NTFS: Remove all the make_bad_inode() calls. This should only be called
[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) 2001-2006 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                         u8 *kaddr;
90                         int ofs;
91
92                         ofs = 0;
93                         if (file_ofs < init_size)
94                                 ofs = init_size - file_ofs;
95                         kaddr = kmap_atomic(page, KM_BIO_SRC_IRQ);
96                         memset(kaddr + bh_offset(bh) + ofs, 0,
97                                         bh->b_size - ofs);
98                         kunmap_atomic(kaddr, KM_BIO_SRC_IRQ);
99                         flush_dcache_page(page);
100                 }
101         } else {
102                 clear_buffer_uptodate(bh);
103                 SetPageError(page);
104                 ntfs_error(ni->vol->sb, "Buffer I/O error, logical block "
105                                 "0x%llx.", (unsigned long long)bh->b_blocknr);
106         }
107         first = page_buffers(page);
108         local_irq_save(flags);
109         bit_spin_lock(BH_Uptodate_Lock, &first->b_state);
110         clear_buffer_async_read(bh);
111         unlock_buffer(bh);
112         tmp = bh;
113         do {
114                 if (!buffer_uptodate(tmp))
115                         page_uptodate = 0;
116                 if (buffer_async_read(tmp)) {
117                         if (likely(buffer_locked(tmp)))
118                                 goto still_busy;
119                         /* Async buffers must be locked. */
120                         BUG();
121                 }
122                 tmp = tmp->b_this_page;
123         } while (tmp != bh);
124         bit_spin_unlock(BH_Uptodate_Lock, &first->b_state);
125         local_irq_restore(flags);
126         /*
127          * If none of the buffers had errors then we can set the page uptodate,
128          * but we first have to perform the post read mst fixups, if the
129          * attribute is mst protected, i.e. if NInoMstProteced(ni) is true.
130          * Note we ignore fixup errors as those are detected when
131          * map_mft_record() is called which gives us per record granularity
132          * rather than per page granularity.
133          */
134         if (!NInoMstProtected(ni)) {
135                 if (likely(page_uptodate && !PageError(page)))
136                         SetPageUptodate(page);
137         } else {
138                 u8 *kaddr;
139                 unsigned int i, recs;
140                 u32 rec_size;
141
142                 rec_size = ni->itype.index.block_size;
143                 recs = PAGE_CACHE_SIZE / rec_size;
144                 /* Should have been verified before we got here... */
145                 BUG_ON(!recs);
146                 kaddr = kmap_atomic(page, KM_BIO_SRC_IRQ);
147                 for (i = 0; i < recs; i++)
148                         post_read_mst_fixup((NTFS_RECORD*)(kaddr +
149                                         i * rec_size), rec_size);
150                 kunmap_atomic(kaddr, KM_BIO_SRC_IRQ);
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                 u8 *kaddr;
245                 int err;
246
247                 if (unlikely(buffer_uptodate(bh)))
248                         continue;
249                 if (unlikely(buffer_mapped(bh))) {
250                         arr[nr++] = bh;
251                         continue;
252                 }
253                 err = 0;
254                 bh->b_bdev = vol->sb->s_bdev;
255                 /* Is the block within the allowed limits? */
256                 if (iblock < lblock) {
257                         BOOL is_retry = FALSE;
258
259                         /* Convert iblock into corresponding vcn and offset. */
260                         vcn = (VCN)iblock << blocksize_bits >>
261                                         vol->cluster_size_bits;
262                         vcn_ofs = ((VCN)iblock << blocksize_bits) &
263                                         vol->cluster_size_mask;
264                         if (!rl) {
265 lock_retry_remap:
266                                 down_read(&ni->runlist.lock);
267                                 rl = ni->runlist.rl;
268                         }
269                         if (likely(rl != NULL)) {
270                                 /* Seek to element containing target vcn. */
271                                 while (rl->length && rl[1].vcn <= vcn)
272                                         rl++;
273                                 lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
274                         } else
275                                 lcn = LCN_RL_NOT_MAPPED;
276                         /* Successful remap. */
277                         if (lcn >= 0) {
278                                 /* Setup buffer head to correct block. */
279                                 bh->b_blocknr = ((lcn << vol->cluster_size_bits)
280                                                 + vcn_ofs) >> blocksize_bits;
281                                 set_buffer_mapped(bh);
282                                 /* Only read initialized data blocks. */
283                                 if (iblock < zblock) {
284                                         arr[nr++] = bh;
285                                         continue;
286                                 }
287                                 /* Fully non-initialized data block, zero it. */
288                                 goto handle_zblock;
289                         }
290                         /* It is a hole, need to zero it. */
291                         if (lcn == LCN_HOLE)
292                                 goto handle_hole;
293                         /* If first try and runlist unmapped, map and retry. */
294                         if (!is_retry && lcn == LCN_RL_NOT_MAPPED) {
295                                 is_retry = TRUE;
296                                 /*
297                                  * Attempt to map runlist, dropping lock for
298                                  * the duration.
299                                  */
300                                 up_read(&ni->runlist.lock);
301                                 err = ntfs_map_runlist(ni, vcn);
302                                 if (likely(!err))
303                                         goto lock_retry_remap;
304                                 rl = NULL;
305                         } else if (!rl)
306                                 up_read(&ni->runlist.lock);
307                         /*
308                          * If buffer is outside the runlist, treat it as a
309                          * hole.  This can happen due to concurrent truncate
310                          * for example.
311                          */
312                         if (err == -ENOENT || lcn == LCN_ENOENT) {
313                                 err = 0;
314                                 goto handle_hole;
315                         }
316                         /* Hard error, zero out region. */
317                         if (!err)
318                                 err = -EIO;
319                         bh->b_blocknr = -1;
320                         SetPageError(page);
321                         ntfs_error(vol->sb, "Failed to read from inode 0x%lx, "
322                                         "attribute type 0x%x, vcn 0x%llx, "
323                                         "offset 0x%x because its location on "
324                                         "disk could not be determined%s "
325                                         "(error code %i).", ni->mft_no,
326                                         ni->type, (unsigned long long)vcn,
327                                         vcn_ofs, is_retry ? " even after "
328                                         "retrying" : "", err);
329                 }
330                 /*
331                  * Either iblock was outside lblock limits or
332                  * ntfs_rl_vcn_to_lcn() returned error.  Just zero that portion
333                  * of the page and set the buffer uptodate.
334                  */
335 handle_hole:
336                 bh->b_blocknr = -1UL;
337                 clear_buffer_mapped(bh);
338 handle_zblock:
339                 kaddr = kmap_atomic(page, KM_USER0);
340                 memset(kaddr + i * blocksize, 0, blocksize);
341                 kunmap_atomic(kaddr, KM_USER0);
342                 flush_dcache_page(page);
343                 if (likely(!err))
344                         set_buffer_uptodate(bh);
345         } while (i++, iblock++, (bh = bh->b_this_page) != head);
346
347         /* Release the lock if we took it. */
348         if (rl)
349                 up_read(&ni->runlist.lock);
350
351         /* Check we have at least one buffer ready for i/o. */
352         if (nr) {
353                 struct buffer_head *tbh;
354
355                 /* Lock the buffers. */
356                 for (i = 0; i < nr; i++) {
357                         tbh = arr[i];
358                         lock_buffer(tbh);
359                         tbh->b_end_io = ntfs_end_buffer_async_read;
360                         set_buffer_async_read(tbh);
361                 }
362                 /* Finally, start i/o on the buffers. */
363                 for (i = 0; i < nr; i++) {
364                         tbh = arr[i];
365                         if (likely(!buffer_uptodate(tbh)))
366                                 submit_bh(READ, tbh);
367                         else
368                                 ntfs_end_buffer_async_read(tbh, 1);
369                 }
370                 return 0;
371         }
372         /* No i/o was scheduled on any of the buffers. */
373         if (likely(!PageError(page)))
374                 SetPageUptodate(page);
375         else /* Signal synchronous i/o error. */
376                 nr = -EIO;
377         unlock_page(page);
378         return nr;
379 }
380
381 /**
382  * ntfs_readpage - fill a @page of a @file with data from the device
383  * @file:       open file to which the page @page belongs or NULL
384  * @page:       page cache page to fill with data
385  *
386  * For non-resident attributes, ntfs_readpage() fills the @page of the open
387  * file @file by calling the ntfs version of the generic block_read_full_page()
388  * function, ntfs_read_block(), which in turn creates and reads in the buffers
389  * associated with the page asynchronously.
390  *
391  * For resident attributes, OTOH, ntfs_readpage() fills @page by copying the
392  * data from the mft record (which at this stage is most likely in memory) and
393  * fills the remainder with zeroes. Thus, in this case, I/O is synchronous, as
394  * even if the mft record is not cached at this point in time, we need to wait
395  * for it to be read in before we can do the copy.
396  *
397  * Return 0 on success and -errno on error.
398  */
399 static int ntfs_readpage(struct file *file, struct page *page)
400 {
401         loff_t i_size;
402         struct inode *vi;
403         ntfs_inode *ni, *base_ni;
404         u8 *kaddr;
405         ntfs_attr_search_ctx *ctx;
406         MFT_RECORD *mrec;
407         unsigned long flags;
408         u32 attr_len;
409         int err = 0;
410
411 retry_readpage:
412         BUG_ON(!PageLocked(page));
413         /*
414          * This can potentially happen because we clear PageUptodate() during
415          * ntfs_writepage() of MstProtected() attributes.
416          */
417         if (PageUptodate(page)) {
418                 unlock_page(page);
419                 return 0;
420         }
421         vi = page->mapping->host;
422         ni = NTFS_I(vi);
423         /*
424          * Only $DATA attributes can be encrypted and only unnamed $DATA
425          * attributes can be compressed.  Index root can have the flags set but
426          * this means to create compressed/encrypted files, not that the
427          * attribute is compressed/encrypted.  Note we need to check for
428          * AT_INDEX_ALLOCATION since this is the type of both directory and
429          * index inodes.
430          */
431         if (ni->type != AT_INDEX_ALLOCATION) {
432                 /* If attribute is encrypted, deny access, just like NT4. */
433                 if (NInoEncrypted(ni)) {
434                         BUG_ON(ni->type != AT_DATA);
435                         err = -EACCES;
436                         goto err_out;
437                 }
438                 /* Compressed data streams are handled in compress.c. */
439                 if (NInoNonResident(ni) && NInoCompressed(ni)) {
440                         BUG_ON(ni->type != AT_DATA);
441                         BUG_ON(ni->name_len);
442                         return ntfs_read_compressed_block(page);
443                 }
444         }
445         /* NInoNonResident() == NInoIndexAllocPresent() */
446         if (NInoNonResident(ni)) {
447                 /* Normal, non-resident data stream. */
448                 return ntfs_read_block(page);
449         }
450         /*
451          * Attribute is resident, implying it is not compressed or encrypted.
452          * This also means the attribute is smaller than an mft record and
453          * hence smaller than a page, so can simply zero out any pages with
454          * index above 0.  Note the attribute can actually be marked compressed
455          * but if it is resident the actual data is not compressed so we are
456          * ok to ignore the compressed flag here.
457          */
458         if (unlikely(page->index > 0)) {
459                 kaddr = kmap_atomic(page, KM_USER0);
460                 memset(kaddr, 0, PAGE_CACHE_SIZE);
461                 flush_dcache_page(page);
462                 kunmap_atomic(kaddr, 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         kaddr = kmap_atomic(page, KM_USER0);
503         /* Copy the data to the page. */
504         memcpy(kaddr, (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(kaddr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
509         flush_dcache_page(page);
510         kunmap_atomic(kaddr, 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                         u8 *kaddr;
790
791                         bh->b_blocknr = -1;
792                         clear_buffer_dirty(bh);
793                         kaddr = kmap_atomic(page, KM_USER0);
794                         memset(kaddr + bh_offset(bh), 0, blocksize);
795                         kunmap_atomic(kaddr, KM_USER0);
796                         flush_dcache_page(page);
797                         set_buffer_uptodate(bh);
798                         err = 0;
799                         continue;
800                 }
801                 /* Failed to map the buffer, even after retrying. */
802                 if (!err)
803                         err = -EIO;
804                 bh->b_blocknr = -1;
805                 ntfs_error(vol->sb, "Failed to write to inode 0x%lx, "
806                                 "attribute type 0x%x, vcn 0x%llx, offset 0x%x "
807                                 "because its location on disk could not be "
808                                 "determined%s (error code %i).", ni->mft_no,
809                                 ni->type, (unsigned long long)vcn,
810                                 vcn_ofs, is_retry ? " even after "
811                                 "retrying" : "", err);
812                 break;
813         } while (block++, (bh = bh->b_this_page) != head);
814
815         /* Release the lock if we took it. */
816         if (rl)
817                 up_read(&ni->runlist.lock);
818
819         /* For the error case, need to reset bh to the beginning. */
820         bh = head;
821
822         /* Just an optimization, so ->readpage() is not called later. */
823         if (unlikely(!PageUptodate(page))) {
824                 int uptodate = 1;
825                 do {
826                         if (!buffer_uptodate(bh)) {
827                                 uptodate = 0;
828                                 bh = head;
829                                 break;
830                         }
831                 } while ((bh = bh->b_this_page) != head);
832                 if (uptodate)
833                         SetPageUptodate(page);
834         }
835
836         /* Setup all mapped, dirty buffers for async write i/o. */
837         do {
838                 if (buffer_mapped(bh) && buffer_dirty(bh)) {
839                         lock_buffer(bh);
840                         if (test_clear_buffer_dirty(bh)) {
841                                 BUG_ON(!buffer_uptodate(bh));
842                                 mark_buffer_async_write(bh);
843                         } else
844                                 unlock_buffer(bh);
845                 } else if (unlikely(err)) {
846                         /*
847                          * For the error case. The buffer may have been set
848                          * dirty during attachment to a dirty page.
849                          */
850                         if (err != -ENOMEM)
851                                 clear_buffer_dirty(bh);
852                 }
853         } while ((bh = bh->b_this_page) != head);
854
855         if (unlikely(err)) {
856                 // TODO: Remove the -EOPNOTSUPP check later on...
857                 if (unlikely(err == -EOPNOTSUPP))
858                         err = 0;
859                 else if (err == -ENOMEM) {
860                         ntfs_warning(vol->sb, "Error allocating memory. "
861                                         "Redirtying page so we try again "
862                                         "later.");
863                         /*
864                          * Put the page back on mapping->dirty_pages, but
865                          * leave its buffer's dirty state as-is.
866                          */
867                         redirty_page_for_writepage(wbc, page);
868                         err = 0;
869                 } else
870                         SetPageError(page);
871         }
872
873         BUG_ON(PageWriteback(page));
874         set_page_writeback(page);       /* Keeps try_to_free_buffers() away. */
875
876         /* Submit the prepared buffers for i/o. */
877         need_end_writeback = TRUE;
878         do {
879                 struct buffer_head *next = bh->b_this_page;
880                 if (buffer_async_write(bh)) {
881                         submit_bh(WRITE, bh);
882                         need_end_writeback = FALSE;
883                 }
884                 bh = next;
885         } while (bh != head);
886         unlock_page(page);
887
888         /* If no i/o was started, need to end_page_writeback(). */
889         if (unlikely(need_end_writeback))
890                 end_page_writeback(page);
891
892         ntfs_debug("Done.");
893         return err;
894 }
895
896 /**
897  * ntfs_write_mst_block - write a @page to the backing store
898  * @page:       page cache page to write out
899  * @wbc:        writeback control structure
900  *
901  * This function is for writing pages belonging to non-resident, mst protected
902  * attributes to their backing store.  The only supported attributes are index
903  * allocation and $MFT/$DATA.  Both directory inodes and index inodes are
904  * supported for the index allocation case.
905  *
906  * The page must remain locked for the duration of the write because we apply
907  * the mst fixups, write, and then undo the fixups, so if we were to unlock the
908  * page before undoing the fixups, any other user of the page will see the
909  * page contents as corrupt.
910  *
911  * We clear the page uptodate flag for the duration of the function to ensure
912  * exclusion for the $MFT/$DATA case against someone mapping an mft record we
913  * are about to apply the mst fixups to.
914  *
915  * Return 0 on success and -errno on error.
916  *
917  * Based on ntfs_write_block(), ntfs_mft_writepage(), and
918  * write_mft_record_nolock().
919  */
920 static int ntfs_write_mst_block(struct page *page,
921                 struct writeback_control *wbc)
922 {
923         sector_t block, dblock, rec_block;
924         struct inode *vi = page->mapping->host;
925         ntfs_inode *ni = NTFS_I(vi);
926         ntfs_volume *vol = ni->vol;
927         u8 *kaddr;
928         unsigned int rec_size = ni->itype.index.block_size;
929         ntfs_inode *locked_nis[PAGE_CACHE_SIZE / rec_size];
930         struct buffer_head *bh, *head, *tbh, *rec_start_bh;
931         struct buffer_head *bhs[MAX_BUF_PER_PAGE];
932         runlist_element *rl;
933         int i, nr_locked_nis, nr_recs, nr_bhs, max_bhs, bhs_per_rec, err, err2;
934         unsigned bh_size, rec_size_bits;
935         BOOL sync, is_mft, page_is_dirty, rec_is_dirty;
936         unsigned char bh_size_bits;
937
938         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
939                         "0x%lx.", vi->i_ino, ni->type, page->index);
940         BUG_ON(!NInoNonResident(ni));
941         BUG_ON(!NInoMstProtected(ni));
942         is_mft = (S_ISREG(vi->i_mode) && !vi->i_ino);
943         /*
944          * NOTE: ntfs_write_mst_block() would be called for $MFTMirr if a page
945          * in its page cache were to be marked dirty.  However this should
946          * never happen with the current driver and considering we do not
947          * handle this case here we do want to BUG(), at least for now.
948          */
949         BUG_ON(!(is_mft || S_ISDIR(vi->i_mode) ||
950                         (NInoAttr(ni) && ni->type == AT_INDEX_ALLOCATION)));
951         bh_size = vol->sb->s_blocksize;
952         bh_size_bits = vol->sb->s_blocksize_bits;
953         max_bhs = PAGE_CACHE_SIZE / bh_size;
954         BUG_ON(!max_bhs);
955         BUG_ON(max_bhs > MAX_BUF_PER_PAGE);
956
957         /* Were we called for sync purposes? */
958         sync = (wbc->sync_mode == WB_SYNC_ALL);
959
960         /* Make sure we have mapped buffers. */
961         bh = head = page_buffers(page);
962         BUG_ON(!bh);
963
964         rec_size_bits = ni->itype.index.block_size_bits;
965         BUG_ON(!(PAGE_CACHE_SIZE >> rec_size_bits));
966         bhs_per_rec = rec_size >> bh_size_bits;
967         BUG_ON(!bhs_per_rec);
968
969         /* The first block in the page. */
970         rec_block = block = (sector_t)page->index <<
971                         (PAGE_CACHE_SHIFT - bh_size_bits);
972
973         /* The first out of bounds block for the data size. */
974         dblock = (i_size_read(vi) + bh_size - 1) >> bh_size_bits;
975
976         rl = NULL;
977         err = err2 = nr_bhs = nr_recs = nr_locked_nis = 0;
978         page_is_dirty = rec_is_dirty = FALSE;
979         rec_start_bh = NULL;
980         do {
981                 BOOL is_retry = FALSE;
982
983                 if (likely(block < rec_block)) {
984                         if (unlikely(block >= dblock)) {
985                                 clear_buffer_dirty(bh);
986                                 set_buffer_uptodate(bh);
987                                 continue;
988                         }
989                         /*
990                          * This block is not the first one in the record.  We
991                          * ignore the buffer's dirty state because we could
992                          * have raced with a parallel mark_ntfs_record_dirty().
993                          */
994                         if (!rec_is_dirty)
995                                 continue;
996                         if (unlikely(err2)) {
997                                 if (err2 != -ENOMEM)
998                                         clear_buffer_dirty(bh);
999                                 continue;
1000                         }
1001                 } else /* if (block == rec_block) */ {
1002                         BUG_ON(block > rec_block);
1003                         /* This block is the first one in the record. */
1004                         rec_block += bhs_per_rec;
1005                         err2 = 0;
1006                         if (unlikely(block >= dblock)) {
1007                                 clear_buffer_dirty(bh);
1008                                 continue;
1009                         }
1010                         if (!buffer_dirty(bh)) {
1011                                 /* Clean records are not written out. */
1012                                 rec_is_dirty = FALSE;
1013                                 continue;
1014                         }
1015                         rec_is_dirty = TRUE;
1016                         rec_start_bh = bh;
1017                 }
1018                 /* Need to map the buffer if it is not mapped already. */
1019                 if (unlikely(!buffer_mapped(bh))) {
1020                         VCN vcn;
1021                         LCN lcn;
1022                         unsigned int vcn_ofs;
1023
1024                         bh->b_bdev = vol->sb->s_bdev;
1025                         /* Obtain the vcn and offset of the current block. */
1026                         vcn = (VCN)block << bh_size_bits;
1027                         vcn_ofs = vcn & vol->cluster_size_mask;
1028                         vcn >>= vol->cluster_size_bits;
1029                         if (!rl) {
1030 lock_retry_remap:
1031                                 down_read(&ni->runlist.lock);
1032                                 rl = ni->runlist.rl;
1033                         }
1034                         if (likely(rl != NULL)) {
1035                                 /* Seek to element containing target vcn. */
1036                                 while (rl->length && rl[1].vcn <= vcn)
1037                                         rl++;
1038                                 lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
1039                         } else
1040                                 lcn = LCN_RL_NOT_MAPPED;
1041                         /* Successful remap. */
1042                         if (likely(lcn >= 0)) {
1043                                 /* Setup buffer head to correct block. */
1044                                 bh->b_blocknr = ((lcn <<
1045                                                 vol->cluster_size_bits) +
1046                                                 vcn_ofs) >> bh_size_bits;
1047                                 set_buffer_mapped(bh);
1048                         } else {
1049                                 /*
1050                                  * Remap failed.  Retry to map the runlist once
1051                                  * unless we are working on $MFT which always
1052                                  * has the whole of its runlist in memory.
1053                                  */
1054                                 if (!is_mft && !is_retry &&
1055                                                 lcn == LCN_RL_NOT_MAPPED) {
1056                                         is_retry = TRUE;
1057                                         /*
1058                                          * Attempt to map runlist, dropping
1059                                          * lock for the duration.
1060                                          */
1061                                         up_read(&ni->runlist.lock);
1062                                         err2 = ntfs_map_runlist(ni, vcn);
1063                                         if (likely(!err2))
1064                                                 goto lock_retry_remap;
1065                                         if (err2 == -ENOMEM)
1066                                                 page_is_dirty = TRUE;
1067                                         lcn = err2;
1068                                 } else {
1069                                         err2 = -EIO;
1070                                         if (!rl)
1071                                                 up_read(&ni->runlist.lock);
1072                                 }
1073                                 /* Hard error.  Abort writing this record. */
1074                                 if (!err || err == -ENOMEM)
1075                                         err = err2;
1076                                 bh->b_blocknr = -1;
1077                                 ntfs_error(vol->sb, "Cannot write ntfs record "
1078                                                 "0x%llx (inode 0x%lx, "
1079                                                 "attribute type 0x%x) because "
1080                                                 "its location on disk could "
1081                                                 "not be determined (error "
1082                                                 "code %lli).",
1083                                                 (long long)block <<
1084                                                 bh_size_bits >>
1085                                                 vol->mft_record_size_bits,
1086                                                 ni->mft_no, ni->type,
1087                                                 (long long)lcn);
1088                                 /*
1089                                  * If this is not the first buffer, remove the
1090                                  * buffers in this record from the list of
1091                                  * buffers to write and clear their dirty bit
1092                                  * if not error -ENOMEM.
1093                                  */
1094                                 if (rec_start_bh != bh) {
1095                                         while (bhs[--nr_bhs] != rec_start_bh)
1096                                                 ;
1097                                         if (err2 != -ENOMEM) {
1098                                                 do {
1099                                                         clear_buffer_dirty(
1100                                                                 rec_start_bh);
1101                                                 } while ((rec_start_bh =
1102                                                                 rec_start_bh->
1103                                                                 b_this_page) !=
1104                                                                 bh);
1105                                         }
1106                                 }
1107                                 continue;
1108                         }
1109                 }
1110                 BUG_ON(!buffer_uptodate(bh));
1111                 BUG_ON(nr_bhs >= max_bhs);
1112                 bhs[nr_bhs++] = bh;
1113         } while (block++, (bh = bh->b_this_page) != head);
1114         if (unlikely(rl))
1115                 up_read(&ni->runlist.lock);
1116         /* If there were no dirty buffers, we are done. */
1117         if (!nr_bhs)
1118                 goto done;
1119         /* Map the page so we can access its contents. */
1120         kaddr = kmap(page);
1121         /* Clear the page uptodate flag whilst the mst fixups are applied. */
1122         BUG_ON(!PageUptodate(page));
1123         ClearPageUptodate(page);
1124         for (i = 0; i < nr_bhs; i++) {
1125                 unsigned int ofs;
1126
1127                 /* Skip buffers which are not at the beginning of records. */
1128                 if (i % bhs_per_rec)
1129                         continue;
1130                 tbh = bhs[i];
1131                 ofs = bh_offset(tbh);
1132                 if (is_mft) {
1133                         ntfs_inode *tni;
1134                         unsigned long mft_no;
1135
1136                         /* Get the mft record number. */
1137                         mft_no = (((s64)page->index << PAGE_CACHE_SHIFT) + ofs)
1138                                         >> rec_size_bits;
1139                         /* Check whether to write this mft record. */
1140                         tni = NULL;
1141                         if (!ntfs_may_write_mft_record(vol, mft_no,
1142                                         (MFT_RECORD*)(kaddr + ofs), &tni)) {
1143                                 /*
1144                                  * The record should not be written.  This
1145                                  * means we need to redirty the page before
1146                                  * returning.
1147                                  */
1148                                 page_is_dirty = TRUE;
1149                                 /*
1150                                  * Remove the buffers in this mft record from
1151                                  * the list of buffers to write.
1152                                  */
1153                                 do {
1154                                         bhs[i] = NULL;
1155                                 } while (++i % bhs_per_rec);
1156                                 continue;
1157                         }
1158                         /*
1159                          * The record should be written.  If a locked ntfs
1160                          * inode was returned, add it to the array of locked
1161                          * ntfs inodes.
1162                          */
1163                         if (tni)
1164                                 locked_nis[nr_locked_nis++] = tni;
1165                 }
1166                 /* Apply the mst protection fixups. */
1167                 err2 = pre_write_mst_fixup((NTFS_RECORD*)(kaddr + ofs),
1168                                 rec_size);
1169                 if (unlikely(err2)) {
1170                         if (!err || err == -ENOMEM)
1171                                 err = -EIO;
1172                         ntfs_error(vol->sb, "Failed to apply mst fixups "
1173                                         "(inode 0x%lx, attribute type 0x%x, "
1174                                         "page index 0x%lx, page offset 0x%x)!"
1175                                         "  Unmount and run chkdsk.", vi->i_ino,
1176                                         ni->type, page->index, ofs);
1177                         /*
1178                          * Mark all the buffers in this record clean as we do
1179                          * not want to write corrupt data to disk.
1180                          */
1181                         do {
1182                                 clear_buffer_dirty(bhs[i]);
1183                                 bhs[i] = NULL;
1184                         } while (++i % bhs_per_rec);
1185                         continue;
1186                 }
1187                 nr_recs++;
1188         }
1189         /* If no records are to be written out, we are done. */
1190         if (!nr_recs)
1191                 goto unm_done;
1192         flush_dcache_page(page);
1193         /* Lock buffers and start synchronous write i/o on them. */
1194         for (i = 0; i < nr_bhs; i++) {
1195                 tbh = bhs[i];
1196                 if (!tbh)
1197                         continue;
1198                 if (unlikely(test_set_buffer_locked(tbh)))
1199                         BUG();
1200                 /* The buffer dirty state is now irrelevant, just clean it. */
1201                 clear_buffer_dirty(tbh);
1202                 BUG_ON(!buffer_uptodate(tbh));
1203                 BUG_ON(!buffer_mapped(tbh));
1204                 get_bh(tbh);
1205                 tbh->b_end_io = end_buffer_write_sync;
1206                 submit_bh(WRITE, tbh);
1207         }
1208         /* Synchronize the mft mirror now if not @sync. */
1209         if (is_mft && !sync)
1210                 goto do_mirror;
1211 do_wait:
1212         /* Wait on i/o completion of buffers. */
1213         for (i = 0; i < nr_bhs; i++) {
1214                 tbh = bhs[i];
1215                 if (!tbh)
1216                         continue;
1217                 wait_on_buffer(tbh);
1218                 if (unlikely(!buffer_uptodate(tbh))) {
1219                         ntfs_error(vol->sb, "I/O error while writing ntfs "
1220                                         "record buffer (inode 0x%lx, "
1221                                         "attribute type 0x%x, page index "
1222                                         "0x%lx, page offset 0x%lx)!  Unmount "
1223                                         "and run chkdsk.", vi->i_ino, ni->type,
1224                                         page->index, bh_offset(tbh));
1225                         if (!err || err == -ENOMEM)
1226                                 err = -EIO;
1227                         /*
1228                          * Set the buffer uptodate so the page and buffer
1229                          * states do not become out of sync.
1230                          */
1231                         set_buffer_uptodate(tbh);
1232                 }
1233         }
1234         /* If @sync, now synchronize the mft mirror. */
1235         if (is_mft && sync) {
1236 do_mirror:
1237                 for (i = 0; i < nr_bhs; i++) {
1238                         unsigned long mft_no;
1239                         unsigned int ofs;
1240
1241                         /*
1242                          * Skip buffers which are not at the beginning of
1243                          * records.
1244                          */
1245                         if (i % bhs_per_rec)
1246                                 continue;
1247                         tbh = bhs[i];
1248                         /* Skip removed buffers (and hence records). */
1249                         if (!tbh)
1250                                 continue;
1251                         ofs = bh_offset(tbh);
1252                         /* Get the mft record number. */
1253                         mft_no = (((s64)page->index << PAGE_CACHE_SHIFT) + ofs)
1254                                         >> rec_size_bits;
1255                         if (mft_no < vol->mftmirr_size)
1256                                 ntfs_sync_mft_mirror(vol, mft_no,
1257                                                 (MFT_RECORD*)(kaddr + ofs),
1258                                                 sync);
1259                 }
1260                 if (!sync)
1261                         goto do_wait;
1262         }
1263         /* Remove the mst protection fixups again. */
1264         for (i = 0; i < nr_bhs; i++) {
1265                 if (!(i % bhs_per_rec)) {
1266                         tbh = bhs[i];
1267                         if (!tbh)
1268                                 continue;
1269                         post_write_mst_fixup((NTFS_RECORD*)(kaddr +
1270                                         bh_offset(tbh)));
1271                 }
1272         }
1273         flush_dcache_page(page);
1274 unm_done:
1275         /* Unlock any locked inodes. */
1276         while (nr_locked_nis-- > 0) {
1277                 ntfs_inode *tni, *base_tni;
1278                 
1279                 tni = locked_nis[nr_locked_nis];
1280                 /* Get the base inode. */
1281                 down(&tni->extent_lock);
1282                 if (tni->nr_extents >= 0)
1283                         base_tni = tni;
1284                 else {
1285                         base_tni = tni->ext.base_ntfs_ino;
1286                         BUG_ON(!base_tni);
1287                 }
1288                 up(&tni->extent_lock);
1289                 ntfs_debug("Unlocking %s inode 0x%lx.",
1290                                 tni == base_tni ? "base" : "extent",
1291                                 tni->mft_no);
1292                 up(&tni->mrec_lock);
1293                 atomic_dec(&tni->count);
1294                 iput(VFS_I(base_tni));
1295         }
1296         SetPageUptodate(page);
1297         kunmap(page);
1298 done:
1299         if (unlikely(err && err != -ENOMEM)) {
1300                 /*
1301                  * Set page error if there is only one ntfs record in the page.
1302                  * Otherwise we would loose per-record granularity.
1303                  */
1304                 if (ni->itype.index.block_size == PAGE_CACHE_SIZE)
1305                         SetPageError(page);
1306                 NVolSetErrors(vol);
1307         }
1308         if (page_is_dirty) {
1309                 ntfs_debug("Page still contains one or more dirty ntfs "
1310                                 "records.  Redirtying the page starting at "
1311                                 "record 0x%lx.", page->index <<
1312                                 (PAGE_CACHE_SHIFT - rec_size_bits));
1313                 redirty_page_for_writepage(wbc, page);
1314                 unlock_page(page);
1315         } else {
1316                 /*
1317                  * Keep the VM happy.  This must be done otherwise the
1318                  * radix-tree tag PAGECACHE_TAG_DIRTY remains set even though
1319                  * the page is clean.
1320                  */
1321                 BUG_ON(PageWriteback(page));
1322                 set_page_writeback(page);
1323                 unlock_page(page);
1324                 end_page_writeback(page);
1325         }
1326         if (likely(!err))
1327                 ntfs_debug("Done.");
1328         return err;
1329 }
1330
1331 /**
1332  * ntfs_writepage - write a @page to the backing store
1333  * @page:       page cache page to write out
1334  * @wbc:        writeback control structure
1335  *
1336  * This is called from the VM when it wants to have a dirty ntfs page cache
1337  * page cleaned.  The VM has already locked the page and marked it clean.
1338  *
1339  * For non-resident attributes, ntfs_writepage() writes the @page by calling
1340  * the ntfs version of the generic block_write_full_page() function,
1341  * ntfs_write_block(), which in turn if necessary creates and writes the
1342  * buffers associated with the page asynchronously.
1343  *
1344  * For resident attributes, OTOH, ntfs_writepage() writes the @page by copying
1345  * the data to the mft record (which at this stage is most likely in memory).
1346  * The mft record is then marked dirty and written out asynchronously via the
1347  * vfs inode dirty code path for the inode the mft record belongs to or via the
1348  * vm page dirty code path for the page the mft record is in.
1349  *
1350  * Based on ntfs_readpage() and fs/buffer.c::block_write_full_page().
1351  *
1352  * Return 0 on success and -errno on error.
1353  */
1354 static int ntfs_writepage(struct page *page, struct writeback_control *wbc)
1355 {
1356         loff_t i_size;
1357         struct inode *vi = page->mapping->host;
1358         ntfs_inode *base_ni = NULL, *ni = NTFS_I(vi);
1359         char *kaddr;
1360         ntfs_attr_search_ctx *ctx = NULL;
1361         MFT_RECORD *m = NULL;
1362         u32 attr_len;
1363         int err;
1364
1365 retry_writepage:
1366         BUG_ON(!PageLocked(page));
1367         i_size = i_size_read(vi);
1368         /* Is the page fully outside i_size? (truncate in progress) */
1369         if (unlikely(page->index >= (i_size + PAGE_CACHE_SIZE - 1) >>
1370                         PAGE_CACHE_SHIFT)) {
1371                 /*
1372                  * The page may have dirty, unmapped buffers.  Make them
1373                  * freeable here, so the page does not leak.
1374                  */
1375                 block_invalidatepage(page, 0);
1376                 unlock_page(page);
1377                 ntfs_debug("Write outside i_size - truncated?");
1378                 return 0;
1379         }
1380         /*
1381          * Only $DATA attributes can be encrypted and only unnamed $DATA
1382          * attributes can be compressed.  Index root can have the flags set but
1383          * this means to create compressed/encrypted files, not that the
1384          * attribute is compressed/encrypted.  Note we need to check for
1385          * AT_INDEX_ALLOCATION since this is the type of both directory and
1386          * index inodes.
1387          */
1388         if (ni->type != AT_INDEX_ALLOCATION) {
1389                 /* If file is encrypted, deny access, just like NT4. */
1390                 if (NInoEncrypted(ni)) {
1391                         unlock_page(page);
1392                         BUG_ON(ni->type != AT_DATA);
1393                         ntfs_debug("Denying write access to encrypted file.");
1394                         return -EACCES;
1395                 }
1396                 /* Compressed data streams are handled in compress.c. */
1397                 if (NInoNonResident(ni) && NInoCompressed(ni)) {
1398                         BUG_ON(ni->type != AT_DATA);
1399                         BUG_ON(ni->name_len);
1400                         // TODO: Implement and replace this with
1401                         // return ntfs_write_compressed_block(page);
1402                         unlock_page(page);
1403                         ntfs_error(vi->i_sb, "Writing to compressed files is "
1404                                         "not supported yet.  Sorry.");
1405                         return -EOPNOTSUPP;
1406                 }
1407                 // TODO: Implement and remove this check.
1408                 if (NInoNonResident(ni) && NInoSparse(ni)) {
1409                         unlock_page(page);
1410                         ntfs_error(vi->i_sb, "Writing to sparse files is not "
1411                                         "supported yet.  Sorry.");
1412                         return -EOPNOTSUPP;
1413                 }
1414         }
1415         /* NInoNonResident() == NInoIndexAllocPresent() */
1416         if (NInoNonResident(ni)) {
1417                 /* We have to zero every time due to mmap-at-end-of-file. */
1418                 if (page->index >= (i_size >> PAGE_CACHE_SHIFT)) {
1419                         /* The page straddles i_size. */
1420                         unsigned int ofs = i_size & ~PAGE_CACHE_MASK;
1421                         kaddr = kmap_atomic(page, KM_USER0);
1422                         memset(kaddr + ofs, 0, PAGE_CACHE_SIZE - ofs);
1423                         kunmap_atomic(kaddr, KM_USER0);
1424                         flush_dcache_page(page);
1425                 }
1426                 /* Handle mst protected attributes. */
1427                 if (NInoMstProtected(ni))
1428                         return ntfs_write_mst_block(page, wbc);
1429                 /* Normal, non-resident data stream. */
1430                 return ntfs_write_block(page, wbc);
1431         }
1432         /*
1433          * Attribute is resident, implying it is not compressed, encrypted, or
1434          * mst protected.  This also means the attribute is smaller than an mft
1435          * record and hence smaller than a page, so can simply return error on
1436          * any pages with index above 0.  Note the attribute can actually be
1437          * marked compressed but if it is resident the actual data is not
1438          * compressed so we are ok to ignore the compressed flag here.
1439          */
1440         BUG_ON(page_has_buffers(page));
1441         BUG_ON(!PageUptodate(page));
1442         if (unlikely(page->index > 0)) {
1443                 ntfs_error(vi->i_sb, "BUG()! page->index (0x%lx) > 0.  "
1444                                 "Aborting write.", page->index);
1445                 BUG_ON(PageWriteback(page));
1446                 set_page_writeback(page);
1447                 unlock_page(page);
1448                 end_page_writeback(page);
1449                 return -EIO;
1450         }
1451         if (!NInoAttr(ni))
1452                 base_ni = ni;
1453         else
1454                 base_ni = ni->ext.base_ntfs_ino;
1455         /* Map, pin, and lock the mft record. */
1456         m = map_mft_record(base_ni);
1457         if (IS_ERR(m)) {
1458                 err = PTR_ERR(m);
1459                 m = NULL;
1460                 ctx = NULL;
1461                 goto err_out;
1462         }
1463         /*
1464          * If a parallel write made the attribute non-resident, drop the mft
1465          * record and retry the writepage.
1466          */
1467         if (unlikely(NInoNonResident(ni))) {
1468                 unmap_mft_record(base_ni);
1469                 goto retry_writepage;
1470         }
1471         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1472         if (unlikely(!ctx)) {
1473                 err = -ENOMEM;
1474                 goto err_out;
1475         }
1476         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1477                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1478         if (unlikely(err))
1479                 goto err_out;
1480         /*
1481          * Keep the VM happy.  This must be done otherwise the radix-tree tag
1482          * PAGECACHE_TAG_DIRTY remains set even though the page is clean.
1483          */
1484         BUG_ON(PageWriteback(page));
1485         set_page_writeback(page);
1486         unlock_page(page);
1487         attr_len = le32_to_cpu(ctx->attr->data.resident.value_length);
1488         i_size = i_size_read(vi);
1489         if (unlikely(attr_len > i_size)) {
1490                 /* Race with shrinking truncate or a failed truncate. */
1491                 attr_len = i_size;
1492                 /*
1493                  * If the truncate failed, fix it up now.  If a concurrent
1494                  * truncate, we do its job, so it does not have to do anything.
1495                  */
1496                 err = ntfs_resident_attr_value_resize(ctx->mrec, ctx->attr,
1497                                 attr_len);
1498                 /* Shrinking cannot fail. */
1499                 BUG_ON(err);
1500         }
1501         kaddr = kmap_atomic(page, KM_USER0);
1502         /* Copy the data from the page to the mft record. */
1503         memcpy((u8*)ctx->attr +
1504                         le16_to_cpu(ctx->attr->data.resident.value_offset),
1505                         kaddr, attr_len);
1506         /* Zero out of bounds area in the page cache page. */
1507         memset(kaddr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
1508         kunmap_atomic(kaddr, KM_USER0);
1509         flush_dcache_page(page);
1510         flush_dcache_mft_record_page(ctx->ntfs_ino);
1511         /* We are done with the page. */
1512         end_page_writeback(page);
1513         /* Finally, mark the mft record dirty, so it gets written back. */
1514         mark_mft_record_dirty(ctx->ntfs_ino);
1515         ntfs_attr_put_search_ctx(ctx);
1516         unmap_mft_record(base_ni);
1517         return 0;
1518 err_out:
1519         if (err == -ENOMEM) {
1520                 ntfs_warning(vi->i_sb, "Error allocating memory. Redirtying "
1521                                 "page so we try again later.");
1522                 /*
1523                  * Put the page back on mapping->dirty_pages, but leave its
1524                  * buffers' dirty state as-is.
1525                  */
1526                 redirty_page_for_writepage(wbc, page);
1527                 err = 0;
1528         } else {
1529                 ntfs_error(vi->i_sb, "Resident attribute write failed with "
1530                                 "error %i.", err);
1531                 SetPageError(page);
1532                 NVolSetErrors(ni->vol);
1533         }
1534         unlock_page(page);
1535         if (ctx)
1536                 ntfs_attr_put_search_ctx(ctx);
1537         if (m)
1538                 unmap_mft_record(base_ni);
1539         return err;
1540 }
1541
1542 #endif  /* NTFS_RW */
1543
1544 /**
1545  * ntfs_aops - general address space operations for inodes and attributes
1546  */
1547 struct address_space_operations ntfs_aops = {
1548         .readpage       = ntfs_readpage,        /* Fill page with data. */
1549         .sync_page      = block_sync_page,      /* Currently, just unplugs the
1550                                                    disk request queue. */
1551 #ifdef NTFS_RW
1552         .writepage      = ntfs_writepage,       /* Write dirty page to disk. */
1553 #endif /* NTFS_RW */
1554         .migratepage    = buffer_migrate_page,  /* Move a page cache page from
1555                                                    one physical page to an
1556                                                    other. */
1557 };
1558
1559 /**
1560  * ntfs_mst_aops - general address space operations for mst protecteed inodes
1561  *                 and attributes
1562  */
1563 struct address_space_operations ntfs_mst_aops = {
1564         .readpage       = ntfs_readpage,        /* Fill page with data. */
1565         .sync_page      = block_sync_page,      /* Currently, just unplugs the
1566                                                    disk request queue. */
1567 #ifdef NTFS_RW
1568         .writepage      = ntfs_writepage,       /* Write dirty page to disk. */
1569         .set_page_dirty = __set_page_dirty_nobuffers,   /* Set the page dirty
1570                                                    without touching the buffers
1571                                                    belonging to the page. */
1572 #endif /* NTFS_RW */
1573         .migratepage    = buffer_migrate_page,  /* Move a page cache page from
1574                                                    one physical page to an
1575                                                    other. */
1576 };
1577
1578 #ifdef NTFS_RW
1579
1580 /**
1581  * mark_ntfs_record_dirty - mark an ntfs record dirty
1582  * @page:       page containing the ntfs record to mark dirty
1583  * @ofs:        byte offset within @page at which the ntfs record begins
1584  *
1585  * Set the buffers and the page in which the ntfs record is located dirty.
1586  *
1587  * The latter also marks the vfs inode the ntfs record belongs to dirty
1588  * (I_DIRTY_PAGES only).
1589  *
1590  * If the page does not have buffers, we create them and set them uptodate.
1591  * The page may not be locked which is why we need to handle the buffers under
1592  * the mapping->private_lock.  Once the buffers are marked dirty we no longer
1593  * need the lock since try_to_free_buffers() does not free dirty buffers.
1594  */
1595 void mark_ntfs_record_dirty(struct page *page, const unsigned int ofs) {
1596         struct address_space *mapping = page->mapping;
1597         ntfs_inode *ni = NTFS_I(mapping->host);
1598         struct buffer_head *bh, *head, *buffers_to_free = NULL;
1599         unsigned int end, bh_size, bh_ofs;
1600
1601         BUG_ON(!PageUptodate(page));
1602         end = ofs + ni->itype.index.block_size;
1603         bh_size = VFS_I(ni)->i_sb->s_blocksize;
1604         spin_lock(&mapping->private_lock);
1605         if (unlikely(!page_has_buffers(page))) {
1606                 spin_unlock(&mapping->private_lock);
1607                 bh = head = alloc_page_buffers(page, bh_size, 1);
1608                 spin_lock(&mapping->private_lock);
1609                 if (likely(!page_has_buffers(page))) {
1610                         struct buffer_head *tail;
1611
1612                         do {
1613                                 set_buffer_uptodate(bh);
1614                                 tail = bh;
1615                                 bh = bh->b_this_page;
1616                         } while (bh);
1617                         tail->b_this_page = head;
1618                         attach_page_buffers(page, head);
1619                 } else
1620                         buffers_to_free = bh;
1621         }
1622         bh = head = page_buffers(page);
1623         BUG_ON(!bh);
1624         do {
1625                 bh_ofs = bh_offset(bh);
1626                 if (bh_ofs + bh_size <= ofs)
1627                         continue;
1628                 if (unlikely(bh_ofs >= end))
1629                         break;
1630                 set_buffer_dirty(bh);
1631         } while ((bh = bh->b_this_page) != head);
1632         spin_unlock(&mapping->private_lock);
1633         __set_page_dirty_nobuffers(page);
1634         if (unlikely(buffers_to_free)) {
1635                 do {
1636                         bh = buffers_to_free->b_this_page;
1637                         free_buffer_head(buffers_to_free);
1638                         buffers_to_free = bh;
1639                 } while (buffers_to_free);
1640         }
1641 }
1642
1643 #endif /* NTFS_RW */