NTFS: Fix a potential overflow by casting (index + 1) to s64 before doing a
[linux-2.6] / fs / ntfs / file.c
1 /*
2  * file.c - NTFS kernel file operations.  Part of the Linux-NTFS project.
3  *
4  * Copyright (c) 2001-2005 Anton Altaparmakov
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/pagemap.h>
24 #include <linux/pagevec.h>
25 #include <linux/sched.h>
26 #include <linux/swap.h>
27 #include <linux/uio.h>
28 #include <linux/writeback.h>
29
30 #include <asm/page.h>
31 #include <asm/uaccess.h>
32
33 #include "attrib.h"
34 #include "bitmap.h"
35 #include "inode.h"
36 #include "debug.h"
37 #include "lcnalloc.h"
38 #include "malloc.h"
39 #include "mft.h"
40 #include "ntfs.h"
41
42 /**
43  * ntfs_file_open - called when an inode is about to be opened
44  * @vi:         inode to be opened
45  * @filp:       file structure describing the inode
46  *
47  * Limit file size to the page cache limit on architectures where unsigned long
48  * is 32-bits. This is the most we can do for now without overflowing the page
49  * cache page index. Doing it this way means we don't run into problems because
50  * of existing too large files. It would be better to allow the user to read
51  * the beginning of the file but I doubt very much anyone is going to hit this
52  * check on a 32-bit architecture, so there is no point in adding the extra
53  * complexity required to support this.
54  *
55  * On 64-bit architectures, the check is hopefully optimized away by the
56  * compiler.
57  *
58  * After the check passes, just call generic_file_open() to do its work.
59  */
60 static int ntfs_file_open(struct inode *vi, struct file *filp)
61 {
62         if (sizeof(unsigned long) < 8) {
63                 if (i_size_read(vi) > MAX_LFS_FILESIZE)
64                         return -EFBIG;
65         }
66         return generic_file_open(vi, filp);
67 }
68
69 #ifdef NTFS_RW
70
71 /**
72  * ntfs_attr_extend_initialized - extend the initialized size of an attribute
73  * @ni:                 ntfs inode of the attribute to extend
74  * @new_init_size:      requested new initialized size in bytes
75  * @cached_page:        store any allocated but unused page here
76  * @lru_pvec:           lru-buffering pagevec of the caller
77  *
78  * Extend the initialized size of an attribute described by the ntfs inode @ni
79  * to @new_init_size bytes.  This involves zeroing any non-sparse space between
80  * the old initialized size and @new_init_size both in the page cache and on
81  * disk (if relevant complete pages are already uptodate in the page cache then
82  * these are simply marked dirty).
83  *
84  * As a side-effect, the file size (vfs inode->i_size) may be incremented as,
85  * in the resident attribute case, it is tied to the initialized size and, in
86  * the non-resident attribute case, it may not fall below the initialized size.
87  *
88  * Note that if the attribute is resident, we do not need to touch the page
89  * cache at all.  This is because if the page cache page is not uptodate we
90  * bring it uptodate later, when doing the write to the mft record since we
91  * then already have the page mapped.  And if the page is uptodate, the
92  * non-initialized region will already have been zeroed when the page was
93  * brought uptodate and the region may in fact already have been overwritten
94  * with new data via mmap() based writes, so we cannot just zero it.  And since
95  * POSIX specifies that the behaviour of resizing a file whilst it is mmap()ped
96  * is unspecified, we choose not to do zeroing and thus we do not need to touch
97  * the page at all.  For a more detailed explanation see ntfs_truncate() in
98  * fs/ntfs/inode.c.
99  *
100  * @cached_page and @lru_pvec are just optimizations for dealing with multiple
101  * pages.
102  *
103  * Return 0 on success and -errno on error.  In the case that an error is
104  * encountered it is possible that the initialized size will already have been
105  * incremented some way towards @new_init_size but it is guaranteed that if
106  * this is the case, the necessary zeroing will also have happened and that all
107  * metadata is self-consistent.
108  *
109  * Locking: i_sem on the vfs inode corrseponsind to the ntfs inode @ni must be
110  *          held by the caller.
111  */
112 static int ntfs_attr_extend_initialized(ntfs_inode *ni, const s64 new_init_size,
113                 struct page **cached_page, struct pagevec *lru_pvec)
114 {
115         s64 old_init_size;
116         loff_t old_i_size;
117         pgoff_t index, end_index;
118         unsigned long flags;
119         struct inode *vi = VFS_I(ni);
120         ntfs_inode *base_ni;
121         MFT_RECORD *m = NULL;
122         ATTR_RECORD *a;
123         ntfs_attr_search_ctx *ctx = NULL;
124         struct address_space *mapping;
125         struct page *page = NULL;
126         u8 *kattr;
127         int err;
128         u32 attr_len;
129
130         read_lock_irqsave(&ni->size_lock, flags);
131         old_init_size = ni->initialized_size;
132         old_i_size = i_size_read(vi);
133         BUG_ON(new_init_size > ni->allocated_size);
134         read_unlock_irqrestore(&ni->size_lock, flags);
135         ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, "
136                         "old_initialized_size 0x%llx, "
137                         "new_initialized_size 0x%llx, i_size 0x%llx.",
138                         vi->i_ino, (unsigned)le32_to_cpu(ni->type),
139                         (unsigned long long)old_init_size,
140                         (unsigned long long)new_init_size, old_i_size);
141         if (!NInoAttr(ni))
142                 base_ni = ni;
143         else
144                 base_ni = ni->ext.base_ntfs_ino;
145         /* Use goto to reduce indentation and we need the label below anyway. */
146         if (NInoNonResident(ni))
147                 goto do_non_resident_extend;
148         BUG_ON(old_init_size != old_i_size);
149         m = map_mft_record(base_ni);
150         if (IS_ERR(m)) {
151                 err = PTR_ERR(m);
152                 m = NULL;
153                 goto err_out;
154         }
155         ctx = ntfs_attr_get_search_ctx(base_ni, m);
156         if (unlikely(!ctx)) {
157                 err = -ENOMEM;
158                 goto err_out;
159         }
160         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
161                         CASE_SENSITIVE, 0, NULL, 0, ctx);
162         if (unlikely(err)) {
163                 if (err == -ENOENT)
164                         err = -EIO;
165                 goto err_out;
166         }
167         m = ctx->mrec;
168         a = ctx->attr;
169         BUG_ON(a->non_resident);
170         /* The total length of the attribute value. */
171         attr_len = le32_to_cpu(a->data.resident.value_length);
172         BUG_ON(old_i_size != (loff_t)attr_len);
173         /*
174          * Do the zeroing in the mft record and update the attribute size in
175          * the mft record.
176          */
177         kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
178         memset(kattr + attr_len, 0, new_init_size - attr_len);
179         a->data.resident.value_length = cpu_to_le32((u32)new_init_size);
180         /* Finally, update the sizes in the vfs and ntfs inodes. */
181         write_lock_irqsave(&ni->size_lock, flags);
182         i_size_write(vi, new_init_size);
183         ni->initialized_size = new_init_size;
184         write_unlock_irqrestore(&ni->size_lock, flags);
185         goto done;
186 do_non_resident_extend:
187         /*
188          * If the new initialized size @new_init_size exceeds the current file
189          * size (vfs inode->i_size), we need to extend the file size to the
190          * new initialized size.
191          */
192         if (new_init_size > old_i_size) {
193                 m = map_mft_record(base_ni);
194                 if (IS_ERR(m)) {
195                         err = PTR_ERR(m);
196                         m = NULL;
197                         goto err_out;
198                 }
199                 ctx = ntfs_attr_get_search_ctx(base_ni, m);
200                 if (unlikely(!ctx)) {
201                         err = -ENOMEM;
202                         goto err_out;
203                 }
204                 err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
205                                 CASE_SENSITIVE, 0, NULL, 0, ctx);
206                 if (unlikely(err)) {
207                         if (err == -ENOENT)
208                                 err = -EIO;
209                         goto err_out;
210                 }
211                 m = ctx->mrec;
212                 a = ctx->attr;
213                 BUG_ON(!a->non_resident);
214                 BUG_ON(old_i_size != (loff_t)
215                                 sle64_to_cpu(a->data.non_resident.data_size));
216                 a->data.non_resident.data_size = cpu_to_sle64(new_init_size);
217                 flush_dcache_mft_record_page(ctx->ntfs_ino);
218                 mark_mft_record_dirty(ctx->ntfs_ino);
219                 /* Update the file size in the vfs inode. */
220                 i_size_write(vi, new_init_size);
221                 ntfs_attr_put_search_ctx(ctx);
222                 ctx = NULL;
223                 unmap_mft_record(base_ni);
224                 m = NULL;
225         }
226         mapping = vi->i_mapping;
227         index = old_init_size >> PAGE_CACHE_SHIFT;
228         end_index = (new_init_size + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
229         do {
230                 /*
231                  * Read the page.  If the page is not present, this will zero
232                  * the uninitialized regions for us.
233                  */
234                 page = read_cache_page(mapping, index,
235                                 (filler_t*)mapping->a_ops->readpage, NULL);
236                 if (IS_ERR(page)) {
237                         err = PTR_ERR(page);
238                         goto init_err_out;
239                 }
240                 wait_on_page_locked(page);
241                 if (unlikely(!PageUptodate(page) || PageError(page))) {
242                         page_cache_release(page);
243                         err = -EIO;
244                         goto init_err_out;
245                 }
246                 /*
247                  * Update the initialized size in the ntfs inode.  This is
248                  * enough to make ntfs_writepage() work.
249                  */
250                 write_lock_irqsave(&ni->size_lock, flags);
251                 ni->initialized_size = (s64)(index + 1) << PAGE_CACHE_SHIFT;
252                 if (ni->initialized_size > new_init_size)
253                         ni->initialized_size = new_init_size;
254                 write_unlock_irqrestore(&ni->size_lock, flags);
255                 /* Set the page dirty so it gets written out. */
256                 set_page_dirty(page);
257                 page_cache_release(page);
258                 /*
259                  * Play nice with the vm and the rest of the system.  This is
260                  * very much needed as we can potentially be modifying the
261                  * initialised size from a very small value to a really huge
262                  * value, e.g.
263                  *      f = open(somefile, O_TRUNC);
264                  *      truncate(f, 10GiB);
265                  *      seek(f, 10GiB);
266                  *      write(f, 1);
267                  * And this would mean we would be marking dirty hundreds of
268                  * thousands of pages or as in the above example more than
269                  * two and a half million pages!
270                  *
271                  * TODO: For sparse pages could optimize this workload by using
272                  * the FsMisc / MiscFs page bit as a "PageIsSparse" bit.  This
273                  * would be set in readpage for sparse pages and here we would
274                  * not need to mark dirty any pages which have this bit set.
275                  * The only caveat is that we have to clear the bit everywhere
276                  * where we allocate any clusters that lie in the page or that
277                  * contain the page.
278                  *
279                  * TODO: An even greater optimization would be for us to only
280                  * call readpage() on pages which are not in sparse regions as
281                  * determined from the runlist.  This would greatly reduce the
282                  * number of pages we read and make dirty in the case of sparse
283                  * files.
284                  */
285                 balance_dirty_pages_ratelimited(mapping);
286                 cond_resched();
287         } while (++index < end_index);
288         read_lock_irqsave(&ni->size_lock, flags);
289         BUG_ON(ni->initialized_size != new_init_size);
290         read_unlock_irqrestore(&ni->size_lock, flags);
291         /* Now bring in sync the initialized_size in the mft record. */
292         m = map_mft_record(base_ni);
293         if (IS_ERR(m)) {
294                 err = PTR_ERR(m);
295                 m = NULL;
296                 goto init_err_out;
297         }
298         ctx = ntfs_attr_get_search_ctx(base_ni, m);
299         if (unlikely(!ctx)) {
300                 err = -ENOMEM;
301                 goto init_err_out;
302         }
303         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
304                         CASE_SENSITIVE, 0, NULL, 0, ctx);
305         if (unlikely(err)) {
306                 if (err == -ENOENT)
307                         err = -EIO;
308                 goto init_err_out;
309         }
310         m = ctx->mrec;
311         a = ctx->attr;
312         BUG_ON(!a->non_resident);
313         a->data.non_resident.initialized_size = cpu_to_sle64(new_init_size);
314 done:
315         flush_dcache_mft_record_page(ctx->ntfs_ino);
316         mark_mft_record_dirty(ctx->ntfs_ino);
317         if (ctx)
318                 ntfs_attr_put_search_ctx(ctx);
319         if (m)
320                 unmap_mft_record(base_ni);
321         ntfs_debug("Done, initialized_size 0x%llx, i_size 0x%llx.",
322                         (unsigned long long)new_init_size, i_size_read(vi));
323         return 0;
324 init_err_out:
325         write_lock_irqsave(&ni->size_lock, flags);
326         ni->initialized_size = old_init_size;
327         write_unlock_irqrestore(&ni->size_lock, flags);
328 err_out:
329         if (ctx)
330                 ntfs_attr_put_search_ctx(ctx);
331         if (m)
332                 unmap_mft_record(base_ni);
333         ntfs_debug("Failed.  Returning error code %i.", err);
334         return err;
335 }
336
337 /**
338  * ntfs_fault_in_pages_readable -
339  *
340  * Fault a number of userspace pages into pagetables.
341  *
342  * Unlike include/linux/pagemap.h::fault_in_pages_readable(), this one copes
343  * with more than two userspace pages as well as handling the single page case
344  * elegantly.
345  *
346  * If you find this difficult to understand, then think of the while loop being
347  * the following code, except that we do without the integer variable ret:
348  *
349  *      do {
350  *              ret = __get_user(c, uaddr);
351  *              uaddr += PAGE_SIZE;
352  *      } while (!ret && uaddr < end);
353  *
354  * Note, the final __get_user() may well run out-of-bounds of the user buffer,
355  * but _not_ out-of-bounds of the page the user buffer belongs to, and since
356  * this is only a read and not a write, and since it is still in the same page,
357  * it should not matter and this makes the code much simpler.
358  */
359 static inline void ntfs_fault_in_pages_readable(const char __user *uaddr,
360                 int bytes)
361 {
362         const char __user *end;
363         volatile char c;
364
365         /* Set @end to the first byte outside the last page we care about. */
366         end = (const char __user*)PAGE_ALIGN((ptrdiff_t __user)uaddr + bytes);
367
368         while (!__get_user(c, uaddr) && (uaddr += PAGE_SIZE, uaddr < end))
369                 ;
370 }
371
372 /**
373  * ntfs_fault_in_pages_readable_iovec -
374  *
375  * Same as ntfs_fault_in_pages_readable() but operates on an array of iovecs.
376  */
377 static inline void ntfs_fault_in_pages_readable_iovec(const struct iovec *iov,
378                 size_t iov_ofs, int bytes)
379 {
380         do {
381                 const char __user *buf;
382                 unsigned len;
383
384                 buf = iov->iov_base + iov_ofs;
385                 len = iov->iov_len - iov_ofs;
386                 if (len > bytes)
387                         len = bytes;
388                 ntfs_fault_in_pages_readable(buf, len);
389                 bytes -= len;
390                 iov++;
391                 iov_ofs = 0;
392         } while (bytes);
393 }
394
395 /**
396  * __ntfs_grab_cache_pages - obtain a number of locked pages
397  * @mapping:    address space mapping from which to obtain page cache pages
398  * @index:      starting index in @mapping at which to begin obtaining pages
399  * @nr_pages:   number of page cache pages to obtain
400  * @pages:      array of pages in which to return the obtained page cache pages
401  * @cached_page: allocated but as yet unused page
402  * @lru_pvec:   lru-buffering pagevec of caller
403  *
404  * Obtain @nr_pages locked page cache pages from the mapping @maping and
405  * starting at index @index.
406  *
407  * If a page is newly created, increment its refcount and add it to the
408  * caller's lru-buffering pagevec @lru_pvec.
409  *
410  * This is the same as mm/filemap.c::__grab_cache_page(), except that @nr_pages
411  * are obtained at once instead of just one page and that 0 is returned on
412  * success and -errno on error.
413  *
414  * Note, the page locks are obtained in ascending page index order.
415  */
416 static inline int __ntfs_grab_cache_pages(struct address_space *mapping,
417                 pgoff_t index, const unsigned nr_pages, struct page **pages,
418                 struct page **cached_page, struct pagevec *lru_pvec)
419 {
420         int err, nr;
421
422         BUG_ON(!nr_pages);
423         err = nr = 0;
424         do {
425                 pages[nr] = find_lock_page(mapping, index);
426                 if (!pages[nr]) {
427                         if (!*cached_page) {
428                                 *cached_page = page_cache_alloc(mapping);
429                                 if (unlikely(!*cached_page)) {
430                                         err = -ENOMEM;
431                                         goto err_out;
432                                 }
433                         }
434                         err = add_to_page_cache(*cached_page, mapping, index,
435                                         GFP_KERNEL);
436                         if (unlikely(err)) {
437                                 if (err == -EEXIST)
438                                         continue;
439                                 goto err_out;
440                         }
441                         pages[nr] = *cached_page;
442                         page_cache_get(*cached_page);
443                         if (unlikely(!pagevec_add(lru_pvec, *cached_page)))
444                                 __pagevec_lru_add(lru_pvec);
445                         *cached_page = NULL;
446                 }
447                 index++;
448                 nr++;
449         } while (nr < nr_pages);
450 out:
451         return err;
452 err_out:
453         while (nr > 0) {
454                 unlock_page(pages[--nr]);
455                 page_cache_release(pages[nr]);
456         }
457         goto out;
458 }
459
460 static inline int ntfs_submit_bh_for_read(struct buffer_head *bh)
461 {
462         lock_buffer(bh);
463         get_bh(bh);
464         bh->b_end_io = end_buffer_read_sync;
465         return submit_bh(READ, bh);
466 }
467
468 /**
469  * ntfs_prepare_pages_for_non_resident_write - prepare pages for receiving data
470  * @pages:      array of destination pages
471  * @nr_pages:   number of pages in @pages
472  * @pos:        byte position in file at which the write begins
473  * @bytes:      number of bytes to be written
474  *
475  * This is called for non-resident attributes from ntfs_file_buffered_write()
476  * with i_sem held on the inode (@pages[0]->mapping->host).  There are
477  * @nr_pages pages in @pages which are locked but not kmap()ped.  The source
478  * data has not yet been copied into the @pages.
479  * 
480  * Need to fill any holes with actual clusters, allocate buffers if necessary,
481  * ensure all the buffers are mapped, and bring uptodate any buffers that are
482  * only partially being written to.
483  *
484  * If @nr_pages is greater than one, we are guaranteed that the cluster size is
485  * greater than PAGE_CACHE_SIZE, that all pages in @pages are entirely inside
486  * the same cluster and that they are the entirety of that cluster, and that
487  * the cluster is sparse, i.e. we need to allocate a cluster to fill the hole.
488  *
489  * i_size is not to be modified yet.
490  *
491  * Return 0 on success or -errno on error.
492  */
493 static int ntfs_prepare_pages_for_non_resident_write(struct page **pages,
494                 unsigned nr_pages, s64 pos, size_t bytes)
495 {
496         VCN vcn, highest_vcn = 0, cpos, cend, bh_cpos, bh_cend;
497         LCN lcn;
498         s64 bh_pos, vcn_len, end, initialized_size;
499         sector_t lcn_block;
500         struct page *page;
501         struct inode *vi;
502         ntfs_inode *ni, *base_ni = NULL;
503         ntfs_volume *vol;
504         runlist_element *rl, *rl2;
505         struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
506         ntfs_attr_search_ctx *ctx = NULL;
507         MFT_RECORD *m = NULL;
508         ATTR_RECORD *a = NULL;
509         unsigned long flags;
510         u32 attr_rec_len = 0;
511         unsigned blocksize, u;
512         int err, mp_size;
513         BOOL rl_write_locked, was_hole, is_retry;
514         unsigned char blocksize_bits;
515         struct {
516                 u8 runlist_merged:1;
517                 u8 mft_attr_mapped:1;
518                 u8 mp_rebuilt:1;
519                 u8 attr_switched:1;
520         } status = { 0, 0, 0, 0 };
521
522         BUG_ON(!nr_pages);
523         BUG_ON(!pages);
524         BUG_ON(!*pages);
525         vi = pages[0]->mapping->host;
526         ni = NTFS_I(vi);
527         vol = ni->vol;
528         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, start page "
529                         "index 0x%lx, nr_pages 0x%x, pos 0x%llx, bytes 0x%zx.",
530                         vi->i_ino, ni->type, pages[0]->index, nr_pages,
531                         (long long)pos, bytes);
532         blocksize_bits = vi->i_blkbits;
533         blocksize = 1 << blocksize_bits;
534         u = 0;
535         do {
536                 struct page *page = pages[u];
537                 /*
538                  * create_empty_buffers() will create uptodate/dirty buffers if
539                  * the page is uptodate/dirty.
540                  */
541                 if (!page_has_buffers(page)) {
542                         create_empty_buffers(page, blocksize, 0);
543                         if (unlikely(!page_has_buffers(page)))
544                                 return -ENOMEM;
545                 }
546         } while (++u < nr_pages);
547         rl_write_locked = FALSE;
548         rl = NULL;
549         err = 0;
550         vcn = lcn = -1;
551         vcn_len = 0;
552         lcn_block = -1;
553         was_hole = FALSE;
554         cpos = pos >> vol->cluster_size_bits;
555         end = pos + bytes;
556         cend = (end + vol->cluster_size - 1) >> vol->cluster_size_bits;
557         /*
558          * Loop over each page and for each page over each buffer.  Use goto to
559          * reduce indentation.
560          */
561         u = 0;
562 do_next_page:
563         page = pages[u];
564         bh_pos = (s64)page->index << PAGE_CACHE_SHIFT;
565         bh = head = page_buffers(page);
566         do {
567                 VCN cdelta;
568                 s64 bh_end;
569                 unsigned bh_cofs;
570
571                 /* Clear buffer_new on all buffers to reinitialise state. */
572                 if (buffer_new(bh))
573                         clear_buffer_new(bh);
574                 bh_end = bh_pos + blocksize;
575                 bh_cpos = bh_pos >> vol->cluster_size_bits;
576                 bh_cofs = bh_pos & vol->cluster_size_mask;
577                 if (buffer_mapped(bh)) {
578                         /*
579                          * The buffer is already mapped.  If it is uptodate,
580                          * ignore it.
581                          */
582                         if (buffer_uptodate(bh))
583                                 continue;
584                         /*
585                          * The buffer is not uptodate.  If the page is uptodate
586                          * set the buffer uptodate and otherwise ignore it.
587                          */
588                         if (PageUptodate(page)) {
589                                 set_buffer_uptodate(bh);
590                                 continue;
591                         }
592                         /*
593                          * Neither the page nor the buffer are uptodate.  If
594                          * the buffer is only partially being written to, we
595                          * need to read it in before the write, i.e. now.
596                          */
597                         if ((bh_pos < pos && bh_end > pos) ||
598                                         (bh_pos < end && bh_end > end)) {
599                                 /*
600                                  * If the buffer is fully or partially within
601                                  * the initialized size, do an actual read.
602                                  * Otherwise, simply zero the buffer.
603                                  */
604                                 read_lock_irqsave(&ni->size_lock, flags);
605                                 initialized_size = ni->initialized_size;
606                                 read_unlock_irqrestore(&ni->size_lock, flags);
607                                 if (bh_pos < initialized_size) {
608                                         ntfs_submit_bh_for_read(bh);
609                                         *wait_bh++ = bh;
610                                 } else {
611                                         u8 *kaddr = kmap_atomic(page, KM_USER0);
612                                         memset(kaddr + bh_offset(bh), 0,
613                                                         blocksize);
614                                         kunmap_atomic(kaddr, KM_USER0);
615                                         flush_dcache_page(page);
616                                         set_buffer_uptodate(bh);
617                                 }
618                         }
619                         continue;
620                 }
621                 /* Unmapped buffer.  Need to map it. */
622                 bh->b_bdev = vol->sb->s_bdev;
623                 /*
624                  * If the current buffer is in the same clusters as the map
625                  * cache, there is no need to check the runlist again.  The
626                  * map cache is made up of @vcn, which is the first cached file
627                  * cluster, @vcn_len which is the number of cached file
628                  * clusters, @lcn is the device cluster corresponding to @vcn,
629                  * and @lcn_block is the block number corresponding to @lcn.
630                  */
631                 cdelta = bh_cpos - vcn;
632                 if (likely(!cdelta || (cdelta > 0 && cdelta < vcn_len))) {
633 map_buffer_cached:
634                         BUG_ON(lcn < 0);
635                         bh->b_blocknr = lcn_block +
636                                         (cdelta << (vol->cluster_size_bits -
637                                         blocksize_bits)) +
638                                         (bh_cofs >> blocksize_bits);
639                         set_buffer_mapped(bh);
640                         /*
641                          * If the page is uptodate so is the buffer.  If the
642                          * buffer is fully outside the write, we ignore it if
643                          * it was already allocated and we mark it dirty so it
644                          * gets written out if we allocated it.  On the other
645                          * hand, if we allocated the buffer but we are not
646                          * marking it dirty we set buffer_new so we can do
647                          * error recovery.
648                          */
649                         if (PageUptodate(page)) {
650                                 if (!buffer_uptodate(bh))
651                                         set_buffer_uptodate(bh);
652                                 if (unlikely(was_hole)) {
653                                         /* We allocated the buffer. */
654                                         unmap_underlying_metadata(bh->b_bdev,
655                                                         bh->b_blocknr);
656                                         if (bh_end <= pos || bh_pos >= end)
657                                                 mark_buffer_dirty(bh);
658                                         else
659                                                 set_buffer_new(bh);
660                                 }
661                                 continue;
662                         }
663                         /* Page is _not_ uptodate. */
664                         if (likely(!was_hole)) {
665                                 /*
666                                  * Buffer was already allocated.  If it is not
667                                  * uptodate and is only partially being written
668                                  * to, we need to read it in before the write,
669                                  * i.e. now.
670                                  */
671                                 if (!buffer_uptodate(bh) && bh_pos < end &&
672                                                 bh_end > pos &&
673                                                 (bh_pos < pos ||
674                                                 bh_end > end)) {
675                                         /*
676                                          * If the buffer is fully or partially
677                                          * within the initialized size, do an
678                                          * actual read.  Otherwise, simply zero
679                                          * the buffer.
680                                          */
681                                         read_lock_irqsave(&ni->size_lock,
682                                                         flags);
683                                         initialized_size = ni->initialized_size;
684                                         read_unlock_irqrestore(&ni->size_lock,
685                                                         flags);
686                                         if (bh_pos < initialized_size) {
687                                                 ntfs_submit_bh_for_read(bh);
688                                                 *wait_bh++ = bh;
689                                         } else {
690                                                 u8 *kaddr = kmap_atomic(page,
691                                                                 KM_USER0);
692                                                 memset(kaddr + bh_offset(bh),
693                                                                 0, blocksize);
694                                                 kunmap_atomic(kaddr, KM_USER0);
695                                                 flush_dcache_page(page);
696                                                 set_buffer_uptodate(bh);
697                                         }
698                                 }
699                                 continue;
700                         }
701                         /* We allocated the buffer. */
702                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
703                         /*
704                          * If the buffer is fully outside the write, zero it,
705                          * set it uptodate, and mark it dirty so it gets
706                          * written out.  If it is partially being written to,
707                          * zero region surrounding the write but leave it to
708                          * commit write to do anything else.  Finally, if the
709                          * buffer is fully being overwritten, do nothing.
710                          */
711                         if (bh_end <= pos || bh_pos >= end) {
712                                 if (!buffer_uptodate(bh)) {
713                                         u8 *kaddr = kmap_atomic(page, KM_USER0);
714                                         memset(kaddr + bh_offset(bh), 0,
715                                                         blocksize);
716                                         kunmap_atomic(kaddr, KM_USER0);
717                                         flush_dcache_page(page);
718                                         set_buffer_uptodate(bh);
719                                 }
720                                 mark_buffer_dirty(bh);
721                                 continue;
722                         }
723                         set_buffer_new(bh);
724                         if (!buffer_uptodate(bh) &&
725                                         (bh_pos < pos || bh_end > end)) {
726                                 u8 *kaddr;
727                                 unsigned pofs;
728                                         
729                                 kaddr = kmap_atomic(page, KM_USER0);
730                                 if (bh_pos < pos) {
731                                         pofs = bh_pos & ~PAGE_CACHE_MASK;
732                                         memset(kaddr + pofs, 0, pos - bh_pos);
733                                 }
734                                 if (bh_end > end) {
735                                         pofs = end & ~PAGE_CACHE_MASK;
736                                         memset(kaddr + pofs, 0, bh_end - end);
737                                 }
738                                 kunmap_atomic(kaddr, KM_USER0);
739                                 flush_dcache_page(page);
740                         }
741                         continue;
742                 }
743                 /*
744                  * Slow path: this is the first buffer in the cluster.  If it
745                  * is outside allocated size and is not uptodate, zero it and
746                  * set it uptodate.
747                  */
748                 read_lock_irqsave(&ni->size_lock, flags);
749                 initialized_size = ni->allocated_size;
750                 read_unlock_irqrestore(&ni->size_lock, flags);
751                 if (bh_pos > initialized_size) {
752                         if (PageUptodate(page)) {
753                                 if (!buffer_uptodate(bh))
754                                         set_buffer_uptodate(bh);
755                         } else if (!buffer_uptodate(bh)) {
756                                 u8 *kaddr = kmap_atomic(page, KM_USER0);
757                                 memset(kaddr + bh_offset(bh), 0, blocksize);
758                                 kunmap_atomic(kaddr, KM_USER0);
759                                 flush_dcache_page(page);
760                                 set_buffer_uptodate(bh);
761                         }
762                         continue;
763                 }
764                 is_retry = FALSE;
765                 if (!rl) {
766                         down_read(&ni->runlist.lock);
767 retry_remap:
768                         rl = ni->runlist.rl;
769                 }
770                 if (likely(rl != NULL)) {
771                         /* Seek to element containing target cluster. */
772                         while (rl->length && rl[1].vcn <= bh_cpos)
773                                 rl++;
774                         lcn = ntfs_rl_vcn_to_lcn(rl, bh_cpos);
775                         if (likely(lcn >= 0)) {
776                                 /*
777                                  * Successful remap, setup the map cache and
778                                  * use that to deal with the buffer.
779                                  */
780                                 was_hole = FALSE;
781                                 vcn = bh_cpos;
782                                 vcn_len = rl[1].vcn - vcn;
783                                 lcn_block = lcn << (vol->cluster_size_bits -
784                                                 blocksize_bits);
785                                 cdelta = 0;
786                                 /*
787                                  * If the number of remaining clusters touched
788                                  * by the write is smaller or equal to the
789                                  * number of cached clusters, unlock the
790                                  * runlist as the map cache will be used from
791                                  * now on.
792                                  */
793                                 if (likely(vcn + vcn_len >= cend)) {
794                                         if (rl_write_locked) {
795                                                 up_write(&ni->runlist.lock);
796                                                 rl_write_locked = FALSE;
797                                         } else
798                                                 up_read(&ni->runlist.lock);
799                                         rl = NULL;
800                                 }
801                                 goto map_buffer_cached;
802                         }
803                 } else
804                         lcn = LCN_RL_NOT_MAPPED;
805                 /*
806                  * If it is not a hole and not out of bounds, the runlist is
807                  * probably unmapped so try to map it now.
808                  */
809                 if (unlikely(lcn != LCN_HOLE && lcn != LCN_ENOENT)) {
810                         if (likely(!is_retry && lcn == LCN_RL_NOT_MAPPED)) {
811                                 /* Attempt to map runlist. */
812                                 if (!rl_write_locked) {
813                                         /*
814                                          * We need the runlist locked for
815                                          * writing, so if it is locked for
816                                          * reading relock it now and retry in
817                                          * case it changed whilst we dropped
818                                          * the lock.
819                                          */
820                                         up_read(&ni->runlist.lock);
821                                         down_write(&ni->runlist.lock);
822                                         rl_write_locked = TRUE;
823                                         goto retry_remap;
824                                 }
825                                 err = ntfs_map_runlist_nolock(ni, bh_cpos,
826                                                 NULL);
827                                 if (likely(!err)) {
828                                         is_retry = TRUE;
829                                         goto retry_remap;
830                                 }
831                                 /*
832                                  * If @vcn is out of bounds, pretend @lcn is
833                                  * LCN_ENOENT.  As long as the buffer is out
834                                  * of bounds this will work fine.
835                                  */
836                                 if (err == -ENOENT) {
837                                         lcn = LCN_ENOENT;
838                                         err = 0;
839                                         goto rl_not_mapped_enoent;
840                                 }
841                         } else
842                                 err = -EIO;
843                         /* Failed to map the buffer, even after retrying. */
844                         bh->b_blocknr = -1;
845                         ntfs_error(vol->sb, "Failed to write to inode 0x%lx, "
846                                         "attribute type 0x%x, vcn 0x%llx, "
847                                         "vcn offset 0x%x, because its "
848                                         "location on disk could not be "
849                                         "determined%s (error code %i).",
850                                         ni->mft_no, ni->type,
851                                         (unsigned long long)bh_cpos,
852                                         (unsigned)bh_pos &
853                                         vol->cluster_size_mask,
854                                         is_retry ? " even after retrying" : "",
855                                         err);
856                         break;
857                 }
858 rl_not_mapped_enoent:
859                 /*
860                  * The buffer is in a hole or out of bounds.  We need to fill
861                  * the hole, unless the buffer is in a cluster which is not
862                  * touched by the write, in which case we just leave the buffer
863                  * unmapped.  This can only happen when the cluster size is
864                  * less than the page cache size.
865                  */
866                 if (unlikely(vol->cluster_size < PAGE_CACHE_SIZE)) {
867                         bh_cend = (bh_end + vol->cluster_size - 1) >>
868                                         vol->cluster_size_bits;
869                         if ((bh_cend <= cpos || bh_cpos >= cend)) {
870                                 bh->b_blocknr = -1;
871                                 /*
872                                  * If the buffer is uptodate we skip it.  If it
873                                  * is not but the page is uptodate, we can set
874                                  * the buffer uptodate.  If the page is not
875                                  * uptodate, we can clear the buffer and set it
876                                  * uptodate.  Whether this is worthwhile is
877                                  * debatable and this could be removed.
878                                  */
879                                 if (PageUptodate(page)) {
880                                         if (!buffer_uptodate(bh))
881                                                 set_buffer_uptodate(bh);
882                                 } else if (!buffer_uptodate(bh)) {
883                                         u8 *kaddr = kmap_atomic(page, KM_USER0);
884                                         memset(kaddr + bh_offset(bh), 0,
885                                                         blocksize);
886                                         kunmap_atomic(kaddr, KM_USER0);
887                                         flush_dcache_page(page);
888                                         set_buffer_uptodate(bh);
889                                 }
890                                 continue;
891                         }
892                 }
893                 /*
894                  * Out of bounds buffer is invalid if it was not really out of
895                  * bounds.
896                  */
897                 BUG_ON(lcn != LCN_HOLE);
898                 /*
899                  * We need the runlist locked for writing, so if it is locked
900                  * for reading relock it now and retry in case it changed
901                  * whilst we dropped the lock.
902                  */
903                 BUG_ON(!rl);
904                 if (!rl_write_locked) {
905                         up_read(&ni->runlist.lock);
906                         down_write(&ni->runlist.lock);
907                         rl_write_locked = TRUE;
908                         goto retry_remap;
909                 }
910                 /* Find the previous last allocated cluster. */
911                 BUG_ON(rl->lcn != LCN_HOLE);
912                 lcn = -1;
913                 rl2 = rl;
914                 while (--rl2 >= ni->runlist.rl) {
915                         if (rl2->lcn >= 0) {
916                                 lcn = rl2->lcn + rl2->length;
917                                 break;
918                         }
919                 }
920                 rl2 = ntfs_cluster_alloc(vol, bh_cpos, 1, lcn, DATA_ZONE,
921                                 FALSE);
922                 if (IS_ERR(rl2)) {
923                         err = PTR_ERR(rl2);
924                         ntfs_debug("Failed to allocate cluster, error code %i.",
925                                         err);
926                         break;
927                 }
928                 lcn = rl2->lcn;
929                 rl = ntfs_runlists_merge(ni->runlist.rl, rl2);
930                 if (IS_ERR(rl)) {
931                         err = PTR_ERR(rl);
932                         if (err != -ENOMEM)
933                                 err = -EIO;
934                         if (ntfs_cluster_free_from_rl(vol, rl2)) {
935                                 ntfs_error(vol->sb, "Failed to release "
936                                                 "allocated cluster in error "
937                                                 "code path.  Run chkdsk to "
938                                                 "recover the lost cluster.");
939                                 NVolSetErrors(vol);
940                         }
941                         ntfs_free(rl2);
942                         break;
943                 }
944                 ni->runlist.rl = rl;
945                 status.runlist_merged = 1;
946                 ntfs_debug("Allocated cluster, lcn 0x%llx.", lcn);
947                 /* Map and lock the mft record and get the attribute record. */
948                 if (!NInoAttr(ni))
949                         base_ni = ni;
950                 else
951                         base_ni = ni->ext.base_ntfs_ino;
952                 m = map_mft_record(base_ni);
953                 if (IS_ERR(m)) {
954                         err = PTR_ERR(m);
955                         break;
956                 }
957                 ctx = ntfs_attr_get_search_ctx(base_ni, m);
958                 if (unlikely(!ctx)) {
959                         err = -ENOMEM;
960                         unmap_mft_record(base_ni);
961                         break;
962                 }
963                 status.mft_attr_mapped = 1;
964                 err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
965                                 CASE_SENSITIVE, bh_cpos, NULL, 0, ctx);
966                 if (unlikely(err)) {
967                         if (err == -ENOENT)
968                                 err = -EIO;
969                         break;
970                 }
971                 m = ctx->mrec;
972                 a = ctx->attr;
973                 /*
974                  * Find the runlist element with which the attribute extent
975                  * starts.  Note, we cannot use the _attr_ version because we
976                  * have mapped the mft record.  That is ok because we know the
977                  * runlist fragment must be mapped already to have ever gotten
978                  * here, so we can just use the _rl_ version.
979                  */
980                 vcn = sle64_to_cpu(a->data.non_resident.lowest_vcn);
981                 rl2 = ntfs_rl_find_vcn_nolock(rl, vcn);
982                 BUG_ON(!rl2);
983                 BUG_ON(!rl2->length);
984                 BUG_ON(rl2->lcn < LCN_HOLE);
985                 highest_vcn = sle64_to_cpu(a->data.non_resident.highest_vcn);
986                 /*
987                  * If @highest_vcn is zero, calculate the real highest_vcn
988                  * (which can really be zero).
989                  */
990                 if (!highest_vcn)
991                         highest_vcn = (sle64_to_cpu(
992                                         a->data.non_resident.allocated_size) >>
993                                         vol->cluster_size_bits) - 1;
994                 /*
995                  * Determine the size of the mapping pairs array for the new
996                  * extent, i.e. the old extent with the hole filled.
997                  */
998                 mp_size = ntfs_get_size_for_mapping_pairs(vol, rl2, vcn,
999                                 highest_vcn);
1000                 if (unlikely(mp_size <= 0)) {
1001                         if (!(err = mp_size))
1002                                 err = -EIO;
1003                         ntfs_debug("Failed to get size for mapping pairs "
1004                                         "array, error code %i.", err);
1005                         break;
1006                 }
1007                 /*
1008                  * Resize the attribute record to fit the new mapping pairs
1009                  * array.
1010                  */
1011                 attr_rec_len = le32_to_cpu(a->length);
1012                 err = ntfs_attr_record_resize(m, a, mp_size + le16_to_cpu(
1013                                 a->data.non_resident.mapping_pairs_offset));
1014                 if (unlikely(err)) {
1015                         BUG_ON(err != -ENOSPC);
1016                         // TODO: Deal with this by using the current attribute
1017                         // and fill it with as much of the mapping pairs
1018                         // array as possible.  Then loop over each attribute
1019                         // extent rewriting the mapping pairs arrays as we go
1020                         // along and if when we reach the end we have not
1021                         // enough space, try to resize the last attribute
1022                         // extent and if even that fails, add a new attribute
1023                         // extent.
1024                         // We could also try to resize at each step in the hope
1025                         // that we will not need to rewrite every single extent.
1026                         // Note, we may need to decompress some extents to fill
1027                         // the runlist as we are walking the extents...
1028                         ntfs_error(vol->sb, "Not enough space in the mft "
1029                                         "record for the extended attribute "
1030                                         "record.  This case is not "
1031                                         "implemented yet.");
1032                         err = -EOPNOTSUPP;
1033                         break ;
1034                 }
1035                 status.mp_rebuilt = 1;
1036                 /*
1037                  * Generate the mapping pairs array directly into the attribute
1038                  * record.
1039                  */
1040                 err = ntfs_mapping_pairs_build(vol, (u8*)a + le16_to_cpu(
1041                                 a->data.non_resident.mapping_pairs_offset),
1042                                 mp_size, rl2, vcn, highest_vcn, NULL);
1043                 if (unlikely(err)) {
1044                         ntfs_error(vol->sb, "Cannot fill hole in inode 0x%lx, "
1045                                         "attribute type 0x%x, because building "
1046                                         "the mapping pairs failed with error "
1047                                         "code %i.", vi->i_ino,
1048                                         (unsigned)le32_to_cpu(ni->type), err);
1049                         err = -EIO;
1050                         break;
1051                 }
1052                 /* Update the highest_vcn but only if it was not set. */
1053                 if (unlikely(!a->data.non_resident.highest_vcn))
1054                         a->data.non_resident.highest_vcn =
1055                                         cpu_to_sle64(highest_vcn);
1056                 /*
1057                  * If the attribute is sparse/compressed, update the compressed
1058                  * size in the ntfs_inode structure and the attribute record.
1059                  */
1060                 if (likely(NInoSparse(ni) || NInoCompressed(ni))) {
1061                         /*
1062                          * If we are not in the first attribute extent, switch
1063                          * to it, but first ensure the changes will make it to
1064                          * disk later.
1065                          */
1066                         if (a->data.non_resident.lowest_vcn) {
1067                                 flush_dcache_mft_record_page(ctx->ntfs_ino);
1068                                 mark_mft_record_dirty(ctx->ntfs_ino);
1069                                 ntfs_attr_reinit_search_ctx(ctx);
1070                                 err = ntfs_attr_lookup(ni->type, ni->name,
1071                                                 ni->name_len, CASE_SENSITIVE,
1072                                                 0, NULL, 0, ctx);
1073                                 if (unlikely(err)) {
1074                                         status.attr_switched = 1;
1075                                         break;
1076                                 }
1077                                 /* @m is not used any more so do not set it. */
1078                                 a = ctx->attr;
1079                         }
1080                         write_lock_irqsave(&ni->size_lock, flags);
1081                         ni->itype.compressed.size += vol->cluster_size;
1082                         a->data.non_resident.compressed_size =
1083                                         cpu_to_sle64(ni->itype.compressed.size);
1084                         write_unlock_irqrestore(&ni->size_lock, flags);
1085                 }
1086                 /* Ensure the changes make it to disk. */
1087                 flush_dcache_mft_record_page(ctx->ntfs_ino);
1088                 mark_mft_record_dirty(ctx->ntfs_ino);
1089                 ntfs_attr_put_search_ctx(ctx);
1090                 unmap_mft_record(base_ni);
1091                 /* Successfully filled the hole. */
1092                 status.runlist_merged = 0;
1093                 status.mft_attr_mapped = 0;
1094                 status.mp_rebuilt = 0;
1095                 /* Setup the map cache and use that to deal with the buffer. */
1096                 was_hole = TRUE;
1097                 vcn = bh_cpos;
1098                 vcn_len = 1;
1099                 lcn_block = lcn << (vol->cluster_size_bits - blocksize_bits);
1100                 cdelta = 0;
1101                 /*
1102                  * If the number of remaining clusters in the @pages is smaller
1103                  * or equal to the number of cached clusters, unlock the
1104                  * runlist as the map cache will be used from now on.
1105                  */
1106                 if (likely(vcn + vcn_len >= cend)) {
1107                         up_write(&ni->runlist.lock);
1108                         rl_write_locked = FALSE;
1109                         rl = NULL;
1110                 }
1111                 goto map_buffer_cached;
1112         } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
1113         /* If there are no errors, do the next page. */
1114         if (likely(!err && ++u < nr_pages))
1115                 goto do_next_page;
1116         /* If there are no errors, release the runlist lock if we took it. */
1117         if (likely(!err)) {
1118                 if (unlikely(rl_write_locked)) {
1119                         up_write(&ni->runlist.lock);
1120                         rl_write_locked = FALSE;
1121                 } else if (unlikely(rl))
1122                         up_read(&ni->runlist.lock);
1123                 rl = NULL;
1124         }
1125         /* If we issued read requests, let them complete. */
1126         read_lock_irqsave(&ni->size_lock, flags);
1127         initialized_size = ni->initialized_size;
1128         read_unlock_irqrestore(&ni->size_lock, flags);
1129         while (wait_bh > wait) {
1130                 bh = *--wait_bh;
1131                 wait_on_buffer(bh);
1132                 if (likely(buffer_uptodate(bh))) {
1133                         page = bh->b_page;
1134                         bh_pos = ((s64)page->index << PAGE_CACHE_SHIFT) +
1135                                         bh_offset(bh);
1136                         /*
1137                          * If the buffer overflows the initialized size, need
1138                          * to zero the overflowing region.
1139                          */
1140                         if (unlikely(bh_pos + blocksize > initialized_size)) {
1141                                 u8 *kaddr;
1142                                 int ofs = 0;
1143
1144                                 if (likely(bh_pos < initialized_size))
1145                                         ofs = initialized_size - bh_pos;
1146                                 kaddr = kmap_atomic(page, KM_USER0);
1147                                 memset(kaddr + bh_offset(bh) + ofs, 0,
1148                                                 blocksize - ofs);
1149                                 kunmap_atomic(kaddr, KM_USER0);
1150                                 flush_dcache_page(page);
1151                         }
1152                 } else /* if (unlikely(!buffer_uptodate(bh))) */
1153                         err = -EIO;
1154         }
1155         if (likely(!err)) {
1156                 /* Clear buffer_new on all buffers. */
1157                 u = 0;
1158                 do {
1159                         bh = head = page_buffers(pages[u]);
1160                         do {
1161                                 if (buffer_new(bh))
1162                                         clear_buffer_new(bh);
1163                         } while ((bh = bh->b_this_page) != head);
1164                 } while (++u < nr_pages);
1165                 ntfs_debug("Done.");
1166                 return err;
1167         }
1168         if (status.attr_switched) {
1169                 /* Get back to the attribute extent we modified. */
1170                 ntfs_attr_reinit_search_ctx(ctx);
1171                 if (ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1172                                 CASE_SENSITIVE, bh_cpos, NULL, 0, ctx)) {
1173                         ntfs_error(vol->sb, "Failed to find required "
1174                                         "attribute extent of attribute in "
1175                                         "error code path.  Run chkdsk to "
1176                                         "recover.");
1177                         write_lock_irqsave(&ni->size_lock, flags);
1178                         ni->itype.compressed.size += vol->cluster_size;
1179                         write_unlock_irqrestore(&ni->size_lock, flags);
1180                         flush_dcache_mft_record_page(ctx->ntfs_ino);
1181                         mark_mft_record_dirty(ctx->ntfs_ino);
1182                         /*
1183                          * The only thing that is now wrong is the compressed
1184                          * size of the base attribute extent which chkdsk
1185                          * should be able to fix.
1186                          */
1187                         NVolSetErrors(vol);
1188                 } else {
1189                         m = ctx->mrec;
1190                         a = ctx->attr;
1191                         status.attr_switched = 0;
1192                 }
1193         }
1194         /*
1195          * If the runlist has been modified, need to restore it by punching a
1196          * hole into it and we then need to deallocate the on-disk cluster as
1197          * well.  Note, we only modify the runlist if we are able to generate a
1198          * new mapping pairs array, i.e. only when the mapped attribute extent
1199          * is not switched.
1200          */
1201         if (status.runlist_merged && !status.attr_switched) {
1202                 BUG_ON(!rl_write_locked);
1203                 /* Make the file cluster we allocated sparse in the runlist. */
1204                 if (ntfs_rl_punch_nolock(vol, &ni->runlist, bh_cpos, 1)) {
1205                         ntfs_error(vol->sb, "Failed to punch hole into "
1206                                         "attribute runlist in error code "
1207                                         "path.  Run chkdsk to recover the "
1208                                         "lost cluster.");
1209                         make_bad_inode(vi);
1210                         make_bad_inode(VFS_I(base_ni));
1211                         NVolSetErrors(vol);
1212                 } else /* if (success) */ {
1213                         status.runlist_merged = 0;
1214                         /*
1215                          * Deallocate the on-disk cluster we allocated but only
1216                          * if we succeeded in punching its vcn out of the
1217                          * runlist.
1218                          */
1219                         down_write(&vol->lcnbmp_lock);
1220                         if (ntfs_bitmap_clear_bit(vol->lcnbmp_ino, lcn)) {
1221                                 ntfs_error(vol->sb, "Failed to release "
1222                                                 "allocated cluster in error "
1223                                                 "code path.  Run chkdsk to "
1224                                                 "recover the lost cluster.");
1225                                 NVolSetErrors(vol);
1226                         }
1227                         up_write(&vol->lcnbmp_lock);
1228                 }
1229         }
1230         /*
1231          * Resize the attribute record to its old size and rebuild the mapping
1232          * pairs array.  Note, we only can do this if the runlist has been
1233          * restored to its old state which also implies that the mapped
1234          * attribute extent is not switched.
1235          */
1236         if (status.mp_rebuilt && !status.runlist_merged) {
1237                 if (ntfs_attr_record_resize(m, a, attr_rec_len)) {
1238                         ntfs_error(vol->sb, "Failed to restore attribute "
1239                                         "record in error code path.  Run "
1240                                         "chkdsk to recover.");
1241                         make_bad_inode(vi);
1242                         make_bad_inode(VFS_I(base_ni));
1243                         NVolSetErrors(vol);
1244                 } else /* if (success) */ {
1245                         if (ntfs_mapping_pairs_build(vol, (u8*)a +
1246                                         le16_to_cpu(a->data.non_resident.
1247                                         mapping_pairs_offset), attr_rec_len -
1248                                         le16_to_cpu(a->data.non_resident.
1249                                         mapping_pairs_offset), ni->runlist.rl,
1250                                         vcn, highest_vcn, NULL)) {
1251                                 ntfs_error(vol->sb, "Failed to restore "
1252                                                 "mapping pairs array in error "
1253                                                 "code path.  Run chkdsk to "
1254                                                 "recover.");
1255                                 make_bad_inode(vi);
1256                                 make_bad_inode(VFS_I(base_ni));
1257                                 NVolSetErrors(vol);
1258                         }
1259                         flush_dcache_mft_record_page(ctx->ntfs_ino);
1260                         mark_mft_record_dirty(ctx->ntfs_ino);
1261                 }
1262         }
1263         /* Release the mft record and the attribute. */
1264         if (status.mft_attr_mapped) {
1265                 ntfs_attr_put_search_ctx(ctx);
1266                 unmap_mft_record(base_ni);
1267         }
1268         /* Release the runlist lock. */
1269         if (rl_write_locked)
1270                 up_write(&ni->runlist.lock);
1271         else if (rl)
1272                 up_read(&ni->runlist.lock);
1273         /*
1274          * Zero out any newly allocated blocks to avoid exposing stale data.
1275          * If BH_New is set, we know that the block was newly allocated above
1276          * and that it has not been fully zeroed and marked dirty yet.
1277          */
1278         nr_pages = u;
1279         u = 0;
1280         end = bh_cpos << vol->cluster_size_bits;
1281         do {
1282                 page = pages[u];
1283                 bh = head = page_buffers(page);
1284                 do {
1285                         if (u == nr_pages &&
1286                                         ((s64)page->index << PAGE_CACHE_SHIFT) +
1287                                         bh_offset(bh) >= end)
1288                                 break;
1289                         if (!buffer_new(bh))
1290                                 continue;
1291                         clear_buffer_new(bh);
1292                         if (!buffer_uptodate(bh)) {
1293                                 if (PageUptodate(page))
1294                                         set_buffer_uptodate(bh);
1295                                 else {
1296                                         u8 *kaddr = kmap_atomic(page, KM_USER0);
1297                                         memset(kaddr + bh_offset(bh), 0,
1298                                                         blocksize);
1299                                         kunmap_atomic(kaddr, KM_USER0);
1300                                         flush_dcache_page(page);
1301                                         set_buffer_uptodate(bh);
1302                                 }
1303                         }
1304                         mark_buffer_dirty(bh);
1305                 } while ((bh = bh->b_this_page) != head);
1306         } while (++u <= nr_pages);
1307         ntfs_error(vol->sb, "Failed.  Returning error code %i.", err);
1308         return err;
1309 }
1310
1311 /*
1312  * Copy as much as we can into the pages and return the number of bytes which
1313  * were sucessfully copied.  If a fault is encountered then clear the pages
1314  * out to (ofs + bytes) and return the number of bytes which were copied.
1315  */
1316 static inline size_t ntfs_copy_from_user(struct page **pages,
1317                 unsigned nr_pages, unsigned ofs, const char __user *buf,
1318                 size_t bytes)
1319 {
1320         struct page **last_page = pages + nr_pages;
1321         char *kaddr;
1322         size_t total = 0;
1323         unsigned len;
1324         int left;
1325
1326         do {
1327                 len = PAGE_CACHE_SIZE - ofs;
1328                 if (len > bytes)
1329                         len = bytes;
1330                 kaddr = kmap_atomic(*pages, KM_USER0);
1331                 left = __copy_from_user_inatomic(kaddr + ofs, buf, len);
1332                 kunmap_atomic(kaddr, KM_USER0);
1333                 if (unlikely(left)) {
1334                         /* Do it the slow way. */
1335                         kaddr = kmap(*pages);
1336                         left = __copy_from_user(kaddr + ofs, buf, len);
1337                         kunmap(*pages);
1338                         if (unlikely(left))
1339                                 goto err_out;
1340                 }
1341                 total += len;
1342                 bytes -= len;
1343                 if (!bytes)
1344                         break;
1345                 buf += len;
1346                 ofs = 0;
1347         } while (++pages < last_page);
1348 out:
1349         return total;
1350 err_out:
1351         total += len - left;
1352         /* Zero the rest of the target like __copy_from_user(). */
1353         while (++pages < last_page) {
1354                 bytes -= len;
1355                 if (!bytes)
1356                         break;
1357                 len = PAGE_CACHE_SIZE;
1358                 if (len > bytes)
1359                         len = bytes;
1360                 kaddr = kmap_atomic(*pages, KM_USER0);
1361                 memset(kaddr, 0, len);
1362                 kunmap_atomic(kaddr, KM_USER0);
1363         }
1364         goto out;
1365 }
1366
1367 static size_t __ntfs_copy_from_user_iovec(char *vaddr,
1368                 const struct iovec *iov, size_t iov_ofs, size_t bytes)
1369 {
1370         size_t total = 0;
1371
1372         while (1) {
1373                 const char __user *buf = iov->iov_base + iov_ofs;
1374                 unsigned len;
1375                 size_t left;
1376
1377                 len = iov->iov_len - iov_ofs;
1378                 if (len > bytes)
1379                         len = bytes;
1380                 left = __copy_from_user_inatomic(vaddr, buf, len);
1381                 total += len;
1382                 bytes -= len;
1383                 vaddr += len;
1384                 if (unlikely(left)) {
1385                         /*
1386                          * Zero the rest of the target like __copy_from_user().
1387                          */
1388                         memset(vaddr, 0, bytes);
1389                         total -= left;
1390                         break;
1391                 }
1392                 if (!bytes)
1393                         break;
1394                 iov++;
1395                 iov_ofs = 0;
1396         }
1397         return total;
1398 }
1399
1400 static inline void ntfs_set_next_iovec(const struct iovec **iovp,
1401                 size_t *iov_ofsp, size_t bytes)
1402 {
1403         const struct iovec *iov = *iovp;
1404         size_t iov_ofs = *iov_ofsp;
1405
1406         while (bytes) {
1407                 unsigned len;
1408
1409                 len = iov->iov_len - iov_ofs;
1410                 if (len > bytes)
1411                         len = bytes;
1412                 bytes -= len;
1413                 iov_ofs += len;
1414                 if (iov->iov_len == iov_ofs) {
1415                         iov++;
1416                         iov_ofs = 0;
1417                 }
1418         }
1419         *iovp = iov;
1420         *iov_ofsp = iov_ofs;
1421 }
1422
1423 /*
1424  * This has the same side-effects and return value as ntfs_copy_from_user().
1425  * The difference is that on a fault we need to memset the remainder of the
1426  * pages (out to offset + bytes), to emulate ntfs_copy_from_user()'s
1427  * single-segment behaviour.
1428  *
1429  * We call the same helper (__ntfs_copy_from_user_iovec()) both when atomic and
1430  * when not atomic.  This is ok because __ntfs_copy_from_user_iovec() calls
1431  * __copy_from_user_inatomic() and it is ok to call this when non-atomic.  In
1432  * fact, the only difference between __copy_from_user_inatomic() and
1433  * __copy_from_user() is that the latter calls might_sleep().  And on many
1434  * architectures __copy_from_user_inatomic() is just defined to
1435  * __copy_from_user() so it makes no difference at all on those architectures.
1436  */
1437 static inline size_t ntfs_copy_from_user_iovec(struct page **pages,
1438                 unsigned nr_pages, unsigned ofs, const struct iovec **iov,
1439                 size_t *iov_ofs, size_t bytes)
1440 {
1441         struct page **last_page = pages + nr_pages;
1442         char *kaddr;
1443         size_t copied, len, total = 0;
1444
1445         do {
1446                 len = PAGE_CACHE_SIZE - ofs;
1447                 if (len > bytes)
1448                         len = bytes;
1449                 kaddr = kmap_atomic(*pages, KM_USER0);
1450                 copied = __ntfs_copy_from_user_iovec(kaddr + ofs,
1451                                 *iov, *iov_ofs, len);
1452                 kunmap_atomic(kaddr, KM_USER0);
1453                 if (unlikely(copied != len)) {
1454                         /* Do it the slow way. */
1455                         kaddr = kmap(*pages);
1456                         copied = __ntfs_copy_from_user_iovec(kaddr + ofs,
1457                                         *iov, *iov_ofs, len);
1458                         kunmap(*pages);
1459                         if (unlikely(copied != len))
1460                                 goto err_out;
1461                 }
1462                 total += len;
1463                 bytes -= len;
1464                 if (!bytes)
1465                         break;
1466                 ntfs_set_next_iovec(iov, iov_ofs, len);
1467                 ofs = 0;
1468         } while (++pages < last_page);
1469 out:
1470         return total;
1471 err_out:
1472         total += copied;
1473         /* Zero the rest of the target like __copy_from_user(). */
1474         while (++pages < last_page) {
1475                 bytes -= len;
1476                 if (!bytes)
1477                         break;
1478                 len = PAGE_CACHE_SIZE;
1479                 if (len > bytes)
1480                         len = bytes;
1481                 kaddr = kmap_atomic(*pages, KM_USER0);
1482                 memset(kaddr, 0, len);
1483                 kunmap_atomic(kaddr, KM_USER0);
1484         }
1485         goto out;
1486 }
1487
1488 static inline void ntfs_flush_dcache_pages(struct page **pages,
1489                 unsigned nr_pages)
1490 {
1491         BUG_ON(!nr_pages);
1492         do {
1493                 /*
1494                  * Warning: Do not do the decrement at the same time as the
1495                  * call because flush_dcache_page() is a NULL macro on i386
1496                  * and hence the decrement never happens.
1497                  */
1498                 flush_dcache_page(pages[nr_pages]);
1499         } while (--nr_pages > 0);
1500 }
1501
1502 /**
1503  * ntfs_commit_pages_after_non_resident_write - commit the received data
1504  * @pages:      array of destination pages
1505  * @nr_pages:   number of pages in @pages
1506  * @pos:        byte position in file at which the write begins
1507  * @bytes:      number of bytes to be written
1508  *
1509  * See description of ntfs_commit_pages_after_write(), below.
1510  */
1511 static inline int ntfs_commit_pages_after_non_resident_write(
1512                 struct page **pages, const unsigned nr_pages,
1513                 s64 pos, size_t bytes)
1514 {
1515         s64 end, initialized_size;
1516         struct inode *vi;
1517         ntfs_inode *ni, *base_ni;
1518         struct buffer_head *bh, *head;
1519         ntfs_attr_search_ctx *ctx;
1520         MFT_RECORD *m;
1521         ATTR_RECORD *a;
1522         unsigned long flags;
1523         unsigned blocksize, u;
1524         int err;
1525
1526         vi = pages[0]->mapping->host;
1527         ni = NTFS_I(vi);
1528         blocksize = 1 << vi->i_blkbits;
1529         end = pos + bytes;
1530         u = 0;
1531         do {
1532                 s64 bh_pos;
1533                 struct page *page;
1534                 BOOL partial;
1535
1536                 page = pages[u];
1537                 bh_pos = (s64)page->index << PAGE_CACHE_SHIFT;
1538                 bh = head = page_buffers(page);
1539                 partial = FALSE;
1540                 do {
1541                         s64 bh_end;
1542
1543                         bh_end = bh_pos + blocksize;
1544                         if (bh_end <= pos || bh_pos >= end) {
1545                                 if (!buffer_uptodate(bh))
1546                                         partial = TRUE;
1547                         } else {
1548                                 set_buffer_uptodate(bh);
1549                                 mark_buffer_dirty(bh);
1550                         }
1551                 } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
1552                 /*
1553                  * If all buffers are now uptodate but the page is not, set the
1554                  * page uptodate.
1555                  */
1556                 if (!partial && !PageUptodate(page))
1557                         SetPageUptodate(page);
1558         } while (++u < nr_pages);
1559         /*
1560          * Finally, if we do not need to update initialized_size or i_size we
1561          * are finished.
1562          */
1563         read_lock_irqsave(&ni->size_lock, flags);
1564         initialized_size = ni->initialized_size;
1565         read_unlock_irqrestore(&ni->size_lock, flags);
1566         if (end <= initialized_size) {
1567                 ntfs_debug("Done.");
1568                 return 0;
1569         }
1570         /*
1571          * Update initialized_size/i_size as appropriate, both in the inode and
1572          * the mft record.
1573          */
1574         if (!NInoAttr(ni))
1575                 base_ni = ni;
1576         else
1577                 base_ni = ni->ext.base_ntfs_ino;
1578         /* Map, pin, and lock the mft record. */
1579         m = map_mft_record(base_ni);
1580         if (IS_ERR(m)) {
1581                 err = PTR_ERR(m);
1582                 m = NULL;
1583                 ctx = NULL;
1584                 goto err_out;
1585         }
1586         BUG_ON(!NInoNonResident(ni));
1587         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1588         if (unlikely(!ctx)) {
1589                 err = -ENOMEM;
1590                 goto err_out;
1591         }
1592         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1593                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1594         if (unlikely(err)) {
1595                 if (err == -ENOENT)
1596                         err = -EIO;
1597                 goto err_out;
1598         }
1599         a = ctx->attr;
1600         BUG_ON(!a->non_resident);
1601         write_lock_irqsave(&ni->size_lock, flags);
1602         BUG_ON(end > ni->allocated_size);
1603         ni->initialized_size = end;
1604         a->data.non_resident.initialized_size = cpu_to_sle64(end);
1605         if (end > i_size_read(vi)) {
1606                 i_size_write(vi, end);
1607                 a->data.non_resident.data_size =
1608                                 a->data.non_resident.initialized_size;
1609         }
1610         write_unlock_irqrestore(&ni->size_lock, flags);
1611         /* Mark the mft record dirty, so it gets written back. */
1612         flush_dcache_mft_record_page(ctx->ntfs_ino);
1613         mark_mft_record_dirty(ctx->ntfs_ino);
1614         ntfs_attr_put_search_ctx(ctx);
1615         unmap_mft_record(base_ni);
1616         ntfs_debug("Done.");
1617         return 0;
1618 err_out:
1619         if (ctx)
1620                 ntfs_attr_put_search_ctx(ctx);
1621         if (m)
1622                 unmap_mft_record(base_ni);
1623         ntfs_error(vi->i_sb, "Failed to update initialized_size/i_size (error "
1624                         "code %i).", err);
1625         if (err != -ENOMEM) {
1626                 NVolSetErrors(ni->vol);
1627                 make_bad_inode(VFS_I(base_ni));
1628                 make_bad_inode(vi);
1629         }
1630         return err;
1631 }
1632
1633 /**
1634  * ntfs_commit_pages_after_write - commit the received data
1635  * @pages:      array of destination pages
1636  * @nr_pages:   number of pages in @pages
1637  * @pos:        byte position in file at which the write begins
1638  * @bytes:      number of bytes to be written
1639  *
1640  * This is called from ntfs_file_buffered_write() with i_sem held on the inode
1641  * (@pages[0]->mapping->host).  There are @nr_pages pages in @pages which are
1642  * locked but not kmap()ped.  The source data has already been copied into the
1643  * @page.  ntfs_prepare_pages_for_non_resident_write() has been called before
1644  * the data was copied (for non-resident attributes only) and it returned
1645  * success.
1646  *
1647  * Need to set uptodate and mark dirty all buffers within the boundary of the
1648  * write.  If all buffers in a page are uptodate we set the page uptodate, too.
1649  *
1650  * Setting the buffers dirty ensures that they get written out later when
1651  * ntfs_writepage() is invoked by the VM.
1652  *
1653  * Finally, we need to update i_size and initialized_size as appropriate both
1654  * in the inode and the mft record.
1655  *
1656  * This is modelled after fs/buffer.c::generic_commit_write(), which marks
1657  * buffers uptodate and dirty, sets the page uptodate if all buffers in the
1658  * page are uptodate, and updates i_size if the end of io is beyond i_size.  In
1659  * that case, it also marks the inode dirty.
1660  *
1661  * If things have gone as outlined in
1662  * ntfs_prepare_pages_for_non_resident_write(), we do not need to do any page
1663  * content modifications here for non-resident attributes.  For resident
1664  * attributes we need to do the uptodate bringing here which we combine with
1665  * the copying into the mft record which means we save one atomic kmap.
1666  *
1667  * Return 0 on success or -errno on error.
1668  */
1669 static int ntfs_commit_pages_after_write(struct page **pages,
1670                 const unsigned nr_pages, s64 pos, size_t bytes)
1671 {
1672         s64 end, initialized_size;
1673         loff_t i_size;
1674         struct inode *vi;
1675         ntfs_inode *ni, *base_ni;
1676         struct page *page;
1677         ntfs_attr_search_ctx *ctx;
1678         MFT_RECORD *m;
1679         ATTR_RECORD *a;
1680         char *kattr, *kaddr;
1681         unsigned long flags;
1682         u32 attr_len;
1683         int err;
1684
1685         BUG_ON(!nr_pages);
1686         BUG_ON(!pages);
1687         page = pages[0];
1688         BUG_ON(!page);
1689         vi = page->mapping->host;
1690         ni = NTFS_I(vi);
1691         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, start page "
1692                         "index 0x%lx, nr_pages 0x%x, pos 0x%llx, bytes 0x%zx.",
1693                         vi->i_ino, ni->type, page->index, nr_pages,
1694                         (long long)pos, bytes);
1695         if (NInoNonResident(ni))
1696                 return ntfs_commit_pages_after_non_resident_write(pages,
1697                                 nr_pages, pos, bytes);
1698         BUG_ON(nr_pages > 1);
1699         /*
1700          * Attribute is resident, implying it is not compressed, encrypted, or
1701          * sparse.
1702          */
1703         if (!NInoAttr(ni))
1704                 base_ni = ni;
1705         else
1706                 base_ni = ni->ext.base_ntfs_ino;
1707         BUG_ON(NInoNonResident(ni));
1708         /* Map, pin, and lock the mft record. */
1709         m = map_mft_record(base_ni);
1710         if (IS_ERR(m)) {
1711                 err = PTR_ERR(m);
1712                 m = NULL;
1713                 ctx = NULL;
1714                 goto err_out;
1715         }
1716         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1717         if (unlikely(!ctx)) {
1718                 err = -ENOMEM;
1719                 goto err_out;
1720         }
1721         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1722                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1723         if (unlikely(err)) {
1724                 if (err == -ENOENT)
1725                         err = -EIO;
1726                 goto err_out;
1727         }
1728         a = ctx->attr;
1729         BUG_ON(a->non_resident);
1730         /* The total length of the attribute value. */
1731         attr_len = le32_to_cpu(a->data.resident.value_length);
1732         i_size = i_size_read(vi);
1733         BUG_ON(attr_len != i_size);
1734         BUG_ON(pos > attr_len);
1735         end = pos + bytes;
1736         BUG_ON(end > le32_to_cpu(a->length) -
1737                         le16_to_cpu(a->data.resident.value_offset));
1738         kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
1739         kaddr = kmap_atomic(page, KM_USER0);
1740         /* Copy the received data from the page to the mft record. */
1741         memcpy(kattr + pos, kaddr + pos, bytes);
1742         /* Update the attribute length if necessary. */
1743         if (end > attr_len) {
1744                 attr_len = end;
1745                 a->data.resident.value_length = cpu_to_le32(attr_len);
1746         }
1747         /*
1748          * If the page is not uptodate, bring the out of bounds area(s)
1749          * uptodate by copying data from the mft record to the page.
1750          */
1751         if (!PageUptodate(page)) {
1752                 if (pos > 0)
1753                         memcpy(kaddr, kattr, pos);
1754                 if (end < attr_len)
1755                         memcpy(kaddr + end, kattr + end, attr_len - end);
1756                 /* Zero the region outside the end of the attribute value. */
1757                 memset(kaddr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
1758                 flush_dcache_page(page);
1759                 SetPageUptodate(page);
1760         }
1761         kunmap_atomic(kaddr, KM_USER0);
1762         /* Update initialized_size/i_size if necessary. */
1763         read_lock_irqsave(&ni->size_lock, flags);
1764         initialized_size = ni->initialized_size;
1765         BUG_ON(end > ni->allocated_size);
1766         read_unlock_irqrestore(&ni->size_lock, flags);
1767         BUG_ON(initialized_size != i_size);
1768         if (end > initialized_size) {
1769                 unsigned long flags;
1770
1771                 write_lock_irqsave(&ni->size_lock, flags);
1772                 ni->initialized_size = end;
1773                 i_size_write(vi, end);
1774                 write_unlock_irqrestore(&ni->size_lock, flags);
1775         }
1776         /* Mark the mft record dirty, so it gets written back. */
1777         flush_dcache_mft_record_page(ctx->ntfs_ino);
1778         mark_mft_record_dirty(ctx->ntfs_ino);
1779         ntfs_attr_put_search_ctx(ctx);
1780         unmap_mft_record(base_ni);
1781         ntfs_debug("Done.");
1782         return 0;
1783 err_out:
1784         if (err == -ENOMEM) {
1785                 ntfs_warning(vi->i_sb, "Error allocating memory required to "
1786                                 "commit the write.");
1787                 if (PageUptodate(page)) {
1788                         ntfs_warning(vi->i_sb, "Page is uptodate, setting "
1789                                         "dirty so the write will be retried "
1790                                         "later on by the VM.");
1791                         /*
1792                          * Put the page on mapping->dirty_pages, but leave its
1793                          * buffers' dirty state as-is.
1794                          */
1795                         __set_page_dirty_nobuffers(page);
1796                         err = 0;
1797                 } else
1798                         ntfs_error(vi->i_sb, "Page is not uptodate.  Written "
1799                                         "data has been lost.");
1800         } else {
1801                 ntfs_error(vi->i_sb, "Resident attribute commit write failed "
1802                                 "with error %i.", err);
1803                 NVolSetErrors(ni->vol);
1804                 make_bad_inode(VFS_I(base_ni));
1805                 make_bad_inode(vi);
1806         }
1807         if (ctx)
1808                 ntfs_attr_put_search_ctx(ctx);
1809         if (m)
1810                 unmap_mft_record(base_ni);
1811         return err;
1812 }
1813
1814 /**
1815  * ntfs_file_buffered_write -
1816  *
1817  * Locking: The vfs is holding ->i_sem on the inode.
1818  */
1819 static ssize_t ntfs_file_buffered_write(struct kiocb *iocb,
1820                 const struct iovec *iov, unsigned long nr_segs,
1821                 loff_t pos, loff_t *ppos, size_t count)
1822 {
1823         struct file *file = iocb->ki_filp;
1824         struct address_space *mapping = file->f_mapping;
1825         struct inode *vi = mapping->host;
1826         ntfs_inode *ni = NTFS_I(vi);
1827         ntfs_volume *vol = ni->vol;
1828         struct page *pages[NTFS_MAX_PAGES_PER_CLUSTER];
1829         struct page *cached_page = NULL;
1830         char __user *buf = NULL;
1831         s64 end, ll;
1832         VCN last_vcn;
1833         LCN lcn;
1834         unsigned long flags;
1835         size_t bytes, iov_ofs = 0;      /* Offset in the current iovec. */
1836         ssize_t status, written;
1837         unsigned nr_pages;
1838         int err;
1839         struct pagevec lru_pvec;
1840
1841         ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, "
1842                         "pos 0x%llx, count 0x%lx.",
1843                         vi->i_ino, (unsigned)le32_to_cpu(ni->type),
1844                         (unsigned long long)pos, (unsigned long)count);
1845         if (unlikely(!count))
1846                 return 0;
1847         BUG_ON(NInoMstProtected(ni));
1848         /*
1849          * If the attribute is not an index root and it is encrypted or
1850          * compressed, we cannot write to it yet.  Note we need to check for
1851          * AT_INDEX_ALLOCATION since this is the type of both directory and
1852          * index inodes.
1853          */
1854         if (ni->type != AT_INDEX_ALLOCATION) {
1855                 /* If file is encrypted, deny access, just like NT4. */
1856                 if (NInoEncrypted(ni)) {
1857                         /*
1858                          * Reminder for later: Encrypted files are _always_
1859                          * non-resident so that the content can always be
1860                          * encrypted.
1861                          */
1862                         ntfs_debug("Denying write access to encrypted file.");
1863                         return -EACCES;
1864                 }
1865                 if (NInoCompressed(ni)) {
1866                         /* Only unnamed $DATA attribute can be compressed. */
1867                         BUG_ON(ni->type != AT_DATA);
1868                         BUG_ON(ni->name_len);
1869                         /*
1870                          * Reminder for later: If resident, the data is not
1871                          * actually compressed.  Only on the switch to non-
1872                          * resident does compression kick in.  This is in
1873                          * contrast to encrypted files (see above).
1874                          */
1875                         ntfs_error(vi->i_sb, "Writing to compressed files is "
1876                                         "not implemented yet.  Sorry.");
1877                         return -EOPNOTSUPP;
1878                 }
1879         }
1880         /*
1881          * If a previous ntfs_truncate() failed, repeat it and abort if it
1882          * fails again.
1883          */
1884         if (unlikely(NInoTruncateFailed(ni))) {
1885                 down_write(&vi->i_alloc_sem);
1886                 err = ntfs_truncate(vi);
1887                 up_write(&vi->i_alloc_sem);
1888                 if (err || NInoTruncateFailed(ni)) {
1889                         if (!err)
1890                                 err = -EIO;
1891                         ntfs_error(vol->sb, "Cannot perform write to inode "
1892                                         "0x%lx, attribute type 0x%x, because "
1893                                         "ntfs_truncate() failed (error code "
1894                                         "%i).", vi->i_ino,
1895                                         (unsigned)le32_to_cpu(ni->type), err);
1896                         return err;
1897                 }
1898         }
1899         /* The first byte after the write. */
1900         end = pos + count;
1901         /*
1902          * If the write goes beyond the allocated size, extend the allocation
1903          * to cover the whole of the write, rounded up to the nearest cluster.
1904          */
1905         read_lock_irqsave(&ni->size_lock, flags);
1906         ll = ni->allocated_size;
1907         read_unlock_irqrestore(&ni->size_lock, flags);
1908         if (end > ll) {
1909                 /* Extend the allocation without changing the data size. */
1910                 ll = ntfs_attr_extend_allocation(ni, end, -1, pos);
1911                 if (likely(ll >= 0)) {
1912                         BUG_ON(pos >= ll);
1913                         /* If the extension was partial truncate the write. */
1914                         if (end > ll) {
1915                                 ntfs_debug("Truncating write to inode 0x%lx, "
1916                                                 "attribute type 0x%x, because "
1917                                                 "the allocation was only "
1918                                                 "partially extended.",
1919                                                 vi->i_ino, (unsigned)
1920                                                 le32_to_cpu(ni->type));
1921                                 end = ll;
1922                                 count = ll - pos;
1923                         }
1924                 } else {
1925                         err = ll;
1926                         read_lock_irqsave(&ni->size_lock, flags);
1927                         ll = ni->allocated_size;
1928                         read_unlock_irqrestore(&ni->size_lock, flags);
1929                         /* Perform a partial write if possible or fail. */
1930                         if (pos < ll) {
1931                                 ntfs_debug("Truncating write to inode 0x%lx, "
1932                                                 "attribute type 0x%x, because "
1933                                                 "extending the allocation "
1934                                                 "failed (error code %i).",
1935                                                 vi->i_ino, (unsigned)
1936                                                 le32_to_cpu(ni->type), err);
1937                                 end = ll;
1938                                 count = ll - pos;
1939                         } else {
1940                                 ntfs_error(vol->sb, "Cannot perform write to "
1941                                                 "inode 0x%lx, attribute type "
1942                                                 "0x%x, because extending the "
1943                                                 "allocation failed (error "
1944                                                 "code %i).", vi->i_ino,
1945                                                 (unsigned)
1946                                                 le32_to_cpu(ni->type), err);
1947                                 return err;
1948                         }
1949                 }
1950         }
1951         pagevec_init(&lru_pvec, 0);
1952         written = 0;
1953         /*
1954          * If the write starts beyond the initialized size, extend it up to the
1955          * beginning of the write and initialize all non-sparse space between
1956          * the old initialized size and the new one.  This automatically also
1957          * increments the vfs inode->i_size to keep it above or equal to the
1958          * initialized_size.
1959          */
1960         read_lock_irqsave(&ni->size_lock, flags);
1961         ll = ni->initialized_size;
1962         read_unlock_irqrestore(&ni->size_lock, flags);
1963         if (pos > ll) {
1964                 err = ntfs_attr_extend_initialized(ni, pos, &cached_page,
1965                                 &lru_pvec);
1966                 if (err < 0) {
1967                         ntfs_error(vol->sb, "Cannot perform write to inode "
1968                                         "0x%lx, attribute type 0x%x, because "
1969                                         "extending the initialized size "
1970                                         "failed (error code %i).", vi->i_ino,
1971                                         (unsigned)le32_to_cpu(ni->type), err);
1972                         status = err;
1973                         goto err_out;
1974                 }
1975         }
1976         /*
1977          * Determine the number of pages per cluster for non-resident
1978          * attributes.
1979          */
1980         nr_pages = 1;
1981         if (vol->cluster_size > PAGE_CACHE_SIZE && NInoNonResident(ni))
1982                 nr_pages = vol->cluster_size >> PAGE_CACHE_SHIFT;
1983         /* Finally, perform the actual write. */
1984         last_vcn = -1;
1985         if (likely(nr_segs == 1))
1986                 buf = iov->iov_base;
1987         do {
1988                 VCN vcn;
1989                 pgoff_t idx, start_idx;
1990                 unsigned ofs, do_pages, u;
1991                 size_t copied;
1992
1993                 start_idx = idx = pos >> PAGE_CACHE_SHIFT;
1994                 ofs = pos & ~PAGE_CACHE_MASK;
1995                 bytes = PAGE_CACHE_SIZE - ofs;
1996                 do_pages = 1;
1997                 if (nr_pages > 1) {
1998                         vcn = pos >> vol->cluster_size_bits;
1999                         if (vcn != last_vcn) {
2000                                 last_vcn = vcn;
2001                                 /*
2002                                  * Get the lcn of the vcn the write is in.  If
2003                                  * it is a hole, need to lock down all pages in
2004                                  * the cluster.
2005                                  */
2006                                 down_read(&ni->runlist.lock);
2007                                 lcn = ntfs_attr_vcn_to_lcn_nolock(ni, pos >>
2008                                                 vol->cluster_size_bits, FALSE);
2009                                 up_read(&ni->runlist.lock);
2010                                 if (unlikely(lcn < LCN_HOLE)) {
2011                                         status = -EIO;
2012                                         if (lcn == LCN_ENOMEM)
2013                                                 status = -ENOMEM;
2014                                         else
2015                                                 ntfs_error(vol->sb, "Cannot "
2016                                                         "perform write to "
2017                                                         "inode 0x%lx, "
2018                                                         "attribute type 0x%x, "
2019                                                         "because the attribute "
2020                                                         "is corrupt.",
2021                                                         vi->i_ino, (unsigned)
2022                                                         le32_to_cpu(ni->type));
2023                                         break;
2024                                 }
2025                                 if (lcn == LCN_HOLE) {
2026                                         start_idx = (pos & ~(s64)
2027                                                         vol->cluster_size_mask)
2028                                                         >> PAGE_CACHE_SHIFT;
2029                                         bytes = vol->cluster_size - (pos &
2030                                                         vol->cluster_size_mask);
2031                                         do_pages = nr_pages;
2032                                 }
2033                         }
2034                 }
2035                 if (bytes > count)
2036                         bytes = count;
2037                 /*
2038                  * Bring in the user page(s) that we will copy from _first_.
2039                  * Otherwise there is a nasty deadlock on copying from the same
2040                  * page(s) as we are writing to, without it/them being marked
2041                  * up-to-date.  Note, at present there is nothing to stop the
2042                  * pages being swapped out between us bringing them into memory
2043                  * and doing the actual copying.
2044                  */
2045                 if (likely(nr_segs == 1))
2046                         ntfs_fault_in_pages_readable(buf, bytes);
2047                 else
2048                         ntfs_fault_in_pages_readable_iovec(iov, iov_ofs, bytes);
2049                 /* Get and lock @do_pages starting at index @start_idx. */
2050                 status = __ntfs_grab_cache_pages(mapping, start_idx, do_pages,
2051                                 pages, &cached_page, &lru_pvec);
2052                 if (unlikely(status))
2053                         break;
2054                 /*
2055                  * For non-resident attributes, we need to fill any holes with
2056                  * actual clusters and ensure all bufferes are mapped.  We also
2057                  * need to bring uptodate any buffers that are only partially
2058                  * being written to.
2059                  */
2060                 if (NInoNonResident(ni)) {
2061                         status = ntfs_prepare_pages_for_non_resident_write(
2062                                         pages, do_pages, pos, bytes);
2063                         if (unlikely(status)) {
2064                                 loff_t i_size;
2065
2066                                 do {
2067                                         unlock_page(pages[--do_pages]);
2068                                         page_cache_release(pages[do_pages]);
2069                                 } while (do_pages);
2070                                 /*
2071                                  * The write preparation may have instantiated
2072                                  * allocated space outside i_size.  Trim this
2073                                  * off again.  We can ignore any errors in this
2074                                  * case as we will just be waisting a bit of
2075                                  * allocated space, which is not a disaster.
2076                                  */
2077                                 i_size = i_size_read(vi);
2078                                 if (pos + bytes > i_size)
2079                                         vmtruncate(vi, i_size);
2080                                 break;
2081                         }
2082                 }
2083                 u = (pos >> PAGE_CACHE_SHIFT) - pages[0]->index;
2084                 if (likely(nr_segs == 1)) {
2085                         copied = ntfs_copy_from_user(pages + u, do_pages - u,
2086                                         ofs, buf, bytes);
2087                         buf += copied;
2088                 } else
2089                         copied = ntfs_copy_from_user_iovec(pages + u,
2090                                         do_pages - u, ofs, &iov, &iov_ofs,
2091                                         bytes);
2092                 ntfs_flush_dcache_pages(pages + u, do_pages - u);
2093                 status = ntfs_commit_pages_after_write(pages, do_pages, pos,
2094                                 bytes);
2095                 if (likely(!status)) {
2096                         written += copied;
2097                         count -= copied;
2098                         pos += copied;
2099                         if (unlikely(copied != bytes))
2100                                 status = -EFAULT;
2101                 }
2102                 do {
2103                         unlock_page(pages[--do_pages]);
2104                         mark_page_accessed(pages[do_pages]);
2105                         page_cache_release(pages[do_pages]);
2106                 } while (do_pages);
2107                 if (unlikely(status))
2108                         break;
2109                 balance_dirty_pages_ratelimited(mapping);
2110                 cond_resched();
2111         } while (count);
2112 err_out:
2113         *ppos = pos;
2114         if (cached_page)
2115                 page_cache_release(cached_page);
2116         /* For now, when the user asks for O_SYNC, we actually give O_DSYNC. */
2117         if (likely(!status)) {
2118                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(vi))) {
2119                         if (!mapping->a_ops->writepage || !is_sync_kiocb(iocb))
2120                                 status = generic_osync_inode(vi, mapping,
2121                                                 OSYNC_METADATA|OSYNC_DATA);
2122                 }
2123         }
2124         pagevec_lru_add(&lru_pvec);
2125         ntfs_debug("Done.  Returning %s (written 0x%lx, status %li).",
2126                         written ? "written" : "status", (unsigned long)written,
2127                         (long)status);
2128         return written ? written : status;
2129 }
2130
2131 /**
2132  * ntfs_file_aio_write_nolock -
2133  */
2134 static ssize_t ntfs_file_aio_write_nolock(struct kiocb *iocb,
2135                 const struct iovec *iov, unsigned long nr_segs, loff_t *ppos)
2136 {
2137         struct file *file = iocb->ki_filp;
2138         struct address_space *mapping = file->f_mapping;
2139         struct inode *inode = mapping->host;
2140         loff_t pos;
2141         unsigned long seg;
2142         size_t count;           /* after file limit checks */
2143         ssize_t written, err;
2144
2145         count = 0;
2146         for (seg = 0; seg < nr_segs; seg++) {
2147                 const struct iovec *iv = &iov[seg];
2148                 /*
2149                  * If any segment has a negative length, or the cumulative
2150                  * length ever wraps negative then return -EINVAL.
2151                  */
2152                 count += iv->iov_len;
2153                 if (unlikely((ssize_t)(count|iv->iov_len) < 0))
2154                         return -EINVAL;
2155                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
2156                         continue;
2157                 if (!seg)
2158                         return -EFAULT;
2159                 nr_segs = seg;
2160                 count -= iv->iov_len;   /* This segment is no good */
2161                 break;
2162         }
2163         pos = *ppos;
2164         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2165         /* We can write back this queue in page reclaim. */
2166         current->backing_dev_info = mapping->backing_dev_info;
2167         written = 0;
2168         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2169         if (err)
2170                 goto out;
2171         if (!count)
2172                 goto out;
2173         err = remove_suid(file->f_dentry);
2174         if (err)
2175                 goto out;
2176         inode_update_time(inode, 1);
2177         written = ntfs_file_buffered_write(iocb, iov, nr_segs, pos, ppos,
2178                         count);
2179 out:
2180         current->backing_dev_info = NULL;
2181         return written ? written : err;
2182 }
2183
2184 /**
2185  * ntfs_file_aio_write -
2186  */
2187 static ssize_t ntfs_file_aio_write(struct kiocb *iocb, const char __user *buf,
2188                 size_t count, loff_t pos)
2189 {
2190         struct file *file = iocb->ki_filp;
2191         struct address_space *mapping = file->f_mapping;
2192         struct inode *inode = mapping->host;
2193         ssize_t ret;
2194         struct iovec local_iov = { .iov_base = (void __user *)buf,
2195                                    .iov_len = count };
2196
2197         BUG_ON(iocb->ki_pos != pos);
2198
2199         down(&inode->i_sem);
2200         ret = ntfs_file_aio_write_nolock(iocb, &local_iov, 1, &iocb->ki_pos);
2201         up(&inode->i_sem);
2202         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2203                 int err = sync_page_range(inode, mapping, pos, ret);
2204                 if (err < 0)
2205                         ret = err;
2206         }
2207         return ret;
2208 }
2209
2210 /**
2211  * ntfs_file_writev -
2212  *
2213  * Basically the same as generic_file_writev() except that it ends up calling
2214  * ntfs_file_aio_write_nolock() instead of __generic_file_aio_write_nolock().
2215  */
2216 static ssize_t ntfs_file_writev(struct file *file, const struct iovec *iov,
2217                 unsigned long nr_segs, loff_t *ppos)
2218 {
2219         struct address_space *mapping = file->f_mapping;
2220         struct inode *inode = mapping->host;
2221         struct kiocb kiocb;
2222         ssize_t ret;
2223
2224         down(&inode->i_sem);
2225         init_sync_kiocb(&kiocb, file);
2226         ret = ntfs_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2227         if (ret == -EIOCBQUEUED)
2228                 ret = wait_on_sync_kiocb(&kiocb);
2229         up(&inode->i_sem);
2230         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2231                 int err = sync_page_range(inode, mapping, *ppos - ret, ret);
2232                 if (err < 0)
2233                         ret = err;
2234         }
2235         return ret;
2236 }
2237
2238 /**
2239  * ntfs_file_write - simple wrapper for ntfs_file_writev()
2240  */
2241 static ssize_t ntfs_file_write(struct file *file, const char __user *buf,
2242                 size_t count, loff_t *ppos)
2243 {
2244         struct iovec local_iov = { .iov_base = (void __user *)buf,
2245                                    .iov_len = count };
2246
2247         return ntfs_file_writev(file, &local_iov, 1, ppos);
2248 }
2249
2250 /**
2251  * ntfs_file_fsync - sync a file to disk
2252  * @filp:       file to be synced
2253  * @dentry:     dentry describing the file to sync
2254  * @datasync:   if non-zero only flush user data and not metadata
2255  *
2256  * Data integrity sync of a file to disk.  Used for fsync, fdatasync, and msync
2257  * system calls.  This function is inspired by fs/buffer.c::file_fsync().
2258  *
2259  * If @datasync is false, write the mft record and all associated extent mft
2260  * records as well as the $DATA attribute and then sync the block device.
2261  *
2262  * If @datasync is true and the attribute is non-resident, we skip the writing
2263  * of the mft record and all associated extent mft records (this might still
2264  * happen due to the write_inode_now() call).
2265  *
2266  * Also, if @datasync is true, we do not wait on the inode to be written out
2267  * but we always wait on the page cache pages to be written out.
2268  *
2269  * Note: In the past @filp could be NULL so we ignore it as we don't need it
2270  * anyway.
2271  *
2272  * Locking: Caller must hold i_sem on the inode.
2273  *
2274  * TODO: We should probably also write all attribute/index inodes associated
2275  * with this inode but since we have no simple way of getting to them we ignore
2276  * this problem for now.
2277  */
2278 static int ntfs_file_fsync(struct file *filp, struct dentry *dentry,
2279                 int datasync)
2280 {
2281         struct inode *vi = dentry->d_inode;
2282         int err, ret = 0;
2283
2284         ntfs_debug("Entering for inode 0x%lx.", vi->i_ino);
2285         BUG_ON(S_ISDIR(vi->i_mode));
2286         if (!datasync || !NInoNonResident(NTFS_I(vi)))
2287                 ret = ntfs_write_inode(vi, 1);
2288         write_inode_now(vi, !datasync);
2289         /*
2290          * NOTE: If we were to use mapping->private_list (see ext2 and
2291          * fs/buffer.c) for dirty blocks then we could optimize the below to be
2292          * sync_mapping_buffers(vi->i_mapping).
2293          */
2294         err = sync_blockdev(vi->i_sb->s_bdev);
2295         if (unlikely(err && !ret))
2296                 ret = err;
2297         if (likely(!ret))
2298                 ntfs_debug("Done.");
2299         else
2300                 ntfs_warning(vi->i_sb, "Failed to f%ssync inode 0x%lx.  Error "
2301                                 "%u.", datasync ? "data" : "", vi->i_ino, -ret);
2302         return ret;
2303 }
2304
2305 #endif /* NTFS_RW */
2306
2307 struct file_operations ntfs_file_ops = {
2308         .llseek         = generic_file_llseek,   /* Seek inside file. */
2309         .read           = generic_file_read,     /* Read from file. */
2310         .aio_read       = generic_file_aio_read, /* Async read from file. */
2311         .readv          = generic_file_readv,    /* Read from file. */
2312 #ifdef NTFS_RW
2313         .write          = ntfs_file_write,       /* Write to file. */
2314         .aio_write      = ntfs_file_aio_write,   /* Async write to file. */
2315         .writev         = ntfs_file_writev,      /* Write to file. */
2316         /*.release      = ,*/                    /* Last file is closed.  See
2317                                                     fs/ext2/file.c::
2318                                                     ext2_release_file() for
2319                                                     how to use this to discard
2320                                                     preallocated space for
2321                                                     write opened files. */
2322         .fsync          = ntfs_file_fsync,       /* Sync a file to disk. */
2323         /*.aio_fsync    = ,*/                    /* Sync all outstanding async
2324                                                     i/o operations on a
2325                                                     kiocb. */
2326 #endif /* NTFS_RW */
2327         /*.ioctl        = ,*/                    /* Perform function on the
2328                                                     mounted filesystem. */
2329         .mmap           = generic_file_mmap,     /* Mmap file. */
2330         .open           = ntfs_file_open,        /* Open file. */
2331         .sendfile       = generic_file_sendfile, /* Zero-copy data send with
2332                                                     the data source being on
2333                                                     the ntfs partition.  We do
2334                                                     not need to care about the
2335                                                     data destination. */
2336         /*.sendpage     = ,*/                    /* Zero-copy data send with
2337                                                     the data destination being
2338                                                     on the ntfs partition.  We
2339                                                     do not need to care about
2340                                                     the data source. */
2341 };
2342
2343 struct inode_operations ntfs_file_inode_ops = {
2344 #ifdef NTFS_RW
2345         .truncate       = ntfs_truncate_vfs,
2346         .setattr        = ntfs_setattr,
2347 #endif /* NTFS_RW */
2348 };
2349
2350 struct file_operations ntfs_empty_file_ops = {};
2351
2352 struct inode_operations ntfs_empty_inode_ops = {};