Merge tag 'powerpc-5.11-1' of git://git.kernel.org/pub/scm/linux/kernel/git/powerpc...
[linux-2.6-microblaze.git] / fs / ntfs / aops.c
1 // SPDX-License-Identifier: GPL-2.0-or-later
2 /**
3  * aops.c - NTFS kernel address space operations and page cache handling.
4  *
5  * Copyright (c) 2001-2014 Anton Altaparmakov and Tuxera Inc.
6  * Copyright (c) 2002 Richard Russon
7  */
8
9 #include <linux/errno.h>
10 #include <linux/fs.h>
11 #include <linux/gfp.h>
12 #include <linux/mm.h>
13 #include <linux/pagemap.h>
14 #include <linux/swap.h>
15 #include <linux/buffer_head.h>
16 #include <linux/writeback.h>
17 #include <linux/bit_spinlock.h>
18 #include <linux/bio.h>
19
20 #include "aops.h"
21 #include "attrib.h"
22 #include "debug.h"
23 #include "inode.h"
24 #include "mft.h"
25 #include "runlist.h"
26 #include "types.h"
27 #include "ntfs.h"
28
29 /**
30  * ntfs_end_buffer_async_read - async io completion for reading attributes
31  * @bh:         buffer head on which io is completed
32  * @uptodate:   whether @bh is now uptodate or not
33  *
34  * Asynchronous I/O completion handler for reading pages belonging to the
35  * attribute address space of an inode.  The inodes can either be files or
36  * directories or they can be fake inodes describing some attribute.
37  *
38  * If NInoMstProtected(), perform the post read mst fixups when all IO on the
39  * page has been completed and mark the page uptodate or set the error bit on
40  * the page.  To determine the size of the records that need fixing up, we
41  * cheat a little bit by setting the index_block_size in ntfs_inode to the ntfs
42  * record size, and index_block_size_bits, to the log(base 2) of the ntfs
43  * record size.
44  */
45 static void ntfs_end_buffer_async_read(struct buffer_head *bh, int uptodate)
46 {
47         unsigned long flags;
48         struct buffer_head *first, *tmp;
49         struct page *page;
50         struct inode *vi;
51         ntfs_inode *ni;
52         int page_uptodate = 1;
53
54         page = bh->b_page;
55         vi = page->mapping->host;
56         ni = NTFS_I(vi);
57
58         if (likely(uptodate)) {
59                 loff_t i_size;
60                 s64 file_ofs, init_size;
61
62                 set_buffer_uptodate(bh);
63
64                 file_ofs = ((s64)page->index << PAGE_SHIFT) +
65                                 bh_offset(bh);
66                 read_lock_irqsave(&ni->size_lock, flags);
67                 init_size = ni->initialized_size;
68                 i_size = i_size_read(vi);
69                 read_unlock_irqrestore(&ni->size_lock, flags);
70                 if (unlikely(init_size > i_size)) {
71                         /* Race with shrinking truncate. */
72                         init_size = i_size;
73                 }
74                 /* Check for the current buffer head overflowing. */
75                 if (unlikely(file_ofs + bh->b_size > init_size)) {
76                         int ofs;
77                         void *kaddr;
78
79                         ofs = 0;
80                         if (file_ofs < init_size)
81                                 ofs = init_size - file_ofs;
82                         kaddr = kmap_atomic(page);
83                         memset(kaddr + bh_offset(bh) + ofs, 0,
84                                         bh->b_size - ofs);
85                         flush_dcache_page(page);
86                         kunmap_atomic(kaddr);
87                 }
88         } else {
89                 clear_buffer_uptodate(bh);
90                 SetPageError(page);
91                 ntfs_error(ni->vol->sb, "Buffer I/O error, logical block "
92                                 "0x%llx.", (unsigned long long)bh->b_blocknr);
93         }
94         first = page_buffers(page);
95         spin_lock_irqsave(&first->b_uptodate_lock, flags);
96         clear_buffer_async_read(bh);
97         unlock_buffer(bh);
98         tmp = bh;
99         do {
100                 if (!buffer_uptodate(tmp))
101                         page_uptodate = 0;
102                 if (buffer_async_read(tmp)) {
103                         if (likely(buffer_locked(tmp)))
104                                 goto still_busy;
105                         /* Async buffers must be locked. */
106                         BUG();
107                 }
108                 tmp = tmp->b_this_page;
109         } while (tmp != bh);
110         spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
111         /*
112          * If none of the buffers had errors then we can set the page uptodate,
113          * but we first have to perform the post read mst fixups, if the
114          * attribute is mst protected, i.e. if NInoMstProteced(ni) is true.
115          * Note we ignore fixup errors as those are detected when
116          * map_mft_record() is called which gives us per record granularity
117          * rather than per page granularity.
118          */
119         if (!NInoMstProtected(ni)) {
120                 if (likely(page_uptodate && !PageError(page)))
121                         SetPageUptodate(page);
122         } else {
123                 u8 *kaddr;
124                 unsigned int i, recs;
125                 u32 rec_size;
126
127                 rec_size = ni->itype.index.block_size;
128                 recs = PAGE_SIZE / rec_size;
129                 /* Should have been verified before we got here... */
130                 BUG_ON(!recs);
131                 kaddr = kmap_atomic(page);
132                 for (i = 0; i < recs; i++)
133                         post_read_mst_fixup((NTFS_RECORD*)(kaddr +
134                                         i * rec_size), rec_size);
135                 kunmap_atomic(kaddr);
136                 flush_dcache_page(page);
137                 if (likely(page_uptodate && !PageError(page)))
138                         SetPageUptodate(page);
139         }
140         unlock_page(page);
141         return;
142 still_busy:
143         spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
144         return;
145 }
146
147 /**
148  * ntfs_read_block - fill a @page of an address space with data
149  * @page:       page cache page to fill with data
150  *
151  * Fill the page @page of the address space belonging to the @page->host inode.
152  * We read each buffer asynchronously and when all buffers are read in, our io
153  * completion handler ntfs_end_buffer_read_async(), if required, automatically
154  * applies the mst fixups to the page before finally marking it uptodate and
155  * unlocking it.
156  *
157  * We only enforce allocated_size limit because i_size is checked for in
158  * generic_file_read().
159  *
160  * Return 0 on success and -errno on error.
161  *
162  * Contains an adapted version of fs/buffer.c::block_read_full_page().
163  */
164 static int ntfs_read_block(struct page *page)
165 {
166         loff_t i_size;
167         VCN vcn;
168         LCN lcn;
169         s64 init_size;
170         struct inode *vi;
171         ntfs_inode *ni;
172         ntfs_volume *vol;
173         runlist_element *rl;
174         struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
175         sector_t iblock, lblock, zblock;
176         unsigned long flags;
177         unsigned int blocksize, vcn_ofs;
178         int i, nr;
179         unsigned char blocksize_bits;
180
181         vi = page->mapping->host;
182         ni = NTFS_I(vi);
183         vol = ni->vol;
184
185         /* $MFT/$DATA must have its complete runlist in memory at all times. */
186         BUG_ON(!ni->runlist.rl && !ni->mft_no && !NInoAttr(ni));
187
188         blocksize = vol->sb->s_blocksize;
189         blocksize_bits = vol->sb->s_blocksize_bits;
190
191         if (!page_has_buffers(page)) {
192                 create_empty_buffers(page, blocksize, 0);
193                 if (unlikely(!page_has_buffers(page))) {
194                         unlock_page(page);
195                         return -ENOMEM;
196                 }
197         }
198         bh = head = page_buffers(page);
199         BUG_ON(!bh);
200
201         /*
202          * We may be racing with truncate.  To avoid some of the problems we
203          * now take a snapshot of the various sizes and use those for the whole
204          * of the function.  In case of an extending truncate it just means we
205          * may leave some buffers unmapped which are now allocated.  This is
206          * not a problem since these buffers will just get mapped when a write
207          * occurs.  In case of a shrinking truncate, we will detect this later
208          * on due to the runlist being incomplete and if the page is being
209          * fully truncated, truncate will throw it away as soon as we unlock
210          * it so no need to worry what we do with it.
211          */
212         iblock = (s64)page->index << (PAGE_SHIFT - blocksize_bits);
213         read_lock_irqsave(&ni->size_lock, flags);
214         lblock = (ni->allocated_size + blocksize - 1) >> blocksize_bits;
215         init_size = ni->initialized_size;
216         i_size = i_size_read(vi);
217         read_unlock_irqrestore(&ni->size_lock, flags);
218         if (unlikely(init_size > i_size)) {
219                 /* Race with shrinking truncate. */
220                 init_size = i_size;
221         }
222         zblock = (init_size + blocksize - 1) >> blocksize_bits;
223
224         /* Loop through all the buffers in the page. */
225         rl = NULL;
226         nr = i = 0;
227         do {
228                 int err = 0;
229
230                 if (unlikely(buffer_uptodate(bh)))
231                         continue;
232                 if (unlikely(buffer_mapped(bh))) {
233                         arr[nr++] = bh;
234                         continue;
235                 }
236                 bh->b_bdev = vol->sb->s_bdev;
237                 /* Is the block within the allowed limits? */
238                 if (iblock < lblock) {
239                         bool is_retry = false;
240
241                         /* Convert iblock into corresponding vcn and offset. */
242                         vcn = (VCN)iblock << blocksize_bits >>
243                                         vol->cluster_size_bits;
244                         vcn_ofs = ((VCN)iblock << blocksize_bits) &
245                                         vol->cluster_size_mask;
246                         if (!rl) {
247 lock_retry_remap:
248                                 down_read(&ni->runlist.lock);
249                                 rl = ni->runlist.rl;
250                         }
251                         if (likely(rl != NULL)) {
252                                 /* Seek to element containing target vcn. */
253                                 while (rl->length && rl[1].vcn <= vcn)
254                                         rl++;
255                                 lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
256                         } else
257                                 lcn = LCN_RL_NOT_MAPPED;
258                         /* Successful remap. */
259                         if (lcn >= 0) {
260                                 /* Setup buffer head to correct block. */
261                                 bh->b_blocknr = ((lcn << vol->cluster_size_bits)
262                                                 + vcn_ofs) >> blocksize_bits;
263                                 set_buffer_mapped(bh);
264                                 /* Only read initialized data blocks. */
265                                 if (iblock < zblock) {
266                                         arr[nr++] = bh;
267                                         continue;
268                                 }
269                                 /* Fully non-initialized data block, zero it. */
270                                 goto handle_zblock;
271                         }
272                         /* It is a hole, need to zero it. */
273                         if (lcn == LCN_HOLE)
274                                 goto handle_hole;
275                         /* If first try and runlist unmapped, map and retry. */
276                         if (!is_retry && lcn == LCN_RL_NOT_MAPPED) {
277                                 is_retry = true;
278                                 /*
279                                  * Attempt to map runlist, dropping lock for
280                                  * the duration.
281                                  */
282                                 up_read(&ni->runlist.lock);
283                                 err = ntfs_map_runlist(ni, vcn);
284                                 if (likely(!err))
285                                         goto lock_retry_remap;
286                                 rl = NULL;
287                         } else if (!rl)
288                                 up_read(&ni->runlist.lock);
289                         /*
290                          * If buffer is outside the runlist, treat it as a
291                          * hole.  This can happen due to concurrent truncate
292                          * for example.
293                          */
294                         if (err == -ENOENT || lcn == LCN_ENOENT) {
295                                 err = 0;
296                                 goto handle_hole;
297                         }
298                         /* Hard error, zero out region. */
299                         if (!err)
300                                 err = -EIO;
301                         bh->b_blocknr = -1;
302                         SetPageError(page);
303                         ntfs_error(vol->sb, "Failed to read from inode 0x%lx, "
304                                         "attribute type 0x%x, vcn 0x%llx, "
305                                         "offset 0x%x because its location on "
306                                         "disk could not be determined%s "
307                                         "(error code %i).", ni->mft_no,
308                                         ni->type, (unsigned long long)vcn,
309                                         vcn_ofs, is_retry ? " even after "
310                                         "retrying" : "", err);
311                 }
312                 /*
313                  * Either iblock was outside lblock limits or
314                  * ntfs_rl_vcn_to_lcn() returned error.  Just zero that portion
315                  * of the page and set the buffer uptodate.
316                  */
317 handle_hole:
318                 bh->b_blocknr = -1UL;
319                 clear_buffer_mapped(bh);
320 handle_zblock:
321                 zero_user(page, i * blocksize, blocksize);
322                 if (likely(!err))
323                         set_buffer_uptodate(bh);
324         } while (i++, iblock++, (bh = bh->b_this_page) != head);
325
326         /* Release the lock if we took it. */
327         if (rl)
328                 up_read(&ni->runlist.lock);
329
330         /* Check we have at least one buffer ready for i/o. */
331         if (nr) {
332                 struct buffer_head *tbh;
333
334                 /* Lock the buffers. */
335                 for (i = 0; i < nr; i++) {
336                         tbh = arr[i];
337                         lock_buffer(tbh);
338                         tbh->b_end_io = ntfs_end_buffer_async_read;
339                         set_buffer_async_read(tbh);
340                 }
341                 /* Finally, start i/o on the buffers. */
342                 for (i = 0; i < nr; i++) {
343                         tbh = arr[i];
344                         if (likely(!buffer_uptodate(tbh)))
345                                 submit_bh(REQ_OP_READ, 0, tbh);
346                         else
347                                 ntfs_end_buffer_async_read(tbh, 1);
348                 }
349                 return 0;
350         }
351         /* No i/o was scheduled on any of the buffers. */
352         if (likely(!PageError(page)))
353                 SetPageUptodate(page);
354         else /* Signal synchronous i/o error. */
355                 nr = -EIO;
356         unlock_page(page);
357         return nr;
358 }
359
360 /**
361  * ntfs_readpage - fill a @page of a @file with data from the device
362  * @file:       open file to which the page @page belongs or NULL
363  * @page:       page cache page to fill with data
364  *
365  * For non-resident attributes, ntfs_readpage() fills the @page of the open
366  * file @file by calling the ntfs version of the generic block_read_full_page()
367  * function, ntfs_read_block(), which in turn creates and reads in the buffers
368  * associated with the page asynchronously.
369  *
370  * For resident attributes, OTOH, ntfs_readpage() fills @page by copying the
371  * data from the mft record (which at this stage is most likely in memory) and
372  * fills the remainder with zeroes. Thus, in this case, I/O is synchronous, as
373  * even if the mft record is not cached at this point in time, we need to wait
374  * for it to be read in before we can do the copy.
375  *
376  * Return 0 on success and -errno on error.
377  */
378 static int ntfs_readpage(struct file *file, struct page *page)
379 {
380         loff_t i_size;
381         struct inode *vi;
382         ntfs_inode *ni, *base_ni;
383         u8 *addr;
384         ntfs_attr_search_ctx *ctx;
385         MFT_RECORD *mrec;
386         unsigned long flags;
387         u32 attr_len;
388         int err = 0;
389
390 retry_readpage:
391         BUG_ON(!PageLocked(page));
392         vi = page->mapping->host;
393         i_size = i_size_read(vi);
394         /* Is the page fully outside i_size? (truncate in progress) */
395         if (unlikely(page->index >= (i_size + PAGE_SIZE - 1) >>
396                         PAGE_SHIFT)) {
397                 zero_user(page, 0, PAGE_SIZE);
398                 ntfs_debug("Read outside i_size - truncated?");
399                 goto done;
400         }
401         /*
402          * This can potentially happen because we clear PageUptodate() during
403          * ntfs_writepage() of MstProtected() attributes.
404          */
405         if (PageUptodate(page)) {
406                 unlock_page(page);
407                 return 0;
408         }
409         ni = NTFS_I(vi);
410         /*
411          * Only $DATA attributes can be encrypted and only unnamed $DATA
412          * attributes can be compressed.  Index root can have the flags set but
413          * this means to create compressed/encrypted files, not that the
414          * attribute is compressed/encrypted.  Note we need to check for
415          * AT_INDEX_ALLOCATION since this is the type of both directory and
416          * index inodes.
417          */
418         if (ni->type != AT_INDEX_ALLOCATION) {
419                 /* If attribute is encrypted, deny access, just like NT4. */
420                 if (NInoEncrypted(ni)) {
421                         BUG_ON(ni->type != AT_DATA);
422                         err = -EACCES;
423                         goto err_out;
424                 }
425                 /* Compressed data streams are handled in compress.c. */
426                 if (NInoNonResident(ni) && NInoCompressed(ni)) {
427                         BUG_ON(ni->type != AT_DATA);
428                         BUG_ON(ni->name_len);
429                         return ntfs_read_compressed_block(page);
430                 }
431         }
432         /* NInoNonResident() == NInoIndexAllocPresent() */
433         if (NInoNonResident(ni)) {
434                 /* Normal, non-resident data stream. */
435                 return ntfs_read_block(page);
436         }
437         /*
438          * Attribute is resident, implying it is not compressed or encrypted.
439          * This also means the attribute is smaller than an mft record and
440          * hence smaller than a page, so can simply zero out any pages with
441          * index above 0.  Note the attribute can actually be marked compressed
442          * but if it is resident the actual data is not compressed so we are
443          * ok to ignore the compressed flag here.
444          */
445         if (unlikely(page->index > 0)) {
446                 zero_user(page, 0, PAGE_SIZE);
447                 goto done;
448         }
449         if (!NInoAttr(ni))
450                 base_ni = ni;
451         else
452                 base_ni = ni->ext.base_ntfs_ino;
453         /* Map, pin, and lock the mft record. */
454         mrec = map_mft_record(base_ni);
455         if (IS_ERR(mrec)) {
456                 err = PTR_ERR(mrec);
457                 goto err_out;
458         }
459         /*
460          * If a parallel write made the attribute non-resident, drop the mft
461          * record and retry the readpage.
462          */
463         if (unlikely(NInoNonResident(ni))) {
464                 unmap_mft_record(base_ni);
465                 goto retry_readpage;
466         }
467         ctx = ntfs_attr_get_search_ctx(base_ni, mrec);
468         if (unlikely(!ctx)) {
469                 err = -ENOMEM;
470                 goto unm_err_out;
471         }
472         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
473                         CASE_SENSITIVE, 0, NULL, 0, ctx);
474         if (unlikely(err))
475                 goto put_unm_err_out;
476         attr_len = le32_to_cpu(ctx->attr->data.resident.value_length);
477         read_lock_irqsave(&ni->size_lock, flags);
478         if (unlikely(attr_len > ni->initialized_size))
479                 attr_len = ni->initialized_size;
480         i_size = i_size_read(vi);
481         read_unlock_irqrestore(&ni->size_lock, flags);
482         if (unlikely(attr_len > i_size)) {
483                 /* Race with shrinking truncate. */
484                 attr_len = i_size;
485         }
486         addr = kmap_atomic(page);
487         /* Copy the data to the page. */
488         memcpy(addr, (u8*)ctx->attr +
489                         le16_to_cpu(ctx->attr->data.resident.value_offset),
490                         attr_len);
491         /* Zero the remainder of the page. */
492         memset(addr + attr_len, 0, PAGE_SIZE - attr_len);
493         flush_dcache_page(page);
494         kunmap_atomic(addr);
495 put_unm_err_out:
496         ntfs_attr_put_search_ctx(ctx);
497 unm_err_out:
498         unmap_mft_record(base_ni);
499 done:
500         SetPageUptodate(page);
501 err_out:
502         unlock_page(page);
503         return err;
504 }
505
506 #ifdef NTFS_RW
507
508 /**
509  * ntfs_write_block - write a @page to the backing store
510  * @page:       page cache page to write out
511  * @wbc:        writeback control structure
512  *
513  * This function is for writing pages belonging to non-resident, non-mst
514  * protected attributes to their backing store.
515  *
516  * For a page with buffers, map and write the dirty buffers asynchronously
517  * under page writeback. For a page without buffers, create buffers for the
518  * page, then proceed as above.
519  *
520  * If a page doesn't have buffers the page dirty state is definitive. If a page
521  * does have buffers, the page dirty state is just a hint, and the buffer dirty
522  * state is definitive. (A hint which has rules: dirty buffers against a clean
523  * page is illegal. Other combinations are legal and need to be handled. In
524  * particular a dirty page containing clean buffers for example.)
525  *
526  * Return 0 on success and -errno on error.
527  *
528  * Based on ntfs_read_block() and __block_write_full_page().
529  */
530 static int ntfs_write_block(struct page *page, struct writeback_control *wbc)
531 {
532         VCN vcn;
533         LCN lcn;
534         s64 initialized_size;
535         loff_t i_size;
536         sector_t block, dblock, iblock;
537         struct inode *vi;
538         ntfs_inode *ni;
539         ntfs_volume *vol;
540         runlist_element *rl;
541         struct buffer_head *bh, *head;
542         unsigned long flags;
543         unsigned int blocksize, vcn_ofs;
544         int err;
545         bool need_end_writeback;
546         unsigned char blocksize_bits;
547
548         vi = page->mapping->host;
549         ni = NTFS_I(vi);
550         vol = ni->vol;
551
552         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
553                         "0x%lx.", ni->mft_no, ni->type, page->index);
554
555         BUG_ON(!NInoNonResident(ni));
556         BUG_ON(NInoMstProtected(ni));
557         blocksize = vol->sb->s_blocksize;
558         blocksize_bits = vol->sb->s_blocksize_bits;
559         if (!page_has_buffers(page)) {
560                 BUG_ON(!PageUptodate(page));
561                 create_empty_buffers(page, blocksize,
562                                 (1 << BH_Uptodate) | (1 << BH_Dirty));
563                 if (unlikely(!page_has_buffers(page))) {
564                         ntfs_warning(vol->sb, "Error allocating page "
565                                         "buffers.  Redirtying page so we try "
566                                         "again later.");
567                         /*
568                          * Put the page back on mapping->dirty_pages, but leave
569                          * its buffers' dirty state as-is.
570                          */
571                         redirty_page_for_writepage(wbc, page);
572                         unlock_page(page);
573                         return 0;
574                 }
575         }
576         bh = head = page_buffers(page);
577         BUG_ON(!bh);
578
579         /* NOTE: Different naming scheme to ntfs_read_block()! */
580
581         /* The first block in the page. */
582         block = (s64)page->index << (PAGE_SHIFT - blocksize_bits);
583
584         read_lock_irqsave(&ni->size_lock, flags);
585         i_size = i_size_read(vi);
586         initialized_size = ni->initialized_size;
587         read_unlock_irqrestore(&ni->size_lock, flags);
588
589         /* The first out of bounds block for the data size. */
590         dblock = (i_size + blocksize - 1) >> blocksize_bits;
591
592         /* The last (fully or partially) initialized block. */
593         iblock = initialized_size >> blocksize_bits;
594
595         /*
596          * Be very careful.  We have no exclusion from __set_page_dirty_buffers
597          * here, and the (potentially unmapped) buffers may become dirty at
598          * any time.  If a buffer becomes dirty here after we've inspected it
599          * then we just miss that fact, and the page stays dirty.
600          *
601          * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
602          * handle that here by just cleaning them.
603          */
604
605         /*
606          * Loop through all the buffers in the page, mapping all the dirty
607          * buffers to disk addresses and handling any aliases from the
608          * underlying block device's mapping.
609          */
610         rl = NULL;
611         err = 0;
612         do {
613                 bool is_retry = false;
614
615                 if (unlikely(block >= dblock)) {
616                         /*
617                          * Mapped buffers outside i_size will occur, because
618                          * this page can be outside i_size when there is a
619                          * truncate in progress. The contents of such buffers
620                          * were zeroed by ntfs_writepage().
621                          *
622                          * FIXME: What about the small race window where
623                          * ntfs_writepage() has not done any clearing because
624                          * the page was within i_size but before we get here,
625                          * vmtruncate() modifies i_size?
626                          */
627                         clear_buffer_dirty(bh);
628                         set_buffer_uptodate(bh);
629                         continue;
630                 }
631
632                 /* Clean buffers are not written out, so no need to map them. */
633                 if (!buffer_dirty(bh))
634                         continue;
635
636                 /* Make sure we have enough initialized size. */
637                 if (unlikely((block >= iblock) &&
638                                 (initialized_size < i_size))) {
639                         /*
640                          * If this page is fully outside initialized size, zero
641                          * out all pages between the current initialized size
642                          * and the current page. Just use ntfs_readpage() to do
643                          * the zeroing transparently.
644                          */
645                         if (block > iblock) {
646                                 // TODO:
647                                 // For each page do:
648                                 // - read_cache_page()
649                                 // Again for each page do:
650                                 // - wait_on_page_locked()
651                                 // - Check (PageUptodate(page) &&
652                                 //                      !PageError(page))
653                                 // Update initialized size in the attribute and
654                                 // in the inode.
655                                 // Again, for each page do:
656                                 //      __set_page_dirty_buffers();
657                                 // put_page()
658                                 // We don't need to wait on the writes.
659                                 // Update iblock.
660                         }
661                         /*
662                          * The current page straddles initialized size. Zero
663                          * all non-uptodate buffers and set them uptodate (and
664                          * dirty?). Note, there aren't any non-uptodate buffers
665                          * if the page is uptodate.
666                          * FIXME: For an uptodate page, the buffers may need to
667                          * be written out because they were not initialized on
668                          * disk before.
669                          */
670                         if (!PageUptodate(page)) {
671                                 // TODO:
672                                 // Zero any non-uptodate buffers up to i_size.
673                                 // Set them uptodate and dirty.
674                         }
675                         // TODO:
676                         // Update initialized size in the attribute and in the
677                         // inode (up to i_size).
678                         // Update iblock.
679                         // FIXME: This is inefficient. Try to batch the two
680                         // size changes to happen in one go.
681                         ntfs_error(vol->sb, "Writing beyond initialized size "
682                                         "is not supported yet. Sorry.");
683                         err = -EOPNOTSUPP;
684                         break;
685                         // Do NOT set_buffer_new() BUT DO clear buffer range
686                         // outside write request range.
687                         // set_buffer_uptodate() on complete buffers as well as
688                         // set_buffer_dirty().
689                 }
690
691                 /* No need to map buffers that are already mapped. */
692                 if (buffer_mapped(bh))
693                         continue;
694
695                 /* Unmapped, dirty buffer. Need to map it. */
696                 bh->b_bdev = vol->sb->s_bdev;
697
698                 /* Convert block into corresponding vcn and offset. */
699                 vcn = (VCN)block << blocksize_bits;
700                 vcn_ofs = vcn & vol->cluster_size_mask;
701                 vcn >>= vol->cluster_size_bits;
702                 if (!rl) {
703 lock_retry_remap:
704                         down_read(&ni->runlist.lock);
705                         rl = ni->runlist.rl;
706                 }
707                 if (likely(rl != NULL)) {
708                         /* Seek to element containing target vcn. */
709                         while (rl->length && rl[1].vcn <= vcn)
710                                 rl++;
711                         lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
712                 } else
713                         lcn = LCN_RL_NOT_MAPPED;
714                 /* Successful remap. */
715                 if (lcn >= 0) {
716                         /* Setup buffer head to point to correct block. */
717                         bh->b_blocknr = ((lcn << vol->cluster_size_bits) +
718                                         vcn_ofs) >> blocksize_bits;
719                         set_buffer_mapped(bh);
720                         continue;
721                 }
722                 /* It is a hole, need to instantiate it. */
723                 if (lcn == LCN_HOLE) {
724                         u8 *kaddr;
725                         unsigned long *bpos, *bend;
726
727                         /* Check if the buffer is zero. */
728                         kaddr = kmap_atomic(page);
729                         bpos = (unsigned long *)(kaddr + bh_offset(bh));
730                         bend = (unsigned long *)((u8*)bpos + blocksize);
731                         do {
732                                 if (unlikely(*bpos))
733                                         break;
734                         } while (likely(++bpos < bend));
735                         kunmap_atomic(kaddr);
736                         if (bpos == bend) {
737                                 /*
738                                  * Buffer is zero and sparse, no need to write
739                                  * it.
740                                  */
741                                 bh->b_blocknr = -1;
742                                 clear_buffer_dirty(bh);
743                                 continue;
744                         }
745                         // TODO: Instantiate the hole.
746                         // clear_buffer_new(bh);
747                         // clean_bdev_bh_alias(bh);
748                         ntfs_error(vol->sb, "Writing into sparse regions is "
749                                         "not supported yet. Sorry.");
750                         err = -EOPNOTSUPP;
751                         break;
752                 }
753                 /* If first try and runlist unmapped, map and retry. */
754                 if (!is_retry && lcn == LCN_RL_NOT_MAPPED) {
755                         is_retry = true;
756                         /*
757                          * Attempt to map runlist, dropping lock for
758                          * the duration.
759                          */
760                         up_read(&ni->runlist.lock);
761                         err = ntfs_map_runlist(ni, vcn);
762                         if (likely(!err))
763                                 goto lock_retry_remap;
764                         rl = NULL;
765                 } else if (!rl)
766                         up_read(&ni->runlist.lock);
767                 /*
768                  * If buffer is outside the runlist, truncate has cut it out
769                  * of the runlist.  Just clean and clear the buffer and set it
770                  * uptodate so it can get discarded by the VM.
771                  */
772                 if (err == -ENOENT || lcn == LCN_ENOENT) {
773                         bh->b_blocknr = -1;
774                         clear_buffer_dirty(bh);
775                         zero_user(page, bh_offset(bh), blocksize);
776                         set_buffer_uptodate(bh);
777                         err = 0;
778                         continue;
779                 }
780                 /* Failed to map the buffer, even after retrying. */
781                 if (!err)
782                         err = -EIO;
783                 bh->b_blocknr = -1;
784                 ntfs_error(vol->sb, "Failed to write to inode 0x%lx, "
785                                 "attribute type 0x%x, vcn 0x%llx, offset 0x%x "
786                                 "because its location on disk could not be "
787                                 "determined%s (error code %i).", ni->mft_no,
788                                 ni->type, (unsigned long long)vcn,
789                                 vcn_ofs, is_retry ? " even after "
790                                 "retrying" : "", err);
791                 break;
792         } while (block++, (bh = bh->b_this_page) != head);
793
794         /* Release the lock if we took it. */
795         if (rl)
796                 up_read(&ni->runlist.lock);
797
798         /* For the error case, need to reset bh to the beginning. */
799         bh = head;
800
801         /* Just an optimization, so ->readpage() is not called later. */
802         if (unlikely(!PageUptodate(page))) {
803                 int uptodate = 1;
804                 do {
805                         if (!buffer_uptodate(bh)) {
806                                 uptodate = 0;
807                                 bh = head;
808                                 break;
809                         }
810                 } while ((bh = bh->b_this_page) != head);
811                 if (uptodate)
812                         SetPageUptodate(page);
813         }
814
815         /* Setup all mapped, dirty buffers for async write i/o. */
816         do {
817                 if (buffer_mapped(bh) && buffer_dirty(bh)) {
818                         lock_buffer(bh);
819                         if (test_clear_buffer_dirty(bh)) {
820                                 BUG_ON(!buffer_uptodate(bh));
821                                 mark_buffer_async_write(bh);
822                         } else
823                                 unlock_buffer(bh);
824                 } else if (unlikely(err)) {
825                         /*
826                          * For the error case. The buffer may have been set
827                          * dirty during attachment to a dirty page.
828                          */
829                         if (err != -ENOMEM)
830                                 clear_buffer_dirty(bh);
831                 }
832         } while ((bh = bh->b_this_page) != head);
833
834         if (unlikely(err)) {
835                 // TODO: Remove the -EOPNOTSUPP check later on...
836                 if (unlikely(err == -EOPNOTSUPP))
837                         err = 0;
838                 else if (err == -ENOMEM) {
839                         ntfs_warning(vol->sb, "Error allocating memory. "
840                                         "Redirtying page so we try again "
841                                         "later.");
842                         /*
843                          * Put the page back on mapping->dirty_pages, but
844                          * leave its buffer's dirty state as-is.
845                          */
846                         redirty_page_for_writepage(wbc, page);
847                         err = 0;
848                 } else
849                         SetPageError(page);
850         }
851
852         BUG_ON(PageWriteback(page));
853         set_page_writeback(page);       /* Keeps try_to_free_buffers() away. */
854
855         /* Submit the prepared buffers for i/o. */
856         need_end_writeback = true;
857         do {
858                 struct buffer_head *next = bh->b_this_page;
859                 if (buffer_async_write(bh)) {
860                         submit_bh(REQ_OP_WRITE, 0, bh);
861                         need_end_writeback = false;
862                 }
863                 bh = next;
864         } while (bh != head);
865         unlock_page(page);
866
867         /* If no i/o was started, need to end_page_writeback(). */
868         if (unlikely(need_end_writeback))
869                 end_page_writeback(page);
870
871         ntfs_debug("Done.");
872         return err;
873 }
874
875 /**
876  * ntfs_write_mst_block - write a @page to the backing store
877  * @page:       page cache page to write out
878  * @wbc:        writeback control structure
879  *
880  * This function is for writing pages belonging to non-resident, mst protected
881  * attributes to their backing store.  The only supported attributes are index
882  * allocation and $MFT/$DATA.  Both directory inodes and index inodes are
883  * supported for the index allocation case.
884  *
885  * The page must remain locked for the duration of the write because we apply
886  * the mst fixups, write, and then undo the fixups, so if we were to unlock the
887  * page before undoing the fixups, any other user of the page will see the
888  * page contents as corrupt.
889  *
890  * We clear the page uptodate flag for the duration of the function to ensure
891  * exclusion for the $MFT/$DATA case against someone mapping an mft record we
892  * are about to apply the mst fixups to.
893  *
894  * Return 0 on success and -errno on error.
895  *
896  * Based on ntfs_write_block(), ntfs_mft_writepage(), and
897  * write_mft_record_nolock().
898  */
899 static int ntfs_write_mst_block(struct page *page,
900                 struct writeback_control *wbc)
901 {
902         sector_t block, dblock, rec_block;
903         struct inode *vi = page->mapping->host;
904         ntfs_inode *ni = NTFS_I(vi);
905         ntfs_volume *vol = ni->vol;
906         u8 *kaddr;
907         unsigned int rec_size = ni->itype.index.block_size;
908         ntfs_inode *locked_nis[PAGE_SIZE / NTFS_BLOCK_SIZE];
909         struct buffer_head *bh, *head, *tbh, *rec_start_bh;
910         struct buffer_head *bhs[MAX_BUF_PER_PAGE];
911         runlist_element *rl;
912         int i, nr_locked_nis, nr_recs, nr_bhs, max_bhs, bhs_per_rec, err, err2;
913         unsigned bh_size, rec_size_bits;
914         bool sync, is_mft, page_is_dirty, rec_is_dirty;
915         unsigned char bh_size_bits;
916
917         if (WARN_ON(rec_size < NTFS_BLOCK_SIZE))
918                 return -EINVAL;
919
920         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
921                         "0x%lx.", vi->i_ino, ni->type, page->index);
922         BUG_ON(!NInoNonResident(ni));
923         BUG_ON(!NInoMstProtected(ni));
924         is_mft = (S_ISREG(vi->i_mode) && !vi->i_ino);
925         /*
926          * NOTE: ntfs_write_mst_block() would be called for $MFTMirr if a page
927          * in its page cache were to be marked dirty.  However this should
928          * never happen with the current driver and considering we do not
929          * handle this case here we do want to BUG(), at least for now.
930          */
931         BUG_ON(!(is_mft || S_ISDIR(vi->i_mode) ||
932                         (NInoAttr(ni) && ni->type == AT_INDEX_ALLOCATION)));
933         bh_size = vol->sb->s_blocksize;
934         bh_size_bits = vol->sb->s_blocksize_bits;
935         max_bhs = PAGE_SIZE / bh_size;
936         BUG_ON(!max_bhs);
937         BUG_ON(max_bhs > MAX_BUF_PER_PAGE);
938
939         /* Were we called for sync purposes? */
940         sync = (wbc->sync_mode == WB_SYNC_ALL);
941
942         /* Make sure we have mapped buffers. */
943         bh = head = page_buffers(page);
944         BUG_ON(!bh);
945
946         rec_size_bits = ni->itype.index.block_size_bits;
947         BUG_ON(!(PAGE_SIZE >> rec_size_bits));
948         bhs_per_rec = rec_size >> bh_size_bits;
949         BUG_ON(!bhs_per_rec);
950
951         /* The first block in the page. */
952         rec_block = block = (sector_t)page->index <<
953                         (PAGE_SHIFT - bh_size_bits);
954
955         /* The first out of bounds block for the data size. */
956         dblock = (i_size_read(vi) + bh_size - 1) >> bh_size_bits;
957
958         rl = NULL;
959         err = err2 = nr_bhs = nr_recs = nr_locked_nis = 0;
960         page_is_dirty = rec_is_dirty = false;
961         rec_start_bh = NULL;
962         do {
963                 bool is_retry = false;
964
965                 if (likely(block < rec_block)) {
966                         if (unlikely(block >= dblock)) {
967                                 clear_buffer_dirty(bh);
968                                 set_buffer_uptodate(bh);
969                                 continue;
970                         }
971                         /*
972                          * This block is not the first one in the record.  We
973                          * ignore the buffer's dirty state because we could
974                          * have raced with a parallel mark_ntfs_record_dirty().
975                          */
976                         if (!rec_is_dirty)
977                                 continue;
978                         if (unlikely(err2)) {
979                                 if (err2 != -ENOMEM)
980                                         clear_buffer_dirty(bh);
981                                 continue;
982                         }
983                 } else /* if (block == rec_block) */ {
984                         BUG_ON(block > rec_block);
985                         /* This block is the first one in the record. */
986                         rec_block += bhs_per_rec;
987                         err2 = 0;
988                         if (unlikely(block >= dblock)) {
989                                 clear_buffer_dirty(bh);
990                                 continue;
991                         }
992                         if (!buffer_dirty(bh)) {
993                                 /* Clean records are not written out. */
994                                 rec_is_dirty = false;
995                                 continue;
996                         }
997                         rec_is_dirty = true;
998                         rec_start_bh = bh;
999                 }
1000                 /* Need to map the buffer if it is not mapped already. */
1001                 if (unlikely(!buffer_mapped(bh))) {
1002                         VCN vcn;
1003                         LCN lcn;
1004                         unsigned int vcn_ofs;
1005
1006                         bh->b_bdev = vol->sb->s_bdev;
1007                         /* Obtain the vcn and offset of the current block. */
1008                         vcn = (VCN)block << bh_size_bits;
1009                         vcn_ofs = vcn & vol->cluster_size_mask;
1010                         vcn >>= vol->cluster_size_bits;
1011                         if (!rl) {
1012 lock_retry_remap:
1013                                 down_read(&ni->runlist.lock);
1014                                 rl = ni->runlist.rl;
1015                         }
1016                         if (likely(rl != NULL)) {
1017                                 /* Seek to element containing target vcn. */
1018                                 while (rl->length && rl[1].vcn <= vcn)
1019                                         rl++;
1020                                 lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
1021                         } else
1022                                 lcn = LCN_RL_NOT_MAPPED;
1023                         /* Successful remap. */
1024                         if (likely(lcn >= 0)) {
1025                                 /* Setup buffer head to correct block. */
1026                                 bh->b_blocknr = ((lcn <<
1027                                                 vol->cluster_size_bits) +
1028                                                 vcn_ofs) >> bh_size_bits;
1029                                 set_buffer_mapped(bh);
1030                         } else {
1031                                 /*
1032                                  * Remap failed.  Retry to map the runlist once
1033                                  * unless we are working on $MFT which always
1034                                  * has the whole of its runlist in memory.
1035                                  */
1036                                 if (!is_mft && !is_retry &&
1037                                                 lcn == LCN_RL_NOT_MAPPED) {
1038                                         is_retry = true;
1039                                         /*
1040                                          * Attempt to map runlist, dropping
1041                                          * lock for the duration.
1042                                          */
1043                                         up_read(&ni->runlist.lock);
1044                                         err2 = ntfs_map_runlist(ni, vcn);
1045                                         if (likely(!err2))
1046                                                 goto lock_retry_remap;
1047                                         if (err2 == -ENOMEM)
1048                                                 page_is_dirty = true;
1049                                         lcn = err2;
1050                                 } else {
1051                                         err2 = -EIO;
1052                                         if (!rl)
1053                                                 up_read(&ni->runlist.lock);
1054                                 }
1055                                 /* Hard error.  Abort writing this record. */
1056                                 if (!err || err == -ENOMEM)
1057                                         err = err2;
1058                                 bh->b_blocknr = -1;
1059                                 ntfs_error(vol->sb, "Cannot write ntfs record "
1060                                                 "0x%llx (inode 0x%lx, "
1061                                                 "attribute type 0x%x) because "
1062                                                 "its location on disk could "
1063                                                 "not be determined (error "
1064                                                 "code %lli).",
1065                                                 (long long)block <<
1066                                                 bh_size_bits >>
1067                                                 vol->mft_record_size_bits,
1068                                                 ni->mft_no, ni->type,
1069                                                 (long long)lcn);
1070                                 /*
1071                                  * If this is not the first buffer, remove the
1072                                  * buffers in this record from the list of
1073                                  * buffers to write and clear their dirty bit
1074                                  * if not error -ENOMEM.
1075                                  */
1076                                 if (rec_start_bh != bh) {
1077                                         while (bhs[--nr_bhs] != rec_start_bh)
1078                                                 ;
1079                                         if (err2 != -ENOMEM) {
1080                                                 do {
1081                                                         clear_buffer_dirty(
1082                                                                 rec_start_bh);
1083                                                 } while ((rec_start_bh =
1084                                                                 rec_start_bh->
1085                                                                 b_this_page) !=
1086                                                                 bh);
1087                                         }
1088                                 }
1089                                 continue;
1090                         }
1091                 }
1092                 BUG_ON(!buffer_uptodate(bh));
1093                 BUG_ON(nr_bhs >= max_bhs);
1094                 bhs[nr_bhs++] = bh;
1095         } while (block++, (bh = bh->b_this_page) != head);
1096         if (unlikely(rl))
1097                 up_read(&ni->runlist.lock);
1098         /* If there were no dirty buffers, we are done. */
1099         if (!nr_bhs)
1100                 goto done;
1101         /* Map the page so we can access its contents. */
1102         kaddr = kmap(page);
1103         /* Clear the page uptodate flag whilst the mst fixups are applied. */
1104         BUG_ON(!PageUptodate(page));
1105         ClearPageUptodate(page);
1106         for (i = 0; i < nr_bhs; i++) {
1107                 unsigned int ofs;
1108
1109                 /* Skip buffers which are not at the beginning of records. */
1110                 if (i % bhs_per_rec)
1111                         continue;
1112                 tbh = bhs[i];
1113                 ofs = bh_offset(tbh);
1114                 if (is_mft) {
1115                         ntfs_inode *tni;
1116                         unsigned long mft_no;
1117
1118                         /* Get the mft record number. */
1119                         mft_no = (((s64)page->index << PAGE_SHIFT) + ofs)
1120                                         >> rec_size_bits;
1121                         /* Check whether to write this mft record. */
1122                         tni = NULL;
1123                         if (!ntfs_may_write_mft_record(vol, mft_no,
1124                                         (MFT_RECORD*)(kaddr + ofs), &tni)) {
1125                                 /*
1126                                  * The record should not be written.  This
1127                                  * means we need to redirty the page before
1128                                  * returning.
1129                                  */
1130                                 page_is_dirty = true;
1131                                 /*
1132                                  * Remove the buffers in this mft record from
1133                                  * the list of buffers to write.
1134                                  */
1135                                 do {
1136                                         bhs[i] = NULL;
1137                                 } while (++i % bhs_per_rec);
1138                                 continue;
1139                         }
1140                         /*
1141                          * The record should be written.  If a locked ntfs
1142                          * inode was returned, add it to the array of locked
1143                          * ntfs inodes.
1144                          */
1145                         if (tni)
1146                                 locked_nis[nr_locked_nis++] = tni;
1147                 }
1148                 /* Apply the mst protection fixups. */
1149                 err2 = pre_write_mst_fixup((NTFS_RECORD*)(kaddr + ofs),
1150                                 rec_size);
1151                 if (unlikely(err2)) {
1152                         if (!err || err == -ENOMEM)
1153                                 err = -EIO;
1154                         ntfs_error(vol->sb, "Failed to apply mst fixups "
1155                                         "(inode 0x%lx, attribute type 0x%x, "
1156                                         "page index 0x%lx, page offset 0x%x)!"
1157                                         "  Unmount and run chkdsk.", vi->i_ino,
1158                                         ni->type, page->index, ofs);
1159                         /*
1160                          * Mark all the buffers in this record clean as we do
1161                          * not want to write corrupt data to disk.
1162                          */
1163                         do {
1164                                 clear_buffer_dirty(bhs[i]);
1165                                 bhs[i] = NULL;
1166                         } while (++i % bhs_per_rec);
1167                         continue;
1168                 }
1169                 nr_recs++;
1170         }
1171         /* If no records are to be written out, we are done. */
1172         if (!nr_recs)
1173                 goto unm_done;
1174         flush_dcache_page(page);
1175         /* Lock buffers and start synchronous write i/o on them. */
1176         for (i = 0; i < nr_bhs; i++) {
1177                 tbh = bhs[i];
1178                 if (!tbh)
1179                         continue;
1180                 if (!trylock_buffer(tbh))
1181                         BUG();
1182                 /* The buffer dirty state is now irrelevant, just clean it. */
1183                 clear_buffer_dirty(tbh);
1184                 BUG_ON(!buffer_uptodate(tbh));
1185                 BUG_ON(!buffer_mapped(tbh));
1186                 get_bh(tbh);
1187                 tbh->b_end_io = end_buffer_write_sync;
1188                 submit_bh(REQ_OP_WRITE, 0, tbh);
1189         }
1190         /* Synchronize the mft mirror now if not @sync. */
1191         if (is_mft && !sync)
1192                 goto do_mirror;
1193 do_wait:
1194         /* Wait on i/o completion of buffers. */
1195         for (i = 0; i < nr_bhs; i++) {
1196                 tbh = bhs[i];
1197                 if (!tbh)
1198                         continue;
1199                 wait_on_buffer(tbh);
1200                 if (unlikely(!buffer_uptodate(tbh))) {
1201                         ntfs_error(vol->sb, "I/O error while writing ntfs "
1202                                         "record buffer (inode 0x%lx, "
1203                                         "attribute type 0x%x, page index "
1204                                         "0x%lx, page offset 0x%lx)!  Unmount "
1205                                         "and run chkdsk.", vi->i_ino, ni->type,
1206                                         page->index, bh_offset(tbh));
1207                         if (!err || err == -ENOMEM)
1208                                 err = -EIO;
1209                         /*
1210                          * Set the buffer uptodate so the page and buffer
1211                          * states do not become out of sync.
1212                          */
1213                         set_buffer_uptodate(tbh);
1214                 }
1215         }
1216         /* If @sync, now synchronize the mft mirror. */
1217         if (is_mft && sync) {
1218 do_mirror:
1219                 for (i = 0; i < nr_bhs; i++) {
1220                         unsigned long mft_no;
1221                         unsigned int ofs;
1222
1223                         /*
1224                          * Skip buffers which are not at the beginning of
1225                          * records.
1226                          */
1227                         if (i % bhs_per_rec)
1228                                 continue;
1229                         tbh = bhs[i];
1230                         /* Skip removed buffers (and hence records). */
1231                         if (!tbh)
1232                                 continue;
1233                         ofs = bh_offset(tbh);
1234                         /* Get the mft record number. */
1235                         mft_no = (((s64)page->index << PAGE_SHIFT) + ofs)
1236                                         >> rec_size_bits;
1237                         if (mft_no < vol->mftmirr_size)
1238                                 ntfs_sync_mft_mirror(vol, mft_no,
1239                                                 (MFT_RECORD*)(kaddr + ofs),
1240                                                 sync);
1241                 }
1242                 if (!sync)
1243                         goto do_wait;
1244         }
1245         /* Remove the mst protection fixups again. */
1246         for (i = 0; i < nr_bhs; i++) {
1247                 if (!(i % bhs_per_rec)) {
1248                         tbh = bhs[i];
1249                         if (!tbh)
1250                                 continue;
1251                         post_write_mst_fixup((NTFS_RECORD*)(kaddr +
1252                                         bh_offset(tbh)));
1253                 }
1254         }
1255         flush_dcache_page(page);
1256 unm_done:
1257         /* Unlock any locked inodes. */
1258         while (nr_locked_nis-- > 0) {
1259                 ntfs_inode *tni, *base_tni;
1260                 
1261                 tni = locked_nis[nr_locked_nis];
1262                 /* Get the base inode. */
1263                 mutex_lock(&tni->extent_lock);
1264                 if (tni->nr_extents >= 0)
1265                         base_tni = tni;
1266                 else {
1267                         base_tni = tni->ext.base_ntfs_ino;
1268                         BUG_ON(!base_tni);
1269                 }
1270                 mutex_unlock(&tni->extent_lock);
1271                 ntfs_debug("Unlocking %s inode 0x%lx.",
1272                                 tni == base_tni ? "base" : "extent",
1273                                 tni->mft_no);
1274                 mutex_unlock(&tni->mrec_lock);
1275                 atomic_dec(&tni->count);
1276                 iput(VFS_I(base_tni));
1277         }
1278         SetPageUptodate(page);
1279         kunmap(page);
1280 done:
1281         if (unlikely(err && err != -ENOMEM)) {
1282                 /*
1283                  * Set page error if there is only one ntfs record in the page.
1284                  * Otherwise we would loose per-record granularity.
1285                  */
1286                 if (ni->itype.index.block_size == PAGE_SIZE)
1287                         SetPageError(page);
1288                 NVolSetErrors(vol);
1289         }
1290         if (page_is_dirty) {
1291                 ntfs_debug("Page still contains one or more dirty ntfs "
1292                                 "records.  Redirtying the page starting at "
1293                                 "record 0x%lx.", page->index <<
1294                                 (PAGE_SHIFT - rec_size_bits));
1295                 redirty_page_for_writepage(wbc, page);
1296                 unlock_page(page);
1297         } else {
1298                 /*
1299                  * Keep the VM happy.  This must be done otherwise the
1300                  * radix-tree tag PAGECACHE_TAG_DIRTY remains set even though
1301                  * the page is clean.
1302                  */
1303                 BUG_ON(PageWriteback(page));
1304                 set_page_writeback(page);
1305                 unlock_page(page);
1306                 end_page_writeback(page);
1307         }
1308         if (likely(!err))
1309                 ntfs_debug("Done.");
1310         return err;
1311 }
1312
1313 /**
1314  * ntfs_writepage - write a @page to the backing store
1315  * @page:       page cache page to write out
1316  * @wbc:        writeback control structure
1317  *
1318  * This is called from the VM when it wants to have a dirty ntfs page cache
1319  * page cleaned.  The VM has already locked the page and marked it clean.
1320  *
1321  * For non-resident attributes, ntfs_writepage() writes the @page by calling
1322  * the ntfs version of the generic block_write_full_page() function,
1323  * ntfs_write_block(), which in turn if necessary creates and writes the
1324  * buffers associated with the page asynchronously.
1325  *
1326  * For resident attributes, OTOH, ntfs_writepage() writes the @page by copying
1327  * the data to the mft record (which at this stage is most likely in memory).
1328  * The mft record is then marked dirty and written out asynchronously via the
1329  * vfs inode dirty code path for the inode the mft record belongs to or via the
1330  * vm page dirty code path for the page the mft record is in.
1331  *
1332  * Based on ntfs_readpage() and fs/buffer.c::block_write_full_page().
1333  *
1334  * Return 0 on success and -errno on error.
1335  */
1336 static int ntfs_writepage(struct page *page, struct writeback_control *wbc)
1337 {
1338         loff_t i_size;
1339         struct inode *vi = page->mapping->host;
1340         ntfs_inode *base_ni = NULL, *ni = NTFS_I(vi);
1341         char *addr;
1342         ntfs_attr_search_ctx *ctx = NULL;
1343         MFT_RECORD *m = NULL;
1344         u32 attr_len;
1345         int err;
1346
1347 retry_writepage:
1348         BUG_ON(!PageLocked(page));
1349         i_size = i_size_read(vi);
1350         /* Is the page fully outside i_size? (truncate in progress) */
1351         if (unlikely(page->index >= (i_size + PAGE_SIZE - 1) >>
1352                         PAGE_SHIFT)) {
1353                 /*
1354                  * The page may have dirty, unmapped buffers.  Make them
1355                  * freeable here, so the page does not leak.
1356                  */
1357                 block_invalidatepage(page, 0, PAGE_SIZE);
1358                 unlock_page(page);
1359                 ntfs_debug("Write outside i_size - truncated?");
1360                 return 0;
1361         }
1362         /*
1363          * Only $DATA attributes can be encrypted and only unnamed $DATA
1364          * attributes can be compressed.  Index root can have the flags set but
1365          * this means to create compressed/encrypted files, not that the
1366          * attribute is compressed/encrypted.  Note we need to check for
1367          * AT_INDEX_ALLOCATION since this is the type of both directory and
1368          * index inodes.
1369          */
1370         if (ni->type != AT_INDEX_ALLOCATION) {
1371                 /* If file is encrypted, deny access, just like NT4. */
1372                 if (NInoEncrypted(ni)) {
1373                         unlock_page(page);
1374                         BUG_ON(ni->type != AT_DATA);
1375                         ntfs_debug("Denying write access to encrypted file.");
1376                         return -EACCES;
1377                 }
1378                 /* Compressed data streams are handled in compress.c. */
1379                 if (NInoNonResident(ni) && NInoCompressed(ni)) {
1380                         BUG_ON(ni->type != AT_DATA);
1381                         BUG_ON(ni->name_len);
1382                         // TODO: Implement and replace this with
1383                         // return ntfs_write_compressed_block(page);
1384                         unlock_page(page);
1385                         ntfs_error(vi->i_sb, "Writing to compressed files is "
1386                                         "not supported yet.  Sorry.");
1387                         return -EOPNOTSUPP;
1388                 }
1389                 // TODO: Implement and remove this check.
1390                 if (NInoNonResident(ni) && NInoSparse(ni)) {
1391                         unlock_page(page);
1392                         ntfs_error(vi->i_sb, "Writing to sparse files is not "
1393                                         "supported yet.  Sorry.");
1394                         return -EOPNOTSUPP;
1395                 }
1396         }
1397         /* NInoNonResident() == NInoIndexAllocPresent() */
1398         if (NInoNonResident(ni)) {
1399                 /* We have to zero every time due to mmap-at-end-of-file. */
1400                 if (page->index >= (i_size >> PAGE_SHIFT)) {
1401                         /* The page straddles i_size. */
1402                         unsigned int ofs = i_size & ~PAGE_MASK;
1403                         zero_user_segment(page, ofs, PAGE_SIZE);
1404                 }
1405                 /* Handle mst protected attributes. */
1406                 if (NInoMstProtected(ni))
1407                         return ntfs_write_mst_block(page, wbc);
1408                 /* Normal, non-resident data stream. */
1409                 return ntfs_write_block(page, wbc);
1410         }
1411         /*
1412          * Attribute is resident, implying it is not compressed, encrypted, or
1413          * mst protected.  This also means the attribute is smaller than an mft
1414          * record and hence smaller than a page, so can simply return error on
1415          * any pages with index above 0.  Note the attribute can actually be
1416          * marked compressed but if it is resident the actual data is not
1417          * compressed so we are ok to ignore the compressed flag here.
1418          */
1419         BUG_ON(page_has_buffers(page));
1420         BUG_ON(!PageUptodate(page));
1421         if (unlikely(page->index > 0)) {
1422                 ntfs_error(vi->i_sb, "BUG()! page->index (0x%lx) > 0.  "
1423                                 "Aborting write.", page->index);
1424                 BUG_ON(PageWriteback(page));
1425                 set_page_writeback(page);
1426                 unlock_page(page);
1427                 end_page_writeback(page);
1428                 return -EIO;
1429         }
1430         if (!NInoAttr(ni))
1431                 base_ni = ni;
1432         else
1433                 base_ni = ni->ext.base_ntfs_ino;
1434         /* Map, pin, and lock the mft record. */
1435         m = map_mft_record(base_ni);
1436         if (IS_ERR(m)) {
1437                 err = PTR_ERR(m);
1438                 m = NULL;
1439                 ctx = NULL;
1440                 goto err_out;
1441         }
1442         /*
1443          * If a parallel write made the attribute non-resident, drop the mft
1444          * record and retry the writepage.
1445          */
1446         if (unlikely(NInoNonResident(ni))) {
1447                 unmap_mft_record(base_ni);
1448                 goto retry_writepage;
1449         }
1450         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1451         if (unlikely(!ctx)) {
1452                 err = -ENOMEM;
1453                 goto err_out;
1454         }
1455         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1456                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1457         if (unlikely(err))
1458                 goto err_out;
1459         /*
1460          * Keep the VM happy.  This must be done otherwise the radix-tree tag
1461          * PAGECACHE_TAG_DIRTY remains set even though the page is clean.
1462          */
1463         BUG_ON(PageWriteback(page));
1464         set_page_writeback(page);
1465         unlock_page(page);
1466         attr_len = le32_to_cpu(ctx->attr->data.resident.value_length);
1467         i_size = i_size_read(vi);
1468         if (unlikely(attr_len > i_size)) {
1469                 /* Race with shrinking truncate or a failed truncate. */
1470                 attr_len = i_size;
1471                 /*
1472                  * If the truncate failed, fix it up now.  If a concurrent
1473                  * truncate, we do its job, so it does not have to do anything.
1474                  */
1475                 err = ntfs_resident_attr_value_resize(ctx->mrec, ctx->attr,
1476                                 attr_len);
1477                 /* Shrinking cannot fail. */
1478                 BUG_ON(err);
1479         }
1480         addr = kmap_atomic(page);
1481         /* Copy the data from the page to the mft record. */
1482         memcpy((u8*)ctx->attr +
1483                         le16_to_cpu(ctx->attr->data.resident.value_offset),
1484                         addr, attr_len);
1485         /* Zero out of bounds area in the page cache page. */
1486         memset(addr + attr_len, 0, PAGE_SIZE - attr_len);
1487         kunmap_atomic(addr);
1488         flush_dcache_page(page);
1489         flush_dcache_mft_record_page(ctx->ntfs_ino);
1490         /* We are done with the page. */
1491         end_page_writeback(page);
1492         /* Finally, mark the mft record dirty, so it gets written back. */
1493         mark_mft_record_dirty(ctx->ntfs_ino);
1494         ntfs_attr_put_search_ctx(ctx);
1495         unmap_mft_record(base_ni);
1496         return 0;
1497 err_out:
1498         if (err == -ENOMEM) {
1499                 ntfs_warning(vi->i_sb, "Error allocating memory. Redirtying "
1500                                 "page so we try again later.");
1501                 /*
1502                  * Put the page back on mapping->dirty_pages, but leave its
1503                  * buffers' dirty state as-is.
1504                  */
1505                 redirty_page_for_writepage(wbc, page);
1506                 err = 0;
1507         } else {
1508                 ntfs_error(vi->i_sb, "Resident attribute write failed with "
1509                                 "error %i.", err);
1510                 SetPageError(page);
1511                 NVolSetErrors(ni->vol);
1512         }
1513         unlock_page(page);
1514         if (ctx)
1515                 ntfs_attr_put_search_ctx(ctx);
1516         if (m)
1517                 unmap_mft_record(base_ni);
1518         return err;
1519 }
1520
1521 #endif  /* NTFS_RW */
1522
1523 /**
1524  * ntfs_bmap - map logical file block to physical device block
1525  * @mapping:    address space mapping to which the block to be mapped belongs
1526  * @block:      logical block to map to its physical device block
1527  *
1528  * For regular, non-resident files (i.e. not compressed and not encrypted), map
1529  * the logical @block belonging to the file described by the address space
1530  * mapping @mapping to its physical device block.
1531  *
1532  * The size of the block is equal to the @s_blocksize field of the super block
1533  * of the mounted file system which is guaranteed to be smaller than or equal
1534  * to the cluster size thus the block is guaranteed to fit entirely inside the
1535  * cluster which means we do not need to care how many contiguous bytes are
1536  * available after the beginning of the block.
1537  *
1538  * Return the physical device block if the mapping succeeded or 0 if the block
1539  * is sparse or there was an error.
1540  *
1541  * Note: This is a problem if someone tries to run bmap() on $Boot system file
1542  * as that really is in block zero but there is nothing we can do.  bmap() is
1543  * just broken in that respect (just like it cannot distinguish sparse from
1544  * not available or error).
1545  */
1546 static sector_t ntfs_bmap(struct address_space *mapping, sector_t block)
1547 {
1548         s64 ofs, size;
1549         loff_t i_size;
1550         LCN lcn;
1551         unsigned long blocksize, flags;
1552         ntfs_inode *ni = NTFS_I(mapping->host);
1553         ntfs_volume *vol = ni->vol;
1554         unsigned delta;
1555         unsigned char blocksize_bits, cluster_size_shift;
1556
1557         ntfs_debug("Entering for mft_no 0x%lx, logical block 0x%llx.",
1558                         ni->mft_no, (unsigned long long)block);
1559         if (ni->type != AT_DATA || !NInoNonResident(ni) || NInoEncrypted(ni)) {
1560                 ntfs_error(vol->sb, "BMAP does not make sense for %s "
1561                                 "attributes, returning 0.",
1562                                 (ni->type != AT_DATA) ? "non-data" :
1563                                 (!NInoNonResident(ni) ? "resident" :
1564                                 "encrypted"));
1565                 return 0;
1566         }
1567         /* None of these can happen. */
1568         BUG_ON(NInoCompressed(ni));
1569         BUG_ON(NInoMstProtected(ni));
1570         blocksize = vol->sb->s_blocksize;
1571         blocksize_bits = vol->sb->s_blocksize_bits;
1572         ofs = (s64)block << blocksize_bits;
1573         read_lock_irqsave(&ni->size_lock, flags);
1574         size = ni->initialized_size;
1575         i_size = i_size_read(VFS_I(ni));
1576         read_unlock_irqrestore(&ni->size_lock, flags);
1577         /*
1578          * If the offset is outside the initialized size or the block straddles
1579          * the initialized size then pretend it is a hole unless the
1580          * initialized size equals the file size.
1581          */
1582         if (unlikely(ofs >= size || (ofs + blocksize > size && size < i_size)))
1583                 goto hole;
1584         cluster_size_shift = vol->cluster_size_bits;
1585         down_read(&ni->runlist.lock);
1586         lcn = ntfs_attr_vcn_to_lcn_nolock(ni, ofs >> cluster_size_shift, false);
1587         up_read(&ni->runlist.lock);
1588         if (unlikely(lcn < LCN_HOLE)) {
1589                 /*
1590                  * Step down to an integer to avoid gcc doing a long long
1591                  * comparision in the switch when we know @lcn is between
1592                  * LCN_HOLE and LCN_EIO (i.e. -1 to -5).
1593                  *
1594                  * Otherwise older gcc (at least on some architectures) will
1595                  * try to use __cmpdi2() which is of course not available in
1596                  * the kernel.
1597                  */
1598                 switch ((int)lcn) {
1599                 case LCN_ENOENT:
1600                         /*
1601                          * If the offset is out of bounds then pretend it is a
1602                          * hole.
1603                          */
1604                         goto hole;
1605                 case LCN_ENOMEM:
1606                         ntfs_error(vol->sb, "Not enough memory to complete "
1607                                         "mapping for inode 0x%lx.  "
1608                                         "Returning 0.", ni->mft_no);
1609                         break;
1610                 default:
1611                         ntfs_error(vol->sb, "Failed to complete mapping for "
1612                                         "inode 0x%lx.  Run chkdsk.  "
1613                                         "Returning 0.", ni->mft_no);
1614                         break;
1615                 }
1616                 return 0;
1617         }
1618         if (lcn < 0) {
1619                 /* It is a hole. */
1620 hole:
1621                 ntfs_debug("Done (returning hole).");
1622                 return 0;
1623         }
1624         /*
1625          * The block is really allocated and fullfils all our criteria.
1626          * Convert the cluster to units of block size and return the result.
1627          */
1628         delta = ofs & vol->cluster_size_mask;
1629         if (unlikely(sizeof(block) < sizeof(lcn))) {
1630                 block = lcn = ((lcn << cluster_size_shift) + delta) >>
1631                                 blocksize_bits;
1632                 /* If the block number was truncated return 0. */
1633                 if (unlikely(block != lcn)) {
1634                         ntfs_error(vol->sb, "Physical block 0x%llx is too "
1635                                         "large to be returned, returning 0.",
1636                                         (long long)lcn);
1637                         return 0;
1638                 }
1639         } else
1640                 block = ((lcn << cluster_size_shift) + delta) >>
1641                                 blocksize_bits;
1642         ntfs_debug("Done (returning block 0x%llx).", (unsigned long long)lcn);
1643         return block;
1644 }
1645
1646 /**
1647  * ntfs_normal_aops - address space operations for normal inodes and attributes
1648  *
1649  * Note these are not used for compressed or mst protected inodes and
1650  * attributes.
1651  */
1652 const struct address_space_operations ntfs_normal_aops = {
1653         .readpage       = ntfs_readpage,
1654 #ifdef NTFS_RW
1655         .writepage      = ntfs_writepage,
1656         .set_page_dirty = __set_page_dirty_buffers,
1657 #endif /* NTFS_RW */
1658         .bmap           = ntfs_bmap,
1659         .migratepage    = buffer_migrate_page,
1660         .is_partially_uptodate = block_is_partially_uptodate,
1661         .error_remove_page = generic_error_remove_page,
1662 };
1663
1664 /**
1665  * ntfs_compressed_aops - address space operations for compressed inodes
1666  */
1667 const struct address_space_operations ntfs_compressed_aops = {
1668         .readpage       = ntfs_readpage,
1669 #ifdef NTFS_RW
1670         .writepage      = ntfs_writepage,
1671         .set_page_dirty = __set_page_dirty_buffers,
1672 #endif /* NTFS_RW */
1673         .migratepage    = buffer_migrate_page,
1674         .is_partially_uptodate = block_is_partially_uptodate,
1675         .error_remove_page = generic_error_remove_page,
1676 };
1677
1678 /**
1679  * ntfs_mst_aops - general address space operations for mst protecteed inodes
1680  *                 and attributes
1681  */
1682 const struct address_space_operations ntfs_mst_aops = {
1683         .readpage       = ntfs_readpage,        /* Fill page with data. */
1684 #ifdef NTFS_RW
1685         .writepage      = ntfs_writepage,       /* Write dirty page to disk. */
1686         .set_page_dirty = __set_page_dirty_nobuffers,   /* Set the page dirty
1687                                                    without touching the buffers
1688                                                    belonging to the page. */
1689 #endif /* NTFS_RW */
1690         .migratepage    = buffer_migrate_page,
1691         .is_partially_uptodate  = block_is_partially_uptodate,
1692         .error_remove_page = generic_error_remove_page,
1693 };
1694
1695 #ifdef NTFS_RW
1696
1697 /**
1698  * mark_ntfs_record_dirty - mark an ntfs record dirty
1699  * @page:       page containing the ntfs record to mark dirty
1700  * @ofs:        byte offset within @page at which the ntfs record begins
1701  *
1702  * Set the buffers and the page in which the ntfs record is located dirty.
1703  *
1704  * The latter also marks the vfs inode the ntfs record belongs to dirty
1705  * (I_DIRTY_PAGES only).
1706  *
1707  * If the page does not have buffers, we create them and set them uptodate.
1708  * The page may not be locked which is why we need to handle the buffers under
1709  * the mapping->private_lock.  Once the buffers are marked dirty we no longer
1710  * need the lock since try_to_free_buffers() does not free dirty buffers.
1711  */
1712 void mark_ntfs_record_dirty(struct page *page, const unsigned int ofs) {
1713         struct address_space *mapping = page->mapping;
1714         ntfs_inode *ni = NTFS_I(mapping->host);
1715         struct buffer_head *bh, *head, *buffers_to_free = NULL;
1716         unsigned int end, bh_size, bh_ofs;
1717
1718         BUG_ON(!PageUptodate(page));
1719         end = ofs + ni->itype.index.block_size;
1720         bh_size = VFS_I(ni)->i_sb->s_blocksize;
1721         spin_lock(&mapping->private_lock);
1722         if (unlikely(!page_has_buffers(page))) {
1723                 spin_unlock(&mapping->private_lock);
1724                 bh = head = alloc_page_buffers(page, bh_size, true);
1725                 spin_lock(&mapping->private_lock);
1726                 if (likely(!page_has_buffers(page))) {
1727                         struct buffer_head *tail;
1728
1729                         do {
1730                                 set_buffer_uptodate(bh);
1731                                 tail = bh;
1732                                 bh = bh->b_this_page;
1733                         } while (bh);
1734                         tail->b_this_page = head;
1735                         attach_page_private(page, head);
1736                 } else
1737                         buffers_to_free = bh;
1738         }
1739         bh = head = page_buffers(page);
1740         BUG_ON(!bh);
1741         do {
1742                 bh_ofs = bh_offset(bh);
1743                 if (bh_ofs + bh_size <= ofs)
1744                         continue;
1745                 if (unlikely(bh_ofs >= end))
1746                         break;
1747                 set_buffer_dirty(bh);
1748         } while ((bh = bh->b_this_page) != head);
1749         spin_unlock(&mapping->private_lock);
1750         __set_page_dirty_nobuffers(page);
1751         if (unlikely(buffers_to_free)) {
1752                 do {
1753                         bh = buffers_to_free->b_this_page;
1754                         free_buffer_head(buffers_to_free);
1755                         buffers_to_free = bh;
1756                 } while (buffers_to_free);
1757         }
1758 }
1759
1760 #endif /* NTFS_RW */