NTFS: Fix read regression.
[linux-2.6.git] / fs / ntfs / aops.c
1 /**
2  * aops.c - NTFS kernel address space operations and page cache handling.
3  *          Part of the Linux-NTFS project.
4  *
5  * Copyright (c) 2001-2007 Anton Altaparmakov
6  * Copyright (c) 2002 Richard Russon
7  *
8  * This program/include file is free software; you can redistribute it and/or
9  * modify it under the terms of the GNU General Public License as published
10  * by the Free Software Foundation; either version 2 of the License, or
11  * (at your option) any later version.
12  *
13  * This program/include file is distributed in the hope that it will be
14  * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
15  * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
16  * GNU General Public License for more details.
17  *
18  * You should have received a copy of the GNU General Public License
19  * along with this program (in the main directory of the Linux-NTFS
20  * distribution in the file COPYING); if not, write to the Free Software
21  * Foundation,Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
22  */
23
24 #include <linux/errno.h>
25 #include <linux/fs.h>
26 #include <linux/mm.h>
27 #include <linux/pagemap.h>
28 #include <linux/swap.h>
29 #include <linux/buffer_head.h>
30 #include <linux/writeback.h>
31 #include <linux/bit_spinlock.h>
32
33 #include "aops.h"
34 #include "attrib.h"
35 #include "debug.h"
36 #include "inode.h"
37 #include "mft.h"
38 #include "runlist.h"
39 #include "types.h"
40 #include "ntfs.h"
41
42 /**
43  * ntfs_end_buffer_async_read - async io completion for reading attributes
44  * @bh:         buffer head on which io is completed
45  * @uptodate:   whether @bh is now uptodate or not
46  *
47  * Asynchronous I/O completion handler for reading pages belonging to the
48  * attribute address space of an inode.  The inodes can either be files or
49  * directories or they can be fake inodes describing some attribute.
50  *
51  * If NInoMstProtected(), perform the post read mst fixups when all IO on the
52  * page has been completed and mark the page uptodate or set the error bit on
53  * the page.  To determine the size of the records that need fixing up, we
54  * cheat a little bit by setting the index_block_size in ntfs_inode to the ntfs
55  * record size, and index_block_size_bits, to the log(base 2) of the ntfs
56  * record size.
57  */
58 static void ntfs_end_buffer_async_read(struct buffer_head *bh, int uptodate)
59 {
60         unsigned long flags;
61         struct buffer_head *first, *tmp;
62         struct page *page;
63         struct inode *vi;
64         ntfs_inode *ni;
65         int page_uptodate = 1;
66
67         page = bh->b_page;
68         vi = page->mapping->host;
69         ni = NTFS_I(vi);
70
71         if (likely(uptodate)) {
72                 loff_t i_size;
73                 s64 file_ofs, init_size;
74
75                 set_buffer_uptodate(bh);
76
77                 file_ofs = ((s64)page->index << PAGE_CACHE_SHIFT) +
78                                 bh_offset(bh);
79                 read_lock_irqsave(&ni->size_lock, flags);
80                 init_size = ni->initialized_size;
81                 i_size = i_size_read(vi);
82                 read_unlock_irqrestore(&ni->size_lock, flags);
83                 if (unlikely(init_size > i_size)) {
84                         /* Race with shrinking truncate. */
85                         init_size = i_size;
86                 }
87                 /* Check for the current buffer head overflowing. */
88                 if (unlikely(file_ofs + bh->b_size > init_size)) {
89                         int ofs;
90
91                         ofs = 0;
92                         if (file_ofs < init_size)
93                                 ofs = init_size - file_ofs;
94                         local_irq_save(flags);
95                         zero_user_page(page, bh_offset(bh) + ofs,
96                                          bh->b_size - ofs, KM_BIO_SRC_IRQ);
97                         local_irq_restore(flags);
98                 }
99         } else {
100                 clear_buffer_uptodate(bh);
101                 SetPageError(page);
102                 ntfs_error(ni->vol->sb, "Buffer I/O error, logical block "
103                                 "0x%llx.", (unsigned long long)bh->b_blocknr);
104         }
105         first = page_buffers(page);
106         local_irq_save(flags);
107         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
108         clear_buffer_async_read(bh);
109         unlock_buffer(bh);
110         tmp = bh;
111         do {
112                 if (!buffer_uptodate(tmp))
113                         page_uptodate = 0;
114                 if (buffer_async_read(tmp)) {
115                         if (likely(buffer_locked(tmp)))
116                                 goto still_busy;
117                         /* Async buffers must be locked. */
118                         BUG();
119                 }
120                 tmp = tmp->b_this_page;
121         } while (tmp != bh);
122         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
123         local_irq_restore(flags);
124         /*
125          * If none of the buffers had errors then we can set the page uptodate,
126          * but we first have to perform the post read mst fixups, if the
127          * attribute is mst protected, i.e. if NInoMstProteced(ni) is true.
128          * Note we ignore fixup errors as those are detected when
129          * map_mft_record() is called which gives us per record granularity
130          * rather than per page granularity.
131          */
132         if (!NInoMstProtected(ni)) {
133                 if (likely(page_uptodate && !PageError(page)))
134                         SetPageUptodate(page);
135         } else {
136                 u8 *kaddr;
137                 unsigned int i, recs;
138                 u32 rec_size;
139
140                 rec_size = ni->itype.index.block_size;
141                 recs = PAGE_CACHE_SIZE / rec_size;
142                 /* Should have been verified before we got here... */
143                 BUG_ON(!recs);
144                 local_irq_save(flags);
145                 kaddr = kmap_atomic(page, KM_BIO_SRC_IRQ);
146                 for (i = 0; i < recs; i++)
147                         post_read_mst_fixup((NTFS_RECORD*)(kaddr +
148                                         i * rec_size), rec_size);
149                 kunmap_atomic(kaddr, KM_BIO_SRC_IRQ);
150                 local_irq_restore(flags);
151                 flush_dcache_page(page);
152                 if (likely(page_uptodate && !PageError(page)))
153                         SetPageUptodate(page);
154         }
155         unlock_page(page);
156         return;
157 still_busy:
158         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
159         local_irq_restore(flags);
160         return;
161 }
162
163 /**
164  * ntfs_read_block - fill a @page of an address space with data
165  * @page:       page cache page to fill with data
166  *
167  * Fill the page @page of the address space belonging to the @page->host inode.
168  * We read each buffer asynchronously and when all buffers are read in, our io
169  * completion handler ntfs_end_buffer_read_async(), if required, automatically
170  * applies the mst fixups to the page before finally marking it uptodate and
171  * unlocking it.
172  *
173  * We only enforce allocated_size limit because i_size is checked for in
174  * generic_file_read().
175  *
176  * Return 0 on success and -errno on error.
177  *
178  * Contains an adapted version of fs/buffer.c::block_read_full_page().
179  */
180 static int ntfs_read_block(struct page *page)
181 {
182         loff_t i_size;
183         VCN vcn;
184         LCN lcn;
185         s64 init_size;
186         struct inode *vi;
187         ntfs_inode *ni;
188         ntfs_volume *vol;
189         runlist_element *rl;
190         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
191         sector_t iblock, lblock, zblock;
192         unsigned long flags;
193         unsigned int blocksize, vcn_ofs;
194         int i, nr;
195         unsigned char blocksize_bits;
196
197         vi = page->mapping->host;
198         ni = NTFS_I(vi);
199         vol = ni->vol;
200
201         /* $MFT/$DATA must have its complete runlist in memory at all times. */
202         BUG_ON(!ni->runlist.rl && !ni->mft_no && !NInoAttr(ni));
203
204         blocksize = vol->sb->s_blocksize;
205         blocksize_bits = vol->sb->s_blocksize_bits;
206
207         if (!page_has_buffers(page)) {
208                 create_empty_buffers(page, blocksize, 0);
209                 if (unlikely(!page_has_buffers(page))) {
210                         unlock_page(page);
211                         return -ENOMEM;
212                 }
213         }
214         bh = head = page_buffers(page);
215         BUG_ON(!bh);
216
217         /*
218          * We may be racing with truncate.  To avoid some of the problems we
219          * now take a snapshot of the various sizes and use those for the whole
220          * of the function.  In case of an extending truncate it just means we
221          * may leave some buffers unmapped which are now allocated.  This is
222          * not a problem since these buffers will just get mapped when a write
223          * occurs.  In case of a shrinking truncate, we will detect this later
224          * on due to the runlist being incomplete and if the page is being
225          * fully truncated, truncate will throw it away as soon as we unlock
226          * it so no need to worry what we do with it.
227          */
228         iblock = (s64)page->index << (PAGE_CACHE_SHIFT - blocksize_bits);
229         read_lock_irqsave(&ni->size_lock, flags);
230         lblock = (ni->allocated_size + blocksize - 1) >> blocksize_bits;
231         init_size = ni->initialized_size;
232         i_size = i_size_read(vi);
233         read_unlock_irqrestore(&ni->size_lock, flags);
234         if (unlikely(init_size > i_size)) {
235                 /* Race with shrinking truncate. */
236                 init_size = i_size;
237         }
238         zblock = (init_size + blocksize - 1) >> blocksize_bits;
239
240         /* Loop through all the buffers in the page. */
241         rl = NULL;
242         nr = i = 0;
243         do {
244                 int err = 0;
245
246                 if (unlikely(buffer_uptodate(bh)))
247                         continue;
248                 if (unlikely(buffer_mapped(bh))) {
249                         arr[nr++] = bh;
250                         continue;
251                 }
252                 bh->b_bdev = vol->sb->s_bdev;
253                 /* Is the block within the allowed limits? */
254                 if (iblock < lblock) {
255                         bool is_retry = false;
256
257                         /* Convert iblock into corresponding vcn and offset. */
258                         vcn = (VCN)iblock << blocksize_bits >>
259                                         vol->cluster_size_bits;
260                         vcn_ofs = ((VCN)iblock << blocksize_bits) &
261                                         vol->cluster_size_mask;
262                         if (!rl) {
263 lock_retry_remap:
264                                 down_read(&ni->runlist.lock);
265                                 rl = ni->runlist.rl;
266                         }
267                         if (likely(rl != NULL)) {
268                                 /* Seek to element containing target vcn. */
269                                 while (rl->length && rl[1].vcn <= vcn)
270                                         rl++;
271                                 lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
272                         } else
273                                 lcn = LCN_RL_NOT_MAPPED;
274                         /* Successful remap. */
275                         if (lcn >= 0) {
276                                 /* Setup buffer head to correct block. */
277                                 bh->b_blocknr = ((lcn << vol->cluster_size_bits)
278                                                 + vcn_ofs) >> blocksize_bits;
279                                 set_buffer_mapped(bh);
280                                 /* Only read initialized data blocks. */
281                                 if (iblock < zblock) {
282                                         arr[nr++] = bh;
283                                         continue;
284                                 }
285                                 /* Fully non-initialized data block, zero it. */
286                                 goto handle_zblock;
287                         }
288                         /* It is a hole, need to zero it. */
289                         if (lcn == LCN_HOLE)
290                                 goto handle_hole;
291                         /* If first try and runlist unmapped, map and retry. */
292                         if (!is_retry && lcn == LCN_RL_NOT_MAPPED) {
293                                 is_retry = true;
294                                 /*
295                                  * Attempt to map runlist, dropping lock for
296                                  * the duration.
297                                  */
298                                 up_read(&ni->runlist.lock);
299                                 err = ntfs_map_runlist(ni, vcn);
300                                 if (likely(!err))
301                                         goto lock_retry_remap;
302                                 rl = NULL;
303                         } else if (!rl)
304                                 up_read(&ni->runlist.lock);
305                         /*
306                          * If buffer is outside the runlist, treat it as a
307                          * hole.  This can happen due to concurrent truncate
308                          * for example.
309                          */
310                         if (err == -ENOENT || lcn == LCN_ENOENT) {
311                                 err = 0;
312                                 goto handle_hole;
313                         }
314                         /* Hard error, zero out region. */
315                         if (!err)
316                                 err = -EIO;
317                         bh->b_blocknr = -1;
318                         SetPageError(page);
319                         ntfs_error(vol->sb, "Failed to read from inode 0x%lx, "
320                                         "attribute type 0x%x, vcn 0x%llx, "
321                                         "offset 0x%x because its location on "
322                                         "disk could not be determined%s "
323                                         "(error code %i).", ni->mft_no,
324                                         ni->type, (unsigned long long)vcn,
325                                         vcn_ofs, is_retry ? " even after "
326                                         "retrying" : "", err);
327                 }
328                 /*
329                  * Either iblock was outside lblock limits or
330                  * ntfs_rl_vcn_to_lcn() returned error.  Just zero that portion
331                  * of the page and set the buffer uptodate.
332                  */
333 handle_hole:
334                 bh->b_blocknr = -1UL;
335                 clear_buffer_mapped(bh);
336 handle_zblock:
337                 zero_user_page(page, i * blocksize, blocksize, KM_USER0);
338                 if (likely(!err))
339                         set_buffer_uptodate(bh);
340         } while (i++, iblock++, (bh = bh->b_this_page) != head);
341
342         /* Release the lock if we took it. */
343         if (rl)
344                 up_read(&ni->runlist.lock);
345
346         /* Check we have at least one buffer ready for i/o. */
347         if (nr) {
348                 struct buffer_head *tbh;
349
350                 /* Lock the buffers. */
351                 for (i = 0; i < nr; i++) {
352                         tbh = arr[i];
353                         lock_buffer(tbh);
354                         tbh->b_end_io = ntfs_end_buffer_async_read;
355                         set_buffer_async_read(tbh);
356                 }
357                 /* Finally, start i/o on the buffers. */
358                 for (i = 0; i < nr; i++) {
359                         tbh = arr[i];
360                         if (likely(!buffer_uptodate(tbh)))
361                                 submit_bh(READ, tbh);
362                         else
363                                 ntfs_end_buffer_async_read(tbh, 1);
364                 }
365                 return 0;
366         }
367         /* No i/o was scheduled on any of the buffers. */
368         if (likely(!PageError(page)))
369                 SetPageUptodate(page);
370         else /* Signal synchronous i/o error. */
371                 nr = -EIO;
372         unlock_page(page);
373         return nr;
374 }
375
376 /**
377  * ntfs_readpage - fill a @page of a @file with data from the device
378  * @file:       open file to which the page @page belongs or NULL
379  * @page:       page cache page to fill with data
380  *
381  * For non-resident attributes, ntfs_readpage() fills the @page of the open
382  * file @file by calling the ntfs version of the generic block_read_full_page()
383  * function, ntfs_read_block(), which in turn creates and reads in the buffers
384  * associated with the page asynchronously.
385  *
386  * For resident attributes, OTOH, ntfs_readpage() fills @page by copying the
387  * data from the mft record (which at this stage is most likely in memory) and
388  * fills the remainder with zeroes. Thus, in this case, I/O is synchronous, as
389  * even if the mft record is not cached at this point in time, we need to wait
390  * for it to be read in before we can do the copy.
391  *
392  * Return 0 on success and -errno on error.
393  */
394 static int ntfs_readpage(struct file *file, struct page *page)
395 {
396         loff_t i_size;
397         struct inode *vi;
398         ntfs_inode *ni, *base_ni;
399         u8 *addr;
400         ntfs_attr_search_ctx *ctx;
401         MFT_RECORD *mrec;
402         unsigned long flags;
403         u32 attr_len;
404         int err = 0;
405
406 retry_readpage:
407         BUG_ON(!PageLocked(page));
408         vi = page->mapping->host;
409         i_size = i_size_read(vi);
410         /* Is the page fully outside i_size? (truncate in progress) */
411         if (unlikely(page->index >= (i_size + PAGE_CACHE_SIZE - 1) >>
412                         PAGE_CACHE_SHIFT)) {
413                 zero_user_page(page, 0, PAGE_CACHE_SIZE, KM_USER0);
414                 ntfs_debug("Read outside i_size - truncated?");
415                 goto done;
416         }
417         /*
418          * This can potentially happen because we clear PageUptodate() during
419          * ntfs_writepage() of MstProtected() attributes.
420          */
421         if (PageUptodate(page)) {
422                 unlock_page(page);
423                 return 0;
424         }
425         ni = NTFS_I(vi);
426         /*
427          * Only $DATA attributes can be encrypted and only unnamed $DATA
428          * attributes can be compressed.  Index root can have the flags set but
429          * this means to create compressed/encrypted files, not that the
430          * attribute is compressed/encrypted.  Note we need to check for
431          * AT_INDEX_ALLOCATION since this is the type of both directory and
432          * index inodes.
433          */
434         if (ni->type != AT_INDEX_ALLOCATION) {
435                 /* If attribute is encrypted, deny access, just like NT4. */
436                 if (NInoEncrypted(ni)) {
437                         BUG_ON(ni->type != AT_DATA);
438                         err = -EACCES;
439                         goto err_out;
440                 }
441                 /* Compressed data streams are handled in compress.c. */
442                 if (NInoNonResident(ni) && NInoCompressed(ni)) {
443                         BUG_ON(ni->type != AT_DATA);
444                         BUG_ON(ni->name_len);
445                         return ntfs_read_compressed_block(page);
446                 }
447         }
448         /* NInoNonResident() == NInoIndexAllocPresent() */
449         if (NInoNonResident(ni)) {
450                 /* Normal, non-resident data stream. */
451                 return ntfs_read_block(page);
452         }
453         /*
454          * Attribute is resident, implying it is not compressed or encrypted.
455          * This also means the attribute is smaller than an mft record and
456          * hence smaller than a page, so can simply zero out any pages with
457          * index above 0.  Note the attribute can actually be marked compressed
458          * but if it is resident the actual data is not compressed so we are
459          * ok to ignore the compressed flag here.
460          */
461         if (unlikely(page->index > 0)) {
462                 zero_user_page(page, 0, PAGE_CACHE_SIZE, KM_USER0);
463                 goto done;
464         }
465         if (!NInoAttr(ni))
466                 base_ni = ni;
467         else
468                 base_ni = ni->ext.base_ntfs_ino;
469         /* Map, pin, and lock the mft record. */
470         mrec = map_mft_record(base_ni);
471         if (IS_ERR(mrec)) {
472                 err = PTR_ERR(mrec);
473                 goto err_out;
474         }
475         /*
476          * If a parallel write made the attribute non-resident, drop the mft
477          * record and retry the readpage.
478          */
479         if (unlikely(NInoNonResident(ni))) {
480                 unmap_mft_record(base_ni);
481                 goto retry_readpage;
482         }
483         ctx = ntfs_attr_get_search_ctx(base_ni, mrec);
484         if (unlikely(!ctx)) {
485                 err = -ENOMEM;
486                 goto unm_err_out;
487         }
488         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
489                         CASE_SENSITIVE, 0, NULL, 0, ctx);
490         if (unlikely(err))
491                 goto put_unm_err_out;
492         attr_len = le32_to_cpu(ctx->attr->data.resident.value_length);
493         read_lock_irqsave(&ni->size_lock, flags);
494         if (unlikely(attr_len > ni->initialized_size))
495                 attr_len = ni->initialized_size;
496         i_size = i_size_read(vi);
497         read_unlock_irqrestore(&ni->size_lock, flags);
498         if (unlikely(attr_len > i_size)) {
499                 /* Race with shrinking truncate. */
500                 attr_len = i_size;
501         }
502         addr = kmap_atomic(page, KM_USER0);
503         /* Copy the data to the page. */
504         memcpy(addr, (u8*)ctx->attr +
505                         le16_to_cpu(ctx->attr->data.resident.value_offset),
506                         attr_len);
507         /* Zero the remainder of the page. */
508         memset(addr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
509         flush_dcache_page(page);
510         kunmap_atomic(addr, KM_USER0);
511 put_unm_err_out:
512         ntfs_attr_put_search_ctx(ctx);
513 unm_err_out:
514         unmap_mft_record(base_ni);
515 done:
516         SetPageUptodate(page);
517 err_out:
518         unlock_page(page);
519         return err;
520 }
521
522 #ifdef NTFS_RW
523
524 /**
525  * ntfs_write_block - write a @page to the backing store
526  * @page:       page cache page to write out
527  * @wbc:        writeback control structure
528  *
529  * This function is for writing pages belonging to non-resident, non-mst
530  * protected attributes to their backing store.
531  *
532  * For a page with buffers, map and write the dirty buffers asynchronously
533  * under page writeback. For a page without buffers, create buffers for the
534  * page, then proceed as above.
535  *
536  * If a page doesn't have buffers the page dirty state is definitive. If a page
537  * does have buffers, the page dirty state is just a hint, and the buffer dirty
538  * state is definitive. (A hint which has rules: dirty buffers against a clean
539  * page is illegal. Other combinations are legal and need to be handled. In
540  * particular a dirty page containing clean buffers for example.)
541  *
542  * Return 0 on success and -errno on error.
543  *
544  * Based on ntfs_read_block() and __block_write_full_page().
545  */
546 static int ntfs_write_block(struct page *page, struct writeback_control *wbc)
547 {
548         VCN vcn;
549         LCN lcn;
550         s64 initialized_size;
551         loff_t i_size;
552         sector_t block, dblock, iblock;
553         struct inode *vi;
554         ntfs_inode *ni;
555         ntfs_volume *vol;
556         runlist_element *rl;
557         struct buffer_head *bh, *head;
558         unsigned long flags;
559         unsigned int blocksize, vcn_ofs;
560         int err;
561         bool need_end_writeback;
562         unsigned char blocksize_bits;
563
564         vi = page->mapping->host;
565         ni = NTFS_I(vi);
566         vol = ni->vol;
567
568         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
569                         "0x%lx.", ni->mft_no, ni->type, page->index);
570
571         BUG_ON(!NInoNonResident(ni));
572         BUG_ON(NInoMstProtected(ni));
573         blocksize = vol->sb->s_blocksize;
574         blocksize_bits = vol->sb->s_blocksize_bits;
575         if (!page_has_buffers(page)) {
576                 BUG_ON(!PageUptodate(page));
577                 create_empty_buffers(page, blocksize,
578                                 (1 << BH_Uptodate) | (1 << BH_Dirty));
579                 if (unlikely(!page_has_buffers(page))) {
580                         ntfs_warning(vol->sb, "Error allocating page "
581                                         "buffers.  Redirtying page so we try "
582                                         "again later.");
583                         /*
584                          * Put the page back on mapping->dirty_pages, but leave
585                          * its buffers' dirty state as-is.
586                          */
587                         redirty_page_for_writepage(wbc, page);
588                         unlock_page(page);
589                         return 0;
590                 }
591         }
592         bh = head = page_buffers(page);
593         BUG_ON(!bh);
594
595         /* NOTE: Different naming scheme to ntfs_read_block()! */
596
597         /* The first block in the page. */
598         block = (s64)page->index << (PAGE_CACHE_SHIFT - blocksize_bits);
599
600         read_lock_irqsave(&ni->size_lock, flags);
601         i_size = i_size_read(vi);
602         initialized_size = ni->initialized_size;
603         read_unlock_irqrestore(&ni->size_lock, flags);
604
605         /* The first out of bounds block for the data size. */
606         dblock = (i_size + blocksize - 1) >> blocksize_bits;
607
608         /* The last (fully or partially) initialized block. */
609         iblock = initialized_size >> blocksize_bits;
610
611         /*
612          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
613          * here, and the (potentially unmapped) buffers may become dirty at
614          * any time.  If a buffer becomes dirty here after we've inspected it
615          * then we just miss that fact, and the page stays dirty.
616          *
617          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
618          * handle that here by just cleaning them.
619          */
620
621         /*
622          * Loop through all the buffers in the page, mapping all the dirty
623          * buffers to disk addresses and handling any aliases from the
624          * underlying block device's mapping.
625          */
626         rl = NULL;
627         err = 0;
628         do {
629                 bool is_retry = false;
630
631                 if (unlikely(block >= dblock)) {
632                         /*
633                          * Mapped buffers outside i_size will occur, because
634                          * this page can be outside i_size when there is a
635                          * truncate in progress. The contents of such buffers
636                          * were zeroed by ntfs_writepage().
637                          *
638                          * FIXME: What about the small race window where
639                          * ntfs_writepage() has not done any clearing because
640                          * the page was within i_size but before we get here,
641                          * vmtruncate() modifies i_size?
642                          */
643                         clear_buffer_dirty(bh);
644                         set_buffer_uptodate(bh);
645                         continue;
646                 }
647
648                 /* Clean buffers are not written out, so no need to map them. */
649                 if (!buffer_dirty(bh))
650                         continue;
651
652                 /* Make sure we have enough initialized size. */
653                 if (unlikely((block >= iblock) &&
654                                 (initialized_size < i_size))) {
655                         /*
656                          * If this page is fully outside initialized size, zero
657                          * out all pages between the current initialized size
658                          * and the current page. Just use ntfs_readpage() to do
659                          * the zeroing transparently.
660                          */
661                         if (block > iblock) {
662                                 // TODO:
663                                 // For each page do:
664                                 // - read_cache_page()
665                                 // Again for each page do:
666                                 // - wait_on_page_locked()
667                                 // - Check (PageUptodate(page) &&
668                                 //                      !PageError(page))
669                                 // Update initialized size in the attribute and
670                                 // in the inode.
671                                 // Again, for each page do:
672                                 //      __set_page_dirty_buffers();
673                                 // page_cache_release()
674                                 // We don't need to wait on the writes.
675                                 // Update iblock.
676                         }
677                         /*
678                          * The current page straddles initialized size. Zero
679                          * all non-uptodate buffers and set them uptodate (and
680                          * dirty?). Note, there aren't any non-uptodate buffers
681                          * if the page is uptodate.
682                          * FIXME: For an uptodate page, the buffers may need to
683                          * be written out because they were not initialized on
684                          * disk before.
685                          */
686                         if (!PageUptodate(page)) {
687                                 // TODO:
688                                 // Zero any non-uptodate buffers up to i_size.
689                                 // Set them uptodate and dirty.
690                         }
691                         // TODO:
692                         // Update initialized size in the attribute and in the
693                         // inode (up to i_size).
694                         // Update iblock.
695                         // FIXME: This is inefficient. Try to batch the two
696                         // size changes to happen in one go.
697                         ntfs_error(vol->sb, "Writing beyond initialized size "
698                                         "is not supported yet. Sorry.");
699                         err = -EOPNOTSUPP;
700                         break;
701                         // Do NOT set_buffer_new() BUT DO clear buffer range
702                         // outside write request range.
703                         // set_buffer_uptodate() on complete buffers as well as
704                         // set_buffer_dirty().
705                 }
706
707                 /* No need to map buffers that are already mapped. */
708                 if (buffer_mapped(bh))
709                         continue;
710
711                 /* Unmapped, dirty buffer. Need to map it. */
712                 bh->b_bdev = vol->sb->s_bdev;
713
714                 /* Convert block into corresponding vcn and offset. */
715                 vcn = (VCN)block << blocksize_bits;
716                 vcn_ofs = vcn & vol->cluster_size_mask;
717                 vcn >>= vol->cluster_size_bits;
718                 if (!rl) {
719 lock_retry_remap:
720                         down_read(&ni->runlist.lock);
721                         rl = ni->runlist.rl;
722                 }
723                 if (likely(rl != NULL)) {
724                         /* Seek to element containing target vcn. */
725                         while (rl->length && rl[1].vcn <= vcn)
726                                 rl++;
727                         lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
728                 } else
729                         lcn = LCN_RL_NOT_MAPPED;
730                 /* Successful remap. */
731                 if (lcn >= 0) {
732                         /* Setup buffer head to point to correct block. */
733                         bh->b_blocknr = ((lcn << vol->cluster_size_bits) +
734                                         vcn_ofs) >> blocksize_bits;
735                         set_buffer_mapped(bh);
736                         continue;
737                 }
738                 /* It is a hole, need to instantiate it. */
739                 if (lcn == LCN_HOLE) {
740                         u8 *kaddr;
741                         unsigned long *bpos, *bend;
742
743                         /* Check if the buffer is zero. */
744                         kaddr = kmap_atomic(page, KM_USER0);
745                         bpos = (unsigned long *)(kaddr + bh_offset(bh));
746                         bend = (unsigned long *)((u8*)bpos + blocksize);
747                         do {
748                                 if (unlikely(*bpos))
749                                         break;
750                         } while (likely(++bpos < bend));
751                         kunmap_atomic(kaddr, KM_USER0);
752                         if (bpos == bend) {
753                                 /*
754                                  * Buffer is zero and sparse, no need to write
755                                  * it.
756                                  */
757                                 bh->b_blocknr = -1;
758                                 clear_buffer_dirty(bh);
759                                 continue;
760                         }
761                         // TODO: Instantiate the hole.
762                         // clear_buffer_new(bh);
763                         // unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
764                         ntfs_error(vol->sb, "Writing into sparse regions is "
765                                         "not supported yet. Sorry.");
766                         err = -EOPNOTSUPP;
767                         break;
768                 }
769                 /* If first try and runlist unmapped, map and retry. */
770                 if (!is_retry && lcn == LCN_RL_NOT_MAPPED) {
771                         is_retry = true;
772                         /*
773                          * Attempt to map runlist, dropping lock for
774                          * the duration.
775                          */
776                         up_read(&ni->runlist.lock);
777                         err = ntfs_map_runlist(ni, vcn);
778                         if (likely(!err))
779                                 goto lock_retry_remap;
780                         rl = NULL;
781                 } else if (!rl)
782                         up_read(&ni->runlist.lock);
783                 /*
784                  * If buffer is outside the runlist, truncate has cut it out
785                  * of the runlist.  Just clean and clear the buffer and set it
786                  * uptodate so it can get discarded by the VM.
787                  */
788                 if (err == -ENOENT || lcn == LCN_ENOENT) {
789                         bh->b_blocknr = -1;
790                         clear_buffer_dirty(bh);
791                         zero_user_page(page, bh_offset(bh), blocksize,
792                                         KM_USER0);
793                         set_buffer_uptodate(bh);
794                         err = 0;
795                         continue;
796                 }
797                 /* Failed to map the buffer, even after retrying. */
798                 if (!err)
799                         err = -EIO;
800                 bh->b_blocknr = -1;
801                 ntfs_error(vol->sb, "Failed to write to inode 0x%lx, "
802                                 "attribute type 0x%x, vcn 0x%llx, offset 0x%x "
803                                 "because its location on disk could not be "
804                                 "determined%s (error code %i).", ni->mft_no,
805                                 ni->type, (unsigned long long)vcn,
806                                 vcn_ofs, is_retry ? " even after "
807                                 "retrying" : "", err);
808                 break;
809         } while (block++, (bh = bh->b_this_page) != head);
810
811         /* Release the lock if we took it. */
812         if (rl)
813                 up_read(&ni->runlist.lock);
814
815         /* For the error case, need to reset bh to the beginning. */
816         bh = head;
817
818         /* Just an optimization, so ->readpage() is not called later. */
819         if (unlikely(!PageUptodate(page))) {
820                 int uptodate = 1;
821                 do {
822                         if (!buffer_uptodate(bh)) {
823                                 uptodate = 0;
824                                 bh = head;
825                                 break;
826                         }
827                 } while ((bh = bh->b_this_page) != head);
828                 if (uptodate)
829                         SetPageUptodate(page);
830         }
831
832         /* Setup all mapped, dirty buffers for async write i/o. */
833         do {
834                 if (buffer_mapped(bh) && buffer_dirty(bh)) {
835                         lock_buffer(bh);
836                         if (test_clear_buffer_dirty(bh)) {
837                                 BUG_ON(!buffer_uptodate(bh));
838                                 mark_buffer_async_write(bh);
839                         } else
840                                 unlock_buffer(bh);
841                 } else if (unlikely(err)) {
842                         /*
843                          * For the error case. The buffer may have been set
844                          * dirty during attachment to a dirty page.
845                          */
846                         if (err != -ENOMEM)
847                                 clear_buffer_dirty(bh);
848                 }
849         } while ((bh = bh->b_this_page) != head);
850
851         if (unlikely(err)) {
852                 // TODO: Remove the -EOPNOTSUPP check later on...
853                 if (unlikely(err == -EOPNOTSUPP))
854                         err = 0;
855                 else if (err == -ENOMEM) {
856                         ntfs_warning(vol->sb, "Error allocating memory. "
857                                         "Redirtying page so we try again "
858                                         "later.");
859                         /*
860                          * Put the page back on mapping->dirty_pages, but
861                          * leave its buffer's dirty state as-is.
862                          */
863                         redirty_page_for_writepage(wbc, page);
864                         err = 0;
865                 } else
866                         SetPageError(page);
867         }
868
869         BUG_ON(PageWriteback(page));
870         set_page_writeback(page);       /* Keeps try_to_free_buffers() away. */
871
872         /* Submit the prepared buffers for i/o. */
873         need_end_writeback = true;
874         do {
875                 struct buffer_head *next = bh->b_this_page;
876                 if (buffer_async_write(bh)) {
877                         submit_bh(WRITE, bh);
878                         need_end_writeback = false;
879                 }
880                 bh = next;
881         } while (bh != head);
882         unlock_page(page);
883
884         /* If no i/o was started, need to end_page_writeback(). */
885         if (unlikely(need_end_writeback))
886                 end_page_writeback(page);
887
888         ntfs_debug("Done.");
889         return err;
890 }
891
892 /**
893  * ntfs_write_mst_block - write a @page to the backing store
894  * @page:       page cache page to write out
895  * @wbc:        writeback control structure
896  *
897  * This function is for writing pages belonging to non-resident, mst protected
898  * attributes to their backing store.  The only supported attributes are index
899  * allocation and $MFT/$DATA.  Both directory inodes and index inodes are
900  * supported for the index allocation case.
901  *
902  * The page must remain locked for the duration of the write because we apply
903  * the mst fixups, write, and then undo the fixups, so if we were to unlock the
904  * page before undoing the fixups, any other user of the page will see the
905  * page contents as corrupt.
906  *
907  * We clear the page uptodate flag for the duration of the function to ensure
908  * exclusion for the $MFT/$DATA case against someone mapping an mft record we
909  * are about to apply the mst fixups to.
910  *
911  * Return 0 on success and -errno on error.
912  *
913  * Based on ntfs_write_block(), ntfs_mft_writepage(), and
914  * write_mft_record_nolock().
915  */
916 static int ntfs_write_mst_block(struct page *page,
917                 struct writeback_control *wbc)
918 {
919         sector_t block, dblock, rec_block;
920         struct inode *vi = page->mapping->host;
921         ntfs_inode *ni = NTFS_I(vi);
922         ntfs_volume *vol = ni->vol;
923         u8 *kaddr;
924         unsigned int rec_size = ni->itype.index.block_size;
925         ntfs_inode *locked_nis[PAGE_CACHE_SIZE / rec_size];
926         struct buffer_head *bh, *head, *tbh, *rec_start_bh;
927         struct buffer_head *bhs[MAX_BUF_PER_PAGE];
928         runlist_element *rl;
929         int i, nr_locked_nis, nr_recs, nr_bhs, max_bhs, bhs_per_rec, err, err2;
930         unsigned bh_size, rec_size_bits;
931         bool sync, is_mft, page_is_dirty, rec_is_dirty;
932         unsigned char bh_size_bits;
933
934         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
935                         "0x%lx.", vi->i_ino, ni->type, page->index);
936         BUG_ON(!NInoNonResident(ni));
937         BUG_ON(!NInoMstProtected(ni));
938         is_mft = (S_ISREG(vi->i_mode) && !vi->i_ino);
939         /*
940          * NOTE: ntfs_write_mst_block() would be called for $MFTMirr if a page
941          * in its page cache were to be marked dirty.  However this should
942          * never happen with the current driver and considering we do not
943          * handle this case here we do want to BUG(), at least for now.
944          */
945         BUG_ON(!(is_mft || S_ISDIR(vi->i_mode) ||
946                         (NInoAttr(ni) && ni->type == AT_INDEX_ALLOCATION)));
947         bh_size = vol->sb->s_blocksize;
948         bh_size_bits = vol->sb->s_blocksize_bits;
949         max_bhs = PAGE_CACHE_SIZE / bh_size;
950         BUG_ON(!max_bhs);
951         BUG_ON(max_bhs > MAX_BUF_PER_PAGE);
952
953         /* Were we called for sync purposes? */
954         sync = (wbc->sync_mode == WB_SYNC_ALL);
955
956         /* Make sure we have mapped buffers. */
957         bh = head = page_buffers(page);
958         BUG_ON(!bh);
959
960         rec_size_bits = ni->itype.index.block_size_bits;
961         BUG_ON(!(PAGE_CACHE_SIZE >> rec_size_bits));
962         bhs_per_rec = rec_size >> bh_size_bits;
963         BUG_ON(!bhs_per_rec);
964
965         /* The first block in the page. */
966         rec_block = block = (sector_t)page->index <<
967                         (PAGE_CACHE_SHIFT - bh_size_bits);
968
969         /* The first out of bounds block for the data size. */
970         dblock = (i_size_read(vi) + bh_size - 1) >> bh_size_bits;
971
972         rl = NULL;
973         err = err2 = nr_bhs = nr_recs = nr_locked_nis = 0;
974         page_is_dirty = rec_is_dirty = false;
975         rec_start_bh = NULL;
976         do {
977                 bool is_retry = false;
978
979                 if (likely(block < rec_block)) {
980                         if (unlikely(block >= dblock)) {
981                                 clear_buffer_dirty(bh);
982                                 set_buffer_uptodate(bh);
983                                 continue;
984                         }
985                         /*
986                          * This block is not the first one in the record.  We
987                          * ignore the buffer's dirty state because we could
988                          * have raced with a parallel mark_ntfs_record_dirty().
989                          */
990                         if (!rec_is_dirty)
991                                 continue;
992                         if (unlikely(err2)) {
993                                 if (err2 != -ENOMEM)
994                                         clear_buffer_dirty(bh);
995                                 continue;
996                         }
997                 } else /* if (block == rec_block) */ {
998                         BUG_ON(block > rec_block);
999                         /* This block is the first one in the record. */
1000                         rec_block += bhs_per_rec;
1001                         err2 = 0;
1002                         if (unlikely(block >= dblock)) {
1003                                 clear_buffer_dirty(bh);
1004                                 continue;
1005                         }
1006                         if (!buffer_dirty(bh)) {
1007                                 /* Clean records are not written out. */
1008                                 rec_is_dirty = false;
1009                                 continue;
1010                         }
1011                         rec_is_dirty = true;
1012                         rec_start_bh = bh;
1013                 }
1014                 /* Need to map the buffer if it is not mapped already. */
1015                 if (unlikely(!buffer_mapped(bh))) {
1016                         VCN vcn;
1017                         LCN lcn;
1018                         unsigned int vcn_ofs;
1019
1020                         bh->b_bdev = vol->sb->s_bdev;
1021                         /* Obtain the vcn and offset of the current block. */
1022                         vcn = (VCN)block << bh_size_bits;
1023                         vcn_ofs = vcn & vol->cluster_size_mask;
1024                         vcn >>= vol->cluster_size_bits;
1025                         if (!rl) {
1026 lock_retry_remap:
1027                                 down_read(&ni->runlist.lock);
1028                                 rl = ni->runlist.rl;
1029                         }
1030                         if (likely(rl != NULL)) {
1031                                 /* Seek to element containing target vcn. */
1032                                 while (rl->length && rl[1].vcn <= vcn)
1033                                         rl++;
1034                                 lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
1035                         } else
1036                                 lcn = LCN_RL_NOT_MAPPED;
1037                         /* Successful remap. */
1038                         if (likely(lcn >= 0)) {
1039                                 /* Setup buffer head to correct block. */
1040                                 bh->b_blocknr = ((lcn <<
1041                                                 vol->cluster_size_bits) +
1042                                                 vcn_ofs) >> bh_size_bits;
1043                                 set_buffer_mapped(bh);
1044                         } else {
1045                                 /*
1046                                  * Remap failed.  Retry to map the runlist once
1047                                  * unless we are working on $MFT which always
1048                                  * has the whole of its runlist in memory.
1049                                  */
1050                                 if (!is_mft && !is_retry &&
1051                                                 lcn == LCN_RL_NOT_MAPPED) {
1052                                         is_retry = true;
1053                                         /*
1054                                          * Attempt to map runlist, dropping
1055                                          * lock for the duration.
1056                                          */
1057                                         up_read(&ni->runlist.lock);
1058                                         err2 = ntfs_map_runlist(ni, vcn);
1059                                         if (likely(!err2))
1060                                                 goto lock_retry_remap;
1061                                         if (err2 == -ENOMEM)
1062                                                 page_is_dirty = true;
1063                                         lcn = err2;
1064                                 } else {
1065                                         err2 = -EIO;
1066                                         if (!rl)
1067                                                 up_read(&ni->runlist.lock);
1068                                 }
1069                                 /* Hard error.  Abort writing this record. */
1070                                 if (!err || err == -ENOMEM)
1071                                         err = err2;
1072                                 bh->b_blocknr = -1;
1073                                 ntfs_error(vol->sb, "Cannot write ntfs record "
1074                                                 "0x%llx (inode 0x%lx, "
1075                                                 "attribute type 0x%x) because "
1076                                                 "its location on disk could "
1077                                                 "not be determined (error "
1078                                                 "code %lli).",
1079                                                 (long long)block <<
1080                                                 bh_size_bits >>
1081                                                 vol->mft_record_size_bits,
1082                                                 ni->mft_no, ni->type,
1083                                                 (long long)lcn);
1084                                 /*
1085                                  * If this is not the first buffer, remove the
1086                                  * buffers in this record from the list of
1087                                  * buffers to write and clear their dirty bit
1088                                  * if not error -ENOMEM.
1089                                  */
1090                                 if (rec_start_bh != bh) {
1091                                         while (bhs[--nr_bhs] != rec_start_bh)
1092                                                 ;
1093                                         if (err2 != -ENOMEM) {
1094                                                 do {
1095                                                         clear_buffer_dirty(
1096                                                                 rec_start_bh);
1097                                                 } while ((rec_start_bh =
1098                                                                 rec_start_bh->
1099                                                                 b_this_page) !=
1100                                                                 bh);
1101                                         }
1102                                 }
1103                                 continue;
1104                         }
1105                 }
1106                 BUG_ON(!buffer_uptodate(bh));
1107                 BUG_ON(nr_bhs >= max_bhs);
1108                 bhs[nr_bhs++] = bh;
1109         } while (block++, (bh = bh->b_this_page) != head);
1110         if (unlikely(rl))
1111                 up_read(&ni->runlist.lock);
1112         /* If there were no dirty buffers, we are done. */
1113         if (!nr_bhs)
1114                 goto done;
1115         /* Map the page so we can access its contents. */
1116         kaddr = kmap(page);
1117         /* Clear the page uptodate flag whilst the mst fixups are applied. */
1118         BUG_ON(!PageUptodate(page));
1119         ClearPageUptodate(page);
1120         for (i = 0; i < nr_bhs; i++) {
1121                 unsigned int ofs;
1122
1123                 /* Skip buffers which are not at the beginning of records. */
1124                 if (i % bhs_per_rec)
1125                         continue;
1126                 tbh = bhs[i];
1127                 ofs = bh_offset(tbh);
1128                 if (is_mft) {
1129                         ntfs_inode *tni;
1130                         unsigned long mft_no;
1131
1132                         /* Get the mft record number. */
1133                         mft_no = (((s64)page->index << PAGE_CACHE_SHIFT) + ofs)
1134                                         >> rec_size_bits;
1135                         /* Check whether to write this mft record. */
1136                         tni = NULL;
1137                         if (!ntfs_may_write_mft_record(vol, mft_no,
1138                                         (MFT_RECORD*)(kaddr + ofs), &tni)) {
1139                                 /*
1140                                  * The record should not be written.  This
1141                                  * means we need to redirty the page before
1142                                  * returning.
1143                                  */
1144                                 page_is_dirty = true;
1145                                 /*
1146                                  * Remove the buffers in this mft record from
1147                                  * the list of buffers to write.
1148                                  */
1149                                 do {
1150                                         bhs[i] = NULL;
1151                                 } while (++i % bhs_per_rec);
1152                                 continue;
1153                         }
1154                         /*
1155                          * The record should be written.  If a locked ntfs
1156                          * inode was returned, add it to the array of locked
1157                          * ntfs inodes.
1158                          */
1159                         if (tni)
1160                                 locked_nis[nr_locked_nis++] = tni;
1161                 }
1162                 /* Apply the mst protection fixups. */
1163                 err2 = pre_write_mst_fixup((NTFS_RECORD*)(kaddr + ofs),
1164                                 rec_size);
1165                 if (unlikely(err2)) {
1166                         if (!err || err == -ENOMEM)
1167                                 err = -EIO;
1168                         ntfs_error(vol->sb, "Failed to apply mst fixups "
1169                                         "(inode 0x%lx, attribute type 0x%x, "
1170                                         "page index 0x%lx, page offset 0x%x)!"
1171                                         "  Unmount and run chkdsk.", vi->i_ino,
1172                                         ni->type, page->index, ofs);
1173                         /*
1174                          * Mark all the buffers in this record clean as we do
1175                          * not want to write corrupt data to disk.
1176                          */
1177                         do {
1178                                 clear_buffer_dirty(bhs[i]);
1179                                 bhs[i] = NULL;
1180                         } while (++i % bhs_per_rec);
1181                         continue;
1182                 }
1183                 nr_recs++;
1184         }
1185         /* If no records are to be written out, we are done. */
1186         if (!nr_recs)
1187                 goto unm_done;
1188         flush_dcache_page(page);
1189         /* Lock buffers and start synchronous write i/o on them. */
1190         for (i = 0; i < nr_bhs; i++) {
1191                 tbh = bhs[i];
1192                 if (!tbh)
1193                         continue;
1194                 if (unlikely(test_set_buffer_locked(tbh)))
1195                         BUG();
1196                 /* The buffer dirty state is now irrelevant, just clean it. */
1197                 clear_buffer_dirty(tbh);
1198                 BUG_ON(!buffer_uptodate(tbh));
1199                 BUG_ON(!buffer_mapped(tbh));
1200                 get_bh(tbh);
1201                 tbh->b_end_io = end_buffer_write_sync;
1202                 submit_bh(WRITE, tbh);
1203         }
1204         /* Synchronize the mft mirror now if not @sync. */
1205         if (is_mft && !sync)
1206                 goto do_mirror;
1207 do_wait:
1208         /* Wait on i/o completion of buffers. */
1209         for (i = 0; i < nr_bhs; i++) {
1210                 tbh = bhs[i];
1211                 if (!tbh)
1212                         continue;
1213                 wait_on_buffer(tbh);
1214                 if (unlikely(!buffer_uptodate(tbh))) {
1215                         ntfs_error(vol->sb, "I/O error while writing ntfs "
1216                                         "record buffer (inode 0x%lx, "
1217                                         "attribute type 0x%x, page index "
1218                                         "0x%lx, page offset 0x%lx)!  Unmount "
1219                                         "and run chkdsk.", vi->i_ino, ni->type,
1220                                         page->index, bh_offset(tbh));
1221                         if (!err || err == -ENOMEM)
1222                                 err = -EIO;
1223                         /*
1224                          * Set the buffer uptodate so the page and buffer
1225                          * states do not become out of sync.
1226                          */
1227                         set_buffer_uptodate(tbh);
1228                 }
1229         }
1230         /* If @sync, now synchronize the mft mirror. */
1231         if (is_mft && sync) {
1232 do_mirror:
1233                 for (i = 0; i < nr_bhs; i++) {
1234                         unsigned long mft_no;
1235                         unsigned int ofs;
1236
1237                         /*
1238                          * Skip buffers which are not at the beginning of
1239                          * records.
1240                          */
1241                         if (i % bhs_per_rec)
1242                                 continue;
1243                         tbh = bhs[i];
1244                         /* Skip removed buffers (and hence records). */
1245                         if (!tbh)
1246                                 continue;
1247                         ofs = bh_offset(tbh);
1248                         /* Get the mft record number. */
1249                         mft_no = (((s64)page->index << PAGE_CACHE_SHIFT) + ofs)
1250                                         >> rec_size_bits;
1251                         if (mft_no < vol->mftmirr_size)
1252                                 ntfs_sync_mft_mirror(vol, mft_no,
1253                                                 (MFT_RECORD*)(kaddr + ofs),
1254                                                 sync);
1255                 }
1256                 if (!sync)
1257                         goto do_wait;
1258         }
1259         /* Remove the mst protection fixups again. */
1260         for (i = 0; i < nr_bhs; i++) {
1261                 if (!(i % bhs_per_rec)) {
1262                         tbh = bhs[i];
1263                         if (!tbh)
1264                                 continue;
1265                         post_write_mst_fixup((NTFS_RECORD*)(kaddr +
1266                                         bh_offset(tbh)));
1267                 }
1268         }
1269         flush_dcache_page(page);
1270 unm_done:
1271         /* Unlock any locked inodes. */
1272         while (nr_locked_nis-- > 0) {
1273                 ntfs_inode *tni, *base_tni;
1274                 
1275                 tni = locked_nis[nr_locked_nis];
1276                 /* Get the base inode. */
1277                 mutex_lock(&tni->extent_lock);
1278                 if (tni->nr_extents >= 0)
1279                         base_tni = tni;
1280                 else {
1281                         base_tni = tni->ext.base_ntfs_ino;
1282                         BUG_ON(!base_tni);
1283                 }
1284                 mutex_unlock(&tni->extent_lock);
1285                 ntfs_debug("Unlocking %s inode 0x%lx.",
1286                                 tni == base_tni ? "base" : "extent",
1287                                 tni->mft_no);
1288                 mutex_unlock(&tni->mrec_lock);
1289                 atomic_dec(&tni->count);
1290                 iput(VFS_I(base_tni));
1291         }
1292         SetPageUptodate(page);
1293         kunmap(page);
1294 done:
1295         if (unlikely(err && err != -ENOMEM)) {
1296                 /*
1297                  * Set page error if there is only one ntfs record in the page.
1298                  * Otherwise we would loose per-record granularity.
1299                  */
1300                 if (ni->itype.index.block_size == PAGE_CACHE_SIZE)
1301                         SetPageError(page);
1302                 NVolSetErrors(vol);
1303         }
1304         if (page_is_dirty) {
1305                 ntfs_debug("Page still contains one or more dirty ntfs "
1306                                 "records.  Redirtying the page starting at "
1307                                 "record 0x%lx.", page->index <<
1308                                 (PAGE_CACHE_SHIFT - rec_size_bits));
1309                 redirty_page_for_writepage(wbc, page);
1310                 unlock_page(page);
1311         } else {
1312                 /*
1313                  * Keep the VM happy.  This must be done otherwise the
1314                  * radix-tree tag PAGECACHE_TAG_DIRTY remains set even though
1315                  * the page is clean.
1316                  */
1317                 BUG_ON(PageWriteback(page));
1318                 set_page_writeback(page);
1319                 unlock_page(page);
1320                 end_page_writeback(page);
1321         }
1322         if (likely(!err))
1323                 ntfs_debug("Done.");
1324         return err;
1325 }
1326
1327 /**
1328  * ntfs_writepage - write a @page to the backing store
1329  * @page:       page cache page to write out
1330  * @wbc:        writeback control structure
1331  *
1332  * This is called from the VM when it wants to have a dirty ntfs page cache
1333  * page cleaned.  The VM has already locked the page and marked it clean.
1334  *
1335  * For non-resident attributes, ntfs_writepage() writes the @page by calling
1336  * the ntfs version of the generic block_write_full_page() function,
1337  * ntfs_write_block(), which in turn if necessary creates and writes the
1338  * buffers associated with the page asynchronously.
1339  *
1340  * For resident attributes, OTOH, ntfs_writepage() writes the @page by copying
1341  * the data to the mft record (which at this stage is most likely in memory).
1342  * The mft record is then marked dirty and written out asynchronously via the
1343  * vfs inode dirty code path for the inode the mft record belongs to or via the
1344  * vm page dirty code path for the page the mft record is in.
1345  *
1346  * Based on ntfs_readpage() and fs/buffer.c::block_write_full_page().
1347  *
1348  * Return 0 on success and -errno on error.
1349  */
1350 static int ntfs_writepage(struct page *page, struct writeback_control *wbc)
1351 {
1352         loff_t i_size;
1353         struct inode *vi = page->mapping->host;
1354         ntfs_inode *base_ni = NULL, *ni = NTFS_I(vi);
1355         char *addr;
1356         ntfs_attr_search_ctx *ctx = NULL;
1357         MFT_RECORD *m = NULL;
1358         u32 attr_len;
1359         int err;
1360
1361 retry_writepage:
1362         BUG_ON(!PageLocked(page));
1363         i_size = i_size_read(vi);
1364         /* Is the page fully outside i_size? (truncate in progress) */
1365         if (unlikely(page->index >= (i_size + PAGE_CACHE_SIZE - 1) >>
1366                         PAGE_CACHE_SHIFT)) {
1367                 /*
1368                  * The page may have dirty, unmapped buffers.  Make them
1369                  * freeable here, so the page does not leak.
1370                  */
1371                 block_invalidatepage(page, 0);
1372                 unlock_page(page);
1373                 ntfs_debug("Write outside i_size - truncated?");
1374                 return 0;
1375         }
1376         /*
1377          * Only $DATA attributes can be encrypted and only unnamed $DATA
1378          * attributes can be compressed.  Index root can have the flags set but
1379          * this means to create compressed/encrypted files, not that the
1380          * attribute is compressed/encrypted.  Note we need to check for
1381          * AT_INDEX_ALLOCATION since this is the type of both directory and
1382          * index inodes.
1383          */
1384         if (ni->type != AT_INDEX_ALLOCATION) {
1385                 /* If file is encrypted, deny access, just like NT4. */
1386                 if (NInoEncrypted(ni)) {
1387                         unlock_page(page);
1388                         BUG_ON(ni->type != AT_DATA);
1389                         ntfs_debug("Denying write access to encrypted file.");
1390                         return -EACCES;
1391                 }
1392                 /* Compressed data streams are handled in compress.c. */
1393                 if (NInoNonResident(ni) && NInoCompressed(ni)) {
1394                         BUG_ON(ni->type != AT_DATA);
1395                         BUG_ON(ni->name_len);
1396                         // TODO: Implement and replace this with
1397                         // return ntfs_write_compressed_block(page);
1398                         unlock_page(page);
1399                         ntfs_error(vi->i_sb, "Writing to compressed files is "
1400                                         "not supported yet.  Sorry.");
1401                         return -EOPNOTSUPP;
1402                 }
1403                 // TODO: Implement and remove this check.
1404                 if (NInoNonResident(ni) && NInoSparse(ni)) {
1405                         unlock_page(page);
1406                         ntfs_error(vi->i_sb, "Writing to sparse files is not "
1407                                         "supported yet.  Sorry.");
1408                         return -EOPNOTSUPP;
1409                 }
1410         }
1411         /* NInoNonResident() == NInoIndexAllocPresent() */
1412         if (NInoNonResident(ni)) {
1413                 /* We have to zero every time due to mmap-at-end-of-file. */
1414                 if (page->index >= (i_size >> PAGE_CACHE_SHIFT)) {
1415                         /* The page straddles i_size. */
1416                         unsigned int ofs = i_size & ~PAGE_CACHE_MASK;
1417                         zero_user_page(page, ofs, PAGE_CACHE_SIZE - ofs,
1418                                         KM_USER0);
1419                 }
1420                 /* Handle mst protected attributes. */
1421                 if (NInoMstProtected(ni))
1422                         return ntfs_write_mst_block(page, wbc);
1423                 /* Normal, non-resident data stream. */
1424                 return ntfs_write_block(page, wbc);
1425         }
1426         /*
1427          * Attribute is resident, implying it is not compressed, encrypted, or
1428          * mst protected.  This also means the attribute is smaller than an mft
1429          * record and hence smaller than a page, so can simply return error on
1430          * any pages with index above 0.  Note the attribute can actually be
1431          * marked compressed but if it is resident the actual data is not
1432          * compressed so we are ok to ignore the compressed flag here.
1433          */
1434         BUG_ON(page_has_buffers(page));
1435         BUG_ON(!PageUptodate(page));
1436         if (unlikely(page->index > 0)) {
1437                 ntfs_error(vi->i_sb, "BUG()! page->index (0x%lx) > 0.  "
1438                                 "Aborting write.", page->index);
1439                 BUG_ON(PageWriteback(page));
1440                 set_page_writeback(page);
1441                 unlock_page(page);
1442                 end_page_writeback(page);
1443                 return -EIO;
1444         }
1445         if (!NInoAttr(ni))
1446                 base_ni = ni;
1447         else
1448                 base_ni = ni->ext.base_ntfs_ino;
1449         /* Map, pin, and lock the mft record. */
1450         m = map_mft_record(base_ni);
1451         if (IS_ERR(m)) {
1452                 err = PTR_ERR(m);
1453                 m = NULL;
1454                 ctx = NULL;
1455                 goto err_out;
1456         }
1457         /*
1458          * If a parallel write made the attribute non-resident, drop the mft
1459          * record and retry the writepage.
1460          */
1461         if (unlikely(NInoNonResident(ni))) {
1462                 unmap_mft_record(base_ni);
1463                 goto retry_writepage;
1464         }
1465         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1466         if (unlikely(!ctx)) {
1467                 err = -ENOMEM;
1468                 goto err_out;
1469         }
1470         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1471                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1472         if (unlikely(err))
1473                 goto err_out;
1474         /*
1475          * Keep the VM happy.  This must be done otherwise the radix-tree tag
1476          * PAGECACHE_TAG_DIRTY remains set even though the page is clean.
1477          */
1478         BUG_ON(PageWriteback(page));
1479         set_page_writeback(page);
1480         unlock_page(page);
1481         attr_len = le32_to_cpu(ctx->attr->data.resident.value_length);
1482         i_size = i_size_read(vi);
1483         if (unlikely(attr_len > i_size)) {
1484                 /* Race with shrinking truncate or a failed truncate. */
1485                 attr_len = i_size;
1486                 /*
1487                  * If the truncate failed, fix it up now.  If a concurrent
1488                  * truncate, we do its job, so it does not have to do anything.
1489                  */
1490                 err = ntfs_resident_attr_value_resize(ctx->mrec, ctx->attr,
1491                                 attr_len);
1492                 /* Shrinking cannot fail. */
1493                 BUG_ON(err);
1494         }
1495         addr = kmap_atomic(page, KM_USER0);
1496         /* Copy the data from the page to the mft record. */
1497         memcpy((u8*)ctx->attr +
1498                         le16_to_cpu(ctx->attr->data.resident.value_offset),
1499                         addr, attr_len);
1500         /* Zero out of bounds area in the page cache page. */
1501         memset(addr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
1502         kunmap_atomic(addr, KM_USER0);
1503         flush_dcache_page(page);
1504         flush_dcache_mft_record_page(ctx->ntfs_ino);
1505         /* We are done with the page. */
1506         end_page_writeback(page);
1507         /* Finally, mark the mft record dirty, so it gets written back. */
1508         mark_mft_record_dirty(ctx->ntfs_ino);
1509         ntfs_attr_put_search_ctx(ctx);
1510         unmap_mft_record(base_ni);
1511         return 0;
1512 err_out:
1513         if (err == -ENOMEM) {
1514                 ntfs_warning(vi->i_sb, "Error allocating memory. Redirtying "
1515                                 "page so we try again later.");
1516                 /*
1517                  * Put the page back on mapping->dirty_pages, but leave its
1518                  * buffers' dirty state as-is.
1519                  */
1520                 redirty_page_for_writepage(wbc, page);
1521                 err = 0;
1522         } else {
1523                 ntfs_error(vi->i_sb, "Resident attribute write failed with "
1524                                 "error %i.", err);
1525                 SetPageError(page);
1526                 NVolSetErrors(ni->vol);
1527         }
1528         unlock_page(page);
1529         if (ctx)
1530                 ntfs_attr_put_search_ctx(ctx);
1531         if (m)
1532                 unmap_mft_record(base_ni);
1533         return err;
1534 }
1535
1536 #endif  /* NTFS_RW */
1537
1538 /**
1539  * ntfs_aops - general address space operations for inodes and attributes
1540  */
1541 const struct address_space_operations ntfs_aops = {
1542         .readpage       = ntfs_readpage,        /* Fill page with data. */
1543         .sync_page      = block_sync_page,      /* Currently, just unplugs the
1544                                                    disk request queue. */
1545 #ifdef NTFS_RW
1546         .writepage      = ntfs_writepage,       /* Write dirty page to disk. */
1547 #endif /* NTFS_RW */
1548         .migratepage    = buffer_migrate_page,  /* Move a page cache page from
1549                                                    one physical page to an
1550                                                    other. */
1551 };
1552
1553 /**
1554  * ntfs_mst_aops - general address space operations for mst protecteed inodes
1555  *                 and attributes
1556  */
1557 const struct address_space_operations ntfs_mst_aops = {
1558         .readpage       = ntfs_readpage,        /* Fill page with data. */
1559         .sync_page      = block_sync_page,      /* Currently, just unplugs the
1560                                                    disk request queue. */
1561 #ifdef NTFS_RW
1562         .writepage      = ntfs_writepage,       /* Write dirty page to disk. */
1563         .set_page_dirty = __set_page_dirty_nobuffers,   /* Set the page dirty
1564                                                    without touching the buffers
1565                                                    belonging to the page. */
1566 #endif /* NTFS_RW */
1567         .migratepage    = buffer_migrate_page,  /* Move a page cache page from
1568                                                    one physical page to an
1569                                                    other. */
1570 };
1571
1572 #ifdef NTFS_RW
1573
1574 /**
1575  * mark_ntfs_record_dirty - mark an ntfs record dirty
1576  * @page:       page containing the ntfs record to mark dirty
1577  * @ofs:        byte offset within @page at which the ntfs record begins
1578  *
1579  * Set the buffers and the page in which the ntfs record is located dirty.
1580  *
1581  * The latter also marks the vfs inode the ntfs record belongs to dirty
1582  * (I_DIRTY_PAGES only).
1583  *
1584  * If the page does not have buffers, we create them and set them uptodate.
1585  * The page may not be locked which is why we need to handle the buffers under
1586  * the mapping->private_lock.  Once the buffers are marked dirty we no longer
1587  * need the lock since try_to_free_buffers() does not free dirty buffers.
1588  */
1589 void mark_ntfs_record_dirty(struct page *page, const unsigned int ofs) {
1590         struct address_space *mapping = page->mapping;
1591         ntfs_inode *ni = NTFS_I(mapping->host);
1592         struct buffer_head *bh, *head, *buffers_to_free = NULL;
1593         unsigned int end, bh_size, bh_ofs;
1594
1595         BUG_ON(!PageUptodate(page));
1596         end = ofs + ni->itype.index.block_size;
1597         bh_size = VFS_I(ni)->i_sb->s_blocksize;
1598         spin_lock(&mapping->private_lock);
1599         if (unlikely(!page_has_buffers(page))) {
1600                 spin_unlock(&mapping->private_lock);
1601                 bh = head = alloc_page_buffers(page, bh_size, 1);
1602                 spin_lock(&mapping->private_lock);
1603                 if (likely(!page_has_buffers(page))) {
1604                         struct buffer_head *tail;
1605
1606                         do {
1607                                 set_buffer_uptodate(bh);
1608                                 tail = bh;
1609                                 bh = bh->b_this_page;
1610                         } while (bh);
1611                         tail->b_this_page = head;
1612                         attach_page_buffers(page, head);
1613                 } else
1614                         buffers_to_free = bh;
1615         }
1616         bh = head = page_buffers(page);
1617         BUG_ON(!bh);
1618         do {
1619                 bh_ofs = bh_offset(bh);
1620                 if (bh_ofs + bh_size <= ofs)
1621                         continue;
1622                 if (unlikely(bh_ofs >= end))
1623                         break;
1624                 set_buffer_dirty(bh);
1625         } while ((bh = bh->b_this_page) != head);
1626         spin_unlock(&mapping->private_lock);
1627         __set_page_dirty_nobuffers(page);
1628         if (unlikely(buffers_to_free)) {
1629                 do {
1630                         bh = buffers_to_free->b_this_page;
1631                         free_buffer_head(buffers_to_free);
1632                         buffers_to_free = bh;
1633                 } while (buffers_to_free);
1634         }
1635 }
1636
1637 #endif /* NTFS_RW */