Merge branch 'x86-mem-for-linus' of git://git.kernel.org/pub/scm/linux/kernel/git...
[pandora-kernel.git] / fs / ntfs / file.c
1 /*
2  * file.c - NTFS kernel file operations.  Part of the Linux-NTFS project.
3  *
4  * Copyright (c) 2001-2011 Anton Altaparmakov and Tuxera Inc.
5  *
6  * This program/include file is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public License as published
8  * by the Free Software Foundation; either version 2 of the License, or
9  * (at your option) any later version.
10  *
11  * This program/include file is distributed in the hope that it will be
12  * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
13  * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
14  * GNU General Public License for more details.
15  *
16  * You should have received a copy of the GNU General Public License
17  * along with this program (in the main directory of the Linux-NTFS
18  * distribution in the file COPYING); if not, write to the Free Software
19  * Foundation,Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
20  */
21
22 #include <linux/buffer_head.h>
23 #include <linux/gfp.h>
24 #include <linux/pagemap.h>
25 #include <linux/pagevec.h>
26 #include <linux/sched.h>
27 #include <linux/swap.h>
28 #include <linux/uio.h>
29 #include <linux/writeback.h>
30
31 #include <asm/page.h>
32 #include <asm/uaccess.h>
33
34 #include "attrib.h"
35 #include "bitmap.h"
36 #include "inode.h"
37 #include "debug.h"
38 #include "lcnalloc.h"
39 #include "malloc.h"
40 #include "mft.h"
41 #include "ntfs.h"
42
43 /**
44  * ntfs_file_open - called when an inode is about to be opened
45  * @vi:         inode to be opened
46  * @filp:       file structure describing the inode
47  *
48  * Limit file size to the page cache limit on architectures where unsigned long
49  * is 32-bits. This is the most we can do for now without overflowing the page
50  * cache page index. Doing it this way means we don't run into problems because
51  * of existing too large files. It would be better to allow the user to read
52  * the beginning of the file but I doubt very much anyone is going to hit this
53  * check on a 32-bit architecture, so there is no point in adding the extra
54  * complexity required to support this.
55  *
56  * On 64-bit architectures, the check is hopefully optimized away by the
57  * compiler.
58  *
59  * After the check passes, just call generic_file_open() to do its work.
60  */
61 static int ntfs_file_open(struct inode *vi, struct file *filp)
62 {
63         if (sizeof(unsigned long) < 8) {
64                 if (i_size_read(vi) > MAX_LFS_FILESIZE)
65                         return -EOVERFLOW;
66         }
67         return generic_file_open(vi, filp);
68 }
69
70 #ifdef NTFS_RW
71
72 /**
73  * ntfs_attr_extend_initialized - extend the initialized size of an attribute
74  * @ni:                 ntfs inode of the attribute to extend
75  * @new_init_size:      requested new initialized size in bytes
76  * @cached_page:        store any allocated but unused page here
77  * @lru_pvec:           lru-buffering pagevec of the caller
78  *
79  * Extend the initialized size of an attribute described by the ntfs inode @ni
80  * to @new_init_size bytes.  This involves zeroing any non-sparse space between
81  * the old initialized size and @new_init_size both in the page cache and on
82  * disk (if relevant complete pages are already uptodate in the page cache then
83  * these are simply marked dirty).
84  *
85  * As a side-effect, the file size (vfs inode->i_size) may be incremented as,
86  * in the resident attribute case, it is tied to the initialized size and, in
87  * the non-resident attribute case, it may not fall below the initialized size.
88  *
89  * Note that if the attribute is resident, we do not need to touch the page
90  * cache at all.  This is because if the page cache page is not uptodate we
91  * bring it uptodate later, when doing the write to the mft record since we
92  * then already have the page mapped.  And if the page is uptodate, the
93  * non-initialized region will already have been zeroed when the page was
94  * brought uptodate and the region may in fact already have been overwritten
95  * with new data via mmap() based writes, so we cannot just zero it.  And since
96  * POSIX specifies that the behaviour of resizing a file whilst it is mmap()ped
97  * is unspecified, we choose not to do zeroing and thus we do not need to touch
98  * the page at all.  For a more detailed explanation see ntfs_truncate() in
99  * fs/ntfs/inode.c.
100  *
101  * Return 0 on success and -errno on error.  In the case that an error is
102  * encountered it is possible that the initialized size will already have been
103  * incremented some way towards @new_init_size but it is guaranteed that if
104  * this is the case, the necessary zeroing will also have happened and that all
105  * metadata is self-consistent.
106  *
107  * Locking: i_mutex on the vfs inode corrseponsind to the ntfs inode @ni must be
108  *          held by the caller.
109  */
110 static int ntfs_attr_extend_initialized(ntfs_inode *ni, const s64 new_init_size)
111 {
112         s64 old_init_size;
113         loff_t old_i_size;
114         pgoff_t index, end_index;
115         unsigned long flags;
116         struct inode *vi = VFS_I(ni);
117         ntfs_inode *base_ni;
118         MFT_RECORD *m = NULL;
119         ATTR_RECORD *a;
120         ntfs_attr_search_ctx *ctx = NULL;
121         struct address_space *mapping;
122         struct page *page = NULL;
123         u8 *kattr;
124         int err;
125         u32 attr_len;
126
127         read_lock_irqsave(&ni->size_lock, flags);
128         old_init_size = ni->initialized_size;
129         old_i_size = i_size_read(vi);
130         BUG_ON(new_init_size > ni->allocated_size);
131         read_unlock_irqrestore(&ni->size_lock, flags);
132         ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, "
133                         "old_initialized_size 0x%llx, "
134                         "new_initialized_size 0x%llx, i_size 0x%llx.",
135                         vi->i_ino, (unsigned)le32_to_cpu(ni->type),
136                         (unsigned long long)old_init_size,
137                         (unsigned long long)new_init_size, old_i_size);
138         if (!NInoAttr(ni))
139                 base_ni = ni;
140         else
141                 base_ni = ni->ext.base_ntfs_ino;
142         /* Use goto to reduce indentation and we need the label below anyway. */
143         if (NInoNonResident(ni))
144                 goto do_non_resident_extend;
145         BUG_ON(old_init_size != old_i_size);
146         m = map_mft_record(base_ni);
147         if (IS_ERR(m)) {
148                 err = PTR_ERR(m);
149                 m = NULL;
150                 goto err_out;
151         }
152         ctx = ntfs_attr_get_search_ctx(base_ni, m);
153         if (unlikely(!ctx)) {
154                 err = -ENOMEM;
155                 goto err_out;
156         }
157         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
158                         CASE_SENSITIVE, 0, NULL, 0, ctx);
159         if (unlikely(err)) {
160                 if (err == -ENOENT)
161                         err = -EIO;
162                 goto err_out;
163         }
164         m = ctx->mrec;
165         a = ctx->attr;
166         BUG_ON(a->non_resident);
167         /* The total length of the attribute value. */
168         attr_len = le32_to_cpu(a->data.resident.value_length);
169         BUG_ON(old_i_size != (loff_t)attr_len);
170         /*
171          * Do the zeroing in the mft record and update the attribute size in
172          * the mft record.
173          */
174         kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
175         memset(kattr + attr_len, 0, new_init_size - attr_len);
176         a->data.resident.value_length = cpu_to_le32((u32)new_init_size);
177         /* Finally, update the sizes in the vfs and ntfs inodes. */
178         write_lock_irqsave(&ni->size_lock, flags);
179         i_size_write(vi, new_init_size);
180         ni->initialized_size = new_init_size;
181         write_unlock_irqrestore(&ni->size_lock, flags);
182         goto done;
183 do_non_resident_extend:
184         /*
185          * If the new initialized size @new_init_size exceeds the current file
186          * size (vfs inode->i_size), we need to extend the file size to the
187          * new initialized size.
188          */
189         if (new_init_size > old_i_size) {
190                 m = map_mft_record(base_ni);
191                 if (IS_ERR(m)) {
192                         err = PTR_ERR(m);
193                         m = NULL;
194                         goto err_out;
195                 }
196                 ctx = ntfs_attr_get_search_ctx(base_ni, m);
197                 if (unlikely(!ctx)) {
198                         err = -ENOMEM;
199                         goto err_out;
200                 }
201                 err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
202                                 CASE_SENSITIVE, 0, NULL, 0, ctx);
203                 if (unlikely(err)) {
204                         if (err == -ENOENT)
205                                 err = -EIO;
206                         goto err_out;
207                 }
208                 m = ctx->mrec;
209                 a = ctx->attr;
210                 BUG_ON(!a->non_resident);
211                 BUG_ON(old_i_size != (loff_t)
212                                 sle64_to_cpu(a->data.non_resident.data_size));
213                 a->data.non_resident.data_size = cpu_to_sle64(new_init_size);
214                 flush_dcache_mft_record_page(ctx->ntfs_ino);
215                 mark_mft_record_dirty(ctx->ntfs_ino);
216                 /* Update the file size in the vfs inode. */
217                 i_size_write(vi, new_init_size);
218                 ntfs_attr_put_search_ctx(ctx);
219                 ctx = NULL;
220                 unmap_mft_record(base_ni);
221                 m = NULL;
222         }
223         mapping = vi->i_mapping;
224         index = old_init_size >> PAGE_CACHE_SHIFT;
225         end_index = (new_init_size + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
226         do {
227                 /*
228                  * Read the page.  If the page is not present, this will zero
229                  * the uninitialized regions for us.
230                  */
231                 page = read_mapping_page(mapping, index, NULL);
232                 if (IS_ERR(page)) {
233                         err = PTR_ERR(page);
234                         goto init_err_out;
235                 }
236                 if (unlikely(PageError(page))) {
237                         page_cache_release(page);
238                         err = -EIO;
239                         goto init_err_out;
240                 }
241                 /*
242                  * Update the initialized size in the ntfs inode.  This is
243                  * enough to make ntfs_writepage() work.
244                  */
245                 write_lock_irqsave(&ni->size_lock, flags);
246                 ni->initialized_size = (s64)(index + 1) << PAGE_CACHE_SHIFT;
247                 if (ni->initialized_size > new_init_size)
248                         ni->initialized_size = new_init_size;
249                 write_unlock_irqrestore(&ni->size_lock, flags);
250                 /* Set the page dirty so it gets written out. */
251                 set_page_dirty(page);
252                 page_cache_release(page);
253                 /*
254                  * Play nice with the vm and the rest of the system.  This is
255                  * very much needed as we can potentially be modifying the
256                  * initialised size from a very small value to a really huge
257                  * value, e.g.
258                  *      f = open(somefile, O_TRUNC);
259                  *      truncate(f, 10GiB);
260                  *      seek(f, 10GiB);
261                  *      write(f, 1);
262                  * And this would mean we would be marking dirty hundreds of
263                  * thousands of pages or as in the above example more than
264                  * two and a half million pages!
265                  *
266                  * TODO: For sparse pages could optimize this workload by using
267                  * the FsMisc / MiscFs page bit as a "PageIsSparse" bit.  This
268                  * would be set in readpage for sparse pages and here we would
269                  * not need to mark dirty any pages which have this bit set.
270                  * The only caveat is that we have to clear the bit everywhere
271                  * where we allocate any clusters that lie in the page or that
272                  * contain the page.
273                  *
274                  * TODO: An even greater optimization would be for us to only
275                  * call readpage() on pages which are not in sparse regions as
276                  * determined from the runlist.  This would greatly reduce the
277                  * number of pages we read and make dirty in the case of sparse
278                  * files.
279                  */
280                 balance_dirty_pages_ratelimited(mapping);
281                 cond_resched();
282         } while (++index < end_index);
283         read_lock_irqsave(&ni->size_lock, flags);
284         BUG_ON(ni->initialized_size != new_init_size);
285         read_unlock_irqrestore(&ni->size_lock, flags);
286         /* Now bring in sync the initialized_size in the mft record. */
287         m = map_mft_record(base_ni);
288         if (IS_ERR(m)) {
289                 err = PTR_ERR(m);
290                 m = NULL;
291                 goto init_err_out;
292         }
293         ctx = ntfs_attr_get_search_ctx(base_ni, m);
294         if (unlikely(!ctx)) {
295                 err = -ENOMEM;
296                 goto init_err_out;
297         }
298         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
299                         CASE_SENSITIVE, 0, NULL, 0, ctx);
300         if (unlikely(err)) {
301                 if (err == -ENOENT)
302                         err = -EIO;
303                 goto init_err_out;
304         }
305         m = ctx->mrec;
306         a = ctx->attr;
307         BUG_ON(!a->non_resident);
308         a->data.non_resident.initialized_size = cpu_to_sle64(new_init_size);
309 done:
310         flush_dcache_mft_record_page(ctx->ntfs_ino);
311         mark_mft_record_dirty(ctx->ntfs_ino);
312         if (ctx)
313                 ntfs_attr_put_search_ctx(ctx);
314         if (m)
315                 unmap_mft_record(base_ni);
316         ntfs_debug("Done, initialized_size 0x%llx, i_size 0x%llx.",
317                         (unsigned long long)new_init_size, i_size_read(vi));
318         return 0;
319 init_err_out:
320         write_lock_irqsave(&ni->size_lock, flags);
321         ni->initialized_size = old_init_size;
322         write_unlock_irqrestore(&ni->size_lock, flags);
323 err_out:
324         if (ctx)
325                 ntfs_attr_put_search_ctx(ctx);
326         if (m)
327                 unmap_mft_record(base_ni);
328         ntfs_debug("Failed.  Returning error code %i.", err);
329         return err;
330 }
331
332 /**
333  * ntfs_fault_in_pages_readable -
334  *
335  * Fault a number of userspace pages into pagetables.
336  *
337  * Unlike include/linux/pagemap.h::fault_in_pages_readable(), this one copes
338  * with more than two userspace pages as well as handling the single page case
339  * elegantly.
340  *
341  * If you find this difficult to understand, then think of the while loop being
342  * the following code, except that we do without the integer variable ret:
343  *
344  *      do {
345  *              ret = __get_user(c, uaddr);
346  *              uaddr += PAGE_SIZE;
347  *      } while (!ret && uaddr < end);
348  *
349  * Note, the final __get_user() may well run out-of-bounds of the user buffer,
350  * but _not_ out-of-bounds of the page the user buffer belongs to, and since
351  * this is only a read and not a write, and since it is still in the same page,
352  * it should not matter and this makes the code much simpler.
353  */
354 static inline void ntfs_fault_in_pages_readable(const char __user *uaddr,
355                 int bytes)
356 {
357         const char __user *end;
358         volatile char c;
359
360         /* Set @end to the first byte outside the last page we care about. */
361         end = (const char __user*)PAGE_ALIGN((unsigned long)uaddr + bytes);
362
363         while (!__get_user(c, uaddr) && (uaddr += PAGE_SIZE, uaddr < end))
364                 ;
365 }
366
367 /**
368  * ntfs_fault_in_pages_readable_iovec -
369  *
370  * Same as ntfs_fault_in_pages_readable() but operates on an array of iovecs.
371  */
372 static inline void ntfs_fault_in_pages_readable_iovec(const struct iovec *iov,
373                 size_t iov_ofs, int bytes)
374 {
375         do {
376                 const char __user *buf;
377                 unsigned len;
378
379                 buf = iov->iov_base + iov_ofs;
380                 len = iov->iov_len - iov_ofs;
381                 if (len > bytes)
382                         len = bytes;
383                 ntfs_fault_in_pages_readable(buf, len);
384                 bytes -= len;
385                 iov++;
386                 iov_ofs = 0;
387         } while (bytes);
388 }
389
390 /**
391  * __ntfs_grab_cache_pages - obtain a number of locked pages
392  * @mapping:    address space mapping from which to obtain page cache pages
393  * @index:      starting index in @mapping at which to begin obtaining pages
394  * @nr_pages:   number of page cache pages to obtain
395  * @pages:      array of pages in which to return the obtained page cache pages
396  * @cached_page: allocated but as yet unused page
397  * @lru_pvec:   lru-buffering pagevec of caller
398  *
399  * Obtain @nr_pages locked page cache pages from the mapping @mapping and
400  * starting at index @index.
401  *
402  * If a page is newly created, add it to lru list
403  *
404  * Note, the page locks are obtained in ascending page index order.
405  */
406 static inline int __ntfs_grab_cache_pages(struct address_space *mapping,
407                 pgoff_t index, const unsigned nr_pages, struct page **pages,
408                 struct page **cached_page)
409 {
410         int err, nr;
411
412         BUG_ON(!nr_pages);
413         err = nr = 0;
414         do {
415                 pages[nr] = find_lock_page(mapping, index);
416                 if (!pages[nr]) {
417                         if (!*cached_page) {
418                                 *cached_page = page_cache_alloc(mapping);
419                                 if (unlikely(!*cached_page)) {
420                                         err = -ENOMEM;
421                                         goto err_out;
422                                 }
423                         }
424                         err = add_to_page_cache_lru(*cached_page, mapping, index,
425                                         GFP_KERNEL);
426                         if (unlikely(err)) {
427                                 if (err == -EEXIST)
428                                         continue;
429                                 goto err_out;
430                         }
431                         pages[nr] = *cached_page;
432                         *cached_page = NULL;
433                 }
434                 index++;
435                 nr++;
436         } while (nr < nr_pages);
437 out:
438         return err;
439 err_out:
440         while (nr > 0) {
441                 unlock_page(pages[--nr]);
442                 page_cache_release(pages[nr]);
443         }
444         goto out;
445 }
446
447 static inline int ntfs_submit_bh_for_read(struct buffer_head *bh)
448 {
449         lock_buffer(bh);
450         get_bh(bh);
451         bh->b_end_io = end_buffer_read_sync;
452         return submit_bh(READ, bh);
453 }
454
455 /**
456  * ntfs_prepare_pages_for_non_resident_write - prepare pages for receiving data
457  * @pages:      array of destination pages
458  * @nr_pages:   number of pages in @pages
459  * @pos:        byte position in file at which the write begins
460  * @bytes:      number of bytes to be written
461  *
462  * This is called for non-resident attributes from ntfs_file_buffered_write()
463  * with i_mutex held on the inode (@pages[0]->mapping->host).  There are
464  * @nr_pages pages in @pages which are locked but not kmap()ped.  The source
465  * data has not yet been copied into the @pages.
466  * 
467  * Need to fill any holes with actual clusters, allocate buffers if necessary,
468  * ensure all the buffers are mapped, and bring uptodate any buffers that are
469  * only partially being written to.
470  *
471  * If @nr_pages is greater than one, we are guaranteed that the cluster size is
472  * greater than PAGE_CACHE_SIZE, that all pages in @pages are entirely inside
473  * the same cluster and that they are the entirety of that cluster, and that
474  * the cluster is sparse, i.e. we need to allocate a cluster to fill the hole.
475  *
476  * i_size is not to be modified yet.
477  *
478  * Return 0 on success or -errno on error.
479  */
480 static int ntfs_prepare_pages_for_non_resident_write(struct page **pages,
481                 unsigned nr_pages, s64 pos, size_t bytes)
482 {
483         VCN vcn, highest_vcn = 0, cpos, cend, bh_cpos, bh_cend;
484         LCN lcn;
485         s64 bh_pos, vcn_len, end, initialized_size;
486         sector_t lcn_block;
487         struct page *page;
488         struct inode *vi;
489         ntfs_inode *ni, *base_ni = NULL;
490         ntfs_volume *vol;
491         runlist_element *rl, *rl2;
492         struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
493         ntfs_attr_search_ctx *ctx = NULL;
494         MFT_RECORD *m = NULL;
495         ATTR_RECORD *a = NULL;
496         unsigned long flags;
497         u32 attr_rec_len = 0;
498         unsigned blocksize, u;
499         int err, mp_size;
500         bool rl_write_locked, was_hole, is_retry;
501         unsigned char blocksize_bits;
502         struct {
503                 u8 runlist_merged:1;
504                 u8 mft_attr_mapped:1;
505                 u8 mp_rebuilt:1;
506                 u8 attr_switched:1;
507         } status = { 0, 0, 0, 0 };
508
509         BUG_ON(!nr_pages);
510         BUG_ON(!pages);
511         BUG_ON(!*pages);
512         vi = pages[0]->mapping->host;
513         ni = NTFS_I(vi);
514         vol = ni->vol;
515         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, start page "
516                         "index 0x%lx, nr_pages 0x%x, pos 0x%llx, bytes 0x%zx.",
517                         vi->i_ino, ni->type, pages[0]->index, nr_pages,
518                         (long long)pos, bytes);
519         blocksize = vol->sb->s_blocksize;
520         blocksize_bits = vol->sb->s_blocksize_bits;
521         u = 0;
522         do {
523                 page = pages[u];
524                 BUG_ON(!page);
525                 /*
526                  * create_empty_buffers() will create uptodate/dirty buffers if
527                  * the page is uptodate/dirty.
528                  */
529                 if (!page_has_buffers(page)) {
530                         create_empty_buffers(page, blocksize, 0);
531                         if (unlikely(!page_has_buffers(page)))
532                                 return -ENOMEM;
533                 }
534         } while (++u < nr_pages);
535         rl_write_locked = false;
536         rl = NULL;
537         err = 0;
538         vcn = lcn = -1;
539         vcn_len = 0;
540         lcn_block = -1;
541         was_hole = false;
542         cpos = pos >> vol->cluster_size_bits;
543         end = pos + bytes;
544         cend = (end + vol->cluster_size - 1) >> vol->cluster_size_bits;
545         /*
546          * Loop over each page and for each page over each buffer.  Use goto to
547          * reduce indentation.
548          */
549         u = 0;
550 do_next_page:
551         page = pages[u];
552         bh_pos = (s64)page->index << PAGE_CACHE_SHIFT;
553         bh = head = page_buffers(page);
554         do {
555                 VCN cdelta;
556                 s64 bh_end;
557                 unsigned bh_cofs;
558
559                 /* Clear buffer_new on all buffers to reinitialise state. */
560                 if (buffer_new(bh))
561                         clear_buffer_new(bh);
562                 bh_end = bh_pos + blocksize;
563                 bh_cpos = bh_pos >> vol->cluster_size_bits;
564                 bh_cofs = bh_pos & vol->cluster_size_mask;
565                 if (buffer_mapped(bh)) {
566                         /*
567                          * The buffer is already mapped.  If it is uptodate,
568                          * ignore it.
569                          */
570                         if (buffer_uptodate(bh))
571                                 continue;
572                         /*
573                          * The buffer is not uptodate.  If the page is uptodate
574                          * set the buffer uptodate and otherwise ignore it.
575                          */
576                         if (PageUptodate(page)) {
577                                 set_buffer_uptodate(bh);
578                                 continue;
579                         }
580                         /*
581                          * Neither the page nor the buffer are uptodate.  If
582                          * the buffer is only partially being written to, we
583                          * need to read it in before the write, i.e. now.
584                          */
585                         if ((bh_pos < pos && bh_end > pos) ||
586                                         (bh_pos < end && bh_end > end)) {
587                                 /*
588                                  * If the buffer is fully or partially within
589                                  * the initialized size, do an actual read.
590                                  * Otherwise, simply zero the buffer.
591                                  */
592                                 read_lock_irqsave(&ni->size_lock, flags);
593                                 initialized_size = ni->initialized_size;
594                                 read_unlock_irqrestore(&ni->size_lock, flags);
595                                 if (bh_pos < initialized_size) {
596                                         ntfs_submit_bh_for_read(bh);
597                                         *wait_bh++ = bh;
598                                 } else {
599                                         zero_user(page, bh_offset(bh),
600                                                         blocksize);
601                                         set_buffer_uptodate(bh);
602                                 }
603                         }
604                         continue;
605                 }
606                 /* Unmapped buffer.  Need to map it. */
607                 bh->b_bdev = vol->sb->s_bdev;
608                 /*
609                  * If the current buffer is in the same clusters as the map
610                  * cache, there is no need to check the runlist again.  The
611                  * map cache is made up of @vcn, which is the first cached file
612                  * cluster, @vcn_len which is the number of cached file
613                  * clusters, @lcn is the device cluster corresponding to @vcn,
614                  * and @lcn_block is the block number corresponding to @lcn.
615                  */
616                 cdelta = bh_cpos - vcn;
617                 if (likely(!cdelta || (cdelta > 0 && cdelta < vcn_len))) {
618 map_buffer_cached:
619                         BUG_ON(lcn < 0);
620                         bh->b_blocknr = lcn_block +
621                                         (cdelta << (vol->cluster_size_bits -
622                                         blocksize_bits)) +
623                                         (bh_cofs >> blocksize_bits);
624                         set_buffer_mapped(bh);
625                         /*
626                          * If the page is uptodate so is the buffer.  If the
627                          * buffer is fully outside the write, we ignore it if
628                          * it was already allocated and we mark it dirty so it
629                          * gets written out if we allocated it.  On the other
630                          * hand, if we allocated the buffer but we are not
631                          * marking it dirty we set buffer_new so we can do
632                          * error recovery.
633                          */
634                         if (PageUptodate(page)) {
635                                 if (!buffer_uptodate(bh))
636                                         set_buffer_uptodate(bh);
637                                 if (unlikely(was_hole)) {
638                                         /* We allocated the buffer. */
639                                         unmap_underlying_metadata(bh->b_bdev,
640                                                         bh->b_blocknr);
641                                         if (bh_end <= pos || bh_pos >= end)
642                                                 mark_buffer_dirty(bh);
643                                         else
644                                                 set_buffer_new(bh);
645                                 }
646                                 continue;
647                         }
648                         /* Page is _not_ uptodate. */
649                         if (likely(!was_hole)) {
650                                 /*
651                                  * Buffer was already allocated.  If it is not
652                                  * uptodate and is only partially being written
653                                  * to, we need to read it in before the write,
654                                  * i.e. now.
655                                  */
656                                 if (!buffer_uptodate(bh) && bh_pos < end &&
657                                                 bh_end > pos &&
658                                                 (bh_pos < pos ||
659                                                 bh_end > end)) {
660                                         /*
661                                          * If the buffer is fully or partially
662                                          * within the initialized size, do an
663                                          * actual read.  Otherwise, simply zero
664                                          * the buffer.
665                                          */
666                                         read_lock_irqsave(&ni->size_lock,
667                                                         flags);
668                                         initialized_size = ni->initialized_size;
669                                         read_unlock_irqrestore(&ni->size_lock,
670                                                         flags);
671                                         if (bh_pos < initialized_size) {
672                                                 ntfs_submit_bh_for_read(bh);
673                                                 *wait_bh++ = bh;
674                                         } else {
675                                                 zero_user(page, bh_offset(bh),
676                                                                 blocksize);
677                                                 set_buffer_uptodate(bh);
678                                         }
679                                 }
680                                 continue;
681                         }
682                         /* We allocated the buffer. */
683                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
684                         /*
685                          * If the buffer is fully outside the write, zero it,
686                          * set it uptodate, and mark it dirty so it gets
687                          * written out.  If it is partially being written to,
688                          * zero region surrounding the write but leave it to
689                          * commit write to do anything else.  Finally, if the
690                          * buffer is fully being overwritten, do nothing.
691                          */
692                         if (bh_end <= pos || bh_pos >= end) {
693                                 if (!buffer_uptodate(bh)) {
694                                         zero_user(page, bh_offset(bh),
695                                                         blocksize);
696                                         set_buffer_uptodate(bh);
697                                 }
698                                 mark_buffer_dirty(bh);
699                                 continue;
700                         }
701                         set_buffer_new(bh);
702                         if (!buffer_uptodate(bh) &&
703                                         (bh_pos < pos || bh_end > end)) {
704                                 u8 *kaddr;
705                                 unsigned pofs;
706                                         
707                                 kaddr = kmap_atomic(page, KM_USER0);
708                                 if (bh_pos < pos) {
709                                         pofs = bh_pos & ~PAGE_CACHE_MASK;
710                                         memset(kaddr + pofs, 0, pos - bh_pos);
711                                 }
712                                 if (bh_end > end) {
713                                         pofs = end & ~PAGE_CACHE_MASK;
714                                         memset(kaddr + pofs, 0, bh_end - end);
715                                 }
716                                 kunmap_atomic(kaddr, KM_USER0);
717                                 flush_dcache_page(page);
718                         }
719                         continue;
720                 }
721                 /*
722                  * Slow path: this is the first buffer in the cluster.  If it
723                  * is outside allocated size and is not uptodate, zero it and
724                  * set it uptodate.
725                  */
726                 read_lock_irqsave(&ni->size_lock, flags);
727                 initialized_size = ni->allocated_size;
728                 read_unlock_irqrestore(&ni->size_lock, flags);
729                 if (bh_pos > initialized_size) {
730                         if (PageUptodate(page)) {
731                                 if (!buffer_uptodate(bh))
732                                         set_buffer_uptodate(bh);
733                         } else if (!buffer_uptodate(bh)) {
734                                 zero_user(page, bh_offset(bh), blocksize);
735                                 set_buffer_uptodate(bh);
736                         }
737                         continue;
738                 }
739                 is_retry = false;
740                 if (!rl) {
741                         down_read(&ni->runlist.lock);
742 retry_remap:
743                         rl = ni->runlist.rl;
744                 }
745                 if (likely(rl != NULL)) {
746                         /* Seek to element containing target cluster. */
747                         while (rl->length && rl[1].vcn <= bh_cpos)
748                                 rl++;
749                         lcn = ntfs_rl_vcn_to_lcn(rl, bh_cpos);
750                         if (likely(lcn >= 0)) {
751                                 /*
752                                  * Successful remap, setup the map cache and
753                                  * use that to deal with the buffer.
754                                  */
755                                 was_hole = false;
756                                 vcn = bh_cpos;
757                                 vcn_len = rl[1].vcn - vcn;
758                                 lcn_block = lcn << (vol->cluster_size_bits -
759                                                 blocksize_bits);
760                                 cdelta = 0;
761                                 /*
762                                  * If the number of remaining clusters touched
763                                  * by the write is smaller or equal to the
764                                  * number of cached clusters, unlock the
765                                  * runlist as the map cache will be used from
766                                  * now on.
767                                  */
768                                 if (likely(vcn + vcn_len >= cend)) {
769                                         if (rl_write_locked) {
770                                                 up_write(&ni->runlist.lock);
771                                                 rl_write_locked = false;
772                                         } else
773                                                 up_read(&ni->runlist.lock);
774                                         rl = NULL;
775                                 }
776                                 goto map_buffer_cached;
777                         }
778                 } else
779                         lcn = LCN_RL_NOT_MAPPED;
780                 /*
781                  * If it is not a hole and not out of bounds, the runlist is
782                  * probably unmapped so try to map it now.
783                  */
784                 if (unlikely(lcn != LCN_HOLE && lcn != LCN_ENOENT)) {
785                         if (likely(!is_retry && lcn == LCN_RL_NOT_MAPPED)) {
786                                 /* Attempt to map runlist. */
787                                 if (!rl_write_locked) {
788                                         /*
789                                          * We need the runlist locked for
790                                          * writing, so if it is locked for
791                                          * reading relock it now and retry in
792                                          * case it changed whilst we dropped
793                                          * the lock.
794                                          */
795                                         up_read(&ni->runlist.lock);
796                                         down_write(&ni->runlist.lock);
797                                         rl_write_locked = true;
798                                         goto retry_remap;
799                                 }
800                                 err = ntfs_map_runlist_nolock(ni, bh_cpos,
801                                                 NULL);
802                                 if (likely(!err)) {
803                                         is_retry = true;
804                                         goto retry_remap;
805                                 }
806                                 /*
807                                  * If @vcn is out of bounds, pretend @lcn is
808                                  * LCN_ENOENT.  As long as the buffer is out
809                                  * of bounds this will work fine.
810                                  */
811                                 if (err == -ENOENT) {
812                                         lcn = LCN_ENOENT;
813                                         err = 0;
814                                         goto rl_not_mapped_enoent;
815                                 }
816                         } else
817                                 err = -EIO;
818                         /* Failed to map the buffer, even after retrying. */
819                         bh->b_blocknr = -1;
820                         ntfs_error(vol->sb, "Failed to write to inode 0x%lx, "
821                                         "attribute type 0x%x, vcn 0x%llx, "
822                                         "vcn offset 0x%x, because its "
823                                         "location on disk could not be "
824                                         "determined%s (error code %i).",
825                                         ni->mft_no, ni->type,
826                                         (unsigned long long)bh_cpos,
827                                         (unsigned)bh_pos &
828                                         vol->cluster_size_mask,
829                                         is_retry ? " even after retrying" : "",
830                                         err);
831                         break;
832                 }
833 rl_not_mapped_enoent:
834                 /*
835                  * The buffer is in a hole or out of bounds.  We need to fill
836                  * the hole, unless the buffer is in a cluster which is not
837                  * touched by the write, in which case we just leave the buffer
838                  * unmapped.  This can only happen when the cluster size is
839                  * less than the page cache size.
840                  */
841                 if (unlikely(vol->cluster_size < PAGE_CACHE_SIZE)) {
842                         bh_cend = (bh_end + vol->cluster_size - 1) >>
843                                         vol->cluster_size_bits;
844                         if ((bh_cend <= cpos || bh_cpos >= cend)) {
845                                 bh->b_blocknr = -1;
846                                 /*
847                                  * If the buffer is uptodate we skip it.  If it
848                                  * is not but the page is uptodate, we can set
849                                  * the buffer uptodate.  If the page is not
850                                  * uptodate, we can clear the buffer and set it
851                                  * uptodate.  Whether this is worthwhile is
852                                  * debatable and this could be removed.
853                                  */
854                                 if (PageUptodate(page)) {
855                                         if (!buffer_uptodate(bh))
856                                                 set_buffer_uptodate(bh);
857                                 } else if (!buffer_uptodate(bh)) {
858                                         zero_user(page, bh_offset(bh),
859                                                 blocksize);
860                                         set_buffer_uptodate(bh);
861                                 }
862                                 continue;
863                         }
864                 }
865                 /*
866                  * Out of bounds buffer is invalid if it was not really out of
867                  * bounds.
868                  */
869                 BUG_ON(lcn != LCN_HOLE);
870                 /*
871                  * We need the runlist locked for writing, so if it is locked
872                  * for reading relock it now and retry in case it changed
873                  * whilst we dropped the lock.
874                  */
875                 BUG_ON(!rl);
876                 if (!rl_write_locked) {
877                         up_read(&ni->runlist.lock);
878                         down_write(&ni->runlist.lock);
879                         rl_write_locked = true;
880                         goto retry_remap;
881                 }
882                 /* Find the previous last allocated cluster. */
883                 BUG_ON(rl->lcn != LCN_HOLE);
884                 lcn = -1;
885                 rl2 = rl;
886                 while (--rl2 >= ni->runlist.rl) {
887                         if (rl2->lcn >= 0) {
888                                 lcn = rl2->lcn + rl2->length;
889                                 break;
890                         }
891                 }
892                 rl2 = ntfs_cluster_alloc(vol, bh_cpos, 1, lcn, DATA_ZONE,
893                                 false);
894                 if (IS_ERR(rl2)) {
895                         err = PTR_ERR(rl2);
896                         ntfs_debug("Failed to allocate cluster, error code %i.",
897                                         err);
898                         break;
899                 }
900                 lcn = rl2->lcn;
901                 rl = ntfs_runlists_merge(ni->runlist.rl, rl2);
902                 if (IS_ERR(rl)) {
903                         err = PTR_ERR(rl);
904                         if (err != -ENOMEM)
905                                 err = -EIO;
906                         if (ntfs_cluster_free_from_rl(vol, rl2)) {
907                                 ntfs_error(vol->sb, "Failed to release "
908                                                 "allocated cluster in error "
909                                                 "code path.  Run chkdsk to "
910                                                 "recover the lost cluster.");
911                                 NVolSetErrors(vol);
912                         }
913                         ntfs_free(rl2);
914                         break;
915                 }
916                 ni->runlist.rl = rl;
917                 status.runlist_merged = 1;
918                 ntfs_debug("Allocated cluster, lcn 0x%llx.",
919                                 (unsigned long long)lcn);
920                 /* Map and lock the mft record and get the attribute record. */
921                 if (!NInoAttr(ni))
922                         base_ni = ni;
923                 else
924                         base_ni = ni->ext.base_ntfs_ino;
925                 m = map_mft_record(base_ni);
926                 if (IS_ERR(m)) {
927                         err = PTR_ERR(m);
928                         break;
929                 }
930                 ctx = ntfs_attr_get_search_ctx(base_ni, m);
931                 if (unlikely(!ctx)) {
932                         err = -ENOMEM;
933                         unmap_mft_record(base_ni);
934                         break;
935                 }
936                 status.mft_attr_mapped = 1;
937                 err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
938                                 CASE_SENSITIVE, bh_cpos, NULL, 0, ctx);
939                 if (unlikely(err)) {
940                         if (err == -ENOENT)
941                                 err = -EIO;
942                         break;
943                 }
944                 m = ctx->mrec;
945                 a = ctx->attr;
946                 /*
947                  * Find the runlist element with which the attribute extent
948                  * starts.  Note, we cannot use the _attr_ version because we
949                  * have mapped the mft record.  That is ok because we know the
950                  * runlist fragment must be mapped already to have ever gotten
951                  * here, so we can just use the _rl_ version.
952                  */
953                 vcn = sle64_to_cpu(a->data.non_resident.lowest_vcn);
954                 rl2 = ntfs_rl_find_vcn_nolock(rl, vcn);
955                 BUG_ON(!rl2);
956                 BUG_ON(!rl2->length);
957                 BUG_ON(rl2->lcn < LCN_HOLE);
958                 highest_vcn = sle64_to_cpu(a->data.non_resident.highest_vcn);
959                 /*
960                  * If @highest_vcn is zero, calculate the real highest_vcn
961                  * (which can really be zero).
962                  */
963                 if (!highest_vcn)
964                         highest_vcn = (sle64_to_cpu(
965                                         a->data.non_resident.allocated_size) >>
966                                         vol->cluster_size_bits) - 1;
967                 /*
968                  * Determine the size of the mapping pairs array for the new
969                  * extent, i.e. the old extent with the hole filled.
970                  */
971                 mp_size = ntfs_get_size_for_mapping_pairs(vol, rl2, vcn,
972                                 highest_vcn);
973                 if (unlikely(mp_size <= 0)) {
974                         if (!(err = mp_size))
975                                 err = -EIO;
976                         ntfs_debug("Failed to get size for mapping pairs "
977                                         "array, error code %i.", err);
978                         break;
979                 }
980                 /*
981                  * Resize the attribute record to fit the new mapping pairs
982                  * array.
983                  */
984                 attr_rec_len = le32_to_cpu(a->length);
985                 err = ntfs_attr_record_resize(m, a, mp_size + le16_to_cpu(
986                                 a->data.non_resident.mapping_pairs_offset));
987                 if (unlikely(err)) {
988                         BUG_ON(err != -ENOSPC);
989                         // TODO: Deal with this by using the current attribute
990                         // and fill it with as much of the mapping pairs
991                         // array as possible.  Then loop over each attribute
992                         // extent rewriting the mapping pairs arrays as we go
993                         // along and if when we reach the end we have not
994                         // enough space, try to resize the last attribute
995                         // extent and if even that fails, add a new attribute
996                         // extent.
997                         // We could also try to resize at each step in the hope
998                         // that we will not need to rewrite every single extent.
999                         // Note, we may need to decompress some extents to fill
1000                         // the runlist as we are walking the extents...
1001                         ntfs_error(vol->sb, "Not enough space in the mft "
1002                                         "record for the extended attribute "
1003                                         "record.  This case is not "
1004                                         "implemented yet.");
1005                         err = -EOPNOTSUPP;
1006                         break ;
1007                 }
1008                 status.mp_rebuilt = 1;
1009                 /*
1010                  * Generate the mapping pairs array directly into the attribute
1011                  * record.
1012                  */
1013                 err = ntfs_mapping_pairs_build(vol, (u8*)a + le16_to_cpu(
1014                                 a->data.non_resident.mapping_pairs_offset),
1015                                 mp_size, rl2, vcn, highest_vcn, NULL);
1016                 if (unlikely(err)) {
1017                         ntfs_error(vol->sb, "Cannot fill hole in inode 0x%lx, "
1018                                         "attribute type 0x%x, because building "
1019                                         "the mapping pairs failed with error "
1020                                         "code %i.", vi->i_ino,
1021                                         (unsigned)le32_to_cpu(ni->type), err);
1022                         err = -EIO;
1023                         break;
1024                 }
1025                 /* Update the highest_vcn but only if it was not set. */
1026                 if (unlikely(!a->data.non_resident.highest_vcn))
1027                         a->data.non_resident.highest_vcn =
1028                                         cpu_to_sle64(highest_vcn);
1029                 /*
1030                  * If the attribute is sparse/compressed, update the compressed
1031                  * size in the ntfs_inode structure and the attribute record.
1032                  */
1033                 if (likely(NInoSparse(ni) || NInoCompressed(ni))) {
1034                         /*
1035                          * If we are not in the first attribute extent, switch
1036                          * to it, but first ensure the changes will make it to
1037                          * disk later.
1038                          */
1039                         if (a->data.non_resident.lowest_vcn) {
1040                                 flush_dcache_mft_record_page(ctx->ntfs_ino);
1041                                 mark_mft_record_dirty(ctx->ntfs_ino);
1042                                 ntfs_attr_reinit_search_ctx(ctx);
1043                                 err = ntfs_attr_lookup(ni->type, ni->name,
1044                                                 ni->name_len, CASE_SENSITIVE,
1045                                                 0, NULL, 0, ctx);
1046                                 if (unlikely(err)) {
1047                                         status.attr_switched = 1;
1048                                         break;
1049                                 }
1050                                 /* @m is not used any more so do not set it. */
1051                                 a = ctx->attr;
1052                         }
1053                         write_lock_irqsave(&ni->size_lock, flags);
1054                         ni->itype.compressed.size += vol->cluster_size;
1055                         a->data.non_resident.compressed_size =
1056                                         cpu_to_sle64(ni->itype.compressed.size);
1057                         write_unlock_irqrestore(&ni->size_lock, flags);
1058                 }
1059                 /* Ensure the changes make it to disk. */
1060                 flush_dcache_mft_record_page(ctx->ntfs_ino);
1061                 mark_mft_record_dirty(ctx->ntfs_ino);
1062                 ntfs_attr_put_search_ctx(ctx);
1063                 unmap_mft_record(base_ni);
1064                 /* Successfully filled the hole. */
1065                 status.runlist_merged = 0;
1066                 status.mft_attr_mapped = 0;
1067                 status.mp_rebuilt = 0;
1068                 /* Setup the map cache and use that to deal with the buffer. */
1069                 was_hole = true;
1070                 vcn = bh_cpos;
1071                 vcn_len = 1;
1072                 lcn_block = lcn << (vol->cluster_size_bits - blocksize_bits);
1073                 cdelta = 0;
1074                 /*
1075                  * If the number of remaining clusters in the @pages is smaller
1076                  * or equal to the number of cached clusters, unlock the
1077                  * runlist as the map cache will be used from now on.
1078                  */
1079                 if (likely(vcn + vcn_len >= cend)) {
1080                         up_write(&ni->runlist.lock);
1081                         rl_write_locked = false;
1082                         rl = NULL;
1083                 }
1084                 goto map_buffer_cached;
1085         } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
1086         /* If there are no errors, do the next page. */
1087         if (likely(!err && ++u < nr_pages))
1088                 goto do_next_page;
1089         /* If there are no errors, release the runlist lock if we took it. */
1090         if (likely(!err)) {
1091                 if (unlikely(rl_write_locked)) {
1092                         up_write(&ni->runlist.lock);
1093                         rl_write_locked = false;
1094                 } else if (unlikely(rl))
1095                         up_read(&ni->runlist.lock);
1096                 rl = NULL;
1097         }
1098         /* If we issued read requests, let them complete. */
1099         read_lock_irqsave(&ni->size_lock, flags);
1100         initialized_size = ni->initialized_size;
1101         read_unlock_irqrestore(&ni->size_lock, flags);
1102         while (wait_bh > wait) {
1103                 bh = *--wait_bh;
1104                 wait_on_buffer(bh);
1105                 if (likely(buffer_uptodate(bh))) {
1106                         page = bh->b_page;
1107                         bh_pos = ((s64)page->index << PAGE_CACHE_SHIFT) +
1108                                         bh_offset(bh);
1109                         /*
1110                          * If the buffer overflows the initialized size, need
1111                          * to zero the overflowing region.
1112                          */
1113                         if (unlikely(bh_pos + blocksize > initialized_size)) {
1114                                 int ofs = 0;
1115
1116                                 if (likely(bh_pos < initialized_size))
1117                                         ofs = initialized_size - bh_pos;
1118                                 zero_user_segment(page, bh_offset(bh) + ofs,
1119                                                 blocksize);
1120                         }
1121                 } else /* if (unlikely(!buffer_uptodate(bh))) */
1122                         err = -EIO;
1123         }
1124         if (likely(!err)) {
1125                 /* Clear buffer_new on all buffers. */
1126                 u = 0;
1127                 do {
1128                         bh = head = page_buffers(pages[u]);
1129                         do {
1130                                 if (buffer_new(bh))
1131                                         clear_buffer_new(bh);
1132                         } while ((bh = bh->b_this_page) != head);
1133                 } while (++u < nr_pages);
1134                 ntfs_debug("Done.");
1135                 return err;
1136         }
1137         if (status.attr_switched) {
1138                 /* Get back to the attribute extent we modified. */
1139                 ntfs_attr_reinit_search_ctx(ctx);
1140                 if (ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1141                                 CASE_SENSITIVE, bh_cpos, NULL, 0, ctx)) {
1142                         ntfs_error(vol->sb, "Failed to find required "
1143                                         "attribute extent of attribute in "
1144                                         "error code path.  Run chkdsk to "
1145                                         "recover.");
1146                         write_lock_irqsave(&ni->size_lock, flags);
1147                         ni->itype.compressed.size += vol->cluster_size;
1148                         write_unlock_irqrestore(&ni->size_lock, flags);
1149                         flush_dcache_mft_record_page(ctx->ntfs_ino);
1150                         mark_mft_record_dirty(ctx->ntfs_ino);
1151                         /*
1152                          * The only thing that is now wrong is the compressed
1153                          * size of the base attribute extent which chkdsk
1154                          * should be able to fix.
1155                          */
1156                         NVolSetErrors(vol);
1157                 } else {
1158                         m = ctx->mrec;
1159                         a = ctx->attr;
1160                         status.attr_switched = 0;
1161                 }
1162         }
1163         /*
1164          * If the runlist has been modified, need to restore it by punching a
1165          * hole into it and we then need to deallocate the on-disk cluster as
1166          * well.  Note, we only modify the runlist if we are able to generate a
1167          * new mapping pairs array, i.e. only when the mapped attribute extent
1168          * is not switched.
1169          */
1170         if (status.runlist_merged && !status.attr_switched) {
1171                 BUG_ON(!rl_write_locked);
1172                 /* Make the file cluster we allocated sparse in the runlist. */
1173                 if (ntfs_rl_punch_nolock(vol, &ni->runlist, bh_cpos, 1)) {
1174                         ntfs_error(vol->sb, "Failed to punch hole into "
1175                                         "attribute runlist in error code "
1176                                         "path.  Run chkdsk to recover the "
1177                                         "lost cluster.");
1178                         NVolSetErrors(vol);
1179                 } else /* if (success) */ {
1180                         status.runlist_merged = 0;
1181                         /*
1182                          * Deallocate the on-disk cluster we allocated but only
1183                          * if we succeeded in punching its vcn out of the
1184                          * runlist.
1185                          */
1186                         down_write(&vol->lcnbmp_lock);
1187                         if (ntfs_bitmap_clear_bit(vol->lcnbmp_ino, lcn)) {
1188                                 ntfs_error(vol->sb, "Failed to release "
1189                                                 "allocated cluster in error "
1190                                                 "code path.  Run chkdsk to "
1191                                                 "recover the lost cluster.");
1192                                 NVolSetErrors(vol);
1193                         }
1194                         up_write(&vol->lcnbmp_lock);
1195                 }
1196         }
1197         /*
1198          * Resize the attribute record to its old size and rebuild the mapping
1199          * pairs array.  Note, we only can do this if the runlist has been
1200          * restored to its old state which also implies that the mapped
1201          * attribute extent is not switched.
1202          */
1203         if (status.mp_rebuilt && !status.runlist_merged) {
1204                 if (ntfs_attr_record_resize(m, a, attr_rec_len)) {
1205                         ntfs_error(vol->sb, "Failed to restore attribute "
1206                                         "record in error code path.  Run "
1207                                         "chkdsk to recover.");
1208                         NVolSetErrors(vol);
1209                 } else /* if (success) */ {
1210                         if (ntfs_mapping_pairs_build(vol, (u8*)a +
1211                                         le16_to_cpu(a->data.non_resident.
1212                                         mapping_pairs_offset), attr_rec_len -
1213                                         le16_to_cpu(a->data.non_resident.
1214                                         mapping_pairs_offset), ni->runlist.rl,
1215                                         vcn, highest_vcn, NULL)) {
1216                                 ntfs_error(vol->sb, "Failed to restore "
1217                                                 "mapping pairs array in error "
1218                                                 "code path.  Run chkdsk to "
1219                                                 "recover.");
1220                                 NVolSetErrors(vol);
1221                         }
1222                         flush_dcache_mft_record_page(ctx->ntfs_ino);
1223                         mark_mft_record_dirty(ctx->ntfs_ino);
1224                 }
1225         }
1226         /* Release the mft record and the attribute. */
1227         if (status.mft_attr_mapped) {
1228                 ntfs_attr_put_search_ctx(ctx);
1229                 unmap_mft_record(base_ni);
1230         }
1231         /* Release the runlist lock. */
1232         if (rl_write_locked)
1233                 up_write(&ni->runlist.lock);
1234         else if (rl)
1235                 up_read(&ni->runlist.lock);
1236         /*
1237          * Zero out any newly allocated blocks to avoid exposing stale data.
1238          * If BH_New is set, we know that the block was newly allocated above
1239          * and that it has not been fully zeroed and marked dirty yet.
1240          */
1241         nr_pages = u;
1242         u = 0;
1243         end = bh_cpos << vol->cluster_size_bits;
1244         do {
1245                 page = pages[u];
1246                 bh = head = page_buffers(page);
1247                 do {
1248                         if (u == nr_pages &&
1249                                         ((s64)page->index << PAGE_CACHE_SHIFT) +
1250                                         bh_offset(bh) >= end)
1251                                 break;
1252                         if (!buffer_new(bh))
1253                                 continue;
1254                         clear_buffer_new(bh);
1255                         if (!buffer_uptodate(bh)) {
1256                                 if (PageUptodate(page))
1257                                         set_buffer_uptodate(bh);
1258                                 else {
1259                                         zero_user(page, bh_offset(bh),
1260                                                         blocksize);
1261                                         set_buffer_uptodate(bh);
1262                                 }
1263                         }
1264                         mark_buffer_dirty(bh);
1265                 } while ((bh = bh->b_this_page) != head);
1266         } while (++u <= nr_pages);
1267         ntfs_error(vol->sb, "Failed.  Returning error code %i.", err);
1268         return err;
1269 }
1270
1271 /*
1272  * Copy as much as we can into the pages and return the number of bytes which
1273  * were successfully copied.  If a fault is encountered then clear the pages
1274  * out to (ofs + bytes) and return the number of bytes which were copied.
1275  */
1276 static inline size_t ntfs_copy_from_user(struct page **pages,
1277                 unsigned nr_pages, unsigned ofs, const char __user *buf,
1278                 size_t bytes)
1279 {
1280         struct page **last_page = pages + nr_pages;
1281         char *addr;
1282         size_t total = 0;
1283         unsigned len;
1284         int left;
1285
1286         do {
1287                 len = PAGE_CACHE_SIZE - ofs;
1288                 if (len > bytes)
1289                         len = bytes;
1290                 addr = kmap_atomic(*pages, KM_USER0);
1291                 left = __copy_from_user_inatomic(addr + ofs, buf, len);
1292                 kunmap_atomic(addr, KM_USER0);
1293                 if (unlikely(left)) {
1294                         /* Do it the slow way. */
1295                         addr = kmap(*pages);
1296                         left = __copy_from_user(addr + ofs, buf, len);
1297                         kunmap(*pages);
1298                         if (unlikely(left))
1299                                 goto err_out;
1300                 }
1301                 total += len;
1302                 bytes -= len;
1303                 if (!bytes)
1304                         break;
1305                 buf += len;
1306                 ofs = 0;
1307         } while (++pages < last_page);
1308 out:
1309         return total;
1310 err_out:
1311         total += len - left;
1312         /* Zero the rest of the target like __copy_from_user(). */
1313         while (++pages < last_page) {
1314                 bytes -= len;
1315                 if (!bytes)
1316                         break;
1317                 len = PAGE_CACHE_SIZE;
1318                 if (len > bytes)
1319                         len = bytes;
1320                 zero_user(*pages, 0, len);
1321         }
1322         goto out;
1323 }
1324
1325 static size_t __ntfs_copy_from_user_iovec_inatomic(char *vaddr,
1326                 const struct iovec *iov, size_t iov_ofs, size_t bytes)
1327 {
1328         size_t total = 0;
1329
1330         while (1) {
1331                 const char __user *buf = iov->iov_base + iov_ofs;
1332                 unsigned len;
1333                 size_t left;
1334
1335                 len = iov->iov_len - iov_ofs;
1336                 if (len > bytes)
1337                         len = bytes;
1338                 left = __copy_from_user_inatomic(vaddr, buf, len);
1339                 total += len;
1340                 bytes -= len;
1341                 vaddr += len;
1342                 if (unlikely(left)) {
1343                         total -= left;
1344                         break;
1345                 }
1346                 if (!bytes)
1347                         break;
1348                 iov++;
1349                 iov_ofs = 0;
1350         }
1351         return total;
1352 }
1353
1354 static inline void ntfs_set_next_iovec(const struct iovec **iovp,
1355                 size_t *iov_ofsp, size_t bytes)
1356 {
1357         const struct iovec *iov = *iovp;
1358         size_t iov_ofs = *iov_ofsp;
1359
1360         while (bytes) {
1361                 unsigned len;
1362
1363                 len = iov->iov_len - iov_ofs;
1364                 if (len > bytes)
1365                         len = bytes;
1366                 bytes -= len;
1367                 iov_ofs += len;
1368                 if (iov->iov_len == iov_ofs) {
1369                         iov++;
1370                         iov_ofs = 0;
1371                 }
1372         }
1373         *iovp = iov;
1374         *iov_ofsp = iov_ofs;
1375 }
1376
1377 /*
1378  * This has the same side-effects and return value as ntfs_copy_from_user().
1379  * The difference is that on a fault we need to memset the remainder of the
1380  * pages (out to offset + bytes), to emulate ntfs_copy_from_user()'s
1381  * single-segment behaviour.
1382  *
1383  * We call the same helper (__ntfs_copy_from_user_iovec_inatomic()) both when
1384  * atomic and when not atomic.  This is ok because it calls
1385  * __copy_from_user_inatomic() and it is ok to call this when non-atomic.  In
1386  * fact, the only difference between __copy_from_user_inatomic() and
1387  * __copy_from_user() is that the latter calls might_sleep() and the former
1388  * should not zero the tail of the buffer on error.  And on many architectures
1389  * __copy_from_user_inatomic() is just defined to __copy_from_user() so it
1390  * makes no difference at all on those architectures.
1391  */
1392 static inline size_t ntfs_copy_from_user_iovec(struct page **pages,
1393                 unsigned nr_pages, unsigned ofs, const struct iovec **iov,
1394                 size_t *iov_ofs, size_t bytes)
1395 {
1396         struct page **last_page = pages + nr_pages;
1397         char *addr;
1398         size_t copied, len, total = 0;
1399
1400         do {
1401                 len = PAGE_CACHE_SIZE - ofs;
1402                 if (len > bytes)
1403                         len = bytes;
1404                 addr = kmap_atomic(*pages, KM_USER0);
1405                 copied = __ntfs_copy_from_user_iovec_inatomic(addr + ofs,
1406                                 *iov, *iov_ofs, len);
1407                 kunmap_atomic(addr, KM_USER0);
1408                 if (unlikely(copied != len)) {
1409                         /* Do it the slow way. */
1410                         addr = kmap(*pages);
1411                         copied = __ntfs_copy_from_user_iovec_inatomic(addr +
1412                                         ofs, *iov, *iov_ofs, len);
1413                         if (unlikely(copied != len))
1414                                 goto err_out;
1415                         kunmap(*pages);
1416                 }
1417                 total += len;
1418                 ntfs_set_next_iovec(iov, iov_ofs, len);
1419                 bytes -= len;
1420                 if (!bytes)
1421                         break;
1422                 ofs = 0;
1423         } while (++pages < last_page);
1424 out:
1425         return total;
1426 err_out:
1427         BUG_ON(copied > len);
1428         /* Zero the rest of the target like __copy_from_user(). */
1429         memset(addr + ofs + copied, 0, len - copied);
1430         kunmap(*pages);
1431         total += copied;
1432         ntfs_set_next_iovec(iov, iov_ofs, copied);
1433         while (++pages < last_page) {
1434                 bytes -= len;
1435                 if (!bytes)
1436                         break;
1437                 len = PAGE_CACHE_SIZE;
1438                 if (len > bytes)
1439                         len = bytes;
1440                 zero_user(*pages, 0, len);
1441         }
1442         goto out;
1443 }
1444
1445 static inline void ntfs_flush_dcache_pages(struct page **pages,
1446                 unsigned nr_pages)
1447 {
1448         BUG_ON(!nr_pages);
1449         /*
1450          * Warning: Do not do the decrement at the same time as the call to
1451          * flush_dcache_page() because it is a NULL macro on i386 and hence the
1452          * decrement never happens so the loop never terminates.
1453          */
1454         do {
1455                 --nr_pages;
1456                 flush_dcache_page(pages[nr_pages]);
1457         } while (nr_pages > 0);
1458 }
1459
1460 /**
1461  * ntfs_commit_pages_after_non_resident_write - commit the received data
1462  * @pages:      array of destination pages
1463  * @nr_pages:   number of pages in @pages
1464  * @pos:        byte position in file at which the write begins
1465  * @bytes:      number of bytes to be written
1466  *
1467  * See description of ntfs_commit_pages_after_write(), below.
1468  */
1469 static inline int ntfs_commit_pages_after_non_resident_write(
1470                 struct page **pages, const unsigned nr_pages,
1471                 s64 pos, size_t bytes)
1472 {
1473         s64 end, initialized_size;
1474         struct inode *vi;
1475         ntfs_inode *ni, *base_ni;
1476         struct buffer_head *bh, *head;
1477         ntfs_attr_search_ctx *ctx;
1478         MFT_RECORD *m;
1479         ATTR_RECORD *a;
1480         unsigned long flags;
1481         unsigned blocksize, u;
1482         int err;
1483
1484         vi = pages[0]->mapping->host;
1485         ni = NTFS_I(vi);
1486         blocksize = vi->i_sb->s_blocksize;
1487         end = pos + bytes;
1488         u = 0;
1489         do {
1490                 s64 bh_pos;
1491                 struct page *page;
1492                 bool partial;
1493
1494                 page = pages[u];
1495                 bh_pos = (s64)page->index << PAGE_CACHE_SHIFT;
1496                 bh = head = page_buffers(page);
1497                 partial = false;
1498                 do {
1499                         s64 bh_end;
1500
1501                         bh_end = bh_pos + blocksize;
1502                         if (bh_end <= pos || bh_pos >= end) {
1503                                 if (!buffer_uptodate(bh))
1504                                         partial = true;
1505                         } else {
1506                                 set_buffer_uptodate(bh);
1507                                 mark_buffer_dirty(bh);
1508                         }
1509                 } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
1510                 /*
1511                  * If all buffers are now uptodate but the page is not, set the
1512                  * page uptodate.
1513                  */
1514                 if (!partial && !PageUptodate(page))
1515                         SetPageUptodate(page);
1516         } while (++u < nr_pages);
1517         /*
1518          * Finally, if we do not need to update initialized_size or i_size we
1519          * are finished.
1520          */
1521         read_lock_irqsave(&ni->size_lock, flags);
1522         initialized_size = ni->initialized_size;
1523         read_unlock_irqrestore(&ni->size_lock, flags);
1524         if (end <= initialized_size) {
1525                 ntfs_debug("Done.");
1526                 return 0;
1527         }
1528         /*
1529          * Update initialized_size/i_size as appropriate, both in the inode and
1530          * the mft record.
1531          */
1532         if (!NInoAttr(ni))
1533                 base_ni = ni;
1534         else
1535                 base_ni = ni->ext.base_ntfs_ino;
1536         /* Map, pin, and lock the mft record. */
1537         m = map_mft_record(base_ni);
1538         if (IS_ERR(m)) {
1539                 err = PTR_ERR(m);
1540                 m = NULL;
1541                 ctx = NULL;
1542                 goto err_out;
1543         }
1544         BUG_ON(!NInoNonResident(ni));
1545         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1546         if (unlikely(!ctx)) {
1547                 err = -ENOMEM;
1548                 goto err_out;
1549         }
1550         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1551                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1552         if (unlikely(err)) {
1553                 if (err == -ENOENT)
1554                         err = -EIO;
1555                 goto err_out;
1556         }
1557         a = ctx->attr;
1558         BUG_ON(!a->non_resident);
1559         write_lock_irqsave(&ni->size_lock, flags);
1560         BUG_ON(end > ni->allocated_size);
1561         ni->initialized_size = end;
1562         a->data.non_resident.initialized_size = cpu_to_sle64(end);
1563         if (end > i_size_read(vi)) {
1564                 i_size_write(vi, end);
1565                 a->data.non_resident.data_size =
1566                                 a->data.non_resident.initialized_size;
1567         }
1568         write_unlock_irqrestore(&ni->size_lock, flags);
1569         /* Mark the mft record dirty, so it gets written back. */
1570         flush_dcache_mft_record_page(ctx->ntfs_ino);
1571         mark_mft_record_dirty(ctx->ntfs_ino);
1572         ntfs_attr_put_search_ctx(ctx);
1573         unmap_mft_record(base_ni);
1574         ntfs_debug("Done.");
1575         return 0;
1576 err_out:
1577         if (ctx)
1578                 ntfs_attr_put_search_ctx(ctx);
1579         if (m)
1580                 unmap_mft_record(base_ni);
1581         ntfs_error(vi->i_sb, "Failed to update initialized_size/i_size (error "
1582                         "code %i).", err);
1583         if (err != -ENOMEM)
1584                 NVolSetErrors(ni->vol);
1585         return err;
1586 }
1587
1588 /**
1589  * ntfs_commit_pages_after_write - commit the received data
1590  * @pages:      array of destination pages
1591  * @nr_pages:   number of pages in @pages
1592  * @pos:        byte position in file at which the write begins
1593  * @bytes:      number of bytes to be written
1594  *
1595  * This is called from ntfs_file_buffered_write() with i_mutex held on the inode
1596  * (@pages[0]->mapping->host).  There are @nr_pages pages in @pages which are
1597  * locked but not kmap()ped.  The source data has already been copied into the
1598  * @page.  ntfs_prepare_pages_for_non_resident_write() has been called before
1599  * the data was copied (for non-resident attributes only) and it returned
1600  * success.
1601  *
1602  * Need to set uptodate and mark dirty all buffers within the boundary of the
1603  * write.  If all buffers in a page are uptodate we set the page uptodate, too.
1604  *
1605  * Setting the buffers dirty ensures that they get written out later when
1606  * ntfs_writepage() is invoked by the VM.
1607  *
1608  * Finally, we need to update i_size and initialized_size as appropriate both
1609  * in the inode and the mft record.
1610  *
1611  * This is modelled after fs/buffer.c::generic_commit_write(), which marks
1612  * buffers uptodate and dirty, sets the page uptodate if all buffers in the
1613  * page are uptodate, and updates i_size if the end of io is beyond i_size.  In
1614  * that case, it also marks the inode dirty.
1615  *
1616  * If things have gone as outlined in
1617  * ntfs_prepare_pages_for_non_resident_write(), we do not need to do any page
1618  * content modifications here for non-resident attributes.  For resident
1619  * attributes we need to do the uptodate bringing here which we combine with
1620  * the copying into the mft record which means we save one atomic kmap.
1621  *
1622  * Return 0 on success or -errno on error.
1623  */
1624 static int ntfs_commit_pages_after_write(struct page **pages,
1625                 const unsigned nr_pages, s64 pos, size_t bytes)
1626 {
1627         s64 end, initialized_size;
1628         loff_t i_size;
1629         struct inode *vi;
1630         ntfs_inode *ni, *base_ni;
1631         struct page *page;
1632         ntfs_attr_search_ctx *ctx;
1633         MFT_RECORD *m;
1634         ATTR_RECORD *a;
1635         char *kattr, *kaddr;
1636         unsigned long flags;
1637         u32 attr_len;
1638         int err;
1639
1640         BUG_ON(!nr_pages);
1641         BUG_ON(!pages);
1642         page = pages[0];
1643         BUG_ON(!page);
1644         vi = page->mapping->host;
1645         ni = NTFS_I(vi);
1646         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, start page "
1647                         "index 0x%lx, nr_pages 0x%x, pos 0x%llx, bytes 0x%zx.",
1648                         vi->i_ino, ni->type, page->index, nr_pages,
1649                         (long long)pos, bytes);
1650         if (NInoNonResident(ni))
1651                 return ntfs_commit_pages_after_non_resident_write(pages,
1652                                 nr_pages, pos, bytes);
1653         BUG_ON(nr_pages > 1);
1654         /*
1655          * Attribute is resident, implying it is not compressed, encrypted, or
1656          * sparse.
1657          */
1658         if (!NInoAttr(ni))
1659                 base_ni = ni;
1660         else
1661                 base_ni = ni->ext.base_ntfs_ino;
1662         BUG_ON(NInoNonResident(ni));
1663         /* Map, pin, and lock the mft record. */
1664         m = map_mft_record(base_ni);
1665         if (IS_ERR(m)) {
1666                 err = PTR_ERR(m);
1667                 m = NULL;
1668                 ctx = NULL;
1669                 goto err_out;
1670         }
1671         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1672         if (unlikely(!ctx)) {
1673                 err = -ENOMEM;
1674                 goto err_out;
1675         }
1676         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1677                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1678         if (unlikely(err)) {
1679                 if (err == -ENOENT)
1680                         err = -EIO;
1681                 goto err_out;
1682         }
1683         a = ctx->attr;
1684         BUG_ON(a->non_resident);
1685         /* The total length of the attribute value. */
1686         attr_len = le32_to_cpu(a->data.resident.value_length);
1687         i_size = i_size_read(vi);
1688         BUG_ON(attr_len != i_size);
1689         BUG_ON(pos > attr_len);
1690         end = pos + bytes;
1691         BUG_ON(end > le32_to_cpu(a->length) -
1692                         le16_to_cpu(a->data.resident.value_offset));
1693         kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
1694         kaddr = kmap_atomic(page, KM_USER0);
1695         /* Copy the received data from the page to the mft record. */
1696         memcpy(kattr + pos, kaddr + pos, bytes);
1697         /* Update the attribute length if necessary. */
1698         if (end > attr_len) {
1699                 attr_len = end;
1700                 a->data.resident.value_length = cpu_to_le32(attr_len);
1701         }
1702         /*
1703          * If the page is not uptodate, bring the out of bounds area(s)
1704          * uptodate by copying data from the mft record to the page.
1705          */
1706         if (!PageUptodate(page)) {
1707                 if (pos > 0)
1708                         memcpy(kaddr, kattr, pos);
1709                 if (end < attr_len)
1710                         memcpy(kaddr + end, kattr + end, attr_len - end);
1711                 /* Zero the region outside the end of the attribute value. */
1712                 memset(kaddr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
1713                 flush_dcache_page(page);
1714                 SetPageUptodate(page);
1715         }
1716         kunmap_atomic(kaddr, KM_USER0);
1717         /* Update initialized_size/i_size if necessary. */
1718         read_lock_irqsave(&ni->size_lock, flags);
1719         initialized_size = ni->initialized_size;
1720         BUG_ON(end > ni->allocated_size);
1721         read_unlock_irqrestore(&ni->size_lock, flags);
1722         BUG_ON(initialized_size != i_size);
1723         if (end > initialized_size) {
1724                 write_lock_irqsave(&ni->size_lock, flags);
1725                 ni->initialized_size = end;
1726                 i_size_write(vi, end);
1727                 write_unlock_irqrestore(&ni->size_lock, flags);
1728         }
1729         /* Mark the mft record dirty, so it gets written back. */
1730         flush_dcache_mft_record_page(ctx->ntfs_ino);
1731         mark_mft_record_dirty(ctx->ntfs_ino);
1732         ntfs_attr_put_search_ctx(ctx);
1733         unmap_mft_record(base_ni);
1734         ntfs_debug("Done.");
1735         return 0;
1736 err_out:
1737         if (err == -ENOMEM) {
1738                 ntfs_warning(vi->i_sb, "Error allocating memory required to "
1739                                 "commit the write.");
1740                 if (PageUptodate(page)) {
1741                         ntfs_warning(vi->i_sb, "Page is uptodate, setting "
1742                                         "dirty so the write will be retried "
1743                                         "later on by the VM.");
1744                         /*
1745                          * Put the page on mapping->dirty_pages, but leave its
1746                          * buffers' dirty state as-is.
1747                          */
1748                         __set_page_dirty_nobuffers(page);
1749                         err = 0;
1750                 } else
1751                         ntfs_error(vi->i_sb, "Page is not uptodate.  Written "
1752                                         "data has been lost.");
1753         } else {
1754                 ntfs_error(vi->i_sb, "Resident attribute commit write failed "
1755                                 "with error %i.", err);
1756                 NVolSetErrors(ni->vol);
1757         }
1758         if (ctx)
1759                 ntfs_attr_put_search_ctx(ctx);
1760         if (m)
1761                 unmap_mft_record(base_ni);
1762         return err;
1763 }
1764
1765 /**
1766  * ntfs_file_buffered_write -
1767  *
1768  * Locking: The vfs is holding ->i_mutex on the inode.
1769  */
1770 static ssize_t ntfs_file_buffered_write(struct kiocb *iocb,
1771                 const struct iovec *iov, unsigned long nr_segs,
1772                 loff_t pos, loff_t *ppos, size_t count)
1773 {
1774         struct file *file = iocb->ki_filp;
1775         struct address_space *mapping = file->f_mapping;
1776         struct inode *vi = mapping->host;
1777         ntfs_inode *ni = NTFS_I(vi);
1778         ntfs_volume *vol = ni->vol;
1779         struct page *pages[NTFS_MAX_PAGES_PER_CLUSTER];
1780         struct page *cached_page = NULL;
1781         char __user *buf = NULL;
1782         s64 end, ll;
1783         VCN last_vcn;
1784         LCN lcn;
1785         unsigned long flags;
1786         size_t bytes, iov_ofs = 0;      /* Offset in the current iovec. */
1787         ssize_t status, written;
1788         unsigned nr_pages;
1789         int err;
1790
1791         ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, "
1792                         "pos 0x%llx, count 0x%lx.",
1793                         vi->i_ino, (unsigned)le32_to_cpu(ni->type),
1794                         (unsigned long long)pos, (unsigned long)count);
1795         if (unlikely(!count))
1796                 return 0;
1797         BUG_ON(NInoMstProtected(ni));
1798         /*
1799          * If the attribute is not an index root and it is encrypted or
1800          * compressed, we cannot write to it yet.  Note we need to check for
1801          * AT_INDEX_ALLOCATION since this is the type of both directory and
1802          * index inodes.
1803          */
1804         if (ni->type != AT_INDEX_ALLOCATION) {
1805                 /* If file is encrypted, deny access, just like NT4. */
1806                 if (NInoEncrypted(ni)) {
1807                         /*
1808                          * Reminder for later: Encrypted files are _always_
1809                          * non-resident so that the content can always be
1810                          * encrypted.
1811                          */
1812                         ntfs_debug("Denying write access to encrypted file.");
1813                         return -EACCES;
1814                 }
1815                 if (NInoCompressed(ni)) {
1816                         /* Only unnamed $DATA attribute can be compressed. */
1817                         BUG_ON(ni->type != AT_DATA);
1818                         BUG_ON(ni->name_len);
1819                         /*
1820                          * Reminder for later: If resident, the data is not
1821                          * actually compressed.  Only on the switch to non-
1822                          * resident does compression kick in.  This is in
1823                          * contrast to encrypted files (see above).
1824                          */
1825                         ntfs_error(vi->i_sb, "Writing to compressed files is "
1826                                         "not implemented yet.  Sorry.");
1827                         return -EOPNOTSUPP;
1828                 }
1829         }
1830         /*
1831          * If a previous ntfs_truncate() failed, repeat it and abort if it
1832          * fails again.
1833          */
1834         if (unlikely(NInoTruncateFailed(ni))) {
1835                 down_write(&vi->i_alloc_sem);
1836                 err = ntfs_truncate(vi);
1837                 up_write(&vi->i_alloc_sem);
1838                 if (err || NInoTruncateFailed(ni)) {
1839                         if (!err)
1840                                 err = -EIO;
1841                         ntfs_error(vol->sb, "Cannot perform write to inode "
1842                                         "0x%lx, attribute type 0x%x, because "
1843                                         "ntfs_truncate() failed (error code "
1844                                         "%i).", vi->i_ino,
1845                                         (unsigned)le32_to_cpu(ni->type), err);
1846                         return err;
1847                 }
1848         }
1849         /* The first byte after the write. */
1850         end = pos + count;
1851         /*
1852          * If the write goes beyond the allocated size, extend the allocation
1853          * to cover the whole of the write, rounded up to the nearest cluster.
1854          */
1855         read_lock_irqsave(&ni->size_lock, flags);
1856         ll = ni->allocated_size;
1857         read_unlock_irqrestore(&ni->size_lock, flags);
1858         if (end > ll) {
1859                 /* Extend the allocation without changing the data size. */
1860                 ll = ntfs_attr_extend_allocation(ni, end, -1, pos);
1861                 if (likely(ll >= 0)) {
1862                         BUG_ON(pos >= ll);
1863                         /* If the extension was partial truncate the write. */
1864                         if (end > ll) {
1865                                 ntfs_debug("Truncating write to inode 0x%lx, "
1866                                                 "attribute type 0x%x, because "
1867                                                 "the allocation was only "
1868                                                 "partially extended.",
1869                                                 vi->i_ino, (unsigned)
1870                                                 le32_to_cpu(ni->type));
1871                                 end = ll;
1872                                 count = ll - pos;
1873                         }
1874                 } else {
1875                         err = ll;
1876                         read_lock_irqsave(&ni->size_lock, flags);
1877                         ll = ni->allocated_size;
1878                         read_unlock_irqrestore(&ni->size_lock, flags);
1879                         /* Perform a partial write if possible or fail. */
1880                         if (pos < ll) {
1881                                 ntfs_debug("Truncating write to inode 0x%lx, "
1882                                                 "attribute type 0x%x, because "
1883                                                 "extending the allocation "
1884                                                 "failed (error code %i).",
1885                                                 vi->i_ino, (unsigned)
1886                                                 le32_to_cpu(ni->type), err);
1887                                 end = ll;
1888                                 count = ll - pos;
1889                         } else {
1890                                 ntfs_error(vol->sb, "Cannot perform write to "
1891                                                 "inode 0x%lx, attribute type "
1892                                                 "0x%x, because extending the "
1893                                                 "allocation failed (error "
1894                                                 "code %i).", vi->i_ino,
1895                                                 (unsigned)
1896                                                 le32_to_cpu(ni->type), err);
1897                                 return err;
1898                         }
1899                 }
1900         }
1901         written = 0;
1902         /*
1903          * If the write starts beyond the initialized size, extend it up to the
1904          * beginning of the write and initialize all non-sparse space between
1905          * the old initialized size and the new one.  This automatically also
1906          * increments the vfs inode->i_size to keep it above or equal to the
1907          * initialized_size.
1908          */
1909         read_lock_irqsave(&ni->size_lock, flags);
1910         ll = ni->initialized_size;
1911         read_unlock_irqrestore(&ni->size_lock, flags);
1912         if (pos > ll) {
1913                 err = ntfs_attr_extend_initialized(ni, pos);
1914                 if (err < 0) {
1915                         ntfs_error(vol->sb, "Cannot perform write to inode "
1916                                         "0x%lx, attribute type 0x%x, because "
1917                                         "extending the initialized size "
1918                                         "failed (error code %i).", vi->i_ino,
1919                                         (unsigned)le32_to_cpu(ni->type), err);
1920                         status = err;
1921                         goto err_out;
1922                 }
1923         }
1924         /*
1925          * Determine the number of pages per cluster for non-resident
1926          * attributes.
1927          */
1928         nr_pages = 1;
1929         if (vol->cluster_size > PAGE_CACHE_SIZE && NInoNonResident(ni))
1930                 nr_pages = vol->cluster_size >> PAGE_CACHE_SHIFT;
1931         /* Finally, perform the actual write. */
1932         last_vcn = -1;
1933         if (likely(nr_segs == 1))
1934                 buf = iov->iov_base;
1935         do {
1936                 VCN vcn;
1937                 pgoff_t idx, start_idx;
1938                 unsigned ofs, do_pages, u;
1939                 size_t copied;
1940
1941                 start_idx = idx = pos >> PAGE_CACHE_SHIFT;
1942                 ofs = pos & ~PAGE_CACHE_MASK;
1943                 bytes = PAGE_CACHE_SIZE - ofs;
1944                 do_pages = 1;
1945                 if (nr_pages > 1) {
1946                         vcn = pos >> vol->cluster_size_bits;
1947                         if (vcn != last_vcn) {
1948                                 last_vcn = vcn;
1949                                 /*
1950                                  * Get the lcn of the vcn the write is in.  If
1951                                  * it is a hole, need to lock down all pages in
1952                                  * the cluster.
1953                                  */
1954                                 down_read(&ni->runlist.lock);
1955                                 lcn = ntfs_attr_vcn_to_lcn_nolock(ni, pos >>
1956                                                 vol->cluster_size_bits, false);
1957                                 up_read(&ni->runlist.lock);
1958                                 if (unlikely(lcn < LCN_HOLE)) {
1959                                         status = -EIO;
1960                                         if (lcn == LCN_ENOMEM)
1961                                                 status = -ENOMEM;
1962                                         else
1963                                                 ntfs_error(vol->sb, "Cannot "
1964                                                         "perform write to "
1965                                                         "inode 0x%lx, "
1966                                                         "attribute type 0x%x, "
1967                                                         "because the attribute "
1968                                                         "is corrupt.",
1969                                                         vi->i_ino, (unsigned)
1970                                                         le32_to_cpu(ni->type));
1971                                         break;
1972                                 }
1973                                 if (lcn == LCN_HOLE) {
1974                                         start_idx = (pos & ~(s64)
1975                                                         vol->cluster_size_mask)
1976                                                         >> PAGE_CACHE_SHIFT;
1977                                         bytes = vol->cluster_size - (pos &
1978                                                         vol->cluster_size_mask);
1979                                         do_pages = nr_pages;
1980                                 }
1981                         }
1982                 }
1983                 if (bytes > count)
1984                         bytes = count;
1985                 /*
1986                  * Bring in the user page(s) that we will copy from _first_.
1987                  * Otherwise there is a nasty deadlock on copying from the same
1988                  * page(s) as we are writing to, without it/them being marked
1989                  * up-to-date.  Note, at present there is nothing to stop the
1990                  * pages being swapped out between us bringing them into memory
1991                  * and doing the actual copying.
1992                  */
1993                 if (likely(nr_segs == 1))
1994                         ntfs_fault_in_pages_readable(buf, bytes);
1995                 else
1996                         ntfs_fault_in_pages_readable_iovec(iov, iov_ofs, bytes);
1997                 /* Get and lock @do_pages starting at index @start_idx. */
1998                 status = __ntfs_grab_cache_pages(mapping, start_idx, do_pages,
1999                                 pages, &cached_page);
2000                 if (unlikely(status))
2001                         break;
2002                 /*
2003                  * For non-resident attributes, we need to fill any holes with
2004                  * actual clusters and ensure all bufferes are mapped.  We also
2005                  * need to bring uptodate any buffers that are only partially
2006                  * being written to.
2007                  */
2008                 if (NInoNonResident(ni)) {
2009                         status = ntfs_prepare_pages_for_non_resident_write(
2010                                         pages, do_pages, pos, bytes);
2011                         if (unlikely(status)) {
2012                                 loff_t i_size;
2013
2014                                 do {
2015                                         unlock_page(pages[--do_pages]);
2016                                         page_cache_release(pages[do_pages]);
2017                                 } while (do_pages);
2018                                 /*
2019                                  * The write preparation may have instantiated
2020                                  * allocated space outside i_size.  Trim this
2021                                  * off again.  We can ignore any errors in this
2022                                  * case as we will just be waisting a bit of
2023                                  * allocated space, which is not a disaster.
2024                                  */
2025                                 i_size = i_size_read(vi);
2026                                 if (pos + bytes > i_size)
2027                                         vmtruncate(vi, i_size);
2028                                 break;
2029                         }
2030                 }
2031                 u = (pos >> PAGE_CACHE_SHIFT) - pages[0]->index;
2032                 if (likely(nr_segs == 1)) {
2033                         copied = ntfs_copy_from_user(pages + u, do_pages - u,
2034                                         ofs, buf, bytes);
2035                         buf += copied;
2036                 } else
2037                         copied = ntfs_copy_from_user_iovec(pages + u,
2038                                         do_pages - u, ofs, &iov, &iov_ofs,
2039                                         bytes);
2040                 ntfs_flush_dcache_pages(pages + u, do_pages - u);
2041                 status = ntfs_commit_pages_after_write(pages, do_pages, pos,
2042                                 bytes);
2043                 if (likely(!status)) {
2044                         written += copied;
2045                         count -= copied;
2046                         pos += copied;
2047                         if (unlikely(copied != bytes))
2048                                 status = -EFAULT;
2049                 }
2050                 do {
2051                         unlock_page(pages[--do_pages]);
2052                         mark_page_accessed(pages[do_pages]);
2053                         page_cache_release(pages[do_pages]);
2054                 } while (do_pages);
2055                 if (unlikely(status))
2056                         break;
2057                 balance_dirty_pages_ratelimited(mapping);
2058                 cond_resched();
2059         } while (count);
2060 err_out:
2061         *ppos = pos;
2062         if (cached_page)
2063                 page_cache_release(cached_page);
2064         ntfs_debug("Done.  Returning %s (written 0x%lx, status %li).",
2065                         written ? "written" : "status", (unsigned long)written,
2066                         (long)status);
2067         return written ? written : status;
2068 }
2069
2070 /**
2071  * ntfs_file_aio_write_nolock -
2072  */
2073 static ssize_t ntfs_file_aio_write_nolock(struct kiocb *iocb,
2074                 const struct iovec *iov, unsigned long nr_segs, loff_t *ppos)
2075 {
2076         struct file *file = iocb->ki_filp;
2077         struct address_space *mapping = file->f_mapping;
2078         struct inode *inode = mapping->host;
2079         loff_t pos;
2080         size_t count;           /* after file limit checks */
2081         ssize_t written, err;
2082
2083         count = 0;
2084         err = generic_segment_checks(iov, &nr_segs, &count, VERIFY_READ);
2085         if (err)
2086                 return err;
2087         pos = *ppos;
2088         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2089         /* We can write back this queue in page reclaim. */
2090         current->backing_dev_info = mapping->backing_dev_info;
2091         written = 0;
2092         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2093         if (err)
2094                 goto out;
2095         if (!count)
2096                 goto out;
2097         err = file_remove_suid(file);
2098         if (err)
2099                 goto out;
2100         file_update_time(file);
2101         written = ntfs_file_buffered_write(iocb, iov, nr_segs, pos, ppos,
2102                         count);
2103 out:
2104         current->backing_dev_info = NULL;
2105         return written ? written : err;
2106 }
2107
2108 /**
2109  * ntfs_file_aio_write -
2110  */
2111 static ssize_t ntfs_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2112                 unsigned long nr_segs, loff_t pos)
2113 {
2114         struct file *file = iocb->ki_filp;
2115         struct address_space *mapping = file->f_mapping;
2116         struct inode *inode = mapping->host;
2117         ssize_t ret;
2118
2119         BUG_ON(iocb->ki_pos != pos);
2120
2121         mutex_lock(&inode->i_mutex);
2122         ret = ntfs_file_aio_write_nolock(iocb, iov, nr_segs, &iocb->ki_pos);
2123         mutex_unlock(&inode->i_mutex);
2124         if (ret > 0) {
2125                 int err = generic_write_sync(file, pos, ret);
2126                 if (err < 0)
2127                         ret = err;
2128         }
2129         return ret;
2130 }
2131
2132 /**
2133  * ntfs_file_fsync - sync a file to disk
2134  * @filp:       file to be synced
2135  * @datasync:   if non-zero only flush user data and not metadata
2136  *
2137  * Data integrity sync of a file to disk.  Used for fsync, fdatasync, and msync
2138  * system calls.  This function is inspired by fs/buffer.c::file_fsync().
2139  *
2140  * If @datasync is false, write the mft record and all associated extent mft
2141  * records as well as the $DATA attribute and then sync the block device.
2142  *
2143  * If @datasync is true and the attribute is non-resident, we skip the writing
2144  * of the mft record and all associated extent mft records (this might still
2145  * happen due to the write_inode_now() call).
2146  *
2147  * Also, if @datasync is true, we do not wait on the inode to be written out
2148  * but we always wait on the page cache pages to be written out.
2149  *
2150  * Locking: Caller must hold i_mutex on the inode.
2151  *
2152  * TODO: We should probably also write all attribute/index inodes associated
2153  * with this inode but since we have no simple way of getting to them we ignore
2154  * this problem for now.
2155  */
2156 static int ntfs_file_fsync(struct file *filp, int datasync)
2157 {
2158         struct inode *vi = filp->f_mapping->host;
2159         int err, ret = 0;
2160
2161         ntfs_debug("Entering for inode 0x%lx.", vi->i_ino);
2162         BUG_ON(S_ISDIR(vi->i_mode));
2163         if (!datasync || !NInoNonResident(NTFS_I(vi)))
2164                 ret = __ntfs_write_inode(vi, 1);
2165         write_inode_now(vi, !datasync);
2166         /*
2167          * NOTE: If we were to use mapping->private_list (see ext2 and
2168          * fs/buffer.c) for dirty blocks then we could optimize the below to be
2169          * sync_mapping_buffers(vi->i_mapping).
2170          */
2171         err = sync_blockdev(vi->i_sb->s_bdev);
2172         if (unlikely(err && !ret))
2173                 ret = err;
2174         if (likely(!ret))
2175                 ntfs_debug("Done.");
2176         else
2177                 ntfs_warning(vi->i_sb, "Failed to f%ssync inode 0x%lx.  Error "
2178                                 "%u.", datasync ? "data" : "", vi->i_ino, -ret);
2179         return ret;
2180 }
2181
2182 #endif /* NTFS_RW */
2183
2184 const struct file_operations ntfs_file_ops = {
2185         .llseek         = generic_file_llseek,   /* Seek inside file. */
2186         .read           = do_sync_read,          /* Read from file. */
2187         .aio_read       = generic_file_aio_read, /* Async read from file. */
2188 #ifdef NTFS_RW
2189         .write          = do_sync_write,         /* Write to file. */
2190         .aio_write      = ntfs_file_aio_write,   /* Async write to file. */
2191         /*.release      = ,*/                    /* Last file is closed.  See
2192                                                     fs/ext2/file.c::
2193                                                     ext2_release_file() for
2194                                                     how to use this to discard
2195                                                     preallocated space for
2196                                                     write opened files. */
2197         .fsync          = ntfs_file_fsync,       /* Sync a file to disk. */
2198         /*.aio_fsync    = ,*/                    /* Sync all outstanding async
2199                                                     i/o operations on a
2200                                                     kiocb. */
2201 #endif /* NTFS_RW */
2202         /*.ioctl        = ,*/                    /* Perform function on the
2203                                                     mounted filesystem. */
2204         .mmap           = generic_file_mmap,     /* Mmap file. */
2205         .open           = ntfs_file_open,        /* Open file. */
2206         .splice_read    = generic_file_splice_read /* Zero-copy data send with
2207                                                     the data source being on
2208                                                     the ntfs partition.  We do
2209                                                     not need to care about the
2210                                                     data destination. */
2211         /*.sendpage     = ,*/                    /* Zero-copy data send with
2212                                                     the data destination being
2213                                                     on the ntfs partition.  We
2214                                                     do not need to care about
2215                                                     the data source. */
2216 };
2217
2218 const struct inode_operations ntfs_file_inode_ops = {
2219 #ifdef NTFS_RW
2220         .truncate       = ntfs_truncate_vfs,
2221         .setattr        = ntfs_setattr,
2222 #endif /* NTFS_RW */
2223 };
2224
2225 const struct file_operations ntfs_empty_file_ops = {};
2226
2227 const struct inode_operations ntfs_empty_inode_ops = {};