NTFS: Fixup handling of sparse, compressed, and encrypted attributes in
[linux-2.6] / fs / ntfs / aops.c
1 /**
2  * aops.c - NTFS kernel address space operations and page cache handling.
3  *          Part of the Linux-NTFS project.
4  *
5  * Copyright (c) 2001-2005 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                         u8 *kaddr;
674                         unsigned long *bpos, *bend;
675
676                         /* Check if the buffer is zero. */
677                         kaddr = kmap_atomic(page, KM_USER0);
678                         bpos = (unsigned long *)(kaddr + bh_offset(bh));
679                         bend = (unsigned long *)((u8*)bpos + blocksize);
680                         do {
681                                 if (unlikely(*bpos))
682                                         break;
683                         } while (likely(++bpos < bend));
684                         kunmap_atomic(kaddr, KM_USER0);
685                         if (bpos == bend) {
686                                 /*
687                                  * Buffer is zero and sparse, no need to write
688                                  * it.
689                                  */
690                                 bh->b_blocknr = -1;
691                                 clear_buffer_dirty(bh);
692                                 continue;
693                         }
694                         // TODO: Instantiate the hole.
695                         // clear_buffer_new(bh);
696                         // unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
697                         ntfs_error(vol->sb, "Writing into sparse regions is "
698                                         "not supported yet. Sorry.");
699                         err = -EOPNOTSUPP;
700                         break;
701                 }
702                 /* If first try and runlist unmapped, map and retry. */
703                 if (!is_retry && lcn == LCN_RL_NOT_MAPPED) {
704                         is_retry = TRUE;
705                         /*
706                          * Attempt to map runlist, dropping lock for
707                          * the duration.
708                          */
709                         up_read(&ni->runlist.lock);
710                         err = ntfs_map_runlist(ni, vcn);
711                         if (likely(!err))
712                                 goto lock_retry_remap;
713                         rl = NULL;
714                         lcn = err;
715                 } else if (!rl)
716                         up_read(&ni->runlist.lock);
717                 /* Failed to map the buffer, even after retrying. */
718                 bh->b_blocknr = -1;
719                 ntfs_error(vol->sb, "Failed to write to inode 0x%lx, "
720                                 "attribute type 0x%x, vcn 0x%llx, offset 0x%x "
721                                 "because its location on disk could not be "
722                                 "determined%s (error code %lli).", ni->mft_no,
723                                 ni->type, (unsigned long long)vcn,
724                                 vcn_ofs, is_retry ? " even after "
725                                 "retrying" : "", (long long)lcn);
726                 if (!err)
727                         err = -EIO;
728                 break;
729         } while (block++, (bh = bh->b_this_page) != head);
730
731         /* Release the lock if we took it. */
732         if (rl)
733                 up_read(&ni->runlist.lock);
734
735         /* For the error case, need to reset bh to the beginning. */
736         bh = head;
737
738         /* Just an optimization, so ->readpage() isn't called later. */
739         if (unlikely(!PageUptodate(page))) {
740                 int uptodate = 1;
741                 do {
742                         if (!buffer_uptodate(bh)) {
743                                 uptodate = 0;
744                                 bh = head;
745                                 break;
746                         }
747                 } while ((bh = bh->b_this_page) != head);
748                 if (uptodate)
749                         SetPageUptodate(page);
750         }
751
752         /* Setup all mapped, dirty buffers for async write i/o. */
753         do {
754                 get_bh(bh);
755                 if (buffer_mapped(bh) && buffer_dirty(bh)) {
756                         lock_buffer(bh);
757                         if (test_clear_buffer_dirty(bh)) {
758                                 BUG_ON(!buffer_uptodate(bh));
759                                 mark_buffer_async_write(bh);
760                         } else
761                                 unlock_buffer(bh);
762                 } else if (unlikely(err)) {
763                         /*
764                          * For the error case. The buffer may have been set
765                          * dirty during attachment to a dirty page.
766                          */
767                         if (err != -ENOMEM)
768                                 clear_buffer_dirty(bh);
769                 }
770         } while ((bh = bh->b_this_page) != head);
771
772         if (unlikely(err)) {
773                 // TODO: Remove the -EOPNOTSUPP check later on...
774                 if (unlikely(err == -EOPNOTSUPP))
775                         err = 0;
776                 else if (err == -ENOMEM) {
777                         ntfs_warning(vol->sb, "Error allocating memory. "
778                                         "Redirtying page so we try again "
779                                         "later.");
780                         /*
781                          * Put the page back on mapping->dirty_pages, but
782                          * leave its buffer's dirty state as-is.
783                          */
784                         redirty_page_for_writepage(wbc, page);
785                         err = 0;
786                 } else
787                         SetPageError(page);
788         }
789
790         BUG_ON(PageWriteback(page));
791         set_page_writeback(page);       /* Keeps try_to_free_buffers() away. */
792         unlock_page(page);
793
794         /*
795          * Submit the prepared buffers for i/o. Note the page is unlocked,
796          * and the async write i/o completion handler can end_page_writeback()
797          * at any time after the *first* submit_bh(). So the buffers can then
798          * disappear...
799          */
800         need_end_writeback = TRUE;
801         do {
802                 struct buffer_head *next = bh->b_this_page;
803                 if (buffer_async_write(bh)) {
804                         submit_bh(WRITE, bh);
805                         need_end_writeback = FALSE;
806                 }
807                 put_bh(bh);
808                 bh = next;
809         } while (bh != head);
810
811         /* If no i/o was started, need to end_page_writeback(). */
812         if (unlikely(need_end_writeback))
813                 end_page_writeback(page);
814
815         ntfs_debug("Done.");
816         return err;
817 }
818
819 /**
820  * ntfs_write_mst_block - write a @page to the backing store
821  * @page:       page cache page to write out
822  * @wbc:        writeback control structure
823  *
824  * This function is for writing pages belonging to non-resident, mst protected
825  * attributes to their backing store.  The only supported attributes are index
826  * allocation and $MFT/$DATA.  Both directory inodes and index inodes are
827  * supported for the index allocation case.
828  *
829  * The page must remain locked for the duration of the write because we apply
830  * the mst fixups, write, and then undo the fixups, so if we were to unlock the
831  * page before undoing the fixups, any other user of the page will see the
832  * page contents as corrupt.
833  *
834  * We clear the page uptodate flag for the duration of the function to ensure
835  * exclusion for the $MFT/$DATA case against someone mapping an mft record we
836  * are about to apply the mst fixups to.
837  *
838  * Return 0 on success and -errno on error.
839  *
840  * Based on ntfs_write_block(), ntfs_mft_writepage(), and
841  * write_mft_record_nolock().
842  */
843 static int ntfs_write_mst_block(struct page *page,
844                 struct writeback_control *wbc)
845 {
846         sector_t block, dblock, rec_block;
847         struct inode *vi = page->mapping->host;
848         ntfs_inode *ni = NTFS_I(vi);
849         ntfs_volume *vol = ni->vol;
850         u8 *kaddr;
851         unsigned int rec_size = ni->itype.index.block_size;
852         ntfs_inode *locked_nis[PAGE_CACHE_SIZE / rec_size];
853         struct buffer_head *bh, *head, *tbh, *rec_start_bh;
854         struct buffer_head *bhs[MAX_BUF_PER_PAGE];
855         runlist_element *rl;
856         int i, nr_locked_nis, nr_recs, nr_bhs, max_bhs, bhs_per_rec, err, err2;
857         unsigned bh_size, rec_size_bits;
858         BOOL sync, is_mft, page_is_dirty, rec_is_dirty;
859         unsigned char bh_size_bits;
860
861         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
862                         "0x%lx.", vi->i_ino, ni->type, page->index);
863         BUG_ON(!NInoNonResident(ni));
864         BUG_ON(!NInoMstProtected(ni));
865         is_mft = (S_ISREG(vi->i_mode) && !vi->i_ino);
866         /*
867          * NOTE: ntfs_write_mst_block() would be called for $MFTMirr if a page
868          * in its page cache were to be marked dirty.  However this should
869          * never happen with the current driver and considering we do not
870          * handle this case here we do want to BUG(), at least for now.
871          */
872         BUG_ON(!(is_mft || S_ISDIR(vi->i_mode) ||
873                         (NInoAttr(ni) && ni->type == AT_INDEX_ALLOCATION)));
874         bh_size_bits = vi->i_blkbits;
875         bh_size = 1 << bh_size_bits;
876         max_bhs = PAGE_CACHE_SIZE / bh_size;
877         BUG_ON(!max_bhs);
878         BUG_ON(max_bhs > MAX_BUF_PER_PAGE);
879
880         /* Were we called for sync purposes? */
881         sync = (wbc->sync_mode == WB_SYNC_ALL);
882
883         /* Make sure we have mapped buffers. */
884         BUG_ON(!page_has_buffers(page));
885         bh = head = page_buffers(page);
886         BUG_ON(!bh);
887
888         rec_size_bits = ni->itype.index.block_size_bits;
889         BUG_ON(!(PAGE_CACHE_SIZE >> rec_size_bits));
890         bhs_per_rec = rec_size >> bh_size_bits;
891         BUG_ON(!bhs_per_rec);
892
893         /* The first block in the page. */
894         rec_block = block = (sector_t)page->index <<
895                         (PAGE_CACHE_SHIFT - bh_size_bits);
896
897         /* The first out of bounds block for the data size. */
898         dblock = (i_size_read(vi) + bh_size - 1) >> bh_size_bits;
899
900         rl = NULL;
901         err = err2 = nr_bhs = nr_recs = nr_locked_nis = 0;
902         page_is_dirty = rec_is_dirty = FALSE;
903         rec_start_bh = NULL;
904         do {
905                 BOOL is_retry = FALSE;
906
907                 if (likely(block < rec_block)) {
908                         if (unlikely(block >= dblock)) {
909                                 clear_buffer_dirty(bh);
910                                 set_buffer_uptodate(bh);
911                                 continue;
912                         }
913                         /*
914                          * This block is not the first one in the record.  We
915                          * ignore the buffer's dirty state because we could
916                          * have raced with a parallel mark_ntfs_record_dirty().
917                          */
918                         if (!rec_is_dirty)
919                                 continue;
920                         if (unlikely(err2)) {
921                                 if (err2 != -ENOMEM)
922                                         clear_buffer_dirty(bh);
923                                 continue;
924                         }
925                 } else /* if (block == rec_block) */ {
926                         BUG_ON(block > rec_block);
927                         /* This block is the first one in the record. */
928                         rec_block += bhs_per_rec;
929                         err2 = 0;
930                         if (unlikely(block >= dblock)) {
931                                 clear_buffer_dirty(bh);
932                                 continue;
933                         }
934                         if (!buffer_dirty(bh)) {
935                                 /* Clean records are not written out. */
936                                 rec_is_dirty = FALSE;
937                                 continue;
938                         }
939                         rec_is_dirty = TRUE;
940                         rec_start_bh = bh;
941                 }
942                 /* Need to map the buffer if it is not mapped already. */
943                 if (unlikely(!buffer_mapped(bh))) {
944                         VCN vcn;
945                         LCN lcn;
946                         unsigned int vcn_ofs;
947
948                         bh->b_bdev = vol->sb->s_bdev;
949                         /* Obtain the vcn and offset of the current block. */
950                         vcn = (VCN)block << bh_size_bits;
951                         vcn_ofs = vcn & vol->cluster_size_mask;
952                         vcn >>= vol->cluster_size_bits;
953                         if (!rl) {
954 lock_retry_remap:
955                                 down_read(&ni->runlist.lock);
956                                 rl = ni->runlist.rl;
957                         }
958                         if (likely(rl != NULL)) {
959                                 /* Seek to element containing target vcn. */
960                                 while (rl->length && rl[1].vcn <= vcn)
961                                         rl++;
962                                 lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
963                         } else
964                                 lcn = LCN_RL_NOT_MAPPED;
965                         /* Successful remap. */
966                         if (likely(lcn >= 0)) {
967                                 /* Setup buffer head to correct block. */
968                                 bh->b_blocknr = ((lcn <<
969                                                 vol->cluster_size_bits) +
970                                                 vcn_ofs) >> bh_size_bits;
971                                 set_buffer_mapped(bh);
972                         } else {
973                                 /*
974                                  * Remap failed.  Retry to map the runlist once
975                                  * unless we are working on $MFT which always
976                                  * has the whole of its runlist in memory.
977                                  */
978                                 if (!is_mft && !is_retry &&
979                                                 lcn == LCN_RL_NOT_MAPPED) {
980                                         is_retry = TRUE;
981                                         /*
982                                          * Attempt to map runlist, dropping
983                                          * lock for the duration.
984                                          */
985                                         up_read(&ni->runlist.lock);
986                                         err2 = ntfs_map_runlist(ni, vcn);
987                                         if (likely(!err2))
988                                                 goto lock_retry_remap;
989                                         if (err2 == -ENOMEM)
990                                                 page_is_dirty = TRUE;
991                                         lcn = err2;
992                                 } else {
993                                         err2 = -EIO;
994                                         if (!rl)
995                                                 up_read(&ni->runlist.lock);
996                                 }
997                                 /* Hard error.  Abort writing this record. */
998                                 if (!err || err == -ENOMEM)
999                                         err = err2;
1000                                 bh->b_blocknr = -1;
1001                                 ntfs_error(vol->sb, "Cannot write ntfs record "
1002                                                 "0x%llx (inode 0x%lx, "
1003                                                 "attribute type 0x%x) because "
1004                                                 "its location on disk could "
1005                                                 "not be determined (error "
1006                                                 "code %lli).",
1007                                                 (long long)block <<
1008                                                 bh_size_bits >>
1009                                                 vol->mft_record_size_bits,
1010                                                 ni->mft_no, ni->type,
1011                                                 (long long)lcn);
1012                                 /*
1013                                  * If this is not the first buffer, remove the
1014                                  * buffers in this record from the list of
1015                                  * buffers to write and clear their dirty bit
1016                                  * if not error -ENOMEM.
1017                                  */
1018                                 if (rec_start_bh != bh) {
1019                                         while (bhs[--nr_bhs] != rec_start_bh)
1020                                                 ;
1021                                         if (err2 != -ENOMEM) {
1022                                                 do {
1023                                                         clear_buffer_dirty(
1024                                                                 rec_start_bh);
1025                                                 } while ((rec_start_bh =
1026                                                                 rec_start_bh->
1027                                                                 b_this_page) !=
1028                                                                 bh);
1029                                         }
1030                                 }
1031                                 continue;
1032                         }
1033                 }
1034                 BUG_ON(!buffer_uptodate(bh));
1035                 BUG_ON(nr_bhs >= max_bhs);
1036                 bhs[nr_bhs++] = bh;
1037         } while (block++, (bh = bh->b_this_page) != head);
1038         if (unlikely(rl))
1039                 up_read(&ni->runlist.lock);
1040         /* If there were no dirty buffers, we are done. */
1041         if (!nr_bhs)
1042                 goto done;
1043         /* Map the page so we can access its contents. */
1044         kaddr = kmap(page);
1045         /* Clear the page uptodate flag whilst the mst fixups are applied. */
1046         BUG_ON(!PageUptodate(page));
1047         ClearPageUptodate(page);
1048         for (i = 0; i < nr_bhs; i++) {
1049                 unsigned int ofs;
1050
1051                 /* Skip buffers which are not at the beginning of records. */
1052                 if (i % bhs_per_rec)
1053                         continue;
1054                 tbh = bhs[i];
1055                 ofs = bh_offset(tbh);
1056                 if (is_mft) {
1057                         ntfs_inode *tni;
1058                         unsigned long mft_no;
1059
1060                         /* Get the mft record number. */
1061                         mft_no = (((s64)page->index << PAGE_CACHE_SHIFT) + ofs)
1062                                         >> rec_size_bits;
1063                         /* Check whether to write this mft record. */
1064                         tni = NULL;
1065                         if (!ntfs_may_write_mft_record(vol, mft_no,
1066                                         (MFT_RECORD*)(kaddr + ofs), &tni)) {
1067                                 /*
1068                                  * The record should not be written.  This
1069                                  * means we need to redirty the page before
1070                                  * returning.
1071                                  */
1072                                 page_is_dirty = TRUE;
1073                                 /*
1074                                  * Remove the buffers in this mft record from
1075                                  * the list of buffers to write.
1076                                  */
1077                                 do {
1078                                         bhs[i] = NULL;
1079                                 } while (++i % bhs_per_rec);
1080                                 continue;
1081                         }
1082                         /*
1083                          * The record should be written.  If a locked ntfs
1084                          * inode was returned, add it to the array of locked
1085                          * ntfs inodes.
1086                          */
1087                         if (tni)
1088                                 locked_nis[nr_locked_nis++] = tni;
1089                 }
1090                 /* Apply the mst protection fixups. */
1091                 err2 = pre_write_mst_fixup((NTFS_RECORD*)(kaddr + ofs),
1092                                 rec_size);
1093                 if (unlikely(err2)) {
1094                         if (!err || err == -ENOMEM)
1095                                 err = -EIO;
1096                         ntfs_error(vol->sb, "Failed to apply mst fixups "
1097                                         "(inode 0x%lx, attribute type 0x%x, "
1098                                         "page index 0x%lx, page offset 0x%x)!"
1099                                         "  Unmount and run chkdsk.", vi->i_ino,
1100                                         ni->type, page->index, ofs);
1101                         /*
1102                          * Mark all the buffers in this record clean as we do
1103                          * not want to write corrupt data to disk.
1104                          */
1105                         do {
1106                                 clear_buffer_dirty(bhs[i]);
1107                                 bhs[i] = NULL;
1108                         } while (++i % bhs_per_rec);
1109                         continue;
1110                 }
1111                 nr_recs++;
1112         }
1113         /* If no records are to be written out, we are done. */
1114         if (!nr_recs)
1115                 goto unm_done;
1116         flush_dcache_page(page);
1117         /* Lock buffers and start synchronous write i/o on them. */
1118         for (i = 0; i < nr_bhs; i++) {
1119                 tbh = bhs[i];
1120                 if (!tbh)
1121                         continue;
1122                 if (unlikely(test_set_buffer_locked(tbh)))
1123                         BUG();
1124                 /* The buffer dirty state is now irrelevant, just clean it. */
1125                 clear_buffer_dirty(tbh);
1126                 BUG_ON(!buffer_uptodate(tbh));
1127                 BUG_ON(!buffer_mapped(tbh));
1128                 get_bh(tbh);
1129                 tbh->b_end_io = end_buffer_write_sync;
1130                 submit_bh(WRITE, tbh);
1131         }
1132         /* Synchronize the mft mirror now if not @sync. */
1133         if (is_mft && !sync)
1134                 goto do_mirror;
1135 do_wait:
1136         /* Wait on i/o completion of buffers. */
1137         for (i = 0; i < nr_bhs; i++) {
1138                 tbh = bhs[i];
1139                 if (!tbh)
1140                         continue;
1141                 wait_on_buffer(tbh);
1142                 if (unlikely(!buffer_uptodate(tbh))) {
1143                         ntfs_error(vol->sb, "I/O error while writing ntfs "
1144                                         "record buffer (inode 0x%lx, "
1145                                         "attribute type 0x%x, page index "
1146                                         "0x%lx, page offset 0x%lx)!  Unmount "
1147                                         "and run chkdsk.", vi->i_ino, ni->type,
1148                                         page->index, bh_offset(tbh));
1149                         if (!err || err == -ENOMEM)
1150                                 err = -EIO;
1151                         /*
1152                          * Set the buffer uptodate so the page and buffer
1153                          * states do not become out of sync.
1154                          */
1155                         set_buffer_uptodate(tbh);
1156                 }
1157         }
1158         /* If @sync, now synchronize the mft mirror. */
1159         if (is_mft && sync) {
1160 do_mirror:
1161                 for (i = 0; i < nr_bhs; i++) {
1162                         unsigned long mft_no;
1163                         unsigned int ofs;
1164
1165                         /*
1166                          * Skip buffers which are not at the beginning of
1167                          * records.
1168                          */
1169                         if (i % bhs_per_rec)
1170                                 continue;
1171                         tbh = bhs[i];
1172                         /* Skip removed buffers (and hence records). */
1173                         if (!tbh)
1174                                 continue;
1175                         ofs = bh_offset(tbh);
1176                         /* Get the mft record number. */
1177                         mft_no = (((s64)page->index << PAGE_CACHE_SHIFT) + ofs)
1178                                         >> rec_size_bits;
1179                         if (mft_no < vol->mftmirr_size)
1180                                 ntfs_sync_mft_mirror(vol, mft_no,
1181                                                 (MFT_RECORD*)(kaddr + ofs),
1182                                                 sync);
1183                 }
1184                 if (!sync)
1185                         goto do_wait;
1186         }
1187         /* Remove the mst protection fixups again. */
1188         for (i = 0; i < nr_bhs; i++) {
1189                 if (!(i % bhs_per_rec)) {
1190                         tbh = bhs[i];
1191                         if (!tbh)
1192                                 continue;
1193                         post_write_mst_fixup((NTFS_RECORD*)(kaddr +
1194                                         bh_offset(tbh)));
1195                 }
1196         }
1197         flush_dcache_page(page);
1198 unm_done:
1199         /* Unlock any locked inodes. */
1200         while (nr_locked_nis-- > 0) {
1201                 ntfs_inode *tni, *base_tni;
1202                 
1203                 tni = locked_nis[nr_locked_nis];
1204                 /* Get the base inode. */
1205                 down(&tni->extent_lock);
1206                 if (tni->nr_extents >= 0)
1207                         base_tni = tni;
1208                 else {
1209                         base_tni = tni->ext.base_ntfs_ino;
1210                         BUG_ON(!base_tni);
1211                 }
1212                 up(&tni->extent_lock);
1213                 ntfs_debug("Unlocking %s inode 0x%lx.",
1214                                 tni == base_tni ? "base" : "extent",
1215                                 tni->mft_no);
1216                 up(&tni->mrec_lock);
1217                 atomic_dec(&tni->count);
1218                 iput(VFS_I(base_tni));
1219         }
1220         SetPageUptodate(page);
1221         kunmap(page);
1222 done:
1223         if (unlikely(err && err != -ENOMEM)) {
1224                 /*
1225                  * Set page error if there is only one ntfs record in the page.
1226                  * Otherwise we would loose per-record granularity.
1227                  */
1228                 if (ni->itype.index.block_size == PAGE_CACHE_SIZE)
1229                         SetPageError(page);
1230                 NVolSetErrors(vol);
1231         }
1232         if (page_is_dirty) {
1233                 ntfs_debug("Page still contains one or more dirty ntfs "
1234                                 "records.  Redirtying the page starting at "
1235                                 "record 0x%lx.", page->index <<
1236                                 (PAGE_CACHE_SHIFT - rec_size_bits));
1237                 redirty_page_for_writepage(wbc, page);
1238                 unlock_page(page);
1239         } else {
1240                 /*
1241                  * Keep the VM happy.  This must be done otherwise the
1242                  * radix-tree tag PAGECACHE_TAG_DIRTY remains set even though
1243                  * the page is clean.
1244                  */
1245                 BUG_ON(PageWriteback(page));
1246                 set_page_writeback(page);
1247                 unlock_page(page);
1248                 end_page_writeback(page);
1249         }
1250         if (likely(!err))
1251                 ntfs_debug("Done.");
1252         return err;
1253 }
1254
1255 /**
1256  * ntfs_writepage - write a @page to the backing store
1257  * @page:       page cache page to write out
1258  * @wbc:        writeback control structure
1259  *
1260  * This is called from the VM when it wants to have a dirty ntfs page cache
1261  * page cleaned.  The VM has already locked the page and marked it clean.
1262  *
1263  * For non-resident attributes, ntfs_writepage() writes the @page by calling
1264  * the ntfs version of the generic block_write_full_page() function,
1265  * ntfs_write_block(), which in turn if necessary creates and writes the
1266  * buffers associated with the page asynchronously.
1267  *
1268  * For resident attributes, OTOH, ntfs_writepage() writes the @page by copying
1269  * the data to the mft record (which at this stage is most likely in memory).
1270  * The mft record is then marked dirty and written out asynchronously via the
1271  * vfs inode dirty code path for the inode the mft record belongs to or via the
1272  * vm page dirty code path for the page the mft record is in.
1273  *
1274  * Based on ntfs_readpage() and fs/buffer.c::block_write_full_page().
1275  *
1276  * Return 0 on success and -errno on error.
1277  */
1278 static int ntfs_writepage(struct page *page, struct writeback_control *wbc)
1279 {
1280         loff_t i_size;
1281         struct inode *vi = page->mapping->host;
1282         ntfs_inode *base_ni = NULL, *ni = NTFS_I(vi);
1283         char *kaddr;
1284         ntfs_attr_search_ctx *ctx = NULL;
1285         MFT_RECORD *m = NULL;
1286         u32 attr_len;
1287         int err;
1288
1289 retry_writepage:
1290         BUG_ON(!PageLocked(page));
1291         i_size = i_size_read(vi);
1292         /* Is the page fully outside i_size? (truncate in progress) */
1293         if (unlikely(page->index >= (i_size + PAGE_CACHE_SIZE - 1) >>
1294                         PAGE_CACHE_SHIFT)) {
1295                 /*
1296                  * The page may have dirty, unmapped buffers.  Make them
1297                  * freeable here, so the page does not leak.
1298                  */
1299                 block_invalidatepage(page, 0);
1300                 unlock_page(page);
1301                 ntfs_debug("Write outside i_size - truncated?");
1302                 return 0;
1303         }
1304         /*
1305          * Only $DATA attributes can be encrypted and only unnamed $DATA
1306          * attributes can be compressed.  Index root can have the flags set but
1307          * this means to create compressed/encrypted files, not that the
1308          * attribute is compressed/encrypted.
1309          */
1310         if (ni->type != AT_INDEX_ROOT) {
1311                 /* If file is encrypted, deny access, just like NT4. */
1312                 if (NInoEncrypted(ni)) {
1313                         unlock_page(page);
1314                         BUG_ON(ni->type != AT_DATA);
1315                         ntfs_debug("Denying write access to encrypted "
1316                                         "file.");
1317                         return -EACCES;
1318                 }
1319                 /* Compressed data streams are handled in compress.c. */
1320                 if (NInoNonResident(ni) && NInoCompressed(ni)) {
1321                         BUG_ON(ni->type != AT_DATA);
1322                         BUG_ON(ni->name_len);
1323                         // TODO: Implement and replace this with
1324                         // return ntfs_write_compressed_block(page);
1325                         unlock_page(page);
1326                         ntfs_error(vi->i_sb, "Writing to compressed files is "
1327                                         "not supported yet.  Sorry.");
1328                         return -EOPNOTSUPP;
1329                 }
1330                 // TODO: Implement and remove this check.
1331                 if (NInoNonResident(ni) && NInoSparse(ni)) {
1332                         unlock_page(page);
1333                         ntfs_error(vi->i_sb, "Writing to sparse files is not "
1334                                         "supported yet.  Sorry.");
1335                         return -EOPNOTSUPP;
1336                 }
1337         }
1338         /* NInoNonResident() == NInoIndexAllocPresent() */
1339         if (NInoNonResident(ni)) {
1340                 /* We have to zero every time due to mmap-at-end-of-file. */
1341                 if (page->index >= (i_size >> PAGE_CACHE_SHIFT)) {
1342                         /* The page straddles i_size. */
1343                         unsigned int ofs = i_size & ~PAGE_CACHE_MASK;
1344                         kaddr = kmap_atomic(page, KM_USER0);
1345                         memset(kaddr + ofs, 0, PAGE_CACHE_SIZE - ofs);
1346                         flush_dcache_page(page);
1347                         kunmap_atomic(kaddr, KM_USER0);
1348                 }
1349                 /* Handle mst protected attributes. */
1350                 if (NInoMstProtected(ni))
1351                         return ntfs_write_mst_block(page, wbc);
1352                 /* Normal, non-resident data stream. */
1353                 return ntfs_write_block(page, wbc);
1354         }
1355         /*
1356          * Attribute is resident, implying it is not compressed, encrypted, or
1357          * mst protected.  This also means the attribute is smaller than an mft
1358          * record and hence smaller than a page, so can simply return error on
1359          * any pages with index above 0.  Note the attribute can actually be
1360          * marked compressed but if it is resident the actual data is not
1361          * compressed so we are ok to ignore the compressed flag here.
1362          */
1363         BUG_ON(page_has_buffers(page));
1364         BUG_ON(!PageUptodate(page));
1365         if (unlikely(page->index > 0)) {
1366                 ntfs_error(vi->i_sb, "BUG()! page->index (0x%lx) > 0.  "
1367                                 "Aborting write.", page->index);
1368                 BUG_ON(PageWriteback(page));
1369                 set_page_writeback(page);
1370                 unlock_page(page);
1371                 end_page_writeback(page);
1372                 return -EIO;
1373         }
1374         if (!NInoAttr(ni))
1375                 base_ni = ni;
1376         else
1377                 base_ni = ni->ext.base_ntfs_ino;
1378         /* Map, pin, and lock the mft record. */
1379         m = map_mft_record(base_ni);
1380         if (IS_ERR(m)) {
1381                 err = PTR_ERR(m);
1382                 m = NULL;
1383                 ctx = NULL;
1384                 goto err_out;
1385         }
1386         /*
1387          * If a parallel write made the attribute non-resident, drop the mft
1388          * record and retry the writepage.
1389          */
1390         if (unlikely(NInoNonResident(ni))) {
1391                 unmap_mft_record(base_ni);
1392                 goto retry_writepage;
1393         }
1394         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1395         if (unlikely(!ctx)) {
1396                 err = -ENOMEM;
1397                 goto err_out;
1398         }
1399         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1400                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1401         if (unlikely(err))
1402                 goto err_out;
1403         /*
1404          * Keep the VM happy.  This must be done otherwise the radix-tree tag
1405          * PAGECACHE_TAG_DIRTY remains set even though the page is clean.
1406          */
1407         BUG_ON(PageWriteback(page));
1408         set_page_writeback(page);
1409         unlock_page(page);
1410         /*
1411          * Here, we do not need to zero the out of bounds area everytime
1412          * because the below memcpy() already takes care of the
1413          * mmap-at-end-of-file requirements.  If the file is converted to a
1414          * non-resident one, then the code path use is switched to the
1415          * non-resident one where the zeroing happens on each ntfs_writepage()
1416          * invocation.
1417          */
1418         attr_len = le32_to_cpu(ctx->attr->data.resident.value_length);
1419         i_size = i_size_read(vi);
1420         if (unlikely(attr_len > i_size)) {
1421                 attr_len = i_size;
1422                 ctx->attr->data.resident.value_length = cpu_to_le32(attr_len);
1423         }
1424         kaddr = kmap_atomic(page, KM_USER0);
1425         /* Copy the data from the page to the mft record. */
1426         memcpy((u8*)ctx->attr +
1427                         le16_to_cpu(ctx->attr->data.resident.value_offset),
1428                         kaddr, attr_len);
1429         flush_dcache_mft_record_page(ctx->ntfs_ino);
1430         /* Zero out of bounds area in the page cache page. */
1431         memset(kaddr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
1432         flush_dcache_page(page);
1433         kunmap_atomic(kaddr, KM_USER0);
1434
1435         end_page_writeback(page);
1436
1437         /* Mark the mft record dirty, so it gets written back. */
1438         mark_mft_record_dirty(ctx->ntfs_ino);
1439         ntfs_attr_put_search_ctx(ctx);
1440         unmap_mft_record(base_ni);
1441         return 0;
1442 err_out:
1443         if (err == -ENOMEM) {
1444                 ntfs_warning(vi->i_sb, "Error allocating memory. Redirtying "
1445                                 "page so we try again later.");
1446                 /*
1447                  * Put the page back on mapping->dirty_pages, but leave its
1448                  * buffers' dirty state as-is.
1449                  */
1450                 redirty_page_for_writepage(wbc, page);
1451                 err = 0;
1452         } else {
1453                 ntfs_error(vi->i_sb, "Resident attribute write failed with "
1454                                 "error %i.", err);
1455                 SetPageError(page);
1456                 NVolSetErrors(ni->vol);
1457                 make_bad_inode(vi);
1458         }
1459         unlock_page(page);
1460         if (ctx)
1461                 ntfs_attr_put_search_ctx(ctx);
1462         if (m)
1463                 unmap_mft_record(base_ni);
1464         return err;
1465 }
1466
1467 /**
1468  * ntfs_prepare_nonresident_write -
1469  *
1470  */
1471 static int ntfs_prepare_nonresident_write(struct page *page,
1472                 unsigned from, unsigned to)
1473 {
1474         VCN vcn;
1475         LCN lcn;
1476         s64 initialized_size;
1477         loff_t i_size;
1478         sector_t block, ablock, iblock;
1479         struct inode *vi;
1480         ntfs_inode *ni;
1481         ntfs_volume *vol;
1482         runlist_element *rl;
1483         struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
1484         unsigned long flags;
1485         unsigned int vcn_ofs, block_start, block_end, blocksize;
1486         int err;
1487         BOOL is_retry;
1488         unsigned char blocksize_bits;
1489
1490         vi = page->mapping->host;
1491         ni = NTFS_I(vi);
1492         vol = ni->vol;
1493
1494         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
1495                         "0x%lx, from = %u, to = %u.", ni->mft_no, ni->type,
1496                         page->index, from, to);
1497
1498         BUG_ON(!NInoNonResident(ni));
1499
1500         blocksize_bits = vi->i_blkbits;
1501         blocksize = 1 << blocksize_bits;
1502
1503         /*
1504          * create_empty_buffers() will create uptodate/dirty buffers if the
1505          * page is uptodate/dirty.
1506          */
1507         if (!page_has_buffers(page))
1508                 create_empty_buffers(page, blocksize, 0);
1509         bh = head = page_buffers(page);
1510         if (unlikely(!bh))
1511                 return -ENOMEM;
1512
1513         /* The first block in the page. */
1514         block = (s64)page->index << (PAGE_CACHE_SHIFT - blocksize_bits);
1515
1516         read_lock_irqsave(&ni->size_lock, flags);
1517         /*
1518          * The first out of bounds block for the allocated size.  No need to
1519          * round up as allocated_size is in multiples of cluster size and the
1520          * minimum cluster size is 512 bytes, which is equal to the smallest
1521          * blocksize.
1522          */
1523         ablock = ni->allocated_size >> blocksize_bits;
1524         i_size = i_size_read(vi);
1525         initialized_size = ni->initialized_size;
1526         read_unlock_irqrestore(&ni->size_lock, flags);
1527
1528         /* The last (fully or partially) initialized block. */
1529         iblock = initialized_size >> blocksize_bits;
1530
1531         /* Loop through all the buffers in the page. */
1532         block_start = 0;
1533         rl = NULL;
1534         err = 0;
1535         do {
1536                 block_end = block_start + blocksize;
1537                 /*
1538                  * If buffer @bh is outside the write, just mark it uptodate
1539                  * if the page is uptodate and continue with the next buffer.
1540                  */
1541                 if (block_end <= from || block_start >= to) {
1542                         if (PageUptodate(page)) {
1543                                 if (!buffer_uptodate(bh))
1544                                         set_buffer_uptodate(bh);
1545                         }
1546                         continue;
1547                 }
1548                 /*
1549                  * @bh is at least partially being written to.
1550                  * Make sure it is not marked as new.
1551                  */
1552                 //if (buffer_new(bh))
1553                 //      clear_buffer_new(bh);
1554
1555                 if (block >= ablock) {
1556                         // TODO: block is above allocated_size, need to
1557                         // allocate it. Best done in one go to accommodate not
1558                         // only block but all above blocks up to and including:
1559                         // ((page->index << PAGE_CACHE_SHIFT) + to + blocksize
1560                         // - 1) >> blobksize_bits. Obviously will need to round
1561                         // up to next cluster boundary, too. This should be
1562                         // done with a helper function, so it can be reused.
1563                         ntfs_error(vol->sb, "Writing beyond allocated size "
1564                                         "is not supported yet. Sorry.");
1565                         err = -EOPNOTSUPP;
1566                         goto err_out;
1567                         // Need to update ablock.
1568                         // Need to set_buffer_new() on all block bhs that are
1569                         // newly allocated.
1570                 }
1571                 /*
1572                  * Now we have enough allocated size to fulfill the whole
1573                  * request, i.e. block < ablock is true.
1574                  */
1575                 if (unlikely((block >= iblock) &&
1576                                 (initialized_size < i_size))) {
1577                         /*
1578                          * If this page is fully outside initialized size, zero
1579                          * out all pages between the current initialized size
1580                          * and the current page. Just use ntfs_readpage() to do
1581                          * the zeroing transparently.
1582                          */
1583                         if (block > iblock) {
1584                                 // TODO:
1585                                 // For each page do:
1586                                 // - read_cache_page()
1587                                 // Again for each page do:
1588                                 // - wait_on_page_locked()
1589                                 // - Check (PageUptodate(page) &&
1590                                 //                      !PageError(page))
1591                                 // Update initialized size in the attribute and
1592                                 // in the inode.
1593                                 // Again, for each page do:
1594                                 //      __set_page_dirty_buffers();
1595                                 // page_cache_release()
1596                                 // We don't need to wait on the writes.
1597                                 // Update iblock.
1598                         }
1599                         /*
1600                          * The current page straddles initialized size. Zero
1601                          * all non-uptodate buffers and set them uptodate (and
1602                          * dirty?). Note, there aren't any non-uptodate buffers
1603                          * if the page is uptodate.
1604                          * FIXME: For an uptodate page, the buffers may need to
1605                          * be written out because they were not initialized on
1606                          * disk before.
1607                          */
1608                         if (!PageUptodate(page)) {
1609                                 // TODO:
1610                                 // Zero any non-uptodate buffers up to i_size.
1611                                 // Set them uptodate and dirty.
1612                         }
1613                         // TODO:
1614                         // Update initialized size in the attribute and in the
1615                         // inode (up to i_size).
1616                         // Update iblock.
1617                         // FIXME: This is inefficient. Try to batch the two
1618                         // size changes to happen in one go.
1619                         ntfs_error(vol->sb, "Writing beyond initialized size "
1620                                         "is not supported yet. Sorry.");
1621                         err = -EOPNOTSUPP;
1622                         goto err_out;
1623                         // Do NOT set_buffer_new() BUT DO clear buffer range
1624                         // outside write request range.
1625                         // set_buffer_uptodate() on complete buffers as well as
1626                         // set_buffer_dirty().
1627                 }
1628
1629                 /* Need to map unmapped buffers. */
1630                 if (!buffer_mapped(bh)) {
1631                         /* Unmapped buffer. Need to map it. */
1632                         bh->b_bdev = vol->sb->s_bdev;
1633
1634                         /* Convert block into corresponding vcn and offset. */
1635                         vcn = (VCN)block << blocksize_bits >>
1636                                         vol->cluster_size_bits;
1637                         vcn_ofs = ((VCN)block << blocksize_bits) &
1638                                         vol->cluster_size_mask;
1639
1640                         is_retry = FALSE;
1641                         if (!rl) {
1642 lock_retry_remap:
1643                                 down_read(&ni->runlist.lock);
1644                                 rl = ni->runlist.rl;
1645                         }
1646                         if (likely(rl != NULL)) {
1647                                 /* Seek to element containing target vcn. */
1648                                 while (rl->length && rl[1].vcn <= vcn)
1649                                         rl++;
1650                                 lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
1651                         } else
1652                                 lcn = LCN_RL_NOT_MAPPED;
1653                         if (unlikely(lcn < 0)) {
1654                                 /*
1655                                  * We extended the attribute allocation above.
1656                                  * If we hit an ENOENT here it means that the
1657                                  * allocation was insufficient which is a bug.
1658                                  */
1659                                 BUG_ON(lcn == LCN_ENOENT);
1660
1661                                 /* It is a hole, need to instantiate it. */
1662                                 if (lcn == LCN_HOLE) {
1663                                         // TODO: Instantiate the hole.
1664                                         // clear_buffer_new(bh);
1665                                         // unmap_underlying_metadata(bh->b_bdev,
1666                                         //              bh->b_blocknr);
1667                                         // For non-uptodate buffers, need to
1668                                         // zero out the region outside the
1669                                         // request in this bh or all bhs,
1670                                         // depending on what we implemented
1671                                         // above.
1672                                         // Need to flush_dcache_page().
1673                                         // Or could use set_buffer_new()
1674                                         // instead?
1675                                         ntfs_error(vol->sb, "Writing into "
1676                                                         "sparse regions is "
1677                                                         "not supported yet. "
1678                                                         "Sorry.");
1679                                         err = -EOPNOTSUPP;
1680                                         if (!rl)
1681                                                 up_read(&ni->runlist.lock);
1682                                         goto err_out;
1683                                 } else if (!is_retry &&
1684                                                 lcn == LCN_RL_NOT_MAPPED) {
1685                                         is_retry = TRUE;
1686                                         /*
1687                                          * Attempt to map runlist, dropping
1688                                          * lock for the duration.
1689                                          */
1690                                         up_read(&ni->runlist.lock);
1691                                         err = ntfs_map_runlist(ni, vcn);
1692                                         if (likely(!err))
1693                                                 goto lock_retry_remap;
1694                                         rl = NULL;
1695                                         lcn = err;
1696                                 } else if (!rl)
1697                                         up_read(&ni->runlist.lock);
1698                                 /*
1699                                  * Failed to map the buffer, even after
1700                                  * retrying.
1701                                  */
1702                                 bh->b_blocknr = -1;
1703                                 ntfs_error(vol->sb, "Failed to write to inode "
1704                                                 "0x%lx, attribute type 0x%x, "
1705                                                 "vcn 0x%llx, offset 0x%x "
1706                                                 "because its location on disk "
1707                                                 "could not be determined%s "
1708                                                 "(error code %lli).",
1709                                                 ni->mft_no, ni->type,
1710                                                 (unsigned long long)vcn,
1711                                                 vcn_ofs, is_retry ? " even "
1712                                                 "after retrying" : "",
1713                                                 (long long)lcn);
1714                                 if (!err)
1715                                         err = -EIO;
1716                                 goto err_out;
1717                         }
1718                         /* We now have a successful remap, i.e. lcn >= 0. */
1719
1720                         /* Setup buffer head to correct block. */
1721                         bh->b_blocknr = ((lcn << vol->cluster_size_bits)
1722                                         + vcn_ofs) >> blocksize_bits;
1723                         set_buffer_mapped(bh);
1724
1725                         // FIXME: Something analogous to this is needed for
1726                         // each newly allocated block, i.e. BH_New.
1727                         // FIXME: Might need to take this out of the
1728                         // if (!buffer_mapped(bh)) {}, depending on how we
1729                         // implement things during the allocated_size and
1730                         // initialized_size extension code above.
1731                         if (buffer_new(bh)) {
1732                                 clear_buffer_new(bh);
1733                                 unmap_underlying_metadata(bh->b_bdev,
1734                                                 bh->b_blocknr);
1735                                 if (PageUptodate(page)) {
1736                                         set_buffer_uptodate(bh);
1737                                         continue;
1738                                 }
1739                                 /*
1740                                  * Page is _not_ uptodate, zero surrounding
1741                                  * region. NOTE: This is how we decide if to
1742                                  * zero or not!
1743                                  */
1744                                 if (block_end > to || block_start < from) {
1745                                         void *kaddr;
1746
1747                                         kaddr = kmap_atomic(page, KM_USER0);
1748                                         if (block_end > to)
1749                                                 memset(kaddr + to, 0,
1750                                                                 block_end - to);
1751                                         if (block_start < from)
1752                                                 memset(kaddr + block_start, 0,
1753                                                                 from -
1754                                                                 block_start);
1755                                         flush_dcache_page(page);
1756                                         kunmap_atomic(kaddr, KM_USER0);
1757                                 }
1758                                 continue;
1759                         }
1760                 }
1761                 /* @bh is mapped, set it uptodate if the page is uptodate. */
1762                 if (PageUptodate(page)) {
1763                         if (!buffer_uptodate(bh))
1764                                 set_buffer_uptodate(bh);
1765                         continue;
1766                 }
1767                 /*
1768                  * The page is not uptodate. The buffer is mapped. If it is not
1769                  * uptodate, and it is only partially being written to, we need
1770                  * to read the buffer in before the write, i.e. right now.
1771                  */
1772                 if (!buffer_uptodate(bh) &&
1773                                 (block_start < from || block_end > to)) {
1774                         ll_rw_block(READ, 1, &bh);
1775                         *wait_bh++ = bh;
1776                 }
1777         } while (block++, block_start = block_end,
1778                         (bh = bh->b_this_page) != head);
1779
1780         /* Release the lock if we took it. */
1781         if (rl) {
1782                 up_read(&ni->runlist.lock);
1783                 rl = NULL;
1784         }
1785
1786         /* If we issued read requests, let them complete. */
1787         while (wait_bh > wait) {
1788                 wait_on_buffer(*--wait_bh);
1789                 if (!buffer_uptodate(*wait_bh))
1790                         return -EIO;
1791         }
1792
1793         ntfs_debug("Done.");
1794         return 0;
1795 err_out:
1796         /*
1797          * Zero out any newly allocated blocks to avoid exposing stale data.
1798          * If BH_New is set, we know that the block was newly allocated in the
1799          * above loop.
1800          * FIXME: What about initialized_size increments? Have we done all the
1801          * required zeroing above? If not this error handling is broken, and
1802          * in particular the if (block_end <= from) check is completely bogus.
1803          */
1804         bh = head;
1805         block_start = 0;
1806         is_retry = FALSE;
1807         do {
1808                 block_end = block_start + blocksize;
1809                 if (block_end <= from)
1810                         continue;
1811                 if (block_start >= to)
1812                         break;
1813                 if (buffer_new(bh)) {
1814                         void *kaddr;
1815
1816                         clear_buffer_new(bh);
1817                         kaddr = kmap_atomic(page, KM_USER0);
1818                         memset(kaddr + block_start, 0, bh->b_size);
1819                         kunmap_atomic(kaddr, KM_USER0);
1820                         set_buffer_uptodate(bh);
1821                         mark_buffer_dirty(bh);
1822                         is_retry = TRUE;
1823                 }
1824         } while (block_start = block_end, (bh = bh->b_this_page) != head);
1825         if (is_retry)
1826                 flush_dcache_page(page);
1827         if (rl)
1828                 up_read(&ni->runlist.lock);
1829         return err;
1830 }
1831
1832 /**
1833  * ntfs_prepare_write - prepare a page for receiving data
1834  *
1835  * This is called from generic_file_write() with i_sem held on the inode
1836  * (@page->mapping->host).  The @page is locked but not kmap()ped.  The source
1837  * data has not yet been copied into the @page.
1838  *
1839  * Need to extend the attribute/fill in holes if necessary, create blocks and
1840  * make partially overwritten blocks uptodate,
1841  *
1842  * i_size is not to be modified yet.
1843  *
1844  * Return 0 on success or -errno on error.
1845  *
1846  * Should be using block_prepare_write() [support for sparse files] or
1847  * cont_prepare_write() [no support for sparse files].  Cannot do that due to
1848  * ntfs specifics but can look at them for implementation guidance.
1849  *
1850  * Note: In the range, @from is inclusive and @to is exclusive, i.e. @from is
1851  * the first byte in the page that will be written to and @to is the first byte
1852  * after the last byte that will be written to.
1853  */
1854 static int ntfs_prepare_write(struct file *file, struct page *page,
1855                 unsigned from, unsigned to)
1856 {
1857         s64 new_size;
1858         loff_t i_size;
1859         struct inode *vi = page->mapping->host;
1860         ntfs_inode *base_ni = NULL, *ni = NTFS_I(vi);
1861         ntfs_volume *vol = ni->vol;
1862         ntfs_attr_search_ctx *ctx = NULL;
1863         MFT_RECORD *m = NULL;
1864         ATTR_RECORD *a;
1865         u8 *kaddr;
1866         u32 attr_len;
1867         int err;
1868
1869         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
1870                         "0x%lx, from = %u, to = %u.", vi->i_ino, ni->type,
1871                         page->index, from, to);
1872         BUG_ON(!PageLocked(page));
1873         BUG_ON(from > PAGE_CACHE_SIZE);
1874         BUG_ON(to > PAGE_CACHE_SIZE);
1875         BUG_ON(from > to);
1876         BUG_ON(NInoMstProtected(ni));
1877         /*
1878          * If a previous ntfs_truncate() failed, repeat it and abort if it
1879          * fails again.
1880          */
1881         if (unlikely(NInoTruncateFailed(ni))) {
1882                 down_write(&vi->i_alloc_sem);
1883                 err = ntfs_truncate(vi);
1884                 up_write(&vi->i_alloc_sem);
1885                 if (err || NInoTruncateFailed(ni)) {
1886                         if (!err)
1887                                 err = -EIO;
1888                         goto err_out;
1889                 }
1890         }
1891         /* If the attribute is not resident, deal with it elsewhere. */
1892         if (NInoNonResident(ni)) {
1893                 /*
1894                  * Only unnamed $DATA attributes can be compressed, encrypted,
1895                  * and/or sparse.
1896                  */
1897                 if (ni->type == AT_DATA && !ni->name_len) {
1898                         /* If file is encrypted, deny access, just like NT4. */
1899                         if (NInoEncrypted(ni)) {
1900                                 ntfs_debug("Denying write access to encrypted "
1901                                                 "file.");
1902                                 return -EACCES;
1903                         }
1904                         /* Compressed data streams are handled in compress.c. */
1905                         if (NInoCompressed(ni)) {
1906                                 // TODO: Implement and replace this check with
1907                                 // return ntfs_write_compressed_block(page);
1908                                 ntfs_error(vi->i_sb, "Writing to compressed "
1909                                                 "files is not supported yet. "
1910                                                 "Sorry.");
1911                                 return -EOPNOTSUPP;
1912                         }
1913                         // TODO: Implement and remove this check.
1914                         if (NInoSparse(ni)) {
1915                                 ntfs_error(vi->i_sb, "Writing to sparse files "
1916                                                 "is not supported yet. Sorry.");
1917                                 return -EOPNOTSUPP;
1918                         }
1919                 }
1920                 /* Normal data stream. */
1921                 return ntfs_prepare_nonresident_write(page, from, to);
1922         }
1923         /*
1924          * Attribute is resident, implying it is not compressed, encrypted, or
1925          * sparse.
1926          */
1927         BUG_ON(page_has_buffers(page));
1928         new_size = ((s64)page->index << PAGE_CACHE_SHIFT) + to;
1929         /* If we do not need to resize the attribute allocation we are done. */
1930         if (new_size <= i_size_read(vi))
1931                 goto done;
1932         /* Map, pin, and lock the (base) mft record. */
1933         if (!NInoAttr(ni))
1934                 base_ni = ni;
1935         else
1936                 base_ni = ni->ext.base_ntfs_ino;
1937         m = map_mft_record(base_ni);
1938         if (IS_ERR(m)) {
1939                 err = PTR_ERR(m);
1940                 m = NULL;
1941                 ctx = NULL;
1942                 goto err_out;
1943         }
1944         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1945         if (unlikely(!ctx)) {
1946                 err = -ENOMEM;
1947                 goto err_out;
1948         }
1949         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1950                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1951         if (unlikely(err)) {
1952                 if (err == -ENOENT)
1953                         err = -EIO;
1954                 goto err_out;
1955         }
1956         m = ctx->mrec;
1957         a = ctx->attr;
1958         /* The total length of the attribute value. */
1959         attr_len = le32_to_cpu(a->data.resident.value_length);
1960         /* Fix an eventual previous failure of ntfs_commit_write(). */
1961         i_size = i_size_read(vi);
1962         if (unlikely(attr_len > i_size)) {
1963                 attr_len = i_size;
1964                 a->data.resident.value_length = cpu_to_le32(attr_len);
1965         }
1966         /* If we do not need to resize the attribute allocation we are done. */
1967         if (new_size <= attr_len)
1968                 goto done_unm;
1969         /* Check if new size is allowed in $AttrDef. */
1970         err = ntfs_attr_size_bounds_check(vol, ni->type, new_size);
1971         if (unlikely(err)) {
1972                 if (err == -ERANGE) {
1973                         ntfs_error(vol->sb, "Write would cause the inode "
1974                                         "0x%lx to exceed the maximum size for "
1975                                         "its attribute type (0x%x).  Aborting "
1976                                         "write.", vi->i_ino,
1977                                         le32_to_cpu(ni->type));
1978                 } else {
1979                         ntfs_error(vol->sb, "Inode 0x%lx has unknown "
1980                                         "attribute type 0x%x.  Aborting "
1981                                         "write.", vi->i_ino,
1982                                         le32_to_cpu(ni->type));
1983                         err = -EIO;
1984                 }
1985                 goto err_out2;
1986         }
1987         /*
1988          * Extend the attribute record to be able to store the new attribute
1989          * size.
1990          */
1991         if (new_size >= vol->mft_record_size || ntfs_attr_record_resize(m, a,
1992                         le16_to_cpu(a->data.resident.value_offset) +
1993                         new_size)) {
1994                 /* Not enough space in the mft record. */
1995                 ntfs_error(vol->sb, "Not enough space in the mft record for "
1996                                 "the resized attribute value.  This is not "
1997                                 "supported yet.  Aborting write.");
1998                 err = -EOPNOTSUPP;
1999                 goto err_out2;
2000         }
2001         /*
2002          * We have enough space in the mft record to fit the write.  This
2003          * implies the attribute is smaller than the mft record and hence the
2004          * attribute must be in a single page and hence page->index must be 0.
2005          */
2006         BUG_ON(page->index);
2007         /*
2008          * If the beginning of the write is past the old size, enlarge the
2009          * attribute value up to the beginning of the write and fill it with
2010          * zeroes.
2011          */
2012         if (from > attr_len) {
2013                 memset((u8*)a + le16_to_cpu(a->data.resident.value_offset) +
2014                                 attr_len, 0, from - attr_len);
2015                 a->data.resident.value_length = cpu_to_le32(from);
2016                 /* Zero the corresponding area in the page as well. */
2017                 if (PageUptodate(page)) {
2018                         kaddr = kmap_atomic(page, KM_USER0);
2019                         memset(kaddr + attr_len, 0, from - attr_len);
2020                         kunmap_atomic(kaddr, KM_USER0);
2021                         flush_dcache_page(page);
2022                 }
2023         }
2024         flush_dcache_mft_record_page(ctx->ntfs_ino);
2025         mark_mft_record_dirty(ctx->ntfs_ino);
2026 done_unm:
2027         ntfs_attr_put_search_ctx(ctx);
2028         unmap_mft_record(base_ni);
2029         /*
2030          * Because resident attributes are handled by memcpy() to/from the
2031          * corresponding MFT record, and because this form of i/o is byte
2032          * aligned rather than block aligned, there is no need to bring the
2033          * page uptodate here as in the non-resident case where we need to
2034          * bring the buffers straddled by the write uptodate before
2035          * generic_file_write() does the copying from userspace.
2036          *
2037          * We thus defer the uptodate bringing of the page region outside the
2038          * region written to to ntfs_commit_write(), which makes the code
2039          * simpler and saves one atomic kmap which is good.
2040          */
2041 done:
2042         ntfs_debug("Done.");
2043         return 0;
2044 err_out:
2045         if (err == -ENOMEM)
2046                 ntfs_warning(vi->i_sb, "Error allocating memory required to "
2047                                 "prepare the write.");
2048         else {
2049                 ntfs_error(vi->i_sb, "Resident attribute prepare write failed "
2050                                 "with error %i.", err);
2051                 NVolSetErrors(vol);
2052                 make_bad_inode(vi);
2053         }
2054 err_out2:
2055         if (ctx)
2056                 ntfs_attr_put_search_ctx(ctx);
2057         if (m)
2058                 unmap_mft_record(base_ni);
2059         return err;
2060 }
2061
2062 /**
2063  * ntfs_commit_nonresident_write -
2064  *
2065  */
2066 static int ntfs_commit_nonresident_write(struct page *page,
2067                 unsigned from, unsigned to)
2068 {
2069         s64 pos = ((s64)page->index << PAGE_CACHE_SHIFT) + to;
2070         struct inode *vi = page->mapping->host;
2071         struct buffer_head *bh, *head;
2072         unsigned int block_start, block_end, blocksize;
2073         BOOL partial;
2074
2075         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
2076                         "0x%lx, from = %u, to = %u.", vi->i_ino,
2077                         NTFS_I(vi)->type, page->index, from, to);
2078         blocksize = 1 << vi->i_blkbits;
2079
2080         // FIXME: We need a whole slew of special cases in here for compressed
2081         // files for example...
2082         // For now, we know ntfs_prepare_write() would have failed so we can't
2083         // get here in any of the cases which we have to special case, so we
2084         // are just a ripped off, unrolled generic_commit_write().
2085
2086         bh = head = page_buffers(page);
2087         block_start = 0;
2088         partial = FALSE;
2089         do {
2090                 block_end = block_start + blocksize;
2091                 if (block_end <= from || block_start >= to) {
2092                         if (!buffer_uptodate(bh))
2093                                 partial = TRUE;
2094                 } else {
2095                         set_buffer_uptodate(bh);
2096                         mark_buffer_dirty(bh);
2097                 }
2098         } while (block_start = block_end, (bh = bh->b_this_page) != head);
2099         /*
2100          * If this is a partial write which happened to make all buffers
2101          * uptodate then we can optimize away a bogus ->readpage() for the next
2102          * read().  Here we 'discover' whether the page went uptodate as a
2103          * result of this (potentially partial) write.
2104          */
2105         if (!partial)
2106                 SetPageUptodate(page);
2107         /*
2108          * Not convinced about this at all.  See disparity comment above.  For
2109          * now we know ntfs_prepare_write() would have failed in the write
2110          * exceeds i_size case, so this will never trigger which is fine.
2111          */
2112         if (pos > i_size_read(vi)) {
2113                 ntfs_error(vi->i_sb, "Writing beyond the existing file size is "
2114                                 "not supported yet.  Sorry.");
2115                 return -EOPNOTSUPP;
2116                 // vi->i_size = pos;
2117                 // mark_inode_dirty(vi);
2118         }
2119         ntfs_debug("Done.");
2120         return 0;
2121 }
2122
2123 /**
2124  * ntfs_commit_write - commit the received data
2125  *
2126  * This is called from generic_file_write() with i_sem held on the inode
2127  * (@page->mapping->host).  The @page is locked but not kmap()ped.  The source
2128  * data has already been copied into the @page.  ntfs_prepare_write() has been
2129  * called before the data copied and it returned success so we can take the
2130  * results of various BUG checks and some error handling for granted.
2131  *
2132  * Need to mark modified blocks dirty so they get written out later when
2133  * ntfs_writepage() is invoked by the VM.
2134  *
2135  * Return 0 on success or -errno on error.
2136  *
2137  * Should be using generic_commit_write().  This marks buffers uptodate and
2138  * dirty, sets the page uptodate if all buffers in the page are uptodate, and
2139  * updates i_size if the end of io is beyond i_size.  In that case, it also
2140  * marks the inode dirty.
2141  *
2142  * Cannot use generic_commit_write() due to ntfs specialities but can look at
2143  * it for implementation guidance.
2144  *
2145  * If things have gone as outlined in ntfs_prepare_write(), then we do not
2146  * need to do any page content modifications here at all, except in the write
2147  * to resident attribute case, where we need to do the uptodate bringing here
2148  * which we combine with the copying into the mft record which means we save
2149  * one atomic kmap.
2150  */
2151 static int ntfs_commit_write(struct file *file, struct page *page,
2152                 unsigned from, unsigned to)
2153 {
2154         struct inode *vi = page->mapping->host;
2155         ntfs_inode *base_ni, *ni = NTFS_I(vi);
2156         char *kaddr, *kattr;
2157         ntfs_attr_search_ctx *ctx;
2158         MFT_RECORD *m;
2159         ATTR_RECORD *a;
2160         u32 attr_len;
2161         int err;
2162
2163         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
2164                         "0x%lx, from = %u, to = %u.", vi->i_ino, ni->type,
2165                         page->index, from, to);
2166         /* If the attribute is not resident, deal with it elsewhere. */
2167         if (NInoNonResident(ni)) {
2168                 /* Only unnamed $DATA attributes can be compressed/encrypted. */
2169                 if (ni->type == AT_DATA && !ni->name_len) {
2170                         /* Encrypted files need separate handling. */
2171                         if (NInoEncrypted(ni)) {
2172                                 // We never get here at present!
2173                                 BUG();
2174                         }
2175                         /* Compressed data streams are handled in compress.c. */
2176                         if (NInoCompressed(ni)) {
2177                                 // TODO: Implement this!
2178                                 // return ntfs_write_compressed_block(page);
2179                                 // We never get here at present!
2180                                 BUG();
2181                         }
2182                 }
2183                 /* Normal data stream. */
2184                 return ntfs_commit_nonresident_write(page, from, to);
2185         }
2186         /*
2187          * Attribute is resident, implying it is not compressed, encrypted, or
2188          * sparse.
2189          */
2190         if (!NInoAttr(ni))
2191                 base_ni = ni;
2192         else
2193                 base_ni = ni->ext.base_ntfs_ino;
2194         /* Map, pin, and lock the mft record. */
2195         m = map_mft_record(base_ni);
2196         if (IS_ERR(m)) {
2197                 err = PTR_ERR(m);
2198                 m = NULL;
2199                 ctx = NULL;
2200                 goto err_out;
2201         }
2202         ctx = ntfs_attr_get_search_ctx(base_ni, m);
2203         if (unlikely(!ctx)) {
2204                 err = -ENOMEM;
2205                 goto err_out;
2206         }
2207         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
2208                         CASE_SENSITIVE, 0, NULL, 0, ctx);
2209         if (unlikely(err)) {
2210                 if (err == -ENOENT)
2211                         err = -EIO;
2212                 goto err_out;
2213         }
2214         a = ctx->attr;
2215         /* The total length of the attribute value. */
2216         attr_len = le32_to_cpu(a->data.resident.value_length);
2217         BUG_ON(from > attr_len);
2218         kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
2219         kaddr = kmap_atomic(page, KM_USER0);
2220         /* Copy the received data from the page to the mft record. */
2221         memcpy(kattr + from, kaddr + from, to - from);
2222         /* Update the attribute length if necessary. */
2223         if (to > attr_len) {
2224                 attr_len = to;
2225                 a->data.resident.value_length = cpu_to_le32(attr_len);
2226         }
2227         /*
2228          * If the page is not uptodate, bring the out of bounds area(s)
2229          * uptodate by copying data from the mft record to the page.
2230          */
2231         if (!PageUptodate(page)) {
2232                 if (from > 0)
2233                         memcpy(kaddr, kattr, from);
2234                 if (to < attr_len)
2235                         memcpy(kaddr + to, kattr + to, attr_len - to);
2236                 /* Zero the region outside the end of the attribute value. */
2237                 if (attr_len < PAGE_CACHE_SIZE)
2238                         memset(kaddr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
2239                 /*
2240                  * The probability of not having done any of the above is
2241                  * extremely small, so we just flush unconditionally.
2242                  */
2243                 flush_dcache_page(page);
2244                 SetPageUptodate(page);
2245         }
2246         kunmap_atomic(kaddr, KM_USER0);
2247         /* Update i_size if necessary. */
2248         if (i_size_read(vi) < attr_len) {
2249                 unsigned long flags;
2250
2251                 write_lock_irqsave(&ni->size_lock, flags);
2252                 ni->allocated_size = ni->initialized_size = attr_len;
2253                 i_size_write(vi, attr_len);
2254                 write_unlock_irqrestore(&ni->size_lock, flags);
2255         }
2256         /* Mark the mft record dirty, so it gets written back. */
2257         flush_dcache_mft_record_page(ctx->ntfs_ino);
2258         mark_mft_record_dirty(ctx->ntfs_ino);
2259         ntfs_attr_put_search_ctx(ctx);
2260         unmap_mft_record(base_ni);
2261         ntfs_debug("Done.");
2262         return 0;
2263 err_out:
2264         if (err == -ENOMEM) {
2265                 ntfs_warning(vi->i_sb, "Error allocating memory required to "
2266                                 "commit the write.");
2267                 if (PageUptodate(page)) {
2268                         ntfs_warning(vi->i_sb, "Page is uptodate, setting "
2269                                         "dirty so the write will be retried "
2270                                         "later on by the VM.");
2271                         /*
2272                          * Put the page on mapping->dirty_pages, but leave its
2273                          * buffers' dirty state as-is.
2274                          */
2275                         __set_page_dirty_nobuffers(page);
2276                         err = 0;
2277                 } else
2278                         ntfs_error(vi->i_sb, "Page is not uptodate.  Written "
2279                                         "data has been lost.");
2280         } else {
2281                 ntfs_error(vi->i_sb, "Resident attribute commit write failed "
2282                                 "with error %i.", err);
2283                 NVolSetErrors(ni->vol);
2284                 make_bad_inode(vi);
2285         }
2286         if (ctx)
2287                 ntfs_attr_put_search_ctx(ctx);
2288         if (m)
2289                 unmap_mft_record(base_ni);
2290         return err;
2291 }
2292
2293 #endif  /* NTFS_RW */
2294
2295 /**
2296  * ntfs_aops - general address space operations for inodes and attributes
2297  */
2298 struct address_space_operations ntfs_aops = {
2299         .readpage       = ntfs_readpage,        /* Fill page with data. */
2300         .sync_page      = block_sync_page,      /* Currently, just unplugs the
2301                                                    disk request queue. */
2302 #ifdef NTFS_RW
2303         .writepage      = ntfs_writepage,       /* Write dirty page to disk. */
2304         .prepare_write  = ntfs_prepare_write,   /* Prepare page and buffers
2305                                                    ready to receive data. */
2306         .commit_write   = ntfs_commit_write,    /* Commit received data. */
2307 #endif /* NTFS_RW */
2308 };
2309
2310 /**
2311  * ntfs_mst_aops - general address space operations for mst protecteed inodes
2312  *                 and attributes
2313  */
2314 struct address_space_operations ntfs_mst_aops = {
2315         .readpage       = ntfs_readpage,        /* Fill page with data. */
2316         .sync_page      = block_sync_page,      /* Currently, just unplugs the
2317                                                    disk request queue. */
2318 #ifdef NTFS_RW
2319         .writepage      = ntfs_writepage,       /* Write dirty page to disk. */
2320         .set_page_dirty = __set_page_dirty_nobuffers,   /* Set the page dirty
2321                                                    without touching the buffers
2322                                                    belonging to the page. */
2323 #endif /* NTFS_RW */
2324 };
2325
2326 #ifdef NTFS_RW
2327
2328 /**
2329  * mark_ntfs_record_dirty - mark an ntfs record dirty
2330  * @page:       page containing the ntfs record to mark dirty
2331  * @ofs:        byte offset within @page at which the ntfs record begins
2332  *
2333  * Set the buffers and the page in which the ntfs record is located dirty.
2334  *
2335  * The latter also marks the vfs inode the ntfs record belongs to dirty
2336  * (I_DIRTY_PAGES only).
2337  *
2338  * If the page does not have buffers, we create them and set them uptodate.
2339  * The page may not be locked which is why we need to handle the buffers under
2340  * the mapping->private_lock.  Once the buffers are marked dirty we no longer
2341  * need the lock since try_to_free_buffers() does not free dirty buffers.
2342  */
2343 void mark_ntfs_record_dirty(struct page *page, const unsigned int ofs) {
2344         struct address_space *mapping = page->mapping;
2345         ntfs_inode *ni = NTFS_I(mapping->host);
2346         struct buffer_head *bh, *head, *buffers_to_free = NULL;
2347         unsigned int end, bh_size, bh_ofs;
2348
2349         BUG_ON(!PageUptodate(page));
2350         end = ofs + ni->itype.index.block_size;
2351         bh_size = 1 << VFS_I(ni)->i_blkbits;
2352         spin_lock(&mapping->private_lock);
2353         if (unlikely(!page_has_buffers(page))) {
2354                 spin_unlock(&mapping->private_lock);
2355                 bh = head = alloc_page_buffers(page, bh_size, 1);
2356                 spin_lock(&mapping->private_lock);
2357                 if (likely(!page_has_buffers(page))) {
2358                         struct buffer_head *tail;
2359
2360                         do {
2361                                 set_buffer_uptodate(bh);
2362                                 tail = bh;
2363                                 bh = bh->b_this_page;
2364                         } while (bh);
2365                         tail->b_this_page = head;
2366                         attach_page_buffers(page, head);
2367                 } else
2368                         buffers_to_free = bh;
2369         }
2370         bh = head = page_buffers(page);
2371         do {
2372                 bh_ofs = bh_offset(bh);
2373                 if (bh_ofs + bh_size <= ofs)
2374                         continue;
2375                 if (unlikely(bh_ofs >= end))
2376                         break;
2377                 set_buffer_dirty(bh);
2378         } while ((bh = bh->b_this_page) != head);
2379         spin_unlock(&mapping->private_lock);
2380         __set_page_dirty_nobuffers(page);
2381         if (unlikely(buffers_to_free)) {
2382                 do {
2383                         bh = buffers_to_free->b_this_page;
2384                         free_buffer_head(buffers_to_free);
2385                         buffers_to_free = bh;
2386                 } while (buffers_to_free);
2387         }
2388 }
2389
2390 #endif /* NTFS_RW */