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