (from || to) {
556                 if (from > cluster_start)
557                         memset(kaddr + cluster_start, 0, from - cluster_start);
558                 if (to < cluster_end)
559                         memset(kaddr + to, 0, cluster_end - to);
560         } else {
561                 memset(kaddr + cluster_start, 0, cluster_end - cluster_start);
562         }
563 
564         kunmap_atomic(kaddr);
565 }
566 
567 /*
568  * Nonsparse file systems fully allocate before we get to the write
569  * code. This prevents ocfs2_write() from tagging the write as an
570  * allocating one, which means ocfs2_map_page_blocks() might try to
571  * read-in the blocks at the tail of our file. Avoid reading them by
572  * testing i_size against each block offset.
573  */
574 static int ocfs2_should_read_blk(struct inode *inode, struct page *page,
575                                  unsigned int block_start)
576 {
577         u64 offset = page_offset(page) + block_start;
578 
579         if (ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)))
580                 return 1;
581 
582         if (i_size_read(inode) > offset)
583                 return 1;
584 
585         return 0;
586 }
587 
588 /*
589  * Some of this taken from __block_write_begin(). We already have our
590  * mapping by now though, and the entire write will be allocating or
591  * it won't, so not much need to use BH_New.
592  *
593  * This will also skip zeroing, which is handled externally.
594  */
595 int ocfs2_map_page_blocks(struct page *page, u64 *p_blkno,
596                           struct inode *inode, unsigned int from,
597                           unsigned int to, int new)
598 {
599         int ret = 0;
600         struct buffer_head *head, *bh, *wait[2], **wait_bh = wait;
601         unsigned int block_end, block_start;
602         unsigned int bsize = i_blocksize(inode);
603 
604         if (!page_has_buffers(page))
605                 create_empty_buffers(page, bsize, 0);
606 
607         head = page_buffers(page);
608         for (bh = head, block_start = 0; bh != head || !block_start;
609              bh = bh->b_this_page, block_start += bsize) {
610                 block_end = block_start + bsize;
611 
612                 clear_buffer_new(bh);
613 
614                 /*
615                  * Ignore blocks outside of our i/o range -
616                  * they may belong to unallocated clusters.
617                  */
618                 if (block_start >= to || block_end <= from) {
619                         if (PageUptodate(page))
620                                 set_buffer_uptodate(bh);
621                         continue;
622                 }
623 
624                 /*
625                  * For an allocating write with cluster size >= page
626                  * size, we always write the entire page.
627                  */
628                 if (new)
629                         set_buffer_new(bh);
630 
631                 if (!buffer_mapped(bh)) {
632                         map_bh(bh, inode->i_sb, *p_blkno);
633                         clean_bdev_bh_alias(bh);
634                 }
635 
636                 if (PageUptodate(page)) {
637                         if (!buffer_uptodate(bh))
638                                 set_buffer_uptodate(bh);
639                 } else if (!buffer_uptodate(bh) && !buffer_delay(bh) &&
640                            !buffer_new(bh) &&
641                            ocfs2_should_read_blk(inode, page, block_start) &&
642                            (block_start < from || block_end > to)) {
643                         ll_rw_block(REQ_OP_READ, 0, 1, &bh);
644                         *wait_bh++=bh;
645                 }
646 
647                 *p_blkno = *p_blkno + 1;
648         }
649 
650         /*
651          * If we issued read requests - let them complete.
652          */
653         while(wait_bh > wait) {
654                 wait_on_buffer(*--wait_bh);
655                 if (!buffer_uptodate(*wait_bh))
656                         ret = -EIO;
657         }
658 
659         if (ret == 0 || !new)
660                 return ret;
661 
662         /*
663          * If we get -EIO above, zero out any newly allocated blocks
664          * to avoid exposing stale data.
665          */
666         bh = head;
667         block_start = 0;
668         do {
669                 block_end = block_start + bsize;
670                 if (block_end <= from)
671                         goto next_bh;
672                 if (block_start >= to)
673                         break;
674 
675                 zero_user(page, block_start, bh->b_size);
676                 set_buffer_uptodate(bh);
677                 mark_buffer_dirty(bh);
678 
679 next_bh:
680                 block_start = block_end;
681                 bh = bh->b_this_page;
682         } while (bh != head);
683 
684         return ret;
685 }
686 
687 #if (PAGE_SIZE >= OCFS2_MAX_CLUSTERSIZE)
688 #define OCFS2_MAX_CTXT_PAGES    1
689 #else
690 #define OCFS2_MAX_CTXT_PAGES    (OCFS2_MAX_CLUSTERSIZE / PAGE_SIZE)
691 #endif
692 
693 #define OCFS2_MAX_CLUSTERS_PER_PAGE     (PAGE_SIZE / OCFS2_MIN_CLUSTERSIZE)
694 
695 struct ocfs2_unwritten_extent {
696         struct list_head        ue_node;
697         struct list_head        ue_ip_node;
698         u32                     ue_cpos;
699         u32                     ue_phys;
700 };
701 
702 /*
703  * Describe the state of a single cluster to be written to.
704  */
705 struct ocfs2_write_cluster_desc {
706         u32             c_cpos;
707         u32             c_phys;
708         /*
709          * Give this a unique field because c_phys eventually gets
710          * filled.
711          */
712         unsigned        c_new;
713         unsigned        c_clear_unwritten;
714         unsigned        c_needs_zero;
715 };
716 
717 struct ocfs2_write_ctxt {
718         /* Logical cluster position / len of write */
719         u32                             w_cpos;
720         u32                             w_clen;
721 
722         /* First cluster allocated in a nonsparse extend */
723         u32                             w_first_new_cpos;
724 
725         /* Type of caller. Must be one of buffer, mmap, direct.  */
726         ocfs2_write_type_t              w_type;
727 
728         struct ocfs2_write_cluster_desc w_desc[OCFS2_MAX_CLUSTERS_PER_PAGE];
729 
730         /*
731          * This is true if page_size > cluster_size.
732          *
733          * It triggers a set of special cases during write which might
734          * have to deal with allocating writes to partial pages.
735          */
736         unsigned int                    w_large_pages;
737 
738         /*
739          * Pages involved in this write.
740          *
741          * w_target_page is the page being written to by the user.
742          *
743          * w_pages is an array of pages which always contains
744          * w_target_page, and in the case of an allocating write with
745          * page_size < cluster size, it will contain zero'd and mapped
746          * pages adjacent to w_target_page which need to be written
747          * out in so that future reads from that region will get
748          * zero's.
749          */
750         unsigned int                    w_num_pages;
751         struct page                     *w_pages[OCFS2_MAX_CTXT_PAGES];
752         struct page                     *w_target_page;
753 
754         /*
755          * w_target_locked is used for page_mkwrite path indicating no unlocking
756          * against w_target_page in ocfs2_write_end_nolock.
757          */
758         unsigned int                    w_target_locked:1;
759 
760         /*
761          * ocfs2_write_end() uses this to know what the real range to
762          * write in the target should be.
763          */
764         unsigned int                    w_target_from;
765         unsigned int                    w_target_to;
766 
767         /*
768          * We could use journal_current_handle() but this is cleaner,
769          * IMHO -Mark
770          */
771         handle_t                        *w_handle;
772 
773         struct buffer_head              *w_di_bh;
774 
775         struct ocfs2_cached_dealloc_ctxt w_dealloc;
776 
777         struct list_head                w_unwritten_list;
778         unsigned int                    w_unwritten_count;
779 };
780 
781 void ocfs2_unlock_and_free_pages(struct page **pages, int num_pages)
782 {
783         int i;
784 
785         for(i = 0; i < num_pages; i++) {
786                 if (pages[i]) {
787                         unlock_page(pages[i]);
788                         mark_page_accessed(pages[i]);
789                         put_page(pages[i]);
790                 }
791         }
792 }
793 
794 static void ocfs2_unlock_pages(struct ocfs2_write_ctxt *wc)
795 {
796         int i;
797 
798         /*
799          * w_target_locked is only set to true in the page_mkwrite() case.
800          * The intent is to allow us to lock the target page from write_begin()
801          * to write_end(). The caller must hold a ref on w_target_page.
802          */
803         if (wc->w_target_locked) {
804                 BUG_ON(!wc->w_target_page);
805                 for (i = 0; i < wc->w_num_pages; i++) {
806                         if (wc->w_target_page == wc->w_pages[i]) {
807                                 wc->w_pages[i] = NULL;
808                                 break;
809                         }
810                 }
811                 mark_page_accessed(wc->w_target_page);
812                 put_page(wc->w_target_page);
813         }
814         ocfs2_unlock_and_free_pages(wc->w_pages, wc->w_num_pages);
815 }
816 
817 static void ocfs2_free_unwritten_list(struct inode *inode,
818                                  struct list_head *head)
819 {
820         struct ocfs2_inode_info *oi = OCFS2_I(inode);
821         struct ocfs2_unwritten_extent *ue = NULL, *tmp = NULL;
822 
823         list_for_each_entry_safe(ue, tmp, head, ue_node) {
824                 list_del(&ue->ue_node);
825                 spin_lock(&oi->ip_lock);
826                 list_del(&ue->ue_ip_node);
827                 spin_unlock(&oi->ip_lock);
828                 kfree(ue);
829         }
830 }
831 
832 static void ocfs2_free_write_ctxt(struct inode *inode,
833                                   struct ocfs2_write_ctxt *wc)
834 {
835         ocfs2_free_unwritten_list(inode, &wc->w_unwritten_list);
836         ocfs2_unlock_pages(wc);
837         brelse(wc->w_di_bh);
838         kfree(wc);
839 }
840 
841 static int ocfs2_alloc_write_ctxt(struct ocfs2_write_ctxt **wcp,
842                                   struct ocfs2_super *osb, loff_t pos,
843                                   unsigned len, ocfs2_write_type_t type,
844                                   struct buffer_head *di_bh)
845 {
846         u32 cend;
847         struct ocfs2_write_ctxt *wc;
848 
849         wc = kzalloc(sizeof(struct ocfs2_write_ctxt), GFP_NOFS);
850         if (!wc)
851                 return -ENOMEM;
852 
853         wc->w_cpos = pos >> osb->s_clustersize_bits;
854         wc->w_first_new_cpos = UINT_MAX;
855         cend = (pos + len - 1) >> osb->s_clustersize_bits;
856         wc->w_clen = cend - wc->w_cpos + 1;
857         get_bh(di_bh);
858         wc->w_di_bh = di_bh;
859         wc->w_type = type;
860 
861         if (unlikely(PAGE_SHIFT > osb->s_clustersize_bits))
862                 wc->w_large_pages = 1;
863         else
864                 wc->w_large_pages = 0;
865 
866         ocfs2_init_dealloc_ctxt(&wc->w_dealloc);
867         INIT_LIST_HEAD(&wc->w_unwritten_list);
868 
869         *wcp = wc;
870 
871         return 0;
872 }
873 
874 /*
875  * If a page has any new buffers, zero them out here, and mark them uptodate
876  * and dirty so they'll be written out (in order to prevent uninitialised
877  * block data from leaking). And clear the new bit.
878  */
879 static void ocfs2_zero_new_buffers(struct page *page, unsigned from, unsigned to)
880 {
881         unsigned int block_start, block_end;
882         struct buffer_head *head, *bh;
883 
884         BUG_ON(!PageLocked(page));
885         if (!page_has_buffers(page))
886                 return;
887 
888         bh = head = page_buffers(page);
889         block_start = 0;
890         do {
891                 block_end = block_start + bh->b_size;
892 
893                 if (buffer_new(bh)) {
894                         if (block_end > from && block_start < to) {
895                                 if (!PageUptodate(page)) {
896                                         unsigned start, end;
897 
898                                         start = max(from, block_start);
899                                         end = min(to, block_end);
900 
901                                         zero_user_segment(page, start, end);
902                                         set_buffer_uptodate(bh);
903                                 }
904 
905                                 clear_buffer_new(bh);
906                                 mark_buffer_dirty(bh);
907                         }
908                 }
909 
910                 block_start = block_end;
911                 bh = bh->b_this_page;
912         } while (bh != head);
913 }
914 
915 /*
916  * Only called when we have a failure during allocating write to write
917  * zero's to the newly allocated region.
918  */
919 static void ocfs2_write_failure(struct inode *inode,
920                                 struct ocfs2_write_ctxt *wc,
921                                 loff_t user_pos, unsigned user_len)
922 {
923         int i;
924         unsigned from = user_pos & (PAGE_SIZE - 1),
925                 to = user_pos + user_len;
926         struct page *tmppage;
927 
928         if (wc->w_target_page)
929                 ocfs2_zero_new_buffers(wc->w_target_page, from, to);
930 
931         for(i = 0; i < wc->w_num_pages; i++) {
932                 tmppage = wc->w_pages[i];
933 
934                 if (tmppage && page_has_buffers(tmppage)) {
935                         if (ocfs2_should_order_data(inode))
936                                 ocfs2_jbd2_inode_add_write(wc->w_handle, inode,
937                                                            user_pos, user_len);
938 
939                         block_commit_write(tmppage, from, to);
940                 }
941         }
942 }
943 
944 static int ocfs2_prepare_page_for_write(struct inode *inode, u64 *p_blkno,
945                                         struct ocfs2_write_ctxt *wc,
946                                         struct page *page, u32 cpos,
947                                         loff_t user_pos, unsigned user_len,
948                                         int new)
949 {
950         int ret;
951         unsigned int map_from = 0, map_to = 0;
952         unsigned int cluster_start, cluster_end;
953         unsigned int user_data_from = 0, user_data_to = 0;
954 
955         ocfs2_figure_cluster_boundaries(OCFS2_SB(inode->i_sb), cpos,
956                                         &cluster_start, &cluster_end);
957 
958         /* treat the write as new if the a hole/lseek spanned across
959          * the page boundary.
960          */
961         new = new | ((i_size_read(inode) <= page_offset(page)) &&
962                         (page_offset(page) <= user_pos));
963 
964         if (page == wc->w_target_page) {
965                 map_from = user_pos & (PAGE_SIZE - 1);
966                 map_to = map_from + user_len;
967 
968                 if (new)
969                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
970                                                     cluster_start, cluster_end,
971                                                     new);
972                 else
973                         ret = ocfs2_map_page_blocks(page, p_blkno, inode,
974                                                     map_from, map_to, new);
975                 if (ret) {
976                         mlog_errno(ret);
977                         goto out;
978                 }
979 
980                 user_data_from = map_from;
981                 user_data_to = map_to;
982                 if (new) {
983                         map_from = cluster_start;
984                         map_to = cluster_end;
985                 }
986         } else {
987                 /*
988                  * If we haven't allocated the new page yet, we
989                  * shouldn't be writing it out without copying user
990                  * data. This is likely a math error from the caller.
991                  */
992                 BUG_ON(!new);
993 
994                 map_from = cluster_start;
995                 map_to = cluster_end;
996 
997                 ret = ocfs2_map_page_blocks(page, p_blkno, inode,
998                                             cluster_start, cluster_end, new);
999                 if (ret) {
1000                         mlog_errno(ret);
1001                         goto out;
1002                 }
1003         }
1004 
1005         /*
1006          * Parts of newly allocated pages need to be zero'd.
1007          *
1008          * Above, we have also rewritten 'to' and 'from' - as far as
1009          * the rest of the function is concerned, the entire cluster
1010          * range inside of a page needs to be written.
1011          *
1012          * We can skip this if the page is up to date - it's already
1013          * been zero'd from being read in as a hole.
1014          */
1015         if (new && !PageUptodate(page))
1016                 ocfs2_clear_page_regions(page, OCFS2_SB(inode->i_sb),
1017                                          cpos, user_data_from, user_data_to);
1018 
1019         flush_dcache_page(page);
1020 
1021 out:
1022         return ret;
1023 }
1024 
1025 /*
1026  * This function will only grab one clusters worth of pages.
1027  */
1028 static int ocfs2_grab_pages_for_write(struct address_space *mapping,
1029                                       struct ocfs2_write_ctxt *wc,
1030                                       u32 cpos, loff_t user_pos,
1031                                       unsigned user_len, int new,
1032                                       struct page *mmap_page)
1033 {
1034         int ret = 0, i;
1035         unsigned long start, target_index, end_index, index;
1036         struct inode *inode = mapping->host;
1037         loff_t last_byte;
1038 
1039         target_index = user_pos >> PAGE_SHIFT;
1040 
1041         /*
1042          * Figure out how many pages we'll be manipulating here. For
1043          * non allocating write, we just change the one
1044          * page. Otherwise, we'll need a whole clusters worth.  If we're
1045          * writing past i_size, we only need enough pages to cover the
1046          * last page of the write.
1047          */
1048         if (new) {
1049                 wc->w_num_pages = ocfs2_pages_per_cluster(inode->i_sb);
1050                 start = ocfs2_align_clusters_to_page_index(inode->i_sb, cpos);
1051                 /*
1052                  * We need the index *past* the last page we could possibly
1053                  * touch.  This is the page past the end of the write or
1054                  * i_size, whichever is greater.
1055                  */
1056                 last_byte = max(user_pos + user_len, i_size_read(inode));
1057                 BUG_ON(last_byte < 1);
1058                 end_index = ((last_byte - 1) >> PAGE_SHIFT) + 1;
1059                 if ((start + wc->w_num_pages) > end_index)
1060                         wc->w_num_pages = end_index - start;
1061         } else {
1062                 wc->w_num_pages = 1;
1063                 start = target_index;
1064         }
1065         end_index = (user_pos + user_len - 1) >> PAGE_SHIFT;
1066 
1067         for(i = 0; i < wc->w_num_pages; i++) {
1068                 index = start + i;
1069 
1070                 if (index >= target_index && index <= end_index &&
1071                     wc->w_type == OCFS2_WRITE_MMAP) {
1072                         /*
1073                          * ocfs2_pagemkwrite() is a little different
1074                          * and wants us to directly use the page
1075                          * passed in.
1076                          */
1077                         lock_page(mmap_page);
1078 
1079                         /* Exit and let the caller retry */
1080                         if (mmap_page->mapping != mapping) {
1081                                 WARN_ON(mmap_page->mapping);
1082                                 unlock_page(mmap_page);
1083                                 ret = -EAGAIN;
1084                                 goto out;
1085                         }
1086 
1087                         get_page(mmap_page);
1088                         wc->w_pages[i] = mmap_page;
1089                         wc->w_target_locked = true;
1090                 } else if (index >= target_index && index <= end_index &&
1091                            wc->w_type == OCFS2_WRITE_DIRECT) {
1092                         /* Direct write has no mapping page. */
1093                         wc->w_pages[i] = NULL;
1094                         continue;
1095                 } else {
1096                         wc->w_pages[i] = find_or_create_page(mapping, index,
1097                                                              GFP_NOFS);
1098                         if (!wc->w_pages[i]) {
1099                                 ret = -ENOMEM;
1100                                 mlog_errno(ret);
1101                                 goto out;
1102                         }
1103                 }
1104                 wait_for_stable_page(wc->w_pages[i]);
1105 
1106                 if (index == target_index)
1107                         wc->w_target_page = wc->w_pages[i];
1108         }
1109 out:
1110         if (ret)
1111                 wc->w_target_locked = false;
1112         return ret;
1113 }
1114 
1115 /*
1116  * Prepare a single cluster for write one cluster into the file.
1117  */
1118 static int ocfs2_write_cluster(struct address_space *mapping,
1119                                u32 *phys, unsigned int new,
1120                                unsigned int clear_unwritten,
1121                                unsigned int should_zero,
1122                                struct ocfs2_alloc_context *data_ac,
1123                                struct ocfs2_alloc_context *meta_ac,
1124                                struct ocfs2_write_ctxt *wc, u32 cpos,
1125                                loff_t user_pos, unsigned user_len)
1126 {
1127         int ret, i;
1128         u64 p_blkno;
1129         struct inode *inode = mapping->host;
1130         struct ocfs2_extent_tree et;
1131         int bpc = ocfs2_clusters_to_blocks(inode->i_sb, 1);
1132 
1133         if (new) {
1134                 u32 tmp_pos;
1135 
1136                 /*
1137                  * This is safe to call with the page locks - it won't take
1138                  * any additional semaphores or cluster locks.
1139                  */
1140                 tmp_pos = cpos;
1141                 ret = ocfs2_add_inode_data(OCFS2_SB(inode->i_sb), inode,
1142                                            &tmp_pos, 1, !clear_unwritten,
1143                                            wc->w_di_bh, wc->w_handle,
1144                                            data_ac, meta_ac, NULL);
1145                 /*
1146                  * This shouldn't happen because we must have already
1147                  * calculated the correct meta data allocation required. The
1148                  * internal tree allocation code should know how to increase
1149                  * transaction credits itself.
1150                  *
1151                  * If need be, we could handle -EAGAIN for a
1152                  * RESTART_TRANS here.
1153                  */
1154                 mlog_bug_on_msg(ret == -EAGAIN,
1155                                 "Inode %llu: EAGAIN return during allocation.\n",
1156                                 (unsigned long long)OCFS2_I(inode)->ip_blkno);
1157                 if (ret < 0) {
1158                         mlog_errno(ret);
1159                         goto out;
1160                 }
1161         } else if (clear_unwritten) {
1162                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1163                                               wc->w_di_bh);
1164                 ret = ocfs2_mark_extent_written(inode, &et,
1165                                                 wc->w_handle, cpos, 1, *phys,
1166                                                 meta_ac, &wc->w_dealloc);
1167                 if (ret < 0) {
1168                         mlog_errno(ret);
1169                         goto out;
1170                 }
1171         }
1172 
1173         /*
1174          * The only reason this should fail is due to an inability to
1175          * find the extent added.
1176          */
1177         ret = ocfs2_get_clusters(inode, cpos, phys, NULL, NULL);
1178         if (ret < 0) {
1179                 mlog(ML_ERROR, "Get physical blkno failed for inode %llu, "
1180                             "at logical cluster %u",
1181                             (unsigned long long)OCFS2_I(inode)->ip_blkno, cpos);
1182                 goto out;
1183         }
1184 
1185         BUG_ON(*phys == 0);
1186 
1187         p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, *phys);
1188         if (!should_zero)
1189                 p_blkno += (user_pos >> inode->i_sb->s_blocksize_bits) & (u64)(bpc - 1);
1190 
1191         for(i = 0; i < wc->w_num_pages; i++) {
1192                 int tmpret;
1193 
1194                 /* This is the direct io target page. */
1195                 if (wc->w_pages[i] == NULL) {
1196                         p_blkno++;
1197                         continue;
1198                 }
1199 
1200                 tmpret = ocfs2_prepare_page_for_write(inode, &p_blkno, wc,
1201                                                       wc->w_pages[i], cpos,
1202                                                       user_pos, user_len,
1203                                                       should_zero);
1204                 if (tmpret) {
1205                         mlog_errno(tmpret);
1206                         if (ret == 0)
1207                                 ret = tmpret;
1208                 }
1209         }
1210 
1211         /*
1212          * We only have cleanup to do in case of allocating write.
1213          */
1214         if (ret && new)
1215                 ocfs2_write_failure(inode, wc, user_pos, user_len);
1216 
1217 out:
1218 
1219         return ret;
1220 }
1221 
1222 static int ocfs2_write_cluster_by_desc(struct address_space *mapping,
1223                                        struct ocfs2_alloc_context *data_ac,
1224                                        struct ocfs2_alloc_context *meta_ac,
1225                                        struct ocfs2_write_ctxt *wc,
1226                                        loff_t pos, unsigned len)
1227 {
1228         int ret, i;
1229         loff_t cluster_off;
1230         unsigned int local_len = len;
1231         struct ocfs2_write_cluster_desc *desc;
1232         struct ocfs2_super *osb = OCFS2_SB(mapping->host->i_sb);
1233 
1234         for (i = 0; i < wc->w_clen; i++) {
1235                 desc = &wc->w_desc[i];
1236 
1237                 /*
1238                  * We have to make sure that the total write passed in
1239                  * doesn't extend past a single cluster.
1240                  */
1241                 local_len = len;
1242                 cluster_off = pos & (osb->s_clustersize - 1);
1243                 if ((cluster_off + local_len) > osb->s_clustersize)
1244                         local_len = osb->s_clustersize - cluster_off;
1245 
1246                 ret = ocfs2_write_cluster(mapping, &desc->c_phys,
1247                                           desc->c_new,
1248                                           desc->c_clear_unwritten,
1249                                           desc->c_needs_zero,
1250                                           data_ac, meta_ac,
1251                                           wc, desc->c_cpos, pos, local_len);
1252                 if (ret) {
1253                         mlog_errno(ret);
1254                         goto out;
1255                 }
1256 
1257                 len -= local_len;
1258                 pos += local_len;
1259         }
1260 
1261         ret = 0;
1262 out:
1263         return ret;
1264 }
1265 
1266 /*
1267  * ocfs2_write_end() wants to know which parts of the target page it
1268  * should complete the write on. It's easiest to compute them ahead of
1269  * time when a more complete view of the write is available.
1270  */
1271 static void ocfs2_set_target_boundaries(struct ocfs2_super *osb,
1272                                         struct ocfs2_write_ctxt *wc,
1273                                         loff_t pos, unsigned len, int alloc)
1274 {
1275         struct ocfs2_write_cluster_desc *desc;
1276 
1277         wc->w_target_from = pos & (PAGE_SIZE - 1);
1278         wc->w_target_to = wc->w_target_from + len;
1279 
1280         if (alloc == 0)
1281                 return;
1282 
1283         /*
1284          * Allocating write - we may have different boundaries based
1285          * on page size and cluster size.
1286          *
1287          * NOTE: We can no longer compute one value from the other as
1288          * the actual write length and user provided length may be
1289          * different.
1290          */
1291 
1292         if (wc->w_large_pages) {
1293                 /*
1294                  * We only care about the 1st and last cluster within
1295                  * our range and whether they should be zero'd or not. Either
1296                  * value may be extended out to the start/end of a
1297                  * newly allocated cluster.
1298                  */
1299                 desc = &wc->w_desc[0];
1300                 if (desc->c_needs_zero)
1301                         ocfs2_figure_cluster_boundaries(osb,
1302                                                         desc->c_cpos,
1303                                                         &wc->w_target_from,
1304                                                         NULL);
1305 
1306                 desc = &wc->w_desc[wc->w_clen - 1];
1307                 if (desc->c_needs_zero)
1308                         ocfs2_figure_cluster_boundaries(osb,
1309                                                         desc->c_cpos,
1310                                                         NULL,
1311                                                         &wc->w_target_to);
1312         } else {
1313                 wc->w_target_from = 0;
1314                 wc->w_target_to = PAGE_SIZE;
1315         }
1316 }
1317 
1318 /*
1319  * Check if this extent is marked UNWRITTEN by direct io. If so, we need not to
1320  * do the zero work. And should not to clear UNWRITTEN since it will be cleared
1321  * by the direct io procedure.
1322  * If this is a new extent that allocated by direct io, we should mark it in
1323  * the ip_unwritten_list.
1324  */
1325 static int ocfs2_unwritten_check(struct inode *inode,
1326                                  struct ocfs2_write_ctxt *wc,
1327                                  struct ocfs2_write_cluster_desc *desc)
1328 {
1329         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1330         struct ocfs2_unwritten_extent *ue = NULL, *new = NULL;
1331         int ret = 0;
1332 
1333         if (!desc->c_needs_zero)
1334                 return 0;
1335 
1336 retry:
1337         spin_lock(&oi->ip_lock);
1338         /* Needs not to zero no metter buffer or direct. The one who is zero
1339          * the cluster is doing zero. And he will clear unwritten after all
1340          * cluster io finished. */
1341         list_for_each_entry(ue, &oi->ip_unwritten_list, ue_ip_node) {
1342                 if (desc->c_cpos == ue->ue_cpos) {
1343                         BUG_ON(desc->c_new);
1344                         desc->c_needs_zero = 0;
1345                         desc->c_clear_unwritten = 0;
1346                         goto unlock;
1347                 }
1348         }
1349 
1350         if (wc->w_type != OCFS2_WRITE_DIRECT)
1351                 goto unlock;
1352 
1353         if (new == NULL) {
1354                 spin_unlock(&oi->ip_lock);
1355                 new = kmalloc(sizeof(struct ocfs2_unwritten_extent),
1356                              GFP_NOFS);
1357                 if (new == NULL) {
1358                         ret = -ENOMEM;
1359                         goto out;
1360                 }
1361                 goto retry;
1362         }
1363         /* This direct write will doing zero. */
1364         new->ue_cpos = desc->c_cpos;
1365         new->ue_phys = desc->c_phys;
1366         desc->c_clear_unwritten = 0;
1367         list_add_tail(&new->ue_ip_node, &oi->ip_unwritten_list);
1368         list_add_tail(&new->ue_node, &wc->w_unwritten_list);
1369         wc->w_unwritten_count++;
1370         new = NULL;
1371 unlock:
1372         spin_unlock(&oi->ip_lock);
1373 out:
1374         kfree(new);
1375         return ret;
1376 }
1377 
1378 /*
1379  * Populate each single-cluster write descriptor in the write context
1380  * with information about the i/o to be done.
1381  *
1382  * Returns the number of clusters that will have to be allocated, as
1383  * well as a worst case estimate of the number of extent records that
1384  * would have to be created during a write to an unwritten region.
1385  */
1386 static int ocfs2_populate_write_desc(struct inode *inode,
1387                                      struct ocfs2_write_ctxt *wc,
1388                                      unsigned int *clusters_to_alloc,
1389                                      unsigned int *extents_to_split)
1390 {
1391         int ret;
1392         struct ocfs2_write_cluster_desc *desc;
1393         unsigned int num_clusters = 0;
1394         unsigned int ext_flags = 0;
1395         u32 phys = 0;
1396         int i;
1397 
1398         *clusters_to_alloc = 0;
1399         *extents_to_split = 0;
1400 
1401         for (i = 0; i < wc->w_clen; i++) {
1402                 desc = &wc->w_desc[i];
1403                 desc->c_cpos = wc->w_cpos + i;
1404 
1405                 if (num_clusters == 0) {
1406                         /*
1407                          * Need to look up the next extent record.
1408                          */
1409                         ret = ocfs2_get_clusters(inode, desc->c_cpos, &phys,
1410                                                  &num_clusters, &ext_flags);
1411                         if (ret) {
1412                                 mlog_errno(ret);
1413                                 goto out;
1414                         }
1415 
1416                         /* We should already CoW the refcountd extent. */
1417                         BUG_ON(ext_flags & OCFS2_EXT_REFCOUNTED);
1418 
1419                         /*
1420                          * Assume worst case - that we're writing in
1421                          * the middle of the extent.
1422                          *
1423                          * We can assume that the write proceeds from
1424                          * left to right, in which case the extent
1425                          * insert code is smart enough to coalesce the
1426                          * next splits into the previous records created.
1427                          */
1428                         if (ext_flags & OCFS2_EXT_UNWRITTEN)
1429                                 *extents_to_split = *extents_to_split + 2;
1430                 } else if (phys) {
1431                         /*
1432                          * Only increment phys if it doesn't describe
1433                          * a hole.
1434                          */
1435                         phys++;
1436                 }
1437 
1438                 /*
1439                  * If w_first_new_cpos is < UINT_MAX, we have a non-sparse
1440                  * file that got extended.  w_first_new_cpos tells us
1441                  * where the newly allocated clusters are so we can
1442                  * zero them.
1443                  */
1444                 if (desc->c_cpos >= wc->w_first_new_cpos) {
1445                         BUG_ON(phys == 0);
1446                         desc->c_needs_zero = 1;
1447                 }
1448 
1449                 desc->c_phys = phys;
1450                 if (phys == 0) {
1451                         desc->c_new = 1;
1452                         desc->c_needs_zero = 1;
1453                         desc->c_clear_unwritten = 1;
1454                         *clusters_to_alloc = *clusters_to_alloc + 1;
1455                 }
1456 
1457                 if (ext_flags & OCFS2_EXT_UNWRITTEN) {
1458                         desc->c_clear_unwritten = 1;
1459                         desc->c_needs_zero = 1;
1460                 }
1461 
1462                 ret = ocfs2_unwritten_check(inode, wc, desc);
1463                 if (ret) {
1464                         mlog_errno(ret);
1465                         goto out;
1466                 }
1467 
1468                 num_clusters--;
1469         }
1470 
1471         ret = 0;
1472 out:
1473         return ret;
1474 }
1475 
1476 static int ocfs2_write_begin_inline(struct address_space *mapping,
1477                                     struct inode *inode,
1478                                     struct ocfs2_write_ctxt *wc)
1479 {
1480         int ret;
1481         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1482         struct page *page;
1483         handle_t *handle;
1484         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1485 
1486         handle = ocfs2_start_trans(osb, OCFS2_INODE_UPDATE_CREDITS);
1487         if (IS_ERR(handle)) {
1488                 ret = PTR_ERR(handle);
1489                 mlog_errno(ret);
1490                 goto out;
1491         }
1492 
1493         page = find_or_create_page(mapping, 0, GFP_NOFS);
1494         if (!page) {
1495                 ocfs2_commit_trans(osb, handle);
1496                 ret = -ENOMEM;
1497                 mlog_errno(ret);
1498                 goto out;
1499         }
1500         /*
1501          * If we don't set w_num_pages then this page won't get unlocked
1502          * and freed on cleanup of the write context.
1503          */
1504         wc->w_pages[0] = wc->w_target_page = page;
1505         wc->w_num_pages = 1;
1506 
1507         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1508                                       OCFS2_JOURNAL_ACCESS_WRITE);
1509         if (ret) {
1510                 ocfs2_commit_trans(osb, handle);
1511 
1512                 mlog_errno(ret);
1513                 goto out;
1514         }
1515 
1516         if (!(OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL))
1517                 ocfs2_set_inode_data_inline(inode, di);
1518 
1519         if (!PageUptodate(page)) {
1520                 ret = ocfs2_read_inline_data(inode, page, wc->w_di_bh);
1521                 if (ret) {
1522                         ocfs2_commit_trans(osb, handle);
1523 
1524                         goto out;
1525                 }
1526         }
1527 
1528         wc->w_handle = handle;
1529 out:
1530         return ret;
1531 }
1532 
1533 int ocfs2_size_fits_inline_data(struct buffer_head *di_bh, u64 new_size)
1534 {
1535         struct ocfs2_dinode *di = (struct ocfs2_dinode *)di_bh->b_data;
1536 
1537         if (new_size <= le16_to_cpu(di->id2.i_data.id_count))
1538                 return 1;
1539         return 0;
1540 }
1541 
1542 static int ocfs2_try_to_write_inline_data(struct address_space *mapping,
1543                                           struct inode *inode, loff_t pos,
1544                                           unsigned len, struct page *mmap_page,
1545                                           struct ocfs2_write_ctxt *wc)
1546 {
1547         int ret, written = 0;
1548         loff_t end = pos + len;
1549         struct ocfs2_inode_info *oi = OCFS2_I(inode);
1550         struct ocfs2_dinode *di = NULL;
1551 
1552         trace_ocfs2_try_to_write_inline_data((unsigned long long)oi->ip_blkno,
1553                                              len, (unsigned long long)pos,
1554                                              oi->ip_dyn_features);
1555 
1556         /*
1557          * Handle inodes which already have inline data 1st.
1558          */
1559         if (oi->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1560                 if (mmap_page == NULL &&
1561                     ocfs2_size_fits_inline_data(wc->w_di_bh, end))
1562                         goto do_inline_write;
1563 
1564                 /*
1565                  * The write won't fit - we have to give this inode an
1566                  * inline extent list now.
1567                  */
1568                 ret = ocfs2_convert_inline_data_to_extents(inode, wc->w_di_bh);
1569                 if (ret)
1570                         mlog_errno(ret);
1571                 goto out;
1572         }
1573 
1574         /*
1575          * Check whether the inode can accept inline data.
1576          */
1577         if (oi->ip_clusters != 0 || i_size_read(inode) != 0)
1578                 return 0;
1579 
1580         /*
1581          * Check whether the write can fit.
1582          */
1583         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1584         if (mmap_page ||
1585             end > ocfs2_max_inline_data_with_xattr(inode->i_sb, di))
1586                 return 0;
1587 
1588 do_inline_write:
1589         ret = ocfs2_write_begin_inline(mapping, inode, wc);
1590         if (ret) {
1591                 mlog_errno(ret);
1592                 goto out;
1593         }
1594 
1595         /*
1596          * This signals to the caller that the data can be written
1597          * inline.
1598          */
1599         written = 1;
1600 out:
1601         return written ? written : ret;
1602 }
1603 
1604 /*
1605  * This function only does anything for file systems which can't
1606  * handle sparse files.
1607  *
1608  * What we want to do here is fill in any hole between the current end
1609  * of allocation and the end of our write. That way the rest of the
1610  * write path can treat it as an non-allocating write, which has no
1611  * special case code for sparse/nonsparse files.
1612  */
1613 static int ocfs2_expand_nonsparse_inode(struct inode *inode,
1614                                         struct buffer_head *di_bh,
1615                                         loff_t pos, unsigned len,
1616                                         struct ocfs2_write_ctxt *wc)
1617 {
1618         int ret;
1619         loff_t newsize = pos + len;
1620 
1621         BUG_ON(ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1622 
1623         if (newsize <= i_size_read(inode))
1624                 return 0;
1625 
1626         ret = ocfs2_extend_no_holes(inode, di_bh, newsize, pos);
1627         if (ret)
1628                 mlog_errno(ret);
1629 
1630         /* There is no wc if this is call from direct. */
1631         if (wc)
1632                 wc->w_first_new_cpos =
1633                         ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode));
1634 
1635         return ret;
1636 }
1637 
1638 static int ocfs2_zero_tail(struct inode *inode, struct buffer_head *di_bh,
1639                            loff_t pos)
1640 {
1641         int ret = 0;
1642 
1643         BUG_ON(!ocfs2_sparse_alloc(OCFS2_SB(inode->i_sb)));
1644         if (pos > i_size_read(inode))
1645                 ret = ocfs2_zero_extend(inode, di_bh, pos);
1646 
1647         return ret;
1648 }
1649 
1650 int ocfs2_write_begin_nolock(struct address_space *mapping,
1651                              loff_t pos, unsigned len, ocfs2_write_type_t type,
1652                              struct page **pagep, void **fsdata,
1653                              struct buffer_head *di_bh, struct page *mmap_page)
1654 {
1655         int ret, cluster_of_pages, credits = OCFS2_INODE_UPDATE_CREDITS;
1656         unsigned int clusters_to_alloc, extents_to_split, clusters_need = 0;
1657         struct ocfs2_write_ctxt *wc;
1658         struct inode *inode = mapping->host;
1659         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1660         struct ocfs2_dinode *di;
1661         struct ocfs2_alloc_context *data_ac = NULL;
1662         struct ocfs2_alloc_context *meta_ac = NULL;
1663         handle_t *handle;
1664         struct ocfs2_extent_tree et;
1665         int try_free = 1, ret1;
1666 
1667 try_again:
1668         ret = ocfs2_alloc_write_ctxt(&wc, osb, pos, len, type, di_bh);
1669         if (ret) {
1670                 mlog_errno(ret);
1671                 return ret;
1672         }
1673 
1674         if (ocfs2_supports_inline_data(osb)) {
1675                 ret = ocfs2_try_to_write_inline_data(mapping, inode, pos, len,
1676                                                      mmap_page, wc);
1677                 if (ret == 1) {
1678                         ret = 0;
1679                         goto success;
1680                 }
1681                 if (ret < 0) {
1682                         mlog_errno(ret);
1683                         goto out;
1684                 }
1685         }
1686 
1687         /* Direct io change i_size late, should not zero tail here. */
1688         if (type != OCFS2_WRITE_DIRECT) {
1689                 if (ocfs2_sparse_alloc(osb))
1690                         ret = ocfs2_zero_tail(inode, di_bh, pos);
1691                 else
1692                         ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
1693                                                            len, wc);
1694                 if (ret) {
1695                         mlog_errno(ret);
1696                         goto out;
1697                 }
1698         }
1699 
1700         ret = ocfs2_check_range_for_refcount(inode, pos, len);
1701         if (ret < 0) {
1702                 mlog_errno(ret);
1703                 goto out;
1704         } else if (ret == 1) {
1705                 clusters_need = wc->w_clen;
1706                 ret = ocfs2_refcount_cow(inode, di_bh,
1707                                          wc->w_cpos, wc->w_clen, UINT_MAX);
1708                 if (ret) {
1709                         mlog_errno(ret);
1710                         goto out;
1711                 }
1712         }
1713 
1714         ret = ocfs2_populate_write_desc(inode, wc, &clusters_to_alloc,
1715                                         &extents_to_split);
1716         if (ret) {
1717                 mlog_errno(ret);
1718                 goto out;
1719         }
1720         clusters_need += clusters_to_alloc;
1721 
1722         di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1723 
1724         trace_ocfs2_write_begin_nolock(
1725                         (unsigned long long)OCFS2_I(inode)->ip_blkno,
1726                         (long long)i_size_read(inode),
1727                         le32_to_cpu(di->i_clusters),
1728                         pos, len, type, mmap_page,
1729                         clusters_to_alloc, extents_to_split);
1730 
1731         /*
1732          * We set w_target_from, w_target_to here so that
1733          * ocfs2_write_end() knows which range in the target page to
1734          * write out. An allocation requires that we write the entire
1735          * cluster range.
1736          */
1737         if (clusters_to_alloc || extents_to_split) {
1738                 /*
1739                  * XXX: We are stretching the limits of
1740                  * ocfs2_lock_allocators(). It greatly over-estimates
1741                  * the work to be done.
1742                  */
1743                 ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode),
1744                                               wc->w_di_bh);
1745                 ret = ocfs2_lock_allocators(inode, &et,
1746                                             clusters_to_alloc, extents_to_split,
1747                                             &data_ac, &meta_ac);
1748                 if (ret) {
1749                         mlog_errno(ret);
1750                         goto out;
1751                 }
1752 
1753                 if (data_ac)
1754                         data_ac->ac_resv = &OCFS2_I(inode)->ip_la_data_resv;
1755 
1756                 credits = ocfs2_calc_extend_credits(inode->i_sb,
1757                                                     &di->id2.i_list);
1758         } else if (type == OCFS2_WRITE_DIRECT)
1759                 /* direct write needs not to start trans if no extents alloc. */
1760                 goto success;
1761 
1762         /*
1763          * We have to zero sparse allocated clusters, unwritten extent clusters,
1764          * and non-sparse clusters we just extended.  For non-sparse writes,
1765          * we know zeros will only be needed in the first and/or last cluster.
1766          */
1767         if (wc->w_clen && (wc->w_desc[0].c_needs_zero ||
1768                            wc->w_desc[wc->w_clen - 1].c_needs_zero))
1769                 cluster_of_pages = 1;
1770         else
1771                 cluster_of_pages = 0;
1772 
1773         ocfs2_set_target_boundaries(osb, wc, pos, len, cluster_of_pages);
1774 
1775         handle = ocfs2_start_trans(osb, credits);
1776         if (IS_ERR(handle)) {
1777                 ret = PTR_ERR(handle);
1778                 mlog_errno(ret);
1779                 goto out;
1780         }
1781 
1782         wc->w_handle = handle;
1783 
1784         if (clusters_to_alloc) {
1785                 ret = dquot_alloc_space_nodirty(inode,
1786                         ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1787                 if (ret)
1788                         goto out_commit;
1789         }
1790 
1791         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), wc->w_di_bh,
1792                                       OCFS2_JOURNAL_ACCESS_WRITE);
1793         if (ret) {
1794                 mlog_errno(ret);
1795                 goto out_quota;
1796         }
1797 
1798         /*
1799          * Fill our page array first. That way we've grabbed enough so
1800          * that we can zero and flush if we error after adding the
1801          * extent.
1802          */
1803         ret = ocfs2_grab_pages_for_write(mapping, wc, wc->w_cpos, pos, len,
1804                                          cluster_of_pages, mmap_page);
1805         if (ret && ret != -EAGAIN) {
1806                 mlog_errno(ret);
1807                 goto out_quota;
1808         }
1809 
1810         /*
1811          * ocfs2_grab_pages_for_write() returns -EAGAIN if it could not lock
1812          * the target page. In this case, we exit with no error and no target
1813          * page. This will trigger the caller, page_mkwrite(), to re-try
1814          * the operation.
1815          */
1816         if (ret == -EAGAIN) {
1817                 BUG_ON(wc->w_target_page);
1818                 ret = 0;
1819                 goto out_quota;
1820         }
1821 
1822         ret = ocfs2_write_cluster_by_desc(mapping, data_ac, meta_ac, wc, pos,
1823                                           len);
1824         if (ret) {
1825                 mlog_errno(ret);
1826                 goto out_quota;
1827         }
1828 
1829         if (data_ac)
1830                 ocfs2_free_alloc_context(data_ac);
1831         if (meta_ac)
1832                 ocfs2_free_alloc_context(meta_ac);
1833 
1834 success:
1835         if (pagep)
1836                 *pagep = wc->w_target_page;
1837         *fsdata = wc;
1838         return 0;
1839 out_quota:
1840         if (clusters_to_alloc)
1841                 dquot_free_space(inode,
1842                           ocfs2_clusters_to_bytes(osb->sb, clusters_to_alloc));
1843 out_commit:
1844         ocfs2_commit_trans(osb, handle);
1845 
1846 out:
1847         /*
1848          * The mmapped page won't be unlocked in ocfs2_free_write_ctxt(),
1849          * even in case of error here like ENOSPC and ENOMEM. So, we need
1850          * to unlock the target page manually to prevent deadlocks when
1851          * retrying again on ENOSPC, or when returning non-VM_FAULT_LOCKED
1852          * to VM code.
1853          */
1854         if (wc->w_target_locked)
1855                 unlock_page(mmap_page);
1856 
1857         ocfs2_free_write_ctxt(inode, wc);
1858 
1859         if (data_ac) {
1860                 ocfs2_free_alloc_context(data_ac);
1861                 data_ac = NULL;
1862         }
1863         if (meta_ac) {
1864                 ocfs2_free_alloc_context(meta_ac);
1865                 meta_ac = NULL;
1866         }
1867 
1868         if (ret == -ENOSPC && try_free) {
1869                 /*
1870                  * Try to free some truncate log so that we can have enough
1871                  * clusters to allocate.
1872                  */
1873                 try_free = 0;
1874 
1875                 ret1 = ocfs2_try_to_free_truncate_log(osb, clusters_need);
1876                 if (ret1 == 1)
1877                         goto try_again;
1878 
1879                 if (ret1 < 0)
1880                         mlog_errno(ret1);
1881         }
1882 
1883         return ret;
1884 }
1885 
1886 static int ocfs2_write_begin(struct file *file, struct address_space *mapping,
1887                              loff_t pos, unsigned len, unsigned flags,
1888                              struct page **pagep, void **fsdata)
1889 {
1890         int ret;
1891         struct buffer_head *di_bh = NULL;
1892         struct inode *inode = mapping->host;
1893 
1894         ret = ocfs2_inode_lock(inode, &di_bh, 1);
1895         if (ret) {
1896                 mlog_errno(ret);
1897                 return ret;
1898         }
1899 
1900         /*
1901          * Take alloc sem here to prevent concurrent lookups. That way
1902          * the mapping, zeroing and tree manipulation within
1903          * ocfs2_write() will be safe against ->readpage(). This
1904          * should also serve to lock out allocation from a shared
1905          * writeable region.
1906          */
1907         down_write(&OCFS2_I(inode)->ip_alloc_sem);
1908 
1909         ret = ocfs2_write_begin_nolock(mapping, pos, len, OCFS2_WRITE_BUFFER,
1910                                        pagep, fsdata, di_bh, NULL);
1911         if (ret) {
1912                 mlog_errno(ret);
1913                 goto out_fail;
1914         }
1915 
1916         brelse(di_bh);
1917 
1918         return 0;
1919 
1920 out_fail:
1921         up_write(&OCFS2_I(inode)->ip_alloc_sem);
1922 
1923         brelse(di_bh);
1924         ocfs2_inode_unlock(inode, 1);
1925 
1926         return ret;
1927 }
1928 
1929 static void ocfs2_write_end_inline(struct inode *inode, loff_t pos,
1930                                    unsigned len, unsigned *copied,
1931                                    struct ocfs2_dinode *di,
1932                                    struct ocfs2_write_ctxt *wc)
1933 {
1934         void *kaddr;
1935 
1936         if (unlikely(*copied < len)) {
1937                 if (!PageUptodate(wc->w_target_page)) {
1938                         *copied = 0;
1939                         return;
1940                 }
1941         }
1942 
1943         kaddr = kmap_atomic(wc->w_target_page);
1944         memcpy(di->id2.i_data.id_data + pos, kaddr + pos, *copied);
1945         kunmap_atomic(kaddr);
1946 
1947         trace_ocfs2_write_end_inline(
1948              (unsigned long long)OCFS2_I(inode)->ip_blkno,
1949              (unsigned long long)pos, *copied,
1950              le16_to_cpu(di->id2.i_data.id_count),
1951              le16_to_cpu(di->i_dyn_features));
1952 }
1953 
1954 int ocfs2_write_end_nolock(struct address_space *mapping,
1955                            loff_t pos, unsigned len, unsigned copied, void *fsdata)
1956 {
1957         int i, ret;
1958         unsigned from, to, start = pos & (PAGE_SIZE - 1);
1959         struct inode *inode = mapping->host;
1960         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
1961         struct ocfs2_write_ctxt *wc = fsdata;
1962         struct ocfs2_dinode *di = (struct ocfs2_dinode *)wc->w_di_bh->b_data;
1963         handle_t *handle = wc->w_handle;
1964         struct page *tmppage;
1965 
1966         BUG_ON(!list_empty(&wc->w_unwritten_list));
1967 
1968         if (handle) {
1969                 ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode),
1970                                 wc->w_di_bh, OCFS2_JOURNAL_ACCESS_WRITE);
1971                 if (ret) {
1972                         copied = ret;
1973                         mlog_errno(ret);
1974                         goto out;
1975                 }
1976         }
1977 
1978         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL) {
1979                 ocfs2_write_end_inline(inode, pos, len, &copied, di, wc);
1980                 goto out_write_size;
1981         }
1982 
1983         if (unlikely(copied < len) && wc->w_target_page) {
1984                 if (!PageUptodate(wc->w_target_page))
1985                         copied = 0;
1986 
1987                 ocfs2_zero_new_buffers(wc->w_target_page, start+copied,
1988                                        start+len);
1989         }
1990         if (wc->w_target_page)
1991                 flush_dcache_page(wc->w_target_page);
1992 
1993         for(i = 0; i < wc->w_num_pages; i++) {
1994                 tmppage = wc->w_pages[i];
1995 
1996                 /* This is the direct io target page. */
1997                 if (tmppage == NULL)
1998                         continue;
1999 
2000                 if (tmppage == wc->w_target_page) {
2001                         from = wc->w_target_from;
2002                         to = wc->w_target_to;
2003 
2004                         BUG_ON(from > PAGE_SIZE ||
2005                                to > PAGE_SIZE ||
2006                                to < from);
2007                 } else {
2008                         /*
2009                          * Pages adjacent to the target (if any) imply
2010                          * a hole-filling write in which case we want
2011                          * to flush their entire range.
2012                          */
2013                         from = 0;
2014                         to = PAGE_SIZE;
2015                 }
2016 
2017                 if (page_has_buffers(tmppage)) {
2018                         if (handle && ocfs2_should_order_data(inode)) {
2019                                 loff_t start_byte =
2020                                         ((loff_t)tmppage->index << PAGE_SHIFT) +
2021                                         from;
2022                                 loff_t length = to - from;
2023                                 ocfs2_jbd2_inode_add_write(handle, inode,
2024                                                            start_byte, length);
2025                         }
2026                         block_commit_write(tmppage, from, to);
2027                 }
2028         }
2029 
2030 out_write_size:
2031         /* Direct io do not update i_size here. */
2032         if (wc->w_type != OCFS2_WRITE_DIRECT) {
2033                 pos += copied;
2034                 if (pos > i_size_read(inode)) {
2035                         i_size_write(inode, pos);
2036                         mark_inode_dirty(inode);
2037                 }
2038                 inode->i_blocks = ocfs2_inode_sector_count(inode);
2039                 di->i_size = cpu_to_le64((u64)i_size_read(inode));
2040                 inode->i_mtime = inode->i_ctime = current_time(inode);
2041                 di->i_mtime = di->i_ctime = cpu_to_le64(inode->i_mtime.tv_sec);
2042                 di->i_mtime_nsec = di->i_ctime_nsec = cpu_to_le32(inode->i_mtime.tv_nsec);
2043                 if (handle)
2044                         ocfs2_update_inode_fsync_trans(handle, inode, 1);
2045         }
2046         if (handle)
2047                 ocfs2_journal_dirty(handle, wc->w_di_bh);
2048 
2049 out:
2050         /* unlock pages before dealloc since it needs acquiring j_trans_barrier
2051          * lock, or it will cause a deadlock since journal commit threads holds
2052          * this lock and will ask for the page lock when flushing the data.
2053          * put it here to preserve the unlock order.
2054          */
2055         ocfs2_unlock_pages(wc);
2056 
2057         if (handle)
2058                 ocfs2_commit_trans(osb, handle);
2059 
2060         ocfs2_run_deallocs(osb, &wc->w_dealloc);
2061 
2062         brelse(wc->w_di_bh);
2063         kfree(wc);
2064 
2065         return copied;
2066 }
2067 
2068 static int ocfs2_write_end(struct file *file, struct address_space *mapping,
2069                            loff_t pos, unsigned len, unsigned copied,
2070                            struct page *page, void *fsdata)
2071 {
2072         int ret;
2073         struct inode *inode = mapping->host;
2074 
2075         ret = ocfs2_write_end_nolock(mapping, pos, len, copied, fsdata);
2076 
2077         up_write(&OCFS2_I(inode)->ip_alloc_sem);
2078         ocfs2_inode_unlock(inode, 1);
2079 
2080         return ret;
2081 }
2082 
2083 struct ocfs2_dio_write_ctxt {
2084         struct list_head        dw_zero_list;
2085         unsigned                dw_zero_count;
2086         int                     dw_orphaned;
2087         pid_t                   dw_writer_pid;
2088 };
2089 
2090 static struct ocfs2_dio_write_ctxt *
2091 ocfs2_dio_alloc_write_ctx(struct buffer_head *bh, int *alloc)
2092 {
2093         struct ocfs2_dio_write_ctxt *dwc = NULL;
2094 
2095         if (bh->b_private)
2096                 return bh->b_private;
2097 
2098         dwc = kmalloc(sizeof(struct ocfs2_dio_write_ctxt), GFP_NOFS);
2099         if (dwc == NULL)
2100                 return NULL;
2101         INIT_LIST_HEAD(&dwc->dw_zero_list);
2102         dwc->dw_zero_count = 0;
2103         dwc->dw_orphaned = 0;
2104         dwc->dw_writer_pid = task_pid_nr(current);
2105         bh->b_private = dwc;
2106         *alloc = 1;
2107 
2108         return dwc;
2109 }
2110 
2111 static void ocfs2_dio_free_write_ctx(struct inode *inode,
2112                                      struct ocfs2_dio_write_ctxt *dwc)
2113 {
2114         ocfs2_free_unwritten_list(inode, &dwc->dw_zero_list);
2115         kfree(dwc);
2116 }
2117 
2118 /*
2119  * TODO: Make this into a generic get_blocks function.
2120  *
2121  * From do_direct_io in direct-io.c:
2122  *  "So what we do is to permit the ->get_blocks function to populate
2123  *   bh.b_size with the size of IO which is permitted at this offset and
2124  *   this i_blkbits."
2125  *
2126  * This function is called directly from get_more_blocks in direct-io.c.
2127  *
2128  * called like this: dio->get_blocks(dio->inode, fs_startblk,
2129  *                                      fs_count, map_bh, dio->rw == WRITE);
2130  */
2131 static int ocfs2_dio_wr_get_block(struct inode *inode, sector_t iblock,
2132                                struct buffer_head *bh_result, int create)
2133 {
2134         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2135         struct ocfs2_inode_info *oi = OCFS2_I(inode);
2136         struct ocfs2_write_ctxt *wc;
2137         struct ocfs2_write_cluster_desc *desc = NULL;
2138         struct ocfs2_dio_write_ctxt *dwc = NULL;
2139         struct buffer_head *di_bh = NULL;
2140         u64 p_blkno;
2141         unsigned int i_blkbits = inode->i_sb->s_blocksize_bits;
2142         loff_t pos = iblock << i_blkbits;
2143         sector_t endblk = (i_size_read(inode) - 1) >> i_blkbits;
2144         unsigned len, total_len = bh_result->b_size;
2145         int ret = 0, first_get_block = 0;
2146 
2147         len = osb->s_clustersize - (pos & (osb->s_clustersize - 1));
2148         len = min(total_len, len);
2149 
2150         /*
2151          * bh_result->b_size is count in get_more_blocks according to write
2152          * "pos" and "end", we need map twice to return different buffer state:
2153          * 1. area in file size, not set NEW;
2154          * 2. area out file size, set  NEW.
2155          *
2156          *                 iblock    endblk
2157          * |--------|---------|---------|---------
2158          * |<-------area in file------->|
2159          */
2160 
2161         if ((iblock <= endblk) &&
2162             ((iblock + ((len - 1) >> i_blkbits)) > endblk))
2163                 len = (endblk - iblock + 1) << i_blkbits;
2164 
2165         mlog(0, "get block of %lu at %llu:%u req %u\n",
2166                         inode->i_ino, pos, len, total_len);
2167 
2168         /*
2169          * Because we need to change file size in ocfs2_dio_end_io_write(), or
2170          * we may need to add it to orphan dir. So can not fall to fast path
2171          * while file size will be changed.
2172          */
2173         if (pos + total_len <= i_size_read(inode)) {
2174 
2175                 /* This is the fast path for re-write. */
2176                 ret = ocfs2_lock_get_block(inode, iblock, bh_result, create);
2177                 if (buffer_mapped(bh_result) &&
2178                     !buffer_new(bh_result) &&
2179                     ret == 0)
2180                         goto out;
2181 
2182                 /* Clear state set by ocfs2_get_block. */
2183                 bh_result->b_state = 0;
2184         }
2185 
2186         dwc = ocfs2_dio_alloc_write_ctx(bh_result, &first_get_block);
2187         if (unlikely(dwc == NULL)) {
2188                 ret = -ENOMEM;
2189                 mlog_errno(ret);
2190                 goto out;
2191         }
2192 
2193         if (ocfs2_clusters_for_bytes(inode->i_sb, pos + total_len) >
2194             ocfs2_clusters_for_bytes(inode->i_sb, i_size_read(inode)) &&
2195             !dwc->dw_orphaned) {
2196                 /*
2197                  * when we are going to alloc extents beyond file size, add the
2198                  * inode to orphan dir, so we can recall those spaces when
2199                  * system crashed during write.
2200                  */
2201                 ret = ocfs2_add_inode_to_orphan(osb, inode);
2202                 if (ret < 0) {
2203                         mlog_errno(ret);
2204                         goto out;
2205                 }
2206                 dwc->dw_orphaned = 1;
2207         }
2208 
2209         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2210         if (ret) {
2211                 mlog_errno(ret);
2212                 goto out;
2213         }
2214 
2215         down_write(&oi->ip_alloc_sem);
2216 
2217         if (first_get_block) {
2218                 if (ocfs2_sparse_alloc(osb))
2219                         ret = ocfs2_zero_tail(inode, di_bh, pos);
2220                 else
2221                         ret = ocfs2_expand_nonsparse_inode(inode, di_bh, pos,
2222                                                            total_len, NULL);
2223                 if (ret < 0) {
2224                         mlog_errno(ret);
2225                         goto unlock;
2226                 }
2227         }
2228 
2229         ret = ocfs2_write_begin_nolock(inode->i_mapping, pos, len,
2230                                        OCFS2_WRITE_DIRECT, NULL,
2231                                        (void **)&wc, di_bh, NULL);
2232         if (ret) {
2233                 mlog_errno(ret);
2234                 goto unlock;
2235         }
2236 
2237         desc = &wc->w_desc[0];
2238 
2239         p_blkno = ocfs2_clusters_to_blocks(inode->i_sb, desc->c_phys);
2240         BUG_ON(p_blkno == 0);
2241         p_blkno += iblock & (u64)(ocfs2_clusters_to_blocks(inode->i_sb, 1) - 1);
2242 
2243         map_bh(bh_result, inode->i_sb, p_blkno);
2244         bh_result->b_size = len;
2245         if (desc->c_needs_zero)
2246                 set_buffer_new(bh_result);
2247 
2248         if (iblock > endblk)
2249                 set_buffer_new(bh_result);
2250 
2251         /* May sleep in end_io. It should not happen in a irq context. So defer
2252          * it to dio work queue. */
2253         set_buffer_defer_completion(bh_result);
2254 
2255         if (!list_empty(&wc->w_unwritten_list)) {
2256                 struct ocfs2_unwritten_extent *ue = NULL;
2257 
2258                 ue = list_first_entry(&wc->w_unwritten_list,
2259                                       struct ocfs2_unwritten_extent,
2260                                       ue_node);
2261                 BUG_ON(ue->ue_cpos != desc->c_cpos);
2262                 /* The physical address may be 0, fill it. */
2263                 ue->ue_phys = desc->c_phys;
2264 
2265                 list_splice_tail_init(&wc->w_unwritten_list, &dwc->dw_zero_list);
2266                 dwc->dw_zero_count += wc->w_unwritten_count;
2267         }
2268 
2269         ret = ocfs2_write_end_nolock(inode->i_mapping, pos, len, len, wc);
2270         BUG_ON(ret != len);
2271         ret = 0;
2272 unlock:
2273         up_write(&oi->ip_alloc_sem);
2274         ocfs2_inode_unlock(inode, 1);
2275         brelse(di_bh);
2276 out:
2277         if (ret < 0)
2278                 ret = -EIO;
2279         return ret;
2280 }
2281 
2282 static int ocfs2_dio_end_io_write(struct inode *inode,
2283                                   struct ocfs2_dio_write_ctxt *dwc,
2284                                   loff_t offset,
2285                                   ssize_t bytes)
2286 {
2287         struct ocfs2_cached_dealloc_ctxt dealloc;
2288         struct ocfs2_extent_tree et;
2289         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2290         struct ocfs2_inode_info *oi = OCFS2_I(inode);
2291         struct ocfs2_unwritten_extent *ue = NULL;
2292         struct buffer_head *di_bh = NULL;
2293         struct ocfs2_dinode *di;
2294         struct ocfs2_alloc_context *data_ac = NULL;
2295         struct ocfs2_alloc_context *meta_ac = NULL;
2296         handle_t *handle = NULL;
2297         loff_t end = offset + bytes;
2298         int ret = 0, credits = 0;
2299 
2300         ocfs2_init_dealloc_ctxt(&dealloc);
2301 
2302         /* We do clear unwritten, delete orphan, change i_size here. If neither
2303          * of these happen, we can skip all this. */
2304         if (list_empty(&dwc->dw_zero_list) &&
2305             end <= i_size_read(inode) &&
2306             !dwc->dw_orphaned)
2307                 goto out;
2308 
2309         ret = ocfs2_inode_lock(inode, &di_bh, 1);
2310         if (ret < 0) {
2311                 mlog_errno(ret);
2312                 goto out;
2313         }
2314 
2315         down_write(&oi->ip_alloc_sem);
2316 
2317         /* Delete orphan before acquire i_mutex. */
2318         if (dwc->dw_orphaned) {
2319                 BUG_ON(dwc->dw_writer_pid != task_pid_nr(current));
2320 
2321                 end = end > i_size_read(inode) ? end : 0;
2322 
2323                 ret = ocfs2_del_inode_from_orphan(osb, inode, di_bh,
2324                                 !!end, end);
2325                 if (ret < 0)
2326                         mlog_errno(ret);
2327         }
2328 
2329         di = (struct ocfs2_dinode *)di_bh->b_data;
2330 
2331         ocfs2_init_dinode_extent_tree(&et, INODE_CACHE(inode), di_bh);
2332 
2333         /* Attach dealloc with extent tree in case that we may reuse extents
2334          * which are already unlinked from current extent tree due to extent
2335          * rotation and merging.
2336          */
2337         et.et_dealloc = &dealloc;
2338 
2339         ret = ocfs2_lock_allocators(inode, &et, 0, dwc->dw_zero_count*2,
2340                                     &data_ac, &meta_ac);
2341         if (ret) {
2342                 mlog_errno(ret);
2343                 goto unlock;
2344         }
2345 
2346         credits = ocfs2_calc_extend_credits(inode->i_sb, &di->id2.i_list);
2347 
2348         handle = ocfs2_start_trans(osb, credits);
2349         if (IS_ERR(handle)) {
2350                 ret = PTR_ERR(handle);
2351                 mlog_errno(ret);
2352                 goto unlock;
2353         }
2354         ret = ocfs2_journal_access_di(handle, INODE_CACHE(inode), di_bh,
2355                                       OCFS2_JOURNAL_ACCESS_WRITE);
2356         if (ret) {
2357                 mlog_errno(ret);
2358                 goto commit;
2359         }
2360 
2361         list_for_each_entry(ue, &dwc->dw_zero_list, ue_node) {
2362                 ret = ocfs2_mark_extent_written(inode, &et, handle,
2363                                                 ue->ue_cpos, 1,
2364                                                 ue->ue_phys,
2365                                                 meta_ac, &dealloc);
2366                 if (ret < 0) {
2367                         mlog_errno(ret);
2368                         break;
2369                 }
2370         }
2371 
2372         if (end > i_size_read(inode)) {
2373                 ret = ocfs2_set_inode_size(handle, inode, di_bh, end);
2374                 if (ret < 0)
2375                         mlog_errno(ret);
2376         }
2377 commit:
2378         ocfs2_commit_trans(osb, handle);
2379 unlock:
2380         up_write(&oi->ip_alloc_sem);
2381         ocfs2_inode_unlock(inode, 1);
2382         brelse(di_bh);
2383 out:
2384         if (data_ac)
2385                 ocfs2_free_alloc_context(data_ac);
2386         if (meta_ac)
2387                 ocfs2_free_alloc_context(meta_ac);
2388         ocfs2_run_deallocs(osb, &dealloc);
2389         ocfs2_dio_free_write_ctx(inode, dwc);
2390 
2391         return ret;
2392 }
2393 
2394 /*
2395  * ocfs2_dio_end_io is called by the dio core when a dio is finished.  We're
2396  * particularly interested in the aio/dio case.  We use the rw_lock DLM lock
2397  * to protect io on one node from truncation on another.
2398  */
2399 static int ocfs2_dio_end_io(struct kiocb *iocb,
2400                             loff_t offset,
2401                             ssize_t bytes,
2402                             void *private)
2403 {
2404         struct inode *inode = file_inode(iocb->ki_filp);
2405         int level;
2406         int ret = 0;
2407 
2408         /* this io's submitter should not have unlocked this before we could */
2409         BUG_ON(!ocfs2_iocb_is_rw_locked(iocb));
2410 
2411         if (bytes <= 0)
2412                 mlog_ratelimited(ML_ERROR, "Direct IO failed, bytes = %lld",
2413                                  (long long)bytes);
2414         if (private) {
2415                 if (bytes > 0)
2416                         ret = ocfs2_dio_end_io_write(inode, private, offset,
2417                                                      bytes);
2418                 else
2419                         ocfs2_dio_free_write_ctx(inode, private);
2420         }
2421 
2422         ocfs2_iocb_clear_rw_locked(iocb);
2423 
2424         level = ocfs2_iocb_rw_locked_level(iocb);
2425         ocfs2_rw_unlock(inode, level);
2426         return ret;
2427 }
2428 
2429 static ssize_t ocfs2_direct_IO(struct kiocb *iocb, struct iov_iter *iter)
2430 {
2431         struct file *file = iocb->ki_filp;
2432         struct inode *inode = file->f_mapping->host;
2433         struct ocfs2_super *osb = OCFS2_SB(inode->i_sb);
2434         get_block_t *get_block;
2435 
2436         /*
2437          * Fallback to buffered I/O if we see an inode without
2438          * extents.
2439          */
2440         if (OCFS2_I(inode)->ip_dyn_features & OCFS2_INLINE_DATA_FL)
2441                 return 0;
2442 
2443         /* Fallback to buffered I/O if we do not support append dio. */
2444         if (iocb->ki_pos + iter->count > i_size_read(inode) &&
2445             !ocfs2_supports_append_dio(osb))
2446                 return 0;
2447 
2448         if (iov_iter_rw(iter) == READ)
2449                 get_block = ocfs2_lock_get_block;
2450         else
2451                 get_block = ocfs2_dio_wr_get_block;
2452 
2453         return __blockdev_direct_IO(iocb, inode, inode->i_sb->s_bdev,
2454                                     iter, get_block,
2455                                     ocfs2_dio_end_io, NULL, 0);
2456 }
2457 
2458 const struct address_space_operations ocfs2_aops = {
2459         .readpage               = ocfs2_readpage,
2460         .readahead              = ocfs2_readahead,
2461         .writepage              = ocfs2_writepage,
2462         .write_begin            = ocfs2_write_begin,
2463         .write_end              = ocfs2_write_end,
2464         .bmap                   = ocfs2_bmap,
2465         .direct_IO              = ocfs2_direct_IO,
2466         .invalidatepage         = block_invalidatepage,
2467         .releasepage            = ocfs2_releasepage,
2468         .migratepage            = buffer_migrate_page,
2469         .is_partially_uptodate  = block_is_partially_uptodate,
2470         .error_remove_page      = generic_error_remove_page,
2471 };
2472 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp