NTFS: Fix a bug in address space operations error recovery code paths where
[linux-3.10.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) 2001-2005 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
31 #include "aops.h"
32 #include "attrib.h"
33 #include "debug.h"
34 #include "inode.h"
35 #include "mft.h"
36 #include "runlist.h"
37 #include "types.h"
38 #include "ntfs.h"
39
40 /**
41  * ntfs_end_buffer_async_read - async io completion for reading attributes
42  * @bh:         buffer head on which io is completed
43  * @uptodate:   whether @bh is now uptodate or not
44  *
45  * Asynchronous I/O completion handler for reading pages belonging to the
46  * attribute address space of an inode.  The inodes can either be files or
47  * directories or they can be fake inodes describing some attribute.
48  *
49  * If NInoMstProtected(), perform the post read mst fixups when all IO on the
50  * page has been completed and mark the page uptodate or set the error bit on
51  * the page.  To determine the size of the records that need fixing up, we
52  * cheat a little bit by setting the index_block_size in ntfs_inode to the ntfs
53  * record size, and index_block_size_bits, to the log(base 2) of the ntfs
54  * record size.
55  */
56 static void ntfs_end_buffer_async_read(struct buffer_head *bh, int uptodate)
57 {
58         static DEFINE_SPINLOCK(page_uptodate_lock);
59         unsigned long flags;
60         struct buffer_head *tmp;
61         struct page *page;
62         ntfs_inode *ni;
63         int page_uptodate = 1;
64
65         page = bh->b_page;
66         ni = NTFS_I(page->mapping->host);
67
68         if (likely(uptodate)) {
69                 s64 file_ofs, initialized_size;
70
71                 set_buffer_uptodate(bh);
72
73                 file_ofs = ((s64)page->index << PAGE_CACHE_SHIFT) +
74                                 bh_offset(bh);
75                 read_lock_irqsave(&ni->size_lock, flags);
76                 initialized_size = ni->initialized_size;
77                 read_unlock_irqrestore(&ni->size_lock, flags);
78                 /* Check for the current buffer head overflowing. */
79                 if (file_ofs + bh->b_size > initialized_size) {
80                         char *addr;
81                         int ofs = 0;
82
83                         if (file_ofs < initialized_size)
84                                 ofs = initialized_size - file_ofs;
85                         addr = kmap_atomic(page, KM_BIO_SRC_IRQ);
86                         memset(addr + bh_offset(bh) + ofs, 0, bh->b_size - ofs);
87                         flush_dcache_page(page);
88                         kunmap_atomic(addr, KM_BIO_SRC_IRQ);
89                 }
90         } else {
91                 clear_buffer_uptodate(bh);
92                 ntfs_error(ni->vol->sb, "Buffer I/O error, logical block %llu.",
93                                 (unsigned long long)bh->b_blocknr);
94                 SetPageError(page);
95         }
96         spin_lock_irqsave(&page_uptodate_lock, flags);
97         clear_buffer_async_read(bh);
98         unlock_buffer(bh);
99         tmp = bh;
100         do {
101                 if (!buffer_uptodate(tmp))
102                         page_uptodate = 0;
103                 if (buffer_async_read(tmp)) {
104                         if (likely(buffer_locked(tmp)))
105                                 goto still_busy;
106                         /* Async buffers must be locked. */
107                         BUG();
108                 }
109                 tmp = tmp->b_this_page;
110         } while (tmp != bh);
111         spin_unlock_irqrestore(&page_uptodate_lock, flags);
112         /*
113          * If none of the buffers had errors then we can set the page uptodate,
114          * but we first have to perform the post read mst fixups, if the
115          * attribute is mst protected, i.e. if NInoMstProteced(ni) is true.
116          * Note we ignore fixup errors as those are detected when
117          * map_mft_record() is called which gives us per record granularity
118          * rather than per page granularity.
119          */
120         if (!NInoMstProtected(ni)) {
121                 if (likely(page_uptodate && !PageError(page)))
122                         SetPageUptodate(page);
123         } else {
124                 char *addr;
125                 unsigned int i, recs;
126                 u32 rec_size;
127
128                 rec_size = ni->itype.index.block_size;
129                 recs = PAGE_CACHE_SIZE / rec_size;
130                 /* Should have been verified before we got here... */
131                 BUG_ON(!recs);
132                 addr = kmap_atomic(page, KM_BIO_SRC_IRQ);
133                 for (i = 0; i < recs; i++)
134                         post_read_mst_fixup((NTFS_RECORD*)(addr +
135                                         i * rec_size), rec_size);
136                 flush_dcache_page(page);
137                 kunmap_atomic(addr, KM_BIO_SRC_IRQ);
138                 if (likely(page_uptodate && !PageError(page)))
139                         SetPageUptodate(page);
140         }
141         unlock_page(page);
142         return;
143 still_busy:
144         spin_unlock_irqrestore(&page_uptodate_lock, flags);
145         return;
146 }
147
148 /**
149  * ntfs_read_block - fill a @page of an address space with data
150  * @page:       page cache page to fill with data
151  *
152  * Fill the page @page of the address space belonging to the @page->host inode.
153  * We read each buffer asynchronously and when all buffers are read in, our io
154  * completion handler ntfs_end_buffer_read_async(), if required, automatically
155  * applies the mst fixups to the page before finally marking it uptodate and
156  * unlocking it.
157  *
158  * We only enforce allocated_size limit because i_size is checked for in
159  * generic_file_read().
160  *
161  * Return 0 on success and -errno on error.
162  *
163  * Contains an adapted version of fs/buffer.c::block_read_full_page().
164  */
165 static int ntfs_read_block(struct page *page)
166 {
167         VCN vcn;
168         LCN lcn;
169         ntfs_inode *ni;
170         ntfs_volume *vol;
171         runlist_element *rl;
172         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
173         sector_t iblock, lblock, zblock;
174         unsigned long flags;
175         unsigned int blocksize, vcn_ofs;
176         int i, nr;
177         unsigned char blocksize_bits;
178
179         ni = NTFS_I(page->mapping->host);
180         vol = ni->vol;
181
182         /* $MFT/$DATA must have its complete runlist in memory at all times. */
183         BUG_ON(!ni->runlist.rl && !ni->mft_no && !NInoAttr(ni));
184
185         blocksize_bits = VFS_I(ni)->i_blkbits;
186         blocksize = 1 << blocksize_bits;
187
188         if (!page_has_buffers(page))
189                 create_empty_buffers(page, blocksize, 0);
190         bh = head = page_buffers(page);
191         if (unlikely(!bh)) {
192                 unlock_page(page);
193                 return -ENOMEM;
194         }
195
196         iblock = (s64)page->index << (PAGE_CACHE_SHIFT - blocksize_bits);
197         read_lock_irqsave(&ni->size_lock, flags);
198         lblock = (ni->allocated_size + blocksize - 1) >> blocksize_bits;
199         zblock = (ni->initialized_size + blocksize - 1) >> blocksize_bits;
200         read_unlock_irqrestore(&ni->size_lock, flags);
201
202         /* Loop through all the buffers in the page. */
203         rl = NULL;
204         nr = i = 0;
205         do {
206                 u8 *kaddr;
207
208                 if (unlikely(buffer_uptodate(bh)))
209                         continue;
210                 if (unlikely(buffer_mapped(bh))) {
211                         arr[nr++] = bh;
212                         continue;
213                 }
214                 bh->b_bdev = vol->sb->s_bdev;
215                 /* Is the block within the allowed limits? */
216                 if (iblock < lblock) {
217                         BOOL is_retry = FALSE;
218
219                         /* Convert iblock into corresponding vcn and offset. */
220                         vcn = (VCN)iblock << blocksize_bits >>
221                                         vol->cluster_size_bits;
222                         vcn_ofs = ((VCN)iblock << blocksize_bits) &
223                                         vol->cluster_size_mask;
224                         if (!rl) {
225 lock_retry_remap:
226                                 down_read(&ni->runlist.lock);
227                                 rl = ni->runlist.rl;
228                         }
229                         if (likely(rl != NULL)) {
230                                 /* Seek to element containing target vcn. */
231                                 while (rl->length && rl[1].vcn <= vcn)
232                                         rl++;
233                                 lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
234                         } else
235                                 lcn = LCN_RL_NOT_MAPPED;
236                         /* Successful remap. */
237                         if (lcn >= 0) {
238                                 /* Setup buffer head to correct block. */
239                                 bh->b_blocknr = ((lcn << vol->cluster_size_bits)
240                                                 + vcn_ofs) >> blocksize_bits;
241                                 set_buffer_mapped(bh);
242                                 /* Only read initialized data blocks. */
243                                 if (iblock < zblock) {
244                                         arr[nr++] = bh;
245                                         continue;
246                                 }
247                                 /* Fully non-initialized data block, zero it. */
248                                 goto handle_zblock;
249                         }
250                         /* It is a hole, need to zero it. */
251                         if (lcn == LCN_HOLE)
252                                 goto handle_hole;
253                         /* If first try and runlist unmapped, map and retry. */
254                         if (!is_retry && lcn == LCN_RL_NOT_MAPPED) {
255                                 int err;
256                                 is_retry = TRUE;
257                                 /*
258                                  * Attempt to map runlist, dropping lock for
259                                  * the duration.
260                                  */
261                                 up_read(&ni->runlist.lock);
262                                 err = ntfs_map_runlist(ni, vcn);
263                                 if (likely(!err))
264                                         goto lock_retry_remap;
265                                 rl = NULL;
266                                 lcn = err;
267                         } else if (!rl)
268                                 up_read(&ni->runlist.lock);
269                         /* Hard error, zero out region. */
270                         bh->b_blocknr = -1;
271                         SetPageError(page);
272                         ntfs_error(vol->sb, "Failed to read from inode 0x%lx, "
273                                         "attribute type 0x%x, vcn 0x%llx, "
274                                         "offset 0x%x because its location on "
275                                         "disk could not be determined%s "
276                                         "(error code %lli).", ni->mft_no,
277                                         ni->type, (unsigned long long)vcn,
278                                         vcn_ofs, is_retry ? " even after "
279                                         "retrying" : "", (long long)lcn);
280                 }
281                 /*
282                  * Either iblock was outside lblock limits or
283                  * ntfs_rl_vcn_to_lcn() returned error.  Just zero that portion
284                  * of the page and set the buffer uptodate.
285                  */
286 handle_hole:
287                 bh->b_blocknr = -1UL;
288                 clear_buffer_mapped(bh);
289 handle_zblock:
290                 kaddr = kmap_atomic(page, KM_USER0);
291                 memset(kaddr + i * blocksize, 0, blocksize);
292                 flush_dcache_page(page);
293                 kunmap_atomic(kaddr, KM_USER0);
294                 set_buffer_uptodate(bh);
295         } while (i++, iblock++, (bh = bh->b_this_page) != head);
296
297         /* Release the lock if we took it. */
298         if (rl)
299                 up_read(&ni->runlist.lock);
300
301         /* Check we have at least one buffer ready for i/o. */
302         if (nr) {
303                 struct buffer_head *tbh;
304
305                 /* Lock the buffers. */
306                 for (i = 0; i < nr; i++) {
307                         tbh = arr[i];
308                         lock_buffer(tbh);
309                         tbh->b_end_io = ntfs_end_buffer_async_read;
310                         set_buffer_async_read(tbh);
311                 }
312                 /* Finally, start i/o on the buffers. */
313                 for (i = 0; i < nr; i++) {
314                         tbh = arr[i];
315                         if (likely(!buffer_uptodate(tbh)))
316                                 submit_bh(READ, tbh);
317                         else
318                                 ntfs_end_buffer_async_read(tbh, 1);
319                 }
320                 return 0;
321         }
322         /* No i/o was scheduled on any of the buffers. */
323         if (likely(!PageError(page)))
324                 SetPageUptodate(page);
325         else /* Signal synchronous i/o error. */
326                 nr = -EIO;
327         unlock_page(page);
328         return nr;
329 }
330
331 /**
332  * ntfs_readpage - fill a @page of a @file with data from the device
333  * @file:       open file to which the page @page belongs or NULL
334  * @page:       page cache page to fill with data
335  *
336  * For non-resident attributes, ntfs_readpage() fills the @page of the open
337  * file @file by calling the ntfs version of the generic block_read_full_page()
338  * function, ntfs_read_block(), which in turn creates and reads in the buffers
339  * associated with the page asynchronously.
340  *
341  * For resident attributes, OTOH, ntfs_readpage() fills @page by copying the
342  * data from the mft record (which at this stage is most likely in memory) and
343  * fills the remainder with zeroes. Thus, in this case, I/O is synchronous, as
344  * even if the mft record is not cached at this point in time, we need to wait
345  * for it to be read in before we can do the copy.
346  *
347  * Return 0 on success and -errno on error.
348  */
349 static int ntfs_readpage(struct file *file, struct page *page)
350 {
351         ntfs_inode *ni, *base_ni;
352         u8 *kaddr;
353         ntfs_attr_search_ctx *ctx;
354         MFT_RECORD *mrec;
355         unsigned long flags;
356         u32 attr_len;
357         int err = 0;
358
359 retry_readpage:
360         BUG_ON(!PageLocked(page));
361         /*
362          * This can potentially happen because we clear PageUptodate() during
363          * ntfs_writepage() of MstProtected() attributes.
364          */
365         if (PageUptodate(page)) {
366                 unlock_page(page);
367                 return 0;
368         }
369         ni = NTFS_I(page->mapping->host);
370
371         /* NInoNonResident() == NInoIndexAllocPresent() */
372         if (NInoNonResident(ni)) {
373                 /*
374                  * Only unnamed $DATA attributes can be compressed or
375                  * encrypted.
376                  */
377                 if (ni->type == AT_DATA && !ni->name_len) {
378                         /* If file is encrypted, deny access, just like NT4. */
379                         if (NInoEncrypted(ni)) {
380                                 err = -EACCES;
381                                 goto err_out;
382                         }
383                         /* Compressed data streams are handled in compress.c. */
384                         if (NInoCompressed(ni))
385                                 return ntfs_read_compressed_block(page);
386                 }
387                 /* Normal data stream. */
388                 return ntfs_read_block(page);
389         }
390         /*
391          * Attribute is resident, implying it is not compressed or encrypted.
392          * This also means the attribute is smaller than an mft record and
393          * hence smaller than a page, so can simply zero out any pages with
394          * index above 0.
395          */
396         if (unlikely(page->index > 0)) {
397                 kaddr = kmap_atomic(page, KM_USER0);
398                 memset(kaddr, 0, PAGE_CACHE_SIZE);
399                 flush_dcache_page(page);
400                 kunmap_atomic(kaddr, KM_USER0);
401                 goto done;
402         }
403         if (!NInoAttr(ni))
404                 base_ni = ni;
405         else
406                 base_ni = ni->ext.base_ntfs_ino;
407         /* Map, pin, and lock the mft record. */
408         mrec = map_mft_record(base_ni);
409         if (IS_ERR(mrec)) {
410                 err = PTR_ERR(mrec);
411                 goto err_out;
412         }
413         /*
414          * If a parallel write made the attribute non-resident, drop the mft
415          * record and retry the readpage.
416          */
417         if (unlikely(NInoNonResident(ni))) {
418                 unmap_mft_record(base_ni);
419                 goto retry_readpage;
420         }
421         ctx = ntfs_attr_get_search_ctx(base_ni, mrec);
422         if (unlikely(!ctx)) {
423                 err = -ENOMEM;
424                 goto unm_err_out;
425         }
426         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
427                         CASE_SENSITIVE, 0, NULL, 0, ctx);
428         if (unlikely(err))
429                 goto put_unm_err_out;
430         attr_len = le32_to_cpu(ctx->attr->data.resident.value_length);
431         read_lock_irqsave(&ni->size_lock, flags);
432         if (unlikely(attr_len > ni->initialized_size))
433                 attr_len = ni->initialized_size;
434         read_unlock_irqrestore(&ni->size_lock, flags);
435         kaddr = kmap_atomic(page, KM_USER0);
436         /* Copy the data to the page. */
437         memcpy(kaddr, (u8*)ctx->attr +
438                         le16_to_cpu(ctx->attr->data.resident.value_offset),
439                         attr_len);
440         /* Zero the remainder of the page. */
441         memset(kaddr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
442         flush_dcache_page(page);
443         kunmap_atomic(kaddr, KM_USER0);
444 put_unm_err_out:
445         ntfs_attr_put_search_ctx(ctx);
446 unm_err_out:
447         unmap_mft_record(base_ni);
448 done:
449         SetPageUptodate(page);
450 err_out:
451         unlock_page(page);
452         return err;
453 }
454
455 #ifdef NTFS_RW
456
457 /**
458  * ntfs_write_block - write a @page to the backing store
459  * @page:       page cache page to write out
460  * @wbc:        writeback control structure
461  *
462  * This function is for writing pages belonging to non-resident, non-mst
463  * protected attributes to their backing store.
464  *
465  * For a page with buffers, map and write the dirty buffers asynchronously
466  * under page writeback. For a page without buffers, create buffers for the
467  * page, then proceed as above.
468  *
469  * If a page doesn't have buffers the page dirty state is definitive. If a page
470  * does have buffers, the page dirty state is just a hint, and the buffer dirty
471  * state is definitive. (A hint which has rules: dirty buffers against a clean
472  * page is illegal. Other combinations are legal and need to be handled. In
473  * particular a dirty page containing clean buffers for example.)
474  *
475  * Return 0 on success and -errno on error.
476  *
477  * Based on ntfs_read_block() and __block_write_full_page().
478  */
479 static int ntfs_write_block(struct page *page, struct writeback_control *wbc)
480 {
481         VCN vcn;
482         LCN lcn;
483         s64 initialized_size;
484         loff_t i_size;
485         sector_t block, dblock, iblock;
486         struct inode *vi;
487         ntfs_inode *ni;
488         ntfs_volume *vol;
489         runlist_element *rl;
490         struct buffer_head *bh, *head;
491         unsigned long flags;
492         unsigned int blocksize, vcn_ofs;
493         int err;
494         BOOL need_end_writeback;
495         unsigned char blocksize_bits;
496
497         vi = page->mapping->host;
498         ni = NTFS_I(vi);
499         vol = ni->vol;
500
501         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
502                         "0x%lx.", ni->mft_no, ni->type, page->index);
503
504         BUG_ON(!NInoNonResident(ni));
505         BUG_ON(NInoMstProtected(ni));
506
507         blocksize_bits = vi->i_blkbits;
508         blocksize = 1 << blocksize_bits;
509
510         if (!page_has_buffers(page)) {
511                 BUG_ON(!PageUptodate(page));
512                 create_empty_buffers(page, blocksize,
513                                 (1 << BH_Uptodate) | (1 << BH_Dirty));
514         }
515         bh = head = page_buffers(page);
516         if (unlikely(!bh)) {
517                 ntfs_warning(vol->sb, "Error allocating page buffers. "
518                                 "Redirtying page so we try again later.");
519                 /*
520                  * Put the page back on mapping->dirty_pages, but leave its
521                  * buffer's dirty state as-is.
522                  */
523                 redirty_page_for_writepage(wbc, page);
524                 unlock_page(page);
525                 return 0;
526         }
527
528         /* NOTE: Different naming scheme to ntfs_read_block()! */
529
530         /* The first block in the page. */
531         block = (s64)page->index << (PAGE_CACHE_SHIFT - blocksize_bits);
532
533         read_lock_irqsave(&ni->size_lock, flags);
534         i_size = i_size_read(vi);
535         initialized_size = ni->initialized_size;
536         read_unlock_irqrestore(&ni->size_lock, flags);
537
538         /* The first out of bounds block for the data size. */
539         dblock = (i_size + blocksize - 1) >> blocksize_bits;
540
541         /* The last (fully or partially) initialized block. */
542         iblock = initialized_size >> blocksize_bits;
543
544         /*
545          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
546          * here, and the (potentially unmapped) buffers may become dirty at
547          * any time.  If a buffer becomes dirty here after we've inspected it
548          * then we just miss that fact, and the page stays dirty.
549          *
550          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
551          * handle that here by just cleaning them.
552          */
553
554         /*
555          * Loop through all the buffers in the page, mapping all the dirty
556          * buffers to disk addresses and handling any aliases from the
557          * underlying block device's mapping.
558          */
559         rl = NULL;
560         err = 0;
561         do {
562                 BOOL is_retry = FALSE;
563
564                 if (unlikely(block >= dblock)) {
565                         /*
566                          * Mapped buffers outside i_size will occur, because
567                          * this page can be outside i_size when there is a
568                          * truncate in progress. The contents of such buffers
569                          * were zeroed by ntfs_writepage().
570                          *
571                          * FIXME: What about the small race window where
572                          * ntfs_writepage() has not done any clearing because
573                          * the page was within i_size but before we get here,
574                          * vmtruncate() modifies i_size?
575                          */
576                         clear_buffer_dirty(bh);
577                         set_buffer_uptodate(bh);
578                         continue;
579                 }
580
581                 /* Clean buffers are not written out, so no need to map them. */
582                 if (!buffer_dirty(bh))
583                         continue;
584
585                 /* Make sure we have enough initialized size. */
586                 if (unlikely((block >= iblock) &&
587                                 (initialized_size < i_size))) {
588                         /*
589                          * If this page is fully outside initialized size, zero
590                          * out all pages between the current initialized size
591                          * and the current page. Just use ntfs_readpage() to do
592                          * the zeroing transparently.
593                          */
594                         if (block > iblock) {
595                                 // TODO:
596                                 // For each page do:
597                                 // - read_cache_page()
598                                 // Again for each page do:
599                                 // - wait_on_page_locked()
600                                 // - Check (PageUptodate(page) &&
601                                 //                      !PageError(page))
602                                 // Update initialized size in the attribute and
603                                 // in the inode.
604                                 // Again, for each page do:
605                                 //      __set_page_dirty_buffers();
606                                 // page_cache_release()
607                                 // We don't need to wait on the writes.
608                                 // Update iblock.
609                         }
610                         /*
611                          * The current page straddles initialized size. Zero
612                          * all non-uptodate buffers and set them uptodate (and
613                          * dirty?). Note, there aren't any non-uptodate buffers
614                          * if the page is uptodate.
615                          * FIXME: For an uptodate page, the buffers may need to
616                          * be written out because they were not initialized on
617                          * disk before.
618                          */
619                         if (!PageUptodate(page)) {
620                                 // TODO:
621                                 // Zero any non-uptodate buffers up to i_size.
622                                 // Set them uptodate and dirty.
623                         }
624                         // TODO:
625                         // Update initialized size in the attribute and in the
626                         // inode (up to i_size).
627                         // Update iblock.
628                         // FIXME: This is inefficient. Try to batch the two
629                         // size changes to happen in one go.
630                         ntfs_error(vol->sb, "Writing beyond initialized size "
631                                         "is not supported yet. Sorry.");
632                         err = -EOPNOTSUPP;
633                         break;
634                         // Do NOT set_buffer_new() BUT DO clear buffer range
635                         // outside write request range.
636                         // set_buffer_uptodate() on complete buffers as well as
637                         // set_buffer_dirty().
638                 }
639
640                 /* No need to map buffers that are already mapped. */
641                 if (buffer_mapped(bh))
642                         continue;
643
644                 /* Unmapped, dirty buffer. Need to map it. */
645                 bh->b_bdev = vol->sb->s_bdev;
646
647                 /* Convert block into corresponding vcn and offset. */
648                 vcn = (VCN)block << blocksize_bits;
649                 vcn_ofs = vcn & vol->cluster_size_mask;
650                 vcn >>= vol->cluster_size_bits;
651                 if (!rl) {
652 lock_retry_remap:
653                         down_read(&ni->runlist.lock);
654                         rl = ni->runlist.rl;
655                 }
656                 if (likely(rl != NULL)) {
657                         /* Seek to element containing target vcn. */
658                         while (rl->length && rl[1].vcn <= vcn)
659                                 rl++;
660                         lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
661                 } else
662                         lcn = LCN_RL_NOT_MAPPED;
663                 /* Successful remap. */
664                 if (lcn >= 0) {
665                         /* Setup buffer head to point to correct block. */
666                         bh->b_blocknr = ((lcn << vol->cluster_size_bits) +
667                                         vcn_ofs) >> blocksize_bits;
668                         set_buffer_mapped(bh);
669                         continue;
670                 }
671                 /* It is a hole, need to instantiate it. */
672                 if (lcn == LCN_HOLE) {
673                         // TODO: Instantiate the hole.
674                         // clear_buffer_new(bh);
675                         // unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
676                         ntfs_error(vol->sb, "Writing into sparse regions is "
677                                         "not supported yet. Sorry.");
678                         err = -EOPNOTSUPP;
679                         break;
680                 }
681                 /* If first try and runlist unmapped, map and retry. */
682                 if (!is_retry && lcn == LCN_RL_NOT_MAPPED) {
683                         is_retry = TRUE;
684                         /*
685                          * Attempt to map runlist, dropping lock for
686                          * the duration.
687                          */
688                         up_read(&ni->runlist.lock);
689                         err = ntfs_map_runlist(ni, vcn);
690                         if (likely(!err))
691                                 goto lock_retry_remap;
692                         rl = NULL;
693                         lcn = err;
694                 } else if (!rl)
695                         up_read(&ni->runlist.lock);
696                 /* Failed to map the buffer, even after retrying. */
697                 bh->b_blocknr = -1;
698                 ntfs_error(vol->sb, "Failed to write to inode 0x%lx, "
699                                 "attribute type 0x%x, vcn 0x%llx, offset 0x%x "
700                                 "because its location on disk could not be "
701                                 "determined%s (error code %lli).", ni->mft_no,
702                                 ni->type, (unsigned long long)vcn,
703                                 vcn_ofs, is_retry ? " even after "
704                                 "retrying" : "", (long long)lcn);
705                 if (!err)
706                         err = -EIO;
707                 break;
708         } while (block++, (bh = bh->b_this_page) != head);
709
710         /* Release the lock if we took it. */
711         if (rl)
712                 up_read(&ni->runlist.lock);
713
714         /* For the error case, need to reset bh to the beginning. */
715         bh = head;
716
717         /* Just an optimization, so ->readpage() isn't called later. */
718         if (unlikely(!PageUptodate(page))) {
719                 int uptodate = 1;
720                 do {
721                         if (!buffer_uptodate(bh)) {
722                                 uptodate = 0;
723                                 bh = head;
724                                 break;
725                         }
726                 } while ((bh = bh->b_this_page) != head);
727                 if (uptodate)
728                         SetPageUptodate(page);
729         }
730
731         /* Setup all mapped, dirty buffers for async write i/o. */
732         do {
733                 get_bh(bh);
734                 if (buffer_mapped(bh) && buffer_dirty(bh)) {
735                         lock_buffer(bh);
736                         if (test_clear_buffer_dirty(bh)) {
737                                 BUG_ON(!buffer_uptodate(bh));
738                                 mark_buffer_async_write(bh);
739                         } else
740                                 unlock_buffer(bh);
741                 } else if (unlikely(err)) {
742                         /*
743                          * For the error case. The buffer may have been set
744                          * dirty during attachment to a dirty page.
745                          */
746                         if (err != -ENOMEM)
747                                 clear_buffer_dirty(bh);
748                 }
749         } while ((bh = bh->b_this_page) != head);
750
751         if (unlikely(err)) {
752                 // TODO: Remove the -EOPNOTSUPP check later on...
753                 if (unlikely(err == -EOPNOTSUPP))
754                         err = 0;
755                 else if (err == -ENOMEM) {
756                         ntfs_warning(vol->sb, "Error allocating memory. "
757                                         "Redirtying page so we try again "
758                                         "later.");
759                         /*
760                          * Put the page back on mapping->dirty_pages, but
761                          * leave its buffer's dirty state as-is.
762                          */
763                         redirty_page_for_writepage(wbc, page);
764                         err = 0;
765                 } else
766                         SetPageError(page);
767         }
768
769         BUG_ON(PageWriteback(page));
770         set_page_writeback(page);       /* Keeps try_to_free_buffers() away. */
771         unlock_page(page);
772
773         /*
774          * Submit the prepared buffers for i/o. Note the page is unlocked,
775          * and the async write i/o completion handler can end_page_writeback()
776          * at any time after the *first* submit_bh(). So the buffers can then
777          * disappear...
778          */
779         need_end_writeback = TRUE;
780         do {
781                 struct buffer_head *next = bh->b_this_page;
782                 if (buffer_async_write(bh)) {
783                         submit_bh(WRITE, bh);
784                         need_end_writeback = FALSE;
785                 }
786                 put_bh(bh);
787                 bh = next;
788         } while (bh != head);
789
790         /* If no i/o was started, need to end_page_writeback(). */
791         if (unlikely(need_end_writeback))
792                 end_page_writeback(page);
793
794         ntfs_debug("Done.");
795         return err;
796 }
797
798 /**
799  * ntfs_write_mst_block - write a @page to the backing store
800  * @page:       page cache page to write out
801  * @wbc:        writeback control structure
802  *
803  * This function is for writing pages belonging to non-resident, mst protected
804  * attributes to their backing store.  The only supported attributes are index
805  * allocation and $MFT/$DATA.  Both directory inodes and index inodes are
806  * supported for the index allocation case.
807  *
808  * The page must remain locked for the duration of the write because we apply
809  * the mst fixups, write, and then undo the fixups, so if we were to unlock the
810  * page before undoing the fixups, any other user of the page will see the
811  * page contents as corrupt.
812  *
813  * We clear the page uptodate flag for the duration of the function to ensure
814  * exclusion for the $MFT/$DATA case against someone mapping an mft record we
815  * are about to apply the mst fixups to.
816  *
817  * Return 0 on success and -errno on error.
818  *
819  * Based on ntfs_write_block(), ntfs_mft_writepage(), and
820  * write_mft_record_nolock().
821  */
822 static int ntfs_write_mst_block(struct page *page,
823                 struct writeback_control *wbc)
824 {
825         sector_t block, dblock, rec_block;
826         struct inode *vi = page->mapping->host;
827         ntfs_inode *ni = NTFS_I(vi);
828         ntfs_volume *vol = ni->vol;
829         u8 *kaddr;
830         unsigned int rec_size = ni->itype.index.block_size;
831         ntfs_inode *locked_nis[PAGE_CACHE_SIZE / rec_size];
832         struct buffer_head *bh, *head, *tbh, *rec_start_bh;
833         struct buffer_head *bhs[MAX_BUF_PER_PAGE];
834         runlist_element *rl;
835         int i, nr_locked_nis, nr_recs, nr_bhs, max_bhs, bhs_per_rec, err, err2;
836         unsigned bh_size, rec_size_bits;
837         BOOL sync, is_mft, page_is_dirty, rec_is_dirty;
838         unsigned char bh_size_bits;
839
840         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
841                         "0x%lx.", vi->i_ino, ni->type, page->index);
842         BUG_ON(!NInoNonResident(ni));
843         BUG_ON(!NInoMstProtected(ni));
844         is_mft = (S_ISREG(vi->i_mode) && !vi->i_ino);
845         /*
846          * NOTE: ntfs_write_mst_block() would be called for $MFTMirr if a page
847          * in its page cache were to be marked dirty.  However this should
848          * never happen with the current driver and considering we do not
849          * handle this case here we do want to BUG(), at least for now.
850          */
851         BUG_ON(!(is_mft || S_ISDIR(vi->i_mode) ||
852                         (NInoAttr(ni) && ni->type == AT_INDEX_ALLOCATION)));
853         bh_size_bits = vi->i_blkbits;
854         bh_size = 1 << bh_size_bits;
855         max_bhs = PAGE_CACHE_SIZE / bh_size;
856         BUG_ON(!max_bhs);
857         BUG_ON(max_bhs > MAX_BUF_PER_PAGE);
858
859         /* Were we called for sync purposes? */
860         sync = (wbc->sync_mode == WB_SYNC_ALL);
861
862         /* Make sure we have mapped buffers. */
863         BUG_ON(!page_has_buffers(page));
864         bh = head = page_buffers(page);
865         BUG_ON(!bh);
866
867         rec_size_bits = ni->itype.index.block_size_bits;
868         BUG_ON(!(PAGE_CACHE_SIZE >> rec_size_bits));
869         bhs_per_rec = rec_size >> bh_size_bits;
870         BUG_ON(!bhs_per_rec);
871
872         /* The first block in the page. */
873         rec_block = block = (sector_t)page->index <<
874                         (PAGE_CACHE_SHIFT - bh_size_bits);
875
876         /* The first out of bounds block for the data size. */
877         dblock = (i_size_read(vi) + bh_size - 1) >> bh_size_bits;
878
879         rl = NULL;
880         err = err2 = nr_bhs = nr_recs = nr_locked_nis = 0;
881         page_is_dirty = rec_is_dirty = FALSE;
882         rec_start_bh = NULL;
883         do {
884                 BOOL is_retry = FALSE;
885
886                 if (likely(block < rec_block)) {
887                         if (unlikely(block >= dblock)) {
888                                 clear_buffer_dirty(bh);
889                                 set_buffer_uptodate(bh);
890                                 continue;
891                         }
892                         /*
893                          * This block is not the first one in the record.  We
894                          * ignore the buffer's dirty state because we could
895                          * have raced with a parallel mark_ntfs_record_dirty().
896                          */
897                         if (!rec_is_dirty)
898                                 continue;
899                         if (unlikely(err2)) {
900                                 if (err2 != -ENOMEM)
901                                         clear_buffer_dirty(bh);
902                                 continue;
903                         }
904                 } else /* if (block == rec_block) */ {
905                         BUG_ON(block > rec_block);
906                         /* This block is the first one in the record. */
907                         rec_block += bhs_per_rec;
908                         err2 = 0;
909                         if (unlikely(block >= dblock)) {
910                                 clear_buffer_dirty(bh);
911                                 continue;
912                         }
913                         if (!buffer_dirty(bh)) {
914                                 /* Clean records are not written out. */
915                                 rec_is_dirty = FALSE;
916                                 continue;
917                         }
918                         rec_is_dirty = TRUE;
919                         rec_start_bh = bh;
920                 }
921                 /* Need to map the buffer if it is not mapped already. */
922                 if (unlikely(!buffer_mapped(bh))) {
923                         VCN vcn;
924                         LCN lcn;
925                         unsigned int vcn_ofs;
926
927                         /* Obtain the vcn and offset of the current block. */
928                         vcn = (VCN)block << bh_size_bits;
929                         vcn_ofs = vcn & vol->cluster_size_mask;
930                         vcn >>= vol->cluster_size_bits;
931                         if (!rl) {
932 lock_retry_remap:
933                                 down_read(&ni->runlist.lock);
934                                 rl = ni->runlist.rl;
935                         }
936                         if (likely(rl != NULL)) {
937                                 /* Seek to element containing target vcn. */
938                                 while (rl->length && rl[1].vcn <= vcn)
939                                         rl++;
940                                 lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
941                         } else
942                                 lcn = LCN_RL_NOT_MAPPED;
943                         /* Successful remap. */
944                         if (likely(lcn >= 0)) {
945                                 /* Setup buffer head to correct block. */
946                                 bh->b_blocknr = ((lcn <<
947                                                 vol->cluster_size_bits) +
948                                                 vcn_ofs) >> bh_size_bits;
949                                 set_buffer_mapped(bh);
950                         } else {
951                                 /*
952                                  * Remap failed.  Retry to map the runlist once
953                                  * unless we are working on $MFT which always
954                                  * has the whole of its runlist in memory.
955                                  */
956                                 if (!is_mft && !is_retry &&
957                                                 lcn == LCN_RL_NOT_MAPPED) {
958                                         is_retry = TRUE;
959                                         /*
960                                          * Attempt to map runlist, dropping
961                                          * lock for the duration.
962                                          */
963                                         up_read(&ni->runlist.lock);
964                                         err2 = ntfs_map_runlist(ni, vcn);
965                                         if (likely(!err2))
966                                                 goto lock_retry_remap;
967                                         if (err2 == -ENOMEM)
968                                                 page_is_dirty = TRUE;
969                                         lcn = err2;
970                                 } else {
971                                         err2 = -EIO;
972                                         if (!rl)
973                                                 up_read(&ni->runlist.lock);
974                                 }
975                                 /* Hard error.  Abort writing this record. */
976                                 if (!err || err == -ENOMEM)
977                                         err = err2;
978                                 bh->b_blocknr = -1;
979                                 ntfs_error(vol->sb, "Cannot write ntfs record "
980                                                 "0x%llx (inode 0x%lx, "
981                                                 "attribute type 0x%x) because "
982                                                 "its location on disk could "
983                                                 "not be determined (error "
984                                                 "code %lli).",
985                                                 (long long)block <<
986                                                 bh_size_bits >>
987                                                 vol->mft_record_size_bits,
988                                                 ni->mft_no, ni->type,
989                                                 (long long)lcn);
990                                 /*
991                                  * If this is not the first buffer, remove the
992                                  * buffers in this record from the list of
993                                  * buffers to write and clear their dirty bit
994                                  * if not error -ENOMEM.
995                                  */
996                                 if (rec_start_bh != bh) {
997                                         while (bhs[--nr_bhs] != rec_start_bh)
998                                                 ;
999                                         if (err2 != -ENOMEM) {
1000                                                 do {
1001                                                         clear_buffer_dirty(
1002                                                                 rec_start_bh);
1003                                                 } while ((rec_start_bh =
1004                                                                 rec_start_bh->
1005                                                                 b_this_page) !=
1006                                                                 bh);
1007                                         }
1008                                 }
1009                                 continue;
1010                         }
1011                 }
1012                 BUG_ON(!buffer_uptodate(bh));
1013                 BUG_ON(nr_bhs >= max_bhs);
1014                 bhs[nr_bhs++] = bh;
1015         } while (block++, (bh = bh->b_this_page) != head);
1016         if (unlikely(rl))
1017                 up_read(&ni->runlist.lock);
1018         /* If there were no dirty buffers, we are done. */
1019         if (!nr_bhs)
1020                 goto done;
1021         /* Map the page so we can access its contents. */
1022         kaddr = kmap(page);
1023         /* Clear the page uptodate flag whilst the mst fixups are applied. */
1024         BUG_ON(!PageUptodate(page));
1025         ClearPageUptodate(page);
1026         for (i = 0; i < nr_bhs; i++) {
1027                 unsigned int ofs;
1028
1029                 /* Skip buffers which are not at the beginning of records. */
1030                 if (i % bhs_per_rec)
1031                         continue;
1032                 tbh = bhs[i];
1033                 ofs = bh_offset(tbh);
1034                 if (is_mft) {
1035                         ntfs_inode *tni;
1036                         unsigned long mft_no;
1037
1038                         /* Get the mft record number. */
1039                         mft_no = (((s64)page->index << PAGE_CACHE_SHIFT) + ofs)
1040                                         >> rec_size_bits;
1041                         /* Check whether to write this mft record. */
1042                         tni = NULL;
1043                         if (!ntfs_may_write_mft_record(vol, mft_no,
1044                                         (MFT_RECORD*)(kaddr + ofs), &tni)) {
1045                                 /*
1046                                  * The record should not be written.  This
1047                                  * means we need to redirty the page before
1048                                  * returning.
1049                                  */
1050                                 page_is_dirty = TRUE;
1051                                 /*
1052                                  * Remove the buffers in this mft record from
1053                                  * the list of buffers to write.
1054                                  */
1055                                 do {
1056                                         bhs[i] = NULL;
1057                                 } while (++i % bhs_per_rec);
1058                                 continue;
1059                         }
1060                         /*
1061                          * The record should be written.  If a locked ntfs
1062                          * inode was returned, add it to the array of locked
1063                          * ntfs inodes.
1064                          */
1065                         if (tni)
1066                                 locked_nis[nr_locked_nis++] = tni;
1067                 }
1068                 /* Apply the mst protection fixups. */
1069                 err2 = pre_write_mst_fixup((NTFS_RECORD*)(kaddr + ofs),
1070                                 rec_size);
1071                 if (unlikely(err2)) {
1072                         if (!err || err == -ENOMEM)
1073                                 err = -EIO;
1074                         ntfs_error(vol->sb, "Failed to apply mst fixups "
1075                                         "(inode 0x%lx, attribute type 0x%x, "
1076                                         "page index 0x%lx, page offset 0x%x)!"
1077                                         "  Unmount and run chkdsk.", vi->i_ino,
1078                                         ni->type, page->index, ofs);
1079                         /*
1080                          * Mark all the buffers in this record clean as we do
1081                          * not want to write corrupt data to disk.
1082                          */
1083                         do {
1084                                 clear_buffer_dirty(bhs[i]);
1085                                 bhs[i] = NULL;
1086                         } while (++i % bhs_per_rec);
1087                         continue;
1088                 }
1089                 nr_recs++;
1090         }
1091         /* If no records are to be written out, we are done. */
1092         if (!nr_recs)
1093                 goto unm_done;
1094         flush_dcache_page(page);
1095         /* Lock buffers and start synchronous write i/o on them. */
1096         for (i = 0; i < nr_bhs; i++) {
1097                 tbh = bhs[i];
1098                 if (!tbh)
1099                         continue;
1100                 if (unlikely(test_set_buffer_locked(tbh)))
1101                         BUG();
1102                 /* The buffer dirty state is now irrelevant, just clean it. */
1103                 clear_buffer_dirty(tbh);
1104                 BUG_ON(!buffer_uptodate(tbh));
1105                 BUG_ON(!buffer_mapped(tbh));
1106                 get_bh(tbh);
1107                 tbh->b_end_io = end_buffer_write_sync;
1108                 submit_bh(WRITE, tbh);
1109         }
1110         /* Synchronize the mft mirror now if not @sync. */
1111         if (is_mft && !sync)
1112                 goto do_mirror;
1113 do_wait:
1114         /* Wait on i/o completion of buffers. */
1115         for (i = 0; i < nr_bhs; i++) {
1116                 tbh = bhs[i];
1117                 if (!tbh)
1118                         continue;
1119                 wait_on_buffer(tbh);
1120                 if (unlikely(!buffer_uptodate(tbh))) {
1121                         ntfs_error(vol->sb, "I/O error while writing ntfs "
1122                                         "record buffer (inode 0x%lx, "
1123                                         "attribute type 0x%x, page index "
1124                                         "0x%lx, page offset 0x%lx)!  Unmount "
1125                                         "and run chkdsk.", vi->i_ino, ni->type,
1126                                         page->index, bh_offset(tbh));
1127                         if (!err || err == -ENOMEM)
1128                                 err = -EIO;
1129                         /*
1130                          * Set the buffer uptodate so the page and buffer
1131                          * states do not become out of sync.
1132                          */
1133                         set_buffer_uptodate(tbh);
1134                 }
1135         }
1136         /* If @sync, now synchronize the mft mirror. */
1137         if (is_mft && sync) {
1138 do_mirror:
1139                 for (i = 0; i < nr_bhs; i++) {
1140                         unsigned long mft_no;
1141                         unsigned int ofs;
1142
1143                         /*
1144                          * Skip buffers which are not at the beginning of
1145                          * records.
1146                          */
1147                         if (i % bhs_per_rec)
1148                                 continue;
1149                         tbh = bhs[i];
1150                         /* Skip removed buffers (and hence records). */
1151                         if (!tbh)
1152                                 continue;
1153                         ofs = bh_offset(tbh);
1154                         /* Get the mft record number. */
1155                         mft_no = (((s64)page->index << PAGE_CACHE_SHIFT) + ofs)
1156                                         >> rec_size_bits;
1157                         if (mft_no < vol->mftmirr_size)
1158                                 ntfs_sync_mft_mirror(vol, mft_no,
1159                                                 (MFT_RECORD*)(kaddr + ofs),
1160                                                 sync);
1161                 }
1162                 if (!sync)
1163                         goto do_wait;
1164         }
1165         /* Remove the mst protection fixups again. */
1166         for (i = 0; i < nr_bhs; i++) {
1167                 if (!(i % bhs_per_rec)) {
1168                         tbh = bhs[i];
1169                         if (!tbh)
1170                                 continue;
1171                         post_write_mst_fixup((NTFS_RECORD*)(kaddr +
1172                                         bh_offset(tbh)));
1173                 }
1174         }
1175         flush_dcache_page(page);
1176 unm_done:
1177         /* Unlock any locked inodes. */
1178         while (nr_locked_nis-- > 0) {
1179                 ntfs_inode *tni, *base_tni;
1180                 
1181                 tni = locked_nis[nr_locked_nis];
1182                 /* Get the base inode. */
1183                 down(&tni->extent_lock);
1184                 if (tni->nr_extents >= 0)
1185                         base_tni = tni;
1186                 else {
1187                         base_tni = tni->ext.base_ntfs_ino;
1188                         BUG_ON(!base_tni);
1189                 }
1190                 up(&tni->extent_lock);
1191                 ntfs_debug("Unlocking %s inode 0x%lx.",
1192                                 tni == base_tni ? "base" : "extent",
1193                                 tni->mft_no);
1194                 up(&tni->mrec_lock);
1195                 atomic_dec(&tni->count);
1196                 iput(VFS_I(base_tni));
1197         }
1198         SetPageUptodate(page);
1199         kunmap(page);
1200 done:
1201         if (unlikely(err && err != -ENOMEM)) {
1202                 /*
1203                  * Set page error if there is only one ntfs record in the page.
1204                  * Otherwise we would loose per-record granularity.
1205                  */
1206                 if (ni->itype.index.block_size == PAGE_CACHE_SIZE)
1207                         SetPageError(page);
1208                 NVolSetErrors(vol);
1209         }
1210         if (page_is_dirty) {
1211                 ntfs_debug("Page still contains one or more dirty ntfs "
1212                                 "records.  Redirtying the page starting at "
1213                                 "record 0x%lx.", page->index <<
1214                                 (PAGE_CACHE_SHIFT - rec_size_bits));
1215                 redirty_page_for_writepage(wbc, page);
1216                 unlock_page(page);
1217         } else {
1218                 /*
1219                  * Keep the VM happy.  This must be done otherwise the
1220                  * radix-tree tag PAGECACHE_TAG_DIRTY remains set even though
1221                  * the page is clean.
1222                  */
1223                 BUG_ON(PageWriteback(page));
1224                 set_page_writeback(page);
1225                 unlock_page(page);
1226                 end_page_writeback(page);
1227         }
1228         if (likely(!err))
1229                 ntfs_debug("Done.");
1230         return err;
1231 }
1232
1233 /**
1234  * ntfs_writepage - write a @page to the backing store
1235  * @page:       page cache page to write out
1236  * @wbc:        writeback control structure
1237  *
1238  * This is called from the VM when it wants to have a dirty ntfs page cache
1239  * page cleaned.  The VM has already locked the page and marked it clean.
1240  *
1241  * For non-resident attributes, ntfs_writepage() writes the @page by calling
1242  * the ntfs version of the generic block_write_full_page() function,
1243  * ntfs_write_block(), which in turn if necessary creates and writes the
1244  * buffers associated with the page asynchronously.
1245  *
1246  * For resident attributes, OTOH, ntfs_writepage() writes the @page by copying
1247  * the data to the mft record (which at this stage is most likely in memory).
1248  * The mft record is then marked dirty and written out asynchronously via the
1249  * vfs inode dirty code path for the inode the mft record belongs to or via the
1250  * vm page dirty code path for the page the mft record is in.
1251  *
1252  * Based on ntfs_readpage() and fs/buffer.c::block_write_full_page().
1253  *
1254  * Return 0 on success and -errno on error.
1255  */
1256 static int ntfs_writepage(struct page *page, struct writeback_control *wbc)
1257 {
1258         loff_t i_size;
1259         struct inode *vi = page->mapping->host;
1260         ntfs_inode *base_ni = NULL, *ni = NTFS_I(vi);
1261         char *kaddr;
1262         ntfs_attr_search_ctx *ctx = NULL;
1263         MFT_RECORD *m = NULL;
1264         u32 attr_len;
1265         int err;
1266
1267 retry_writepage:
1268         BUG_ON(!PageLocked(page));
1269         i_size = i_size_read(vi);
1270         /* Is the page fully outside i_size? (truncate in progress) */
1271         if (unlikely(page->index >= (i_size + PAGE_CACHE_SIZE - 1) >>
1272                         PAGE_CACHE_SHIFT)) {
1273                 /*
1274                  * The page may have dirty, unmapped buffers.  Make them
1275                  * freeable here, so the page does not leak.
1276                  */
1277                 block_invalidatepage(page, 0);
1278                 unlock_page(page);
1279                 ntfs_debug("Write outside i_size - truncated?");
1280                 return 0;
1281         }
1282         /* NInoNonResident() == NInoIndexAllocPresent() */
1283         if (NInoNonResident(ni)) {
1284                 /*
1285                  * Only unnamed $DATA attributes can be compressed, encrypted,
1286                  * and/or sparse.
1287                  */
1288                 if (ni->type == AT_DATA && !ni->name_len) {
1289                         /* If file is encrypted, deny access, just like NT4. */
1290                         if (NInoEncrypted(ni)) {
1291                                 unlock_page(page);
1292                                 ntfs_debug("Denying write access to encrypted "
1293                                                 "file.");
1294                                 return -EACCES;
1295                         }
1296                         /* Compressed data streams are handled in compress.c. */
1297                         if (NInoCompressed(ni)) {
1298                                 // TODO: Implement and replace this check with
1299                                 // return ntfs_write_compressed_block(page);
1300                                 unlock_page(page);
1301                                 ntfs_error(vi->i_sb, "Writing to compressed "
1302                                                 "files is not supported yet. "
1303                                                 "Sorry.");
1304                                 return -EOPNOTSUPP;
1305                         }
1306                         // TODO: Implement and remove this check.
1307                         if (NInoSparse(ni)) {
1308                                 unlock_page(page);
1309                                 ntfs_error(vi->i_sb, "Writing to sparse files "
1310                                                 "is not supported yet. Sorry.");
1311                                 return -EOPNOTSUPP;
1312                         }
1313                 }
1314                 /* We have to zero every time due to mmap-at-end-of-file. */
1315                 if (page->index >= (i_size >> PAGE_CACHE_SHIFT)) {
1316                         /* The page straddles i_size. */
1317                         unsigned int ofs = i_size & ~PAGE_CACHE_MASK;
1318                         kaddr = kmap_atomic(page, KM_USER0);
1319                         memset(kaddr + ofs, 0, PAGE_CACHE_SIZE - ofs);
1320                         flush_dcache_page(page);
1321                         kunmap_atomic(kaddr, KM_USER0);
1322                 }
1323                 /* Handle mst protected attributes. */
1324                 if (NInoMstProtected(ni))
1325                         return ntfs_write_mst_block(page, wbc);
1326                 /* Normal data stream. */
1327                 return ntfs_write_block(page, wbc);
1328         }
1329         /*
1330          * Attribute is resident, implying it is not compressed, encrypted,
1331          * sparse, or mst protected.  This also means the attribute is smaller
1332          * than an mft record and hence smaller than a page, so can simply
1333          * return error on any pages with index above 0.
1334          */
1335         BUG_ON(page_has_buffers(page));
1336         BUG_ON(!PageUptodate(page));
1337         if (unlikely(page->index > 0)) {
1338                 ntfs_error(vi->i_sb, "BUG()! page->index (0x%lx) > 0.  "
1339                                 "Aborting write.", page->index);
1340                 BUG_ON(PageWriteback(page));
1341                 set_page_writeback(page);
1342                 unlock_page(page);
1343                 end_page_writeback(page);
1344                 return -EIO;
1345         }
1346         if (!NInoAttr(ni))
1347                 base_ni = ni;
1348         else
1349                 base_ni = ni->ext.base_ntfs_ino;
1350         /* Map, pin, and lock the mft record. */
1351         m = map_mft_record(base_ni);
1352         if (IS_ERR(m)) {
1353                 err = PTR_ERR(m);
1354                 m = NULL;
1355                 ctx = NULL;
1356                 goto err_out;
1357         }
1358         /*
1359          * If a parallel write made the attribute non-resident, drop the mft
1360          * record and retry the writepage.
1361          */
1362         if (unlikely(NInoNonResident(ni))) {
1363                 unmap_mft_record(base_ni);
1364                 goto retry_writepage;
1365         }
1366         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1367         if (unlikely(!ctx)) {
1368                 err = -ENOMEM;
1369                 goto err_out;
1370         }
1371         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1372                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1373         if (unlikely(err))
1374                 goto err_out;
1375         /*
1376          * Keep the VM happy.  This must be done otherwise the radix-tree tag
1377          * PAGECACHE_TAG_DIRTY remains set even though the page is clean.
1378          */
1379         BUG_ON(PageWriteback(page));
1380         set_page_writeback(page);
1381         unlock_page(page);
1382
1383         /*
1384          * Here, we don't need to zero the out of bounds area everytime because
1385          * the below memcpy() already takes care of the mmap-at-end-of-file
1386          * requirements. If the file is converted to a non-resident one, then
1387          * the code path use is switched to the non-resident one where the
1388          * zeroing happens on each ntfs_writepage() invocation.
1389          *
1390          * The above also applies nicely when i_size is decreased.
1391          *
1392          * When i_size is increased, the memory between the old and new i_size
1393          * _must_ be zeroed (or overwritten with new data). Otherwise we will
1394          * expose data to userspace/disk which should never have been exposed.
1395          *
1396          * FIXME: Ensure that i_size increases do the zeroing/overwriting and
1397          * if we cannot guarantee that, then enable the zeroing below.  If the
1398          * zeroing below is enabled, we MUST move the unlock_page() from above
1399          * to after the kunmap_atomic(), i.e. just before the
1400          * end_page_writeback().
1401          * UPDATE: ntfs_prepare/commit_write() do the zeroing on i_size
1402          * increases for resident attributes so those are ok.
1403          * TODO: ntfs_truncate(), others?
1404          */
1405
1406         attr_len = le32_to_cpu(ctx->attr->data.resident.value_length);
1407         i_size = i_size_read(vi);
1408         if (unlikely(attr_len > i_size)) {
1409                 attr_len = i_size;
1410                 ctx->attr->data.resident.value_length = cpu_to_le32(attr_len);
1411         }
1412         kaddr = kmap_atomic(page, KM_USER0);
1413         /* Copy the data from the page to the mft record. */
1414         memcpy((u8*)ctx->attr +
1415                         le16_to_cpu(ctx->attr->data.resident.value_offset),
1416                         kaddr, attr_len);
1417         flush_dcache_mft_record_page(ctx->ntfs_ino);
1418         /* Zero out of bounds area in the page cache page. */
1419         memset(kaddr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
1420         flush_dcache_page(page);
1421         kunmap_atomic(kaddr, KM_USER0);
1422
1423         end_page_writeback(page);
1424
1425         /* Mark the mft record dirty, so it gets written back. */
1426         mark_mft_record_dirty(ctx->ntfs_ino);
1427         ntfs_attr_put_search_ctx(ctx);
1428         unmap_mft_record(base_ni);
1429         return 0;
1430 err_out:
1431         if (err == -ENOMEM) {
1432                 ntfs_warning(vi->i_sb, "Error allocating memory. Redirtying "
1433                                 "page so we try again later.");
1434                 /*
1435                  * Put the page back on mapping->dirty_pages, but leave its
1436                  * buffers' dirty state as-is.
1437                  */
1438                 redirty_page_for_writepage(wbc, page);
1439                 err = 0;
1440         } else {
1441                 ntfs_error(vi->i_sb, "Resident attribute write failed with "
1442                                 "error %i.", err);
1443                 SetPageError(page);
1444                 NVolSetErrors(ni->vol);
1445                 make_bad_inode(vi);
1446         }
1447         unlock_page(page);
1448         if (ctx)
1449                 ntfs_attr_put_search_ctx(ctx);
1450         if (m)
1451                 unmap_mft_record(base_ni);
1452         return err;
1453 }
1454
1455 /**
1456  * ntfs_prepare_nonresident_write -
1457  *
1458  */
1459 static int ntfs_prepare_nonresident_write(struct page *page,
1460                 unsigned from, unsigned to)
1461 {
1462         VCN vcn;
1463         LCN lcn;
1464         s64 initialized_size;
1465         loff_t i_size;
1466         sector_t block, ablock, iblock;
1467         struct inode *vi;
1468         ntfs_inode *ni;
1469         ntfs_volume *vol;
1470         runlist_element *rl;
1471         struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
1472         unsigned long flags;
1473         unsigned int vcn_ofs, block_start, block_end, blocksize;
1474         int err;
1475         BOOL is_retry;
1476         unsigned char blocksize_bits;
1477
1478         vi = page->mapping->host;
1479         ni = NTFS_I(vi);
1480         vol = ni->vol;
1481
1482         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
1483                         "0x%lx, from = %u, to = %u.", ni->mft_no, ni->type,
1484                         page->index, from, to);
1485
1486         BUG_ON(!NInoNonResident(ni));
1487
1488         blocksize_bits = vi->i_blkbits;
1489         blocksize = 1 << blocksize_bits;
1490
1491         /*
1492          * create_empty_buffers() will create uptodate/dirty buffers if the
1493          * page is uptodate/dirty.
1494          */
1495         if (!page_has_buffers(page))
1496                 create_empty_buffers(page, blocksize, 0);
1497         bh = head = page_buffers(page);
1498         if (unlikely(!bh))
1499                 return -ENOMEM;
1500
1501         /* The first block in the page. */
1502         block = (s64)page->index << (PAGE_CACHE_SHIFT - blocksize_bits);
1503
1504         read_lock_irqsave(&ni->size_lock, flags);
1505         /*
1506          * The first out of bounds block for the allocated size.  No need to
1507          * round up as allocated_size is in multiples of cluster size and the
1508          * minimum cluster size is 512 bytes, which is equal to the smallest
1509          * blocksize.
1510          */
1511         ablock = ni->allocated_size >> blocksize_bits;
1512         i_size = i_size_read(vi);
1513         initialized_size = ni->initialized_size;
1514         read_unlock_irqrestore(&ni->size_lock, flags);
1515
1516         /* The last (fully or partially) initialized block. */
1517         iblock = initialized_size >> blocksize_bits;
1518
1519         /* Loop through all the buffers in the page. */
1520         block_start = 0;
1521         rl = NULL;
1522         err = 0;
1523         do {
1524                 block_end = block_start + blocksize;
1525                 /*
1526                  * If buffer @bh is outside the write, just mark it uptodate
1527                  * if the page is uptodate and continue with the next buffer.
1528                  */
1529                 if (block_end <= from || block_start >= to) {
1530                         if (PageUptodate(page)) {
1531                                 if (!buffer_uptodate(bh))
1532                                         set_buffer_uptodate(bh);
1533                         }
1534                         continue;
1535                 }
1536                 /*
1537                  * @bh is at least partially being written to.
1538                  * Make sure it is not marked as new.
1539                  */
1540                 //if (buffer_new(bh))
1541                 //      clear_buffer_new(bh);
1542
1543                 if (block >= ablock) {
1544                         // TODO: block is above allocated_size, need to
1545                         // allocate it. Best done in one go to accommodate not
1546                         // only block but all above blocks up to and including:
1547                         // ((page->index << PAGE_CACHE_SHIFT) + to + blocksize
1548                         // - 1) >> blobksize_bits. Obviously will need to round
1549                         // up to next cluster boundary, too. This should be
1550                         // done with a helper function, so it can be reused.
1551                         ntfs_error(vol->sb, "Writing beyond allocated size "
1552                                         "is not supported yet. Sorry.");
1553                         err = -EOPNOTSUPP;
1554                         goto err_out;
1555                         // Need to update ablock.
1556                         // Need to set_buffer_new() on all block bhs that are
1557                         // newly allocated.
1558                 }
1559                 /*
1560                  * Now we have enough allocated size to fulfill the whole
1561                  * request, i.e. block < ablock is true.
1562                  */
1563                 if (unlikely((block >= iblock) &&
1564                                 (initialized_size < i_size))) {
1565                         /*
1566                          * If this page is fully outside initialized size, zero
1567                          * out all pages between the current initialized size
1568                          * and the current page. Just use ntfs_readpage() to do
1569                          * the zeroing transparently.
1570                          */
1571                         if (block > iblock) {
1572                                 // TODO:
1573                                 // For each page do:
1574                                 // - read_cache_page()
1575                                 // Again for each page do:
1576                                 // - wait_on_page_locked()
1577                                 // - Check (PageUptodate(page) &&
1578                                 //                      !PageError(page))
1579                                 // Update initialized size in the attribute and
1580                                 // in the inode.
1581                                 // Again, for each page do:
1582                                 //      __set_page_dirty_buffers();
1583                                 // page_cache_release()
1584                                 // We don't need to wait on the writes.
1585                                 // Update iblock.
1586                         }
1587                         /*
1588                          * The current page straddles initialized size. Zero
1589                          * all non-uptodate buffers and set them uptodate (and
1590                          * dirty?). Note, there aren't any non-uptodate buffers
1591                          * if the page is uptodate.
1592                          * FIXME: For an uptodate page, the buffers may need to
1593                          * be written out because they were not initialized on
1594                          * disk before.
1595                          */
1596                         if (!PageUptodate(page)) {
1597                                 // TODO:
1598                                 // Zero any non-uptodate buffers up to i_size.
1599                                 // Set them uptodate and dirty.
1600                         }
1601                         // TODO:
1602                         // Update initialized size in the attribute and in the
1603                         // inode (up to i_size).
1604                         // Update iblock.
1605                         // FIXME: This is inefficient. Try to batch the two
1606                         // size changes to happen in one go.
1607                         ntfs_error(vol->sb, "Writing beyond initialized size "
1608                                         "is not supported yet. Sorry.");
1609                         err = -EOPNOTSUPP;
1610                         goto err_out;
1611                         // Do NOT set_buffer_new() BUT DO clear buffer range
1612                         // outside write request range.
1613                         // set_buffer_uptodate() on complete buffers as well as
1614                         // set_buffer_dirty().
1615                 }
1616
1617                 /* Need to map unmapped buffers. */
1618                 if (!buffer_mapped(bh)) {
1619                         /* Unmapped buffer. Need to map it. */
1620                         bh->b_bdev = vol->sb->s_bdev;
1621
1622                         /* Convert block into corresponding vcn and offset. */
1623                         vcn = (VCN)block << blocksize_bits >>
1624                                         vol->cluster_size_bits;
1625                         vcn_ofs = ((VCN)block << blocksize_bits) &
1626                                         vol->cluster_size_mask;
1627
1628                         is_retry = FALSE;
1629                         if (!rl) {
1630 lock_retry_remap:
1631                                 down_read(&ni->runlist.lock);
1632                                 rl = ni->runlist.rl;
1633                         }
1634                         if (likely(rl != NULL)) {
1635                                 /* Seek to element containing target vcn. */
1636                                 while (rl->length && rl[1].vcn <= vcn)
1637                                         rl++;
1638                                 lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
1639                         } else
1640                                 lcn = LCN_RL_NOT_MAPPED;
1641                         if (unlikely(lcn < 0)) {
1642                                 /*
1643                                  * We extended the attribute allocation above.
1644                                  * If we hit an ENOENT here it means that the
1645                                  * allocation was insufficient which is a bug.
1646                                  */
1647                                 BUG_ON(lcn == LCN_ENOENT);
1648
1649                                 /* It is a hole, need to instantiate it. */
1650                                 if (lcn == LCN_HOLE) {
1651                                         // TODO: Instantiate the hole.
1652                                         // clear_buffer_new(bh);
1653                                         // unmap_underlying_metadata(bh->b_bdev,
1654                                         //              bh->b_blocknr);
1655                                         // For non-uptodate buffers, need to
1656                                         // zero out the region outside the
1657                                         // request in this bh or all bhs,
1658                                         // depending on what we implemented
1659                                         // above.
1660                                         // Need to flush_dcache_page().
1661                                         // Or could use set_buffer_new()
1662                                         // instead?
1663                                         ntfs_error(vol->sb, "Writing into "
1664                                                         "sparse regions is "
1665                                                         "not supported yet. "
1666                                                         "Sorry.");
1667                                         err = -EOPNOTSUPP;
1668                                         if (!rl)
1669                                                 up_read(&ni->runlist.lock);
1670                                         goto err_out;
1671                                 } else if (!is_retry &&
1672                                                 lcn == LCN_RL_NOT_MAPPED) {
1673                                         is_retry = TRUE;
1674                                         /*
1675                                          * Attempt to map runlist, dropping
1676                                          * lock for the duration.
1677                                          */
1678                                         up_read(&ni->runlist.lock);
1679                                         err = ntfs_map_runlist(ni, vcn);
1680                                         if (likely(!err))
1681                                                 goto lock_retry_remap;
1682                                         rl = NULL;
1683                                         lcn = err;
1684                                 } else if (!rl)
1685                                         up_read(&ni->runlist.lock);
1686                                 /*
1687                                  * Failed to map the buffer, even after
1688                                  * retrying.
1689                                  */
1690                                 bh->b_blocknr = -1;
1691                                 ntfs_error(vol->sb, "Failed to write to inode "
1692                                                 "0x%lx, attribute type 0x%x, "
1693                                                 "vcn 0x%llx, offset 0x%x "
1694                                                 "because its location on disk "
1695                                                 "could not be determined%s "
1696                                                 "(error code %lli).",
1697                                                 ni->mft_no, ni->type,
1698                                                 (unsigned long long)vcn,
1699                                                 vcn_ofs, is_retry ? " even "
1700                                                 "after retrying" : "",
1701                                                 (long long)lcn);
1702                                 if (!err)
1703                                         err = -EIO;
1704                                 goto err_out;
1705                         }
1706                         /* We now have a successful remap, i.e. lcn >= 0. */
1707
1708                         /* Setup buffer head to correct block. */
1709                         bh->b_blocknr = ((lcn << vol->cluster_size_bits)
1710                                         + vcn_ofs) >> blocksize_bits;
1711                         set_buffer_mapped(bh);
1712
1713                         // FIXME: Something analogous to this is needed for
1714                         // each newly allocated block, i.e. BH_New.
1715                         // FIXME: Might need to take this out of the
1716                         // if (!buffer_mapped(bh)) {}, depending on how we
1717                         // implement things during the allocated_size and
1718                         // initialized_size extension code above.
1719                         if (buffer_new(bh)) {
1720                                 clear_buffer_new(bh);
1721                                 unmap_underlying_metadata(bh->b_bdev,
1722                                                 bh->b_blocknr);
1723                                 if (PageUptodate(page)) {
1724                                         set_buffer_uptodate(bh);
1725                                         continue;
1726                                 }
1727                                 /*
1728                                  * Page is _not_ uptodate, zero surrounding
1729                                  * region. NOTE: This is how we decide if to
1730                                  * zero or not!
1731                                  */
1732                                 if (block_end > to || block_start < from) {
1733                                         void *kaddr;
1734
1735                                         kaddr = kmap_atomic(page, KM_USER0);
1736                                         if (block_end > to)
1737                                                 memset(kaddr + to, 0,
1738                                                                 block_end - to);
1739                                         if (block_start < from)
1740                                                 memset(kaddr + block_start, 0,
1741                                                                 from -
1742                                                                 block_start);
1743                                         flush_dcache_page(page);
1744                                         kunmap_atomic(kaddr, KM_USER0);
1745                                 }
1746                                 continue;
1747                         }
1748                 }
1749                 /* @bh is mapped, set it uptodate if the page is uptodate. */
1750                 if (PageUptodate(page)) {
1751                         if (!buffer_uptodate(bh))
1752                                 set_buffer_uptodate(bh);
1753                         continue;
1754                 }
1755                 /*
1756                  * The page is not uptodate. The buffer is mapped. If it is not
1757                  * uptodate, and it is only partially being written to, we need
1758                  * to read the buffer in before the write, i.e. right now.
1759                  */
1760                 if (!buffer_uptodate(bh) &&
1761                                 (block_start < from || block_end > to)) {
1762                         ll_rw_block(READ, 1, &bh);
1763                         *wait_bh++ = bh;
1764                 }
1765         } while (block++, block_start = block_end,
1766                         (bh = bh->b_this_page) != head);
1767
1768         /* Release the lock if we took it. */
1769         if (rl) {
1770                 up_read(&ni->runlist.lock);
1771                 rl = NULL;
1772         }
1773
1774         /* If we issued read requests, let them complete. */
1775         while (wait_bh > wait) {
1776                 wait_on_buffer(*--wait_bh);
1777                 if (!buffer_uptodate(*wait_bh))
1778                         return -EIO;
1779         }
1780
1781         ntfs_debug("Done.");
1782         return 0;
1783 err_out:
1784         /*
1785          * Zero out any newly allocated blocks to avoid exposing stale data.
1786          * If BH_New is set, we know that the block was newly allocated in the
1787          * above loop.
1788          * FIXME: What about initialized_size increments? Have we done all the
1789          * required zeroing above? If not this error handling is broken, and
1790          * in particular the if (block_end <= from) check is completely bogus.
1791          */
1792         bh = head;
1793         block_start = 0;
1794         is_retry = FALSE;
1795         do {
1796                 block_end = block_start + blocksize;
1797                 if (block_end <= from)
1798                         continue;
1799                 if (block_start >= to)
1800                         break;
1801                 if (buffer_new(bh)) {
1802                         void *kaddr;
1803
1804                         clear_buffer_new(bh);
1805                         kaddr = kmap_atomic(page, KM_USER0);
1806                         memset(kaddr + block_start, 0, bh->b_size);
1807                         kunmap_atomic(kaddr, KM_USER0);
1808                         set_buffer_uptodate(bh);
1809                         mark_buffer_dirty(bh);
1810                         is_retry = TRUE;
1811                 }
1812         } while (block_start = block_end, (bh = bh->b_this_page) != head);
1813         if (is_retry)
1814                 flush_dcache_page(page);
1815         if (rl)
1816                 up_read(&ni->runlist.lock);
1817         return err;
1818 }
1819
1820 /**
1821  * ntfs_prepare_write - prepare a page for receiving data
1822  *
1823  * This is called from generic_file_write() with i_sem held on the inode
1824  * (@page->mapping->host).  The @page is locked but not kmap()ped.  The source
1825  * data has not yet been copied into the @page.
1826  *
1827  * Need to extend the attribute/fill in holes if necessary, create blocks and
1828  * make partially overwritten blocks uptodate,
1829  *
1830  * i_size is not to be modified yet.
1831  *
1832  * Return 0 on success or -errno on error.
1833  *
1834  * Should be using block_prepare_write() [support for sparse files] or
1835  * cont_prepare_write() [no support for sparse files].  Cannot do that due to
1836  * ntfs specifics but can look at them for implementation guidance.
1837  *
1838  * Note: In the range, @from is inclusive and @to is exclusive, i.e. @from is
1839  * the first byte in the page that will be written to and @to is the first byte
1840  * after the last byte that will be written to.
1841  */
1842 static int ntfs_prepare_write(struct file *file, struct page *page,
1843                 unsigned from, unsigned to)
1844 {
1845         s64 new_size;
1846         loff_t i_size;
1847         struct inode *vi = page->mapping->host;
1848         ntfs_inode *base_ni = NULL, *ni = NTFS_I(vi);
1849         ntfs_volume *vol = ni->vol;
1850         ntfs_attr_search_ctx *ctx = NULL;
1851         MFT_RECORD *m = NULL;
1852         ATTR_RECORD *a;
1853         u8 *kaddr;
1854         u32 attr_len;
1855         int err;
1856
1857         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
1858                         "0x%lx, from = %u, to = %u.", vi->i_ino, ni->type,
1859                         page->index, from, to);
1860         BUG_ON(!PageLocked(page));
1861         BUG_ON(from > PAGE_CACHE_SIZE);
1862         BUG_ON(to > PAGE_CACHE_SIZE);
1863         BUG_ON(from > to);
1864         BUG_ON(NInoMstProtected(ni));
1865         /*
1866          * If a previous ntfs_truncate() failed, repeat it and abort if it
1867          * fails again.
1868          */
1869         if (unlikely(NInoTruncateFailed(ni))) {
1870                 down_write(&vi->i_alloc_sem);
1871                 err = ntfs_truncate(vi);
1872                 up_write(&vi->i_alloc_sem);
1873                 if (err || NInoTruncateFailed(ni)) {
1874                         if (!err)
1875                                 err = -EIO;
1876                         goto err_out;
1877                 }
1878         }
1879         /* If the attribute is not resident, deal with it elsewhere. */
1880         if (NInoNonResident(ni)) {
1881                 /*
1882                  * Only unnamed $DATA attributes can be compressed, encrypted,
1883                  * and/or sparse.
1884                  */
1885                 if (ni->type == AT_DATA && !ni->name_len) {
1886                         /* If file is encrypted, deny access, just like NT4. */
1887                         if (NInoEncrypted(ni)) {
1888                                 ntfs_debug("Denying write access to encrypted "
1889                                                 "file.");
1890                                 return -EACCES;
1891                         }
1892                         /* Compressed data streams are handled in compress.c. */
1893                         if (NInoCompressed(ni)) {
1894                                 // TODO: Implement and replace this check with
1895                                 // return ntfs_write_compressed_block(page);
1896                                 ntfs_error(vi->i_sb, "Writing to compressed "
1897                                                 "files is not supported yet. "
1898                                                 "Sorry.");
1899                                 return -EOPNOTSUPP;
1900                         }
1901                         // TODO: Implement and remove this check.
1902                         if (NInoSparse(ni)) {
1903                                 ntfs_error(vi->i_sb, "Writing to sparse files "
1904                                                 "is not supported yet. Sorry.");
1905                                 return -EOPNOTSUPP;
1906                         }
1907                 }
1908                 /* Normal data stream. */
1909                 return ntfs_prepare_nonresident_write(page, from, to);
1910         }
1911         /*
1912          * Attribute is resident, implying it is not compressed, encrypted, or
1913          * sparse.
1914          */
1915         BUG_ON(page_has_buffers(page));
1916         new_size = ((s64)page->index << PAGE_CACHE_SHIFT) + to;
1917         /* If we do not need to resize the attribute allocation we are done. */
1918         if (new_size <= i_size_read(vi))
1919                 goto done;
1920         /* Map, pin, and lock the (base) mft record. */
1921         if (!NInoAttr(ni))
1922                 base_ni = ni;
1923         else
1924                 base_ni = ni->ext.base_ntfs_ino;
1925         m = map_mft_record(base_ni);
1926         if (IS_ERR(m)) {
1927                 err = PTR_ERR(m);
1928                 m = NULL;
1929                 ctx = NULL;
1930                 goto err_out;
1931         }
1932         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1933         if (unlikely(!ctx)) {
1934                 err = -ENOMEM;
1935                 goto err_out;
1936         }
1937         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1938                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1939         if (unlikely(err)) {
1940                 if (err == -ENOENT)
1941                         err = -EIO;
1942                 goto err_out;
1943         }
1944         m = ctx->mrec;
1945         a = ctx->attr;
1946         /* The total length of the attribute value. */
1947         attr_len = le32_to_cpu(a->data.resident.value_length);
1948         /* Fix an eventual previous failure of ntfs_commit_write(). */
1949         i_size = i_size_read(vi);
1950         if (unlikely(attr_len > i_size)) {
1951                 attr_len = i_size;
1952                 a->data.resident.value_length = cpu_to_le32(attr_len);
1953         }
1954         /* If we do not need to resize the attribute allocation we are done. */
1955         if (new_size <= attr_len)
1956                 goto done_unm;
1957         /* Check if new size is allowed in $AttrDef. */
1958         err = ntfs_attr_size_bounds_check(vol, ni->type, new_size);
1959         if (unlikely(err)) {
1960                 if (err == -ERANGE) {
1961                         ntfs_error(vol->sb, "Write would cause the inode "
1962                                         "0x%lx to exceed the maximum size for "
1963                                         "its attribute type (0x%x).  Aborting "
1964                                         "write.", vi->i_ino,
1965                                         le32_to_cpu(ni->type));
1966                 } else {
1967                         ntfs_error(vol->sb, "Inode 0x%lx has unknown "
1968                                         "attribute type 0x%x.  Aborting "
1969                                         "write.", vi->i_ino,
1970                                         le32_to_cpu(ni->type));
1971                         err = -EIO;
1972                 }
1973                 goto err_out2;
1974         }
1975         /*
1976          * Extend the attribute record to be able to store the new attribute
1977          * size.
1978          */
1979         if (new_size >= vol->mft_record_size || ntfs_attr_record_resize(m, a,
1980                         le16_to_cpu(a->data.resident.value_offset) +
1981                         new_size)) {
1982                 /* Not enough space in the mft record. */
1983                 ntfs_error(vol->sb, "Not enough space in the mft record for "
1984                                 "the resized attribute value.  This is not "
1985                                 "supported yet.  Aborting write.");
1986                 err = -EOPNOTSUPP;
1987                 goto err_out2;
1988         }
1989         /*
1990          * We have enough space in the mft record to fit the write.  This
1991          * implies the attribute is smaller than the mft record and hence the
1992          * attribute must be in a single page and hence page->index must be 0.
1993          */
1994         BUG_ON(page->index);
1995         /*
1996          * If the beginning of the write is past the old size, enlarge the
1997          * attribute value up to the beginning of the write and fill it with
1998          * zeroes.
1999          */
2000         if (from > attr_len) {
2001                 memset((u8*)a + le16_to_cpu(a->data.resident.value_offset) +
2002                                 attr_len, 0, from - attr_len);
2003                 a->data.resident.value_length = cpu_to_le32(from);
2004                 /* Zero the corresponding area in the page as well. */
2005                 if (PageUptodate(page)) {
2006                         kaddr = kmap_atomic(page, KM_USER0);
2007                         memset(kaddr + attr_len, 0, from - attr_len);
2008                         kunmap_atomic(kaddr, KM_USER0);
2009                         flush_dcache_page(page);
2010                 }
2011         }
2012         flush_dcache_mft_record_page(ctx->ntfs_ino);
2013         mark_mft_record_dirty(ctx->ntfs_ino);
2014 done_unm:
2015         ntfs_attr_put_search_ctx(ctx);
2016         unmap_mft_record(base_ni);
2017         /*
2018          * Because resident attributes are handled by memcpy() to/from the
2019          * corresponding MFT record, and because this form of i/o is byte
2020          * aligned rather than block aligned, there is no need to bring the
2021          * page uptodate here as in the non-resident case where we need to
2022          * bring the buffers straddled by the write uptodate before
2023          * generic_file_write() does the copying from userspace.
2024          *
2025          * We thus defer the uptodate bringing of the page region outside the
2026          * region written to to ntfs_commit_write(), which makes the code
2027          * simpler and saves one atomic kmap which is good.
2028          */
2029 done:
2030         ntfs_debug("Done.");
2031         return 0;
2032 err_out:
2033         if (err == -ENOMEM)
2034                 ntfs_warning(vi->i_sb, "Error allocating memory required to "
2035                                 "prepare the write.");
2036         else {
2037                 ntfs_error(vi->i_sb, "Resident attribute prepare write failed "
2038                                 "with error %i.", err);
2039                 NVolSetErrors(vol);
2040                 make_bad_inode(vi);
2041         }
2042 err_out2:
2043         if (ctx)
2044                 ntfs_attr_put_search_ctx(ctx);
2045         if (m)
2046                 unmap_mft_record(base_ni);
2047         return err;
2048 }
2049
2050 /**
2051  * ntfs_commit_nonresident_write -
2052  *
2053  */
2054 static int ntfs_commit_nonresident_write(struct page *page,
2055                 unsigned from, unsigned to)
2056 {
2057         s64 pos = ((s64)page->index << PAGE_CACHE_SHIFT) + to;
2058         struct inode *vi = page->mapping->host;
2059         struct buffer_head *bh, *head;
2060         unsigned int block_start, block_end, blocksize;
2061         BOOL partial;
2062
2063         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
2064                         "0x%lx, from = %u, to = %u.", vi->i_ino,
2065                         NTFS_I(vi)->type, page->index, from, to);
2066         blocksize = 1 << vi->i_blkbits;
2067
2068         // FIXME: We need a whole slew of special cases in here for compressed
2069         // files for example...
2070         // For now, we know ntfs_prepare_write() would have failed so we can't
2071         // get here in any of the cases which we have to special case, so we
2072         // are just a ripped off, unrolled generic_commit_write().
2073
2074         bh = head = page_buffers(page);
2075         block_start = 0;
2076         partial = FALSE;
2077         do {
2078                 block_end = block_start + blocksize;
2079                 if (block_end <= from || block_start >= to) {
2080                         if (!buffer_uptodate(bh))
2081                                 partial = TRUE;
2082                 } else {
2083                         set_buffer_uptodate(bh);
2084                         mark_buffer_dirty(bh);
2085                 }
2086         } while (block_start = block_end, (bh = bh->b_this_page) != head);
2087         /*
2088          * If this is a partial write which happened to make all buffers
2089          * uptodate then we can optimize away a bogus ->readpage() for the next
2090          * read().  Here we 'discover' whether the page went uptodate as a
2091          * result of this (potentially partial) write.
2092          */
2093         if (!partial)
2094                 SetPageUptodate(page);
2095         /*
2096          * Not convinced about this at all.  See disparity comment above.  For
2097          * now we know ntfs_prepare_write() would have failed in the write
2098          * exceeds i_size case, so this will never trigger which is fine.
2099          */
2100         if (pos > i_size_read(vi)) {
2101                 ntfs_error(vi->i_sb, "Writing beyond the existing file size is "
2102                                 "not supported yet.  Sorry.");
2103                 return -EOPNOTSUPP;
2104                 // vi->i_size = pos;
2105                 // mark_inode_dirty(vi);
2106         }
2107         ntfs_debug("Done.");
2108         return 0;
2109 }
2110
2111 /**
2112  * ntfs_commit_write - commit the received data
2113  *
2114  * This is called from generic_file_write() with i_sem held on the inode
2115  * (@page->mapping->host).  The @page is locked but not kmap()ped.  The source
2116  * data has already been copied into the @page.  ntfs_prepare_write() has been
2117  * called before the data copied and it returned success so we can take the
2118  * results of various BUG checks and some error handling for granted.
2119  *
2120  * Need to mark modified blocks dirty so they get written out later when
2121  * ntfs_writepage() is invoked by the VM.
2122  *
2123  * Return 0 on success or -errno on error.
2124  *
2125  * Should be using generic_commit_write().  This marks buffers uptodate and
2126  * dirty, sets the page uptodate if all buffers in the page are uptodate, and
2127  * updates i_size if the end of io is beyond i_size.  In that case, it also
2128  * marks the inode dirty.
2129  *
2130  * Cannot use generic_commit_write() due to ntfs specialities but can look at
2131  * it for implementation guidance.
2132  *
2133  * If things have gone as outlined in ntfs_prepare_write(), then we do not
2134  * need to do any page content modifications here at all, except in the write
2135  * to resident attribute case, where we need to do the uptodate bringing here
2136  * which we combine with the copying into the mft record which means we save
2137  * one atomic kmap.
2138  */
2139 static int ntfs_commit_write(struct file *file, struct page *page,
2140                 unsigned from, unsigned to)
2141 {
2142         struct inode *vi = page->mapping->host;
2143         ntfs_inode *base_ni, *ni = NTFS_I(vi);
2144         char *kaddr, *kattr;
2145         ntfs_attr_search_ctx *ctx;
2146         MFT_RECORD *m;
2147         ATTR_RECORD *a;
2148         u32 attr_len;
2149         int err;
2150
2151         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
2152                         "0x%lx, from = %u, to = %u.", vi->i_ino, ni->type,
2153                         page->index, from, to);
2154         /* If the attribute is not resident, deal with it elsewhere. */
2155         if (NInoNonResident(ni)) {
2156                 /* Only unnamed $DATA attributes can be compressed/encrypted. */
2157                 if (ni->type == AT_DATA && !ni->name_len) {
2158                         /* Encrypted files need separate handling. */
2159                         if (NInoEncrypted(ni)) {
2160                                 // We never get here at present!
2161                                 BUG();
2162                         }
2163                         /* Compressed data streams are handled in compress.c. */
2164                         if (NInoCompressed(ni)) {
2165                                 // TODO: Implement this!
2166                                 // return ntfs_write_compressed_block(page);
2167                                 // We never get here at present!
2168                                 BUG();
2169                         }
2170                 }
2171                 /* Normal data stream. */
2172                 return ntfs_commit_nonresident_write(page, from, to);
2173         }
2174         /*
2175          * Attribute is resident, implying it is not compressed, encrypted, or
2176          * sparse.
2177          */
2178         if (!NInoAttr(ni))
2179                 base_ni = ni;
2180         else
2181                 base_ni = ni->ext.base_ntfs_ino;
2182         /* Map, pin, and lock the mft record. */
2183         m = map_mft_record(base_ni);
2184         if (IS_ERR(m)) {
2185                 err = PTR_ERR(m);
2186                 m = NULL;
2187                 ctx = NULL;
2188                 goto err_out;
2189         }
2190         ctx = ntfs_attr_get_search_ctx(base_ni, m);
2191         if (unlikely(!ctx)) {
2192                 err = -ENOMEM;
2193                 goto err_out;
2194         }
2195         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
2196                         CASE_SENSITIVE, 0, NULL, 0, ctx);
2197         if (unlikely(err)) {
2198                 if (err == -ENOENT)
2199                         err = -EIO;
2200                 goto err_out;
2201         }
2202         a = ctx->attr;
2203         /* The total length of the attribute value. */
2204         attr_len = le32_to_cpu(a->data.resident.value_length);
2205         BUG_ON(from > attr_len);
2206         kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
2207         kaddr = kmap_atomic(page, KM_USER0);
2208         /* Copy the received data from the page to the mft record. */
2209         memcpy(kattr + from, kaddr + from, to - from);
2210         /* Update the attribute length if necessary. */
2211         if (to > attr_len) {
2212                 attr_len = to;
2213                 a->data.resident.value_length = cpu_to_le32(attr_len);
2214         }
2215         /*
2216          * If the page is not uptodate, bring the out of bounds area(s)
2217          * uptodate by copying data from the mft record to the page.
2218          */
2219         if (!PageUptodate(page)) {
2220                 if (from > 0)
2221                         memcpy(kaddr, kattr, from);
2222                 if (to < attr_len)
2223                         memcpy(kaddr + to, kattr + to, attr_len - to);
2224                 /* Zero the region outside the end of the attribute value. */
2225                 if (attr_len < PAGE_CACHE_SIZE)
2226                         memset(kaddr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
2227                 /*
2228                  * The probability of not having done any of the above is
2229                  * extremely small, so we just flush unconditionally.
2230                  */
2231                 flush_dcache_page(page);
2232                 SetPageUptodate(page);
2233         }
2234         kunmap_atomic(kaddr, KM_USER0);
2235         /* Update i_size if necessary. */
2236         if (i_size_read(vi) < attr_len) {
2237                 unsigned long flags;
2238
2239                 write_lock_irqsave(&ni->size_lock, flags);
2240                 ni->allocated_size = ni->initialized_size = attr_len;
2241                 i_size_write(vi, attr_len);
2242                 write_unlock_irqrestore(&ni->size_lock, flags);
2243         }
2244         /* Mark the mft record dirty, so it gets written back. */
2245         flush_dcache_mft_record_page(ctx->ntfs_ino);
2246         mark_mft_record_dirty(ctx->ntfs_ino);
2247         ntfs_attr_put_search_ctx(ctx);
2248         unmap_mft_record(base_ni);
2249         ntfs_debug("Done.");
2250         return 0;
2251 err_out:
2252         if (err == -ENOMEM) {
2253                 ntfs_warning(vi->i_sb, "Error allocating memory required to "
2254                                 "commit the write.");
2255                 if (PageUptodate(page)) {
2256                         ntfs_warning(vi->i_sb, "Page is uptodate, setting "
2257                                         "dirty so the write will be retried "
2258                                         "later on by the VM.");
2259                         /*
2260                          * Put the page on mapping->dirty_pages, but leave its
2261                          * buffers' dirty state as-is.
2262                          */
2263                         __set_page_dirty_nobuffers(page);
2264                         err = 0;
2265                 } else
2266                         ntfs_error(vi->i_sb, "Page is not uptodate.  Written "
2267                                         "data has been lost.");
2268         } else {
2269                 ntfs_error(vi->i_sb, "Resident attribute commit write failed "
2270                                 "with error %i.", err);
2271                 NVolSetErrors(ni->vol);
2272                 make_bad_inode(vi);
2273         }
2274         if (ctx)
2275                 ntfs_attr_put_search_ctx(ctx);
2276         if (m)
2277                 unmap_mft_record(base_ni);
2278         return err;
2279 }
2280
2281 #endif  /* NTFS_RW */
2282
2283 /**
2284  * ntfs_aops - general address space operations for inodes and attributes
2285  */
2286 struct address_space_operations ntfs_aops = {
2287         .readpage       = ntfs_readpage,        /* Fill page with data. */
2288         .sync_page      = block_sync_page,      /* Currently, just unplugs the
2289                                                    disk request queue. */
2290 #ifdef NTFS_RW
2291         .writepage      = ntfs_writepage,       /* Write dirty page to disk. */
2292         .prepare_write  = ntfs_prepare_write,   /* Prepare page and buffers
2293                                                    ready to receive data. */
2294         .commit_write   = ntfs_commit_write,    /* Commit received data. */
2295 #endif /* NTFS_RW */
2296 };
2297
2298 /**
2299  * ntfs_mst_aops - general address space operations for mst protecteed inodes
2300  *                 and attributes
2301  */
2302 struct address_space_operations ntfs_mst_aops = {
2303         .readpage       = ntfs_readpage,        /* Fill page with data. */
2304         .sync_page      = block_sync_page,      /* Currently, just unplugs the
2305                                                    disk request queue. */
2306 #ifdef NTFS_RW
2307         .writepage      = ntfs_writepage,       /* Write dirty page to disk. */
2308         .set_page_dirty = __set_page_dirty_nobuffers,   /* Set the page dirty
2309                                                    without touching the buffers
2310                                                    belonging to the page. */
2311 #endif /* NTFS_RW */
2312 };
2313
2314 #ifdef NTFS_RW
2315
2316 /**
2317  * mark_ntfs_record_dirty - mark an ntfs record dirty
2318  * @page:       page containing the ntfs record to mark dirty
2319  * @ofs:        byte offset within @page at which the ntfs record begins
2320  *
2321  * Set the buffers and the page in which the ntfs record is located dirty.
2322  *
2323  * The latter also marks the vfs inode the ntfs record belongs to dirty
2324  * (I_DIRTY_PAGES only).
2325  *
2326  * If the page does not have buffers, we create them and set them uptodate.
2327  * The page may not be locked which is why we need to handle the buffers under
2328  * the mapping->private_lock.  Once the buffers are marked dirty we no longer
2329  * need the lock since try_to_free_buffers() does not free dirty buffers.
2330  */
2331 void mark_ntfs_record_dirty(struct page *page, const unsigned int ofs) {
2332         struct address_space *mapping = page->mapping;
2333         ntfs_inode *ni = NTFS_I(mapping->host);
2334         struct buffer_head *bh, *head, *buffers_to_free = NULL;
2335         unsigned int end, bh_size, bh_ofs;
2336
2337         BUG_ON(!PageUptodate(page));
2338         end = ofs + ni->itype.index.block_size;
2339         bh_size = 1 << VFS_I(ni)->i_blkbits;
2340         spin_lock(&mapping->private_lock);
2341         if (unlikely(!page_has_buffers(page))) {
2342                 spin_unlock(&mapping->private_lock);
2343                 bh = head = alloc_page_buffers(page, bh_size, 1);
2344                 spin_lock(&mapping->private_lock);
2345                 if (likely(!page_has_buffers(page))) {
2346                         struct buffer_head *tail;
2347
2348                         do {
2349                                 set_buffer_uptodate(bh);
2350                                 tail = bh;
2351                                 bh = bh->b_this_page;
2352                         } while (bh);
2353                         tail->b_this_page = head;
2354                         attach_page_buffers(page, head);
2355                 } else
2356                         buffers_to_free = bh;
2357         }
2358         bh = head = page_buffers(page);
2359         do {
2360                 bh_ofs = bh_offset(bh);
2361                 if (bh_ofs + bh_size <= ofs)
2362                         continue;
2363                 if (unlikely(bh_ofs >= end))
2364                         break;
2365                 set_buffer_dirty(bh);
2366         } while ((bh = bh->b_this_page) != head);
2367         spin_unlock(&mapping->private_lock);
2368         __set_page_dirty_nobuffers(page);
2369         if (unlikely(buffers_to_free)) {
2370                 do {
2371                         bh = buffers_to_free->b_this_page;
2372                         free_buffer_head(buffers_to_free);
2373                         buffers_to_free = bh;
2374                 } while (buffers_to_free);
2375         }
2376 }
2377
2378 #endif /* NTFS_RW */