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