Merge branch 'upstream-linus' of master.kernel.org:/pub/scm/linux/kernel/git/jgarzik...
[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) 2001-2006 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_mutex 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_mutex 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 = vol->sb->s_blocksize;
533         blocksize_bits = vol->sb->s_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.",
947                                 (unsigned long long)lcn);
948                 /* Map and lock the mft record and get the attribute record. */
949                 if (!NInoAttr(ni))
950                         base_ni = ni;
951                 else
952                         base_ni = ni->ext.base_ntfs_ino;
953                 m = map_mft_record(base_ni);
954                 if (IS_ERR(m)) {
955                         err = PTR_ERR(m);
956                         break;
957                 }
958                 ctx = ntfs_attr_get_search_ctx(base_ni, m);
959                 if (unlikely(!ctx)) {
960                         err = -ENOMEM;
961                         unmap_mft_record(base_ni);
962                         break;
963                 }
964                 status.mft_attr_mapped = 1;
965                 err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
966                                 CASE_SENSITIVE, bh_cpos, NULL, 0, ctx);
967                 if (unlikely(err)) {
968                         if (err == -ENOENT)
969                                 err = -EIO;
970                         break;
971                 }
972                 m = ctx->mrec;
973                 a = ctx->attr;
974                 /*
975                  * Find the runlist element with which the attribute extent
976                  * starts.  Note, we cannot use the _attr_ version because we
977                  * have mapped the mft record.  That is ok because we know the
978                  * runlist fragment must be mapped already to have ever gotten
979                  * here, so we can just use the _rl_ version.
980                  */
981                 vcn = sle64_to_cpu(a->data.non_resident.lowest_vcn);
982                 rl2 = ntfs_rl_find_vcn_nolock(rl, vcn);
983                 BUG_ON(!rl2);
984                 BUG_ON(!rl2->length);
985                 BUG_ON(rl2->lcn < LCN_HOLE);
986                 highest_vcn = sle64_to_cpu(a->data.non_resident.highest_vcn);
987                 /*
988                  * If @highest_vcn is zero, calculate the real highest_vcn
989                  * (which can really be zero).
990                  */
991                 if (!highest_vcn)
992                         highest_vcn = (sle64_to_cpu(
993                                         a->data.non_resident.allocated_size) >>
994                                         vol->cluster_size_bits) - 1;
995                 /*
996                  * Determine the size of the mapping pairs array for the new
997                  * extent, i.e. the old extent with the hole filled.
998                  */
999                 mp_size = ntfs_get_size_for_mapping_pairs(vol, rl2, vcn,
1000                                 highest_vcn);
1001                 if (unlikely(mp_size <= 0)) {
1002                         if (!(err = mp_size))
1003                                 err = -EIO;
1004                         ntfs_debug("Failed to get size for mapping pairs "
1005                                         "array, error code %i.", err);
1006                         break;
1007                 }
1008                 /*
1009                  * Resize the attribute record to fit the new mapping pairs
1010                  * array.
1011                  */
1012                 attr_rec_len = le32_to_cpu(a->length);
1013                 err = ntfs_attr_record_resize(m, a, mp_size + le16_to_cpu(
1014                                 a->data.non_resident.mapping_pairs_offset));
1015                 if (unlikely(err)) {
1016                         BUG_ON(err != -ENOSPC);
1017                         // TODO: Deal with this by using the current attribute
1018                         // and fill it with as much of the mapping pairs
1019                         // array as possible.  Then loop over each attribute
1020                         // extent rewriting the mapping pairs arrays as we go
1021                         // along and if when we reach the end we have not
1022                         // enough space, try to resize the last attribute
1023                         // extent and if even that fails, add a new attribute
1024                         // extent.
1025                         // We could also try to resize at each step in the hope
1026                         // that we will not need to rewrite every single extent.
1027                         // Note, we may need to decompress some extents to fill
1028                         // the runlist as we are walking the extents...
1029                         ntfs_error(vol->sb, "Not enough space in the mft "
1030                                         "record for the extended attribute "
1031                                         "record.  This case is not "
1032                                         "implemented yet.");
1033                         err = -EOPNOTSUPP;
1034                         break ;
1035                 }
1036                 status.mp_rebuilt = 1;
1037                 /*
1038                  * Generate the mapping pairs array directly into the attribute
1039                  * record.
1040                  */
1041                 err = ntfs_mapping_pairs_build(vol, (u8*)a + le16_to_cpu(
1042                                 a->data.non_resident.mapping_pairs_offset),
1043                                 mp_size, rl2, vcn, highest_vcn, NULL);
1044                 if (unlikely(err)) {
1045                         ntfs_error(vol->sb, "Cannot fill hole in inode 0x%lx, "
1046                                         "attribute type 0x%x, because building "
1047                                         "the mapping pairs failed with error "
1048                                         "code %i.", vi->i_ino,
1049                                         (unsigned)le32_to_cpu(ni->type), err);
1050                         err = -EIO;
1051                         break;
1052                 }
1053                 /* Update the highest_vcn but only if it was not set. */
1054                 if (unlikely(!a->data.non_resident.highest_vcn))
1055                         a->data.non_resident.highest_vcn =
1056                                         cpu_to_sle64(highest_vcn);
1057                 /*
1058                  * If the attribute is sparse/compressed, update the compressed
1059                  * size in the ntfs_inode structure and the attribute record.
1060                  */
1061                 if (likely(NInoSparse(ni) || NInoCompressed(ni))) {
1062                         /*
1063                          * If we are not in the first attribute extent, switch
1064                          * to it, but first ensure the changes will make it to
1065                          * disk later.
1066                          */
1067                         if (a->data.non_resident.lowest_vcn) {
1068                                 flush_dcache_mft_record_page(ctx->ntfs_ino);
1069                                 mark_mft_record_dirty(ctx->ntfs_ino);
1070                                 ntfs_attr_reinit_search_ctx(ctx);
1071                                 err = ntfs_attr_lookup(ni->type, ni->name,
1072                                                 ni->name_len, CASE_SENSITIVE,
1073                                                 0, NULL, 0, ctx);
1074                                 if (unlikely(err)) {
1075                                         status.attr_switched = 1;
1076                                         break;
1077                                 }
1078                                 /* @m is not used any more so do not set it. */
1079                                 a = ctx->attr;
1080                         }
1081                         write_lock_irqsave(&ni->size_lock, flags);
1082                         ni->itype.compressed.size += vol->cluster_size;
1083                         a->data.non_resident.compressed_size =
1084                                         cpu_to_sle64(ni->itype.compressed.size);
1085                         write_unlock_irqrestore(&ni->size_lock, flags);
1086                 }
1087                 /* Ensure the changes make it to disk. */
1088                 flush_dcache_mft_record_page(ctx->ntfs_ino);
1089                 mark_mft_record_dirty(ctx->ntfs_ino);
1090                 ntfs_attr_put_search_ctx(ctx);
1091                 unmap_mft_record(base_ni);
1092                 /* Successfully filled the hole. */
1093                 status.runlist_merged = 0;
1094                 status.mft_attr_mapped = 0;
1095                 status.mp_rebuilt = 0;
1096                 /* Setup the map cache and use that to deal with the buffer. */
1097                 was_hole = TRUE;
1098                 vcn = bh_cpos;
1099                 vcn_len = 1;
1100                 lcn_block = lcn << (vol->cluster_size_bits - blocksize_bits);
1101                 cdelta = 0;
1102                 /*
1103                  * If the number of remaining clusters in the @pages is smaller
1104                  * or equal to the number of cached clusters, unlock the
1105                  * runlist as the map cache will be used from now on.
1106                  */
1107                 if (likely(vcn + vcn_len >= cend)) {
1108                         up_write(&ni->runlist.lock);
1109                         rl_write_locked = FALSE;
1110                         rl = NULL;
1111                 }
1112                 goto map_buffer_cached;
1113         } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
1114         /* If there are no errors, do the next page. */
1115         if (likely(!err && ++u < nr_pages))
1116                 goto do_next_page;
1117         /* If there are no errors, release the runlist lock if we took it. */
1118         if (likely(!err)) {
1119                 if (unlikely(rl_write_locked)) {
1120                         up_write(&ni->runlist.lock);
1121                         rl_write_locked = FALSE;
1122                 } else if (unlikely(rl))
1123                         up_read(&ni->runlist.lock);
1124                 rl = NULL;
1125         }
1126         /* If we issued read requests, let them complete. */
1127         read_lock_irqsave(&ni->size_lock, flags);
1128         initialized_size = ni->initialized_size;
1129         read_unlock_irqrestore(&ni->size_lock, flags);
1130         while (wait_bh > wait) {
1131                 bh = *--wait_bh;
1132                 wait_on_buffer(bh);
1133                 if (likely(buffer_uptodate(bh))) {
1134                         page = bh->b_page;
1135                         bh_pos = ((s64)page->index << PAGE_CACHE_SHIFT) +
1136                                         bh_offset(bh);
1137                         /*
1138                          * If the buffer overflows the initialized size, need
1139                          * to zero the overflowing region.
1140                          */
1141                         if (unlikely(bh_pos + blocksize > initialized_size)) {
1142                                 u8 *kaddr;
1143                                 int ofs = 0;
1144
1145                                 if (likely(bh_pos < initialized_size))
1146                                         ofs = initialized_size - bh_pos;
1147                                 kaddr = kmap_atomic(page, KM_USER0);
1148                                 memset(kaddr + bh_offset(bh) + ofs, 0,
1149                                                 blocksize - ofs);
1150                                 kunmap_atomic(kaddr, KM_USER0);
1151                                 flush_dcache_page(page);
1152                         }
1153                 } else /* if (unlikely(!buffer_uptodate(bh))) */
1154                         err = -EIO;
1155         }
1156         if (likely(!err)) {
1157                 /* Clear buffer_new on all buffers. */
1158                 u = 0;
1159                 do {
1160                         bh = head = page_buffers(pages[u]);
1161                         do {
1162                                 if (buffer_new(bh))
1163                                         clear_buffer_new(bh);
1164                         } while ((bh = bh->b_this_page) != head);
1165                 } while (++u < nr_pages);
1166                 ntfs_debug("Done.");
1167                 return err;
1168         }
1169         if (status.attr_switched) {
1170                 /* Get back to the attribute extent we modified. */
1171                 ntfs_attr_reinit_search_ctx(ctx);
1172                 if (ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1173                                 CASE_SENSITIVE, bh_cpos, NULL, 0, ctx)) {
1174                         ntfs_error(vol->sb, "Failed to find required "
1175                                         "attribute extent of attribute in "
1176                                         "error code path.  Run chkdsk to "
1177                                         "recover.");
1178                         write_lock_irqsave(&ni->size_lock, flags);
1179                         ni->itype.compressed.size += vol->cluster_size;
1180                         write_unlock_irqrestore(&ni->size_lock, flags);
1181                         flush_dcache_mft_record_page(ctx->ntfs_ino);
1182                         mark_mft_record_dirty(ctx->ntfs_ino);
1183                         /*
1184                          * The only thing that is now wrong is the compressed
1185                          * size of the base attribute extent which chkdsk
1186                          * should be able to fix.
1187                          */
1188                         NVolSetErrors(vol);
1189                 } else {
1190                         m = ctx->mrec;
1191                         a = ctx->attr;
1192                         status.attr_switched = 0;
1193                 }
1194         }
1195         /*
1196          * If the runlist has been modified, need to restore it by punching a
1197          * hole into it and we then need to deallocate the on-disk cluster as
1198          * well.  Note, we only modify the runlist if we are able to generate a
1199          * new mapping pairs array, i.e. only when the mapped attribute extent
1200          * is not switched.
1201          */
1202         if (status.runlist_merged && !status.attr_switched) {
1203                 BUG_ON(!rl_write_locked);
1204                 /* Make the file cluster we allocated sparse in the runlist. */
1205                 if (ntfs_rl_punch_nolock(vol, &ni->runlist, bh_cpos, 1)) {
1206                         ntfs_error(vol->sb, "Failed to punch hole into "
1207                                         "attribute runlist in error code "
1208                                         "path.  Run chkdsk to recover the "
1209                                         "lost cluster.");
1210                         NVolSetErrors(vol);
1211                 } else /* if (success) */ {
1212                         status.runlist_merged = 0;
1213                         /*
1214                          * Deallocate the on-disk cluster we allocated but only
1215                          * if we succeeded in punching its vcn out of the
1216                          * runlist.
1217                          */
1218                         down_write(&vol->lcnbmp_lock);
1219                         if (ntfs_bitmap_clear_bit(vol->lcnbmp_ino, lcn)) {
1220                                 ntfs_error(vol->sb, "Failed to release "
1221                                                 "allocated cluster in error "
1222                                                 "code path.  Run chkdsk to "
1223                                                 "recover the lost cluster.");
1224                                 NVolSetErrors(vol);
1225                         }
1226                         up_write(&vol->lcnbmp_lock);
1227                 }
1228         }
1229         /*
1230          * Resize the attribute record to its old size and rebuild the mapping
1231          * pairs array.  Note, we only can do this if the runlist has been
1232          * restored to its old state which also implies that the mapped
1233          * attribute extent is not switched.
1234          */
1235         if (status.mp_rebuilt && !status.runlist_merged) {
1236                 if (ntfs_attr_record_resize(m, a, attr_rec_len)) {
1237                         ntfs_error(vol->sb, "Failed to restore attribute "
1238                                         "record in error code path.  Run "
1239                                         "chkdsk to recover.");
1240                         NVolSetErrors(vol);
1241                 } else /* if (success) */ {
1242                         if (ntfs_mapping_pairs_build(vol, (u8*)a +
1243                                         le16_to_cpu(a->data.non_resident.
1244                                         mapping_pairs_offset), attr_rec_len -
1245                                         le16_to_cpu(a->data.non_resident.
1246                                         mapping_pairs_offset), ni->runlist.rl,
1247                                         vcn, highest_vcn, NULL)) {
1248                                 ntfs_error(vol->sb, "Failed to restore "
1249                                                 "mapping pairs array in error "
1250                                                 "code path.  Run chkdsk to "
1251                                                 "recover.");
1252                                 NVolSetErrors(vol);
1253                         }
1254                         flush_dcache_mft_record_page(ctx->ntfs_ino);
1255                         mark_mft_record_dirty(ctx->ntfs_ino);
1256                 }
1257         }
1258         /* Release the mft record and the attribute. */
1259         if (status.mft_attr_mapped) {
1260                 ntfs_attr_put_search_ctx(ctx);
1261                 unmap_mft_record(base_ni);
1262         }
1263         /* Release the runlist lock. */
1264         if (rl_write_locked)
1265                 up_write(&ni->runlist.lock);
1266         else if (rl)
1267                 up_read(&ni->runlist.lock);
1268         /*
1269          * Zero out any newly allocated blocks to avoid exposing stale data.
1270          * If BH_New is set, we know that the block was newly allocated above
1271          * and that it has not been fully zeroed and marked dirty yet.
1272          */
1273         nr_pages = u;
1274         u = 0;
1275         end = bh_cpos << vol->cluster_size_bits;
1276         do {
1277                 page = pages[u];
1278                 bh = head = page_buffers(page);
1279                 do {
1280                         if (u == nr_pages &&
1281                                         ((s64)page->index << PAGE_CACHE_SHIFT) +
1282                                         bh_offset(bh) >= end)
1283                                 break;
1284                         if (!buffer_new(bh))
1285                                 continue;
1286                         clear_buffer_new(bh);
1287                         if (!buffer_uptodate(bh)) {
1288                                 if (PageUptodate(page))
1289                                         set_buffer_uptodate(bh);
1290                                 else {
1291                                         u8 *kaddr = kmap_atomic(page, KM_USER0);
1292                                         memset(kaddr + bh_offset(bh), 0,
1293                                                         blocksize);
1294                                         kunmap_atomic(kaddr, KM_USER0);
1295                                         flush_dcache_page(page);
1296                                         set_buffer_uptodate(bh);
1297                                 }
1298                         }
1299                         mark_buffer_dirty(bh);
1300                 } while ((bh = bh->b_this_page) != head);
1301         } while (++u <= nr_pages);
1302         ntfs_error(vol->sb, "Failed.  Returning error code %i.", err);
1303         return err;
1304 }
1305
1306 /*
1307  * Copy as much as we can into the pages and return the number of bytes which
1308  * were sucessfully copied.  If a fault is encountered then clear the pages
1309  * out to (ofs + bytes) and return the number of bytes which were copied.
1310  */
1311 static inline size_t ntfs_copy_from_user(struct page **pages,
1312                 unsigned nr_pages, unsigned ofs, const char __user *buf,
1313                 size_t bytes)
1314 {
1315         struct page **last_page = pages + nr_pages;
1316         char *kaddr;
1317         size_t total = 0;
1318         unsigned len;
1319         int left;
1320
1321         do {
1322                 len = PAGE_CACHE_SIZE - ofs;
1323                 if (len > bytes)
1324                         len = bytes;
1325                 kaddr = kmap_atomic(*pages, KM_USER0);
1326                 left = __copy_from_user_inatomic(kaddr + ofs, buf, len);
1327                 kunmap_atomic(kaddr, KM_USER0);
1328                 if (unlikely(left)) {
1329                         /* Do it the slow way. */
1330                         kaddr = kmap(*pages);
1331                         left = __copy_from_user(kaddr + ofs, buf, len);
1332                         kunmap(*pages);
1333                         if (unlikely(left))
1334                                 goto err_out;
1335                 }
1336                 total += len;
1337                 bytes -= len;
1338                 if (!bytes)
1339                         break;
1340                 buf += len;
1341                 ofs = 0;
1342         } while (++pages < last_page);
1343 out:
1344         return total;
1345 err_out:
1346         total += len - left;
1347         /* Zero the rest of the target like __copy_from_user(). */
1348         while (++pages < last_page) {
1349                 bytes -= len;
1350                 if (!bytes)
1351                         break;
1352                 len = PAGE_CACHE_SIZE;
1353                 if (len > bytes)
1354                         len = bytes;
1355                 kaddr = kmap_atomic(*pages, KM_USER0);
1356                 memset(kaddr, 0, len);
1357                 kunmap_atomic(kaddr, KM_USER0);
1358         }
1359         goto out;
1360 }
1361
1362 static size_t __ntfs_copy_from_user_iovec(char *vaddr,
1363                 const struct iovec *iov, size_t iov_ofs, size_t bytes)
1364 {
1365         size_t total = 0;
1366
1367         while (1) {
1368                 const char __user *buf = iov->iov_base + iov_ofs;
1369                 unsigned len;
1370                 size_t left;
1371
1372                 len = iov->iov_len - iov_ofs;
1373                 if (len > bytes)
1374                         len = bytes;
1375                 left = __copy_from_user_inatomic(vaddr, buf, len);
1376                 total += len;
1377                 bytes -= len;
1378                 vaddr += len;
1379                 if (unlikely(left)) {
1380                         /*
1381                          * Zero the rest of the target like __copy_from_user().
1382                          */
1383                         memset(vaddr, 0, bytes);
1384                         total -= left;
1385                         break;
1386                 }
1387                 if (!bytes)
1388                         break;
1389                 iov++;
1390                 iov_ofs = 0;
1391         }
1392         return total;
1393 }
1394
1395 static inline void ntfs_set_next_iovec(const struct iovec **iovp,
1396                 size_t *iov_ofsp, size_t bytes)
1397 {
1398         const struct iovec *iov = *iovp;
1399         size_t iov_ofs = *iov_ofsp;
1400
1401         while (bytes) {
1402                 unsigned len;
1403
1404                 len = iov->iov_len - iov_ofs;
1405                 if (len > bytes)
1406                         len = bytes;
1407                 bytes -= len;
1408                 iov_ofs += len;
1409                 if (iov->iov_len == iov_ofs) {
1410                         iov++;
1411                         iov_ofs = 0;
1412                 }
1413         }
1414         *iovp = iov;
1415         *iov_ofsp = iov_ofs;
1416 }
1417
1418 /*
1419  * This has the same side-effects and return value as ntfs_copy_from_user().
1420  * The difference is that on a fault we need to memset the remainder of the
1421  * pages (out to offset + bytes), to emulate ntfs_copy_from_user()'s
1422  * single-segment behaviour.
1423  *
1424  * We call the same helper (__ntfs_copy_from_user_iovec()) both when atomic and
1425  * when not atomic.  This is ok because __ntfs_copy_from_user_iovec() calls
1426  * __copy_from_user_inatomic() and it is ok to call this when non-atomic.  In
1427  * fact, the only difference between __copy_from_user_inatomic() and
1428  * __copy_from_user() is that the latter calls might_sleep().  And on many
1429  * architectures __copy_from_user_inatomic() is just defined to
1430  * __copy_from_user() so it makes no difference at all on those architectures.
1431  */
1432 static inline size_t ntfs_copy_from_user_iovec(struct page **pages,
1433                 unsigned nr_pages, unsigned ofs, const struct iovec **iov,
1434                 size_t *iov_ofs, size_t bytes)
1435 {
1436         struct page **last_page = pages + nr_pages;
1437         char *kaddr;
1438         size_t copied, len, total = 0;
1439
1440         do {
1441                 len = PAGE_CACHE_SIZE - ofs;
1442                 if (len > bytes)
1443                         len = bytes;
1444                 kaddr = kmap_atomic(*pages, KM_USER0);
1445                 copied = __ntfs_copy_from_user_iovec(kaddr + ofs,
1446                                 *iov, *iov_ofs, len);
1447                 kunmap_atomic(kaddr, KM_USER0);
1448                 if (unlikely(copied != len)) {
1449                         /* Do it the slow way. */
1450                         kaddr = kmap(*pages);
1451                         copied = __ntfs_copy_from_user_iovec(kaddr + ofs,
1452                                         *iov, *iov_ofs, len);
1453                         kunmap(*pages);
1454                         if (unlikely(copied != len))
1455                                 goto err_out;
1456                 }
1457                 total += len;
1458                 bytes -= len;
1459                 if (!bytes)
1460                         break;
1461                 ntfs_set_next_iovec(iov, iov_ofs, len);
1462                 ofs = 0;
1463         } while (++pages < last_page);
1464 out:
1465         return total;
1466 err_out:
1467         total += copied;
1468         /* Zero the rest of the target like __copy_from_user(). */
1469         while (++pages < last_page) {
1470                 bytes -= len;
1471                 if (!bytes)
1472                         break;
1473                 len = PAGE_CACHE_SIZE;
1474                 if (len > bytes)
1475                         len = bytes;
1476                 kaddr = kmap_atomic(*pages, KM_USER0);
1477                 memset(kaddr, 0, len);
1478                 kunmap_atomic(kaddr, KM_USER0);
1479         }
1480         goto out;
1481 }
1482
1483 static inline void ntfs_flush_dcache_pages(struct page **pages,
1484                 unsigned nr_pages)
1485 {
1486         BUG_ON(!nr_pages);
1487         do {
1488                 /*
1489                  * Warning: Do not do the decrement at the same time as the
1490                  * call because flush_dcache_page() is a NULL macro on i386
1491                  * and hence the decrement never happens.
1492                  */
1493                 flush_dcache_page(pages[nr_pages]);
1494         } while (--nr_pages > 0);
1495 }
1496
1497 /**
1498  * ntfs_commit_pages_after_non_resident_write - commit the received data
1499  * @pages:      array of destination pages
1500  * @nr_pages:   number of pages in @pages
1501  * @pos:        byte position in file at which the write begins
1502  * @bytes:      number of bytes to be written
1503  *
1504  * See description of ntfs_commit_pages_after_write(), below.
1505  */
1506 static inline int ntfs_commit_pages_after_non_resident_write(
1507                 struct page **pages, const unsigned nr_pages,
1508                 s64 pos, size_t bytes)
1509 {
1510         s64 end, initialized_size;
1511         struct inode *vi;
1512         ntfs_inode *ni, *base_ni;
1513         struct buffer_head *bh, *head;
1514         ntfs_attr_search_ctx *ctx;
1515         MFT_RECORD *m;
1516         ATTR_RECORD *a;
1517         unsigned long flags;
1518         unsigned blocksize, u;
1519         int err;
1520
1521         vi = pages[0]->mapping->host;
1522         ni = NTFS_I(vi);
1523         blocksize = vi->i_sb->s_blocksize;
1524         end = pos + bytes;
1525         u = 0;
1526         do {
1527                 s64 bh_pos;
1528                 struct page *page;
1529                 BOOL partial;
1530
1531                 page = pages[u];
1532                 bh_pos = (s64)page->index << PAGE_CACHE_SHIFT;
1533                 bh = head = page_buffers(page);
1534                 partial = FALSE;
1535                 do {
1536                         s64 bh_end;
1537
1538                         bh_end = bh_pos + blocksize;
1539                         if (bh_end <= pos || bh_pos >= end) {
1540                                 if (!buffer_uptodate(bh))
1541                                         partial = TRUE;
1542                         } else {
1543                                 set_buffer_uptodate(bh);
1544                                 mark_buffer_dirty(bh);
1545                         }
1546                 } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
1547                 /*
1548                  * If all buffers are now uptodate but the page is not, set the
1549                  * page uptodate.
1550                  */
1551                 if (!partial && !PageUptodate(page))
1552                         SetPageUptodate(page);
1553         } while (++u < nr_pages);
1554         /*
1555          * Finally, if we do not need to update initialized_size or i_size we
1556          * are finished.
1557          */
1558         read_lock_irqsave(&ni->size_lock, flags);
1559         initialized_size = ni->initialized_size;
1560         read_unlock_irqrestore(&ni->size_lock, flags);
1561         if (end <= initialized_size) {
1562                 ntfs_debug("Done.");
1563                 return 0;
1564         }
1565         /*
1566          * Update initialized_size/i_size as appropriate, both in the inode and
1567          * the mft record.
1568          */
1569         if (!NInoAttr(ni))
1570                 base_ni = ni;
1571         else
1572                 base_ni = ni->ext.base_ntfs_ino;
1573         /* Map, pin, and lock the mft record. */
1574         m = map_mft_record(base_ni);
1575         if (IS_ERR(m)) {
1576                 err = PTR_ERR(m);
1577                 m = NULL;
1578                 ctx = NULL;
1579                 goto err_out;
1580         }
1581         BUG_ON(!NInoNonResident(ni));
1582         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1583         if (unlikely(!ctx)) {
1584                 err = -ENOMEM;
1585                 goto err_out;
1586         }
1587         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1588                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1589         if (unlikely(err)) {
1590                 if (err == -ENOENT)
1591                         err = -EIO;
1592                 goto err_out;
1593         }
1594         a = ctx->attr;
1595         BUG_ON(!a->non_resident);
1596         write_lock_irqsave(&ni->size_lock, flags);
1597         BUG_ON(end > ni->allocated_size);
1598         ni->initialized_size = end;
1599         a->data.non_resident.initialized_size = cpu_to_sle64(end);
1600         if (end > i_size_read(vi)) {
1601                 i_size_write(vi, end);
1602                 a->data.non_resident.data_size =
1603                                 a->data.non_resident.initialized_size;
1604         }
1605         write_unlock_irqrestore(&ni->size_lock, flags);
1606         /* Mark the mft record dirty, so it gets written back. */
1607         flush_dcache_mft_record_page(ctx->ntfs_ino);
1608         mark_mft_record_dirty(ctx->ntfs_ino);
1609         ntfs_attr_put_search_ctx(ctx);
1610         unmap_mft_record(base_ni);
1611         ntfs_debug("Done.");
1612         return 0;
1613 err_out:
1614         if (ctx)
1615                 ntfs_attr_put_search_ctx(ctx);
1616         if (m)
1617                 unmap_mft_record(base_ni);
1618         ntfs_error(vi->i_sb, "Failed to update initialized_size/i_size (error "
1619                         "code %i).", err);
1620         if (err != -ENOMEM)
1621                 NVolSetErrors(ni->vol);
1622         return err;
1623 }
1624
1625 /**
1626  * ntfs_commit_pages_after_write - commit the received data
1627  * @pages:      array of destination pages
1628  * @nr_pages:   number of pages in @pages
1629  * @pos:        byte position in file at which the write begins
1630  * @bytes:      number of bytes to be written
1631  *
1632  * This is called from ntfs_file_buffered_write() with i_mutex held on the inode
1633  * (@pages[0]->mapping->host).  There are @nr_pages pages in @pages which are
1634  * locked but not kmap()ped.  The source data has already been copied into the
1635  * @page.  ntfs_prepare_pages_for_non_resident_write() has been called before
1636  * the data was copied (for non-resident attributes only) and it returned
1637  * success.
1638  *
1639  * Need to set uptodate and mark dirty all buffers within the boundary of the
1640  * write.  If all buffers in a page are uptodate we set the page uptodate, too.
1641  *
1642  * Setting the buffers dirty ensures that they get written out later when
1643  * ntfs_writepage() is invoked by the VM.
1644  *
1645  * Finally, we need to update i_size and initialized_size as appropriate both
1646  * in the inode and the mft record.
1647  *
1648  * This is modelled after fs/buffer.c::generic_commit_write(), which marks
1649  * buffers uptodate and dirty, sets the page uptodate if all buffers in the
1650  * page are uptodate, and updates i_size if the end of io is beyond i_size.  In
1651  * that case, it also marks the inode dirty.
1652  *
1653  * If things have gone as outlined in
1654  * ntfs_prepare_pages_for_non_resident_write(), we do not need to do any page
1655  * content modifications here for non-resident attributes.  For resident
1656  * attributes we need to do the uptodate bringing here which we combine with
1657  * the copying into the mft record which means we save one atomic kmap.
1658  *
1659  * Return 0 on success or -errno on error.
1660  */
1661 static int ntfs_commit_pages_after_write(struct page **pages,
1662                 const unsigned nr_pages, s64 pos, size_t bytes)
1663 {
1664         s64 end, initialized_size;
1665         loff_t i_size;
1666         struct inode *vi;
1667         ntfs_inode *ni, *base_ni;
1668         struct page *page;
1669         ntfs_attr_search_ctx *ctx;
1670         MFT_RECORD *m;
1671         ATTR_RECORD *a;
1672         char *kattr, *kaddr;
1673         unsigned long flags;
1674         u32 attr_len;
1675         int err;
1676
1677         BUG_ON(!nr_pages);
1678         BUG_ON(!pages);
1679         page = pages[0];
1680         BUG_ON(!page);
1681         vi = page->mapping->host;
1682         ni = NTFS_I(vi);
1683         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, start page "
1684                         "index 0x%lx, nr_pages 0x%x, pos 0x%llx, bytes 0x%zx.",
1685                         vi->i_ino, ni->type, page->index, nr_pages,
1686                         (long long)pos, bytes);
1687         if (NInoNonResident(ni))
1688                 return ntfs_commit_pages_after_non_resident_write(pages,
1689                                 nr_pages, pos, bytes);
1690         BUG_ON(nr_pages > 1);
1691         /*
1692          * Attribute is resident, implying it is not compressed, encrypted, or
1693          * sparse.
1694          */
1695         if (!NInoAttr(ni))
1696                 base_ni = ni;
1697         else
1698                 base_ni = ni->ext.base_ntfs_ino;
1699         BUG_ON(NInoNonResident(ni));
1700         /* Map, pin, and lock the mft record. */
1701         m = map_mft_record(base_ni);
1702         if (IS_ERR(m)) {
1703                 err = PTR_ERR(m);
1704                 m = NULL;
1705                 ctx = NULL;
1706                 goto err_out;
1707         }
1708         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1709         if (unlikely(!ctx)) {
1710                 err = -ENOMEM;
1711                 goto err_out;
1712         }
1713         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1714                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1715         if (unlikely(err)) {
1716                 if (err == -ENOENT)
1717                         err = -EIO;
1718                 goto err_out;
1719         }
1720         a = ctx->attr;
1721         BUG_ON(a->non_resident);
1722         /* The total length of the attribute value. */
1723         attr_len = le32_to_cpu(a->data.resident.value_length);
1724         i_size = i_size_read(vi);
1725         BUG_ON(attr_len != i_size);
1726         BUG_ON(pos > attr_len);
1727         end = pos + bytes;
1728         BUG_ON(end > le32_to_cpu(a->length) -
1729                         le16_to_cpu(a->data.resident.value_offset));
1730         kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
1731         kaddr = kmap_atomic(page, KM_USER0);
1732         /* Copy the received data from the page to the mft record. */
1733         memcpy(kattr + pos, kaddr + pos, bytes);
1734         /* Update the attribute length if necessary. */
1735         if (end > attr_len) {
1736                 attr_len = end;
1737                 a->data.resident.value_length = cpu_to_le32(attr_len);
1738         }
1739         /*
1740          * If the page is not uptodate, bring the out of bounds area(s)
1741          * uptodate by copying data from the mft record to the page.
1742          */
1743         if (!PageUptodate(page)) {
1744                 if (pos > 0)
1745                         memcpy(kaddr, kattr, pos);
1746                 if (end < attr_len)
1747                         memcpy(kaddr + end, kattr + end, attr_len - end);
1748                 /* Zero the region outside the end of the attribute value. */
1749                 memset(kaddr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
1750                 flush_dcache_page(page);
1751                 SetPageUptodate(page);
1752         }
1753         kunmap_atomic(kaddr, KM_USER0);
1754         /* Update initialized_size/i_size if necessary. */
1755         read_lock_irqsave(&ni->size_lock, flags);
1756         initialized_size = ni->initialized_size;
1757         BUG_ON(end > ni->allocated_size);
1758         read_unlock_irqrestore(&ni->size_lock, flags);
1759         BUG_ON(initialized_size != i_size);
1760         if (end > initialized_size) {
1761                 unsigned long flags;
1762
1763                 write_lock_irqsave(&ni->size_lock, flags);
1764                 ni->initialized_size = end;
1765                 i_size_write(vi, end);
1766                 write_unlock_irqrestore(&ni->size_lock, flags);
1767         }
1768         /* Mark the mft record dirty, so it gets written back. */
1769         flush_dcache_mft_record_page(ctx->ntfs_ino);
1770         mark_mft_record_dirty(ctx->ntfs_ino);
1771         ntfs_attr_put_search_ctx(ctx);
1772         unmap_mft_record(base_ni);
1773         ntfs_debug("Done.");
1774         return 0;
1775 err_out:
1776         if (err == -ENOMEM) {
1777                 ntfs_warning(vi->i_sb, "Error allocating memory required to "
1778                                 "commit the write.");
1779                 if (PageUptodate(page)) {
1780                         ntfs_warning(vi->i_sb, "Page is uptodate, setting "
1781                                         "dirty so the write will be retried "
1782                                         "later on by the VM.");
1783                         /*
1784                          * Put the page on mapping->dirty_pages, but leave its
1785                          * buffers' dirty state as-is.
1786                          */
1787                         __set_page_dirty_nobuffers(page);
1788                         err = 0;
1789                 } else
1790                         ntfs_error(vi->i_sb, "Page is not uptodate.  Written "
1791                                         "data has been lost.");
1792         } else {
1793                 ntfs_error(vi->i_sb, "Resident attribute commit write failed "
1794                                 "with error %i.", err);
1795                 NVolSetErrors(ni->vol);
1796         }
1797         if (ctx)
1798                 ntfs_attr_put_search_ctx(ctx);
1799         if (m)
1800                 unmap_mft_record(base_ni);
1801         return err;
1802 }
1803
1804 /**
1805  * ntfs_file_buffered_write -
1806  *
1807  * Locking: The vfs is holding ->i_mutex on the inode.
1808  */
1809 static ssize_t ntfs_file_buffered_write(struct kiocb *iocb,
1810                 const struct iovec *iov, unsigned long nr_segs,
1811                 loff_t pos, loff_t *ppos, size_t count)
1812 {
1813         struct file *file = iocb->ki_filp;
1814         struct address_space *mapping = file->f_mapping;
1815         struct inode *vi = mapping->host;
1816         ntfs_inode *ni = NTFS_I(vi);
1817         ntfs_volume *vol = ni->vol;
1818         struct page *pages[NTFS_MAX_PAGES_PER_CLUSTER];
1819         struct page *cached_page = NULL;
1820         char __user *buf = NULL;
1821         s64 end, ll;
1822         VCN last_vcn;
1823         LCN lcn;
1824         unsigned long flags;
1825         size_t bytes, iov_ofs = 0;      /* Offset in the current iovec. */
1826         ssize_t status, written;
1827         unsigned nr_pages;
1828         int err;
1829         struct pagevec lru_pvec;
1830
1831         ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, "
1832                         "pos 0x%llx, count 0x%lx.",
1833                         vi->i_ino, (unsigned)le32_to_cpu(ni->type),
1834                         (unsigned long long)pos, (unsigned long)count);
1835         if (unlikely(!count))
1836                 return 0;
1837         BUG_ON(NInoMstProtected(ni));
1838         /*
1839          * If the attribute is not an index root and it is encrypted or
1840          * compressed, we cannot write to it yet.  Note we need to check for
1841          * AT_INDEX_ALLOCATION since this is the type of both directory and
1842          * index inodes.
1843          */
1844         if (ni->type != AT_INDEX_ALLOCATION) {
1845                 /* If file is encrypted, deny access, just like NT4. */
1846                 if (NInoEncrypted(ni)) {
1847                         /*
1848                          * Reminder for later: Encrypted files are _always_
1849                          * non-resident so that the content can always be
1850                          * encrypted.
1851                          */
1852                         ntfs_debug("Denying write access to encrypted file.");
1853                         return -EACCES;
1854                 }
1855                 if (NInoCompressed(ni)) {
1856                         /* Only unnamed $DATA attribute can be compressed. */
1857                         BUG_ON(ni->type != AT_DATA);
1858                         BUG_ON(ni->name_len);
1859                         /*
1860                          * Reminder for later: If resident, the data is not
1861                          * actually compressed.  Only on the switch to non-
1862                          * resident does compression kick in.  This is in
1863                          * contrast to encrypted files (see above).
1864                          */
1865                         ntfs_error(vi->i_sb, "Writing to compressed files is "
1866                                         "not implemented yet.  Sorry.");
1867                         return -EOPNOTSUPP;
1868                 }
1869         }
1870         /*
1871          * If a previous ntfs_truncate() failed, repeat it and abort if it
1872          * fails again.
1873          */
1874         if (unlikely(NInoTruncateFailed(ni))) {
1875                 down_write(&vi->i_alloc_sem);
1876                 err = ntfs_truncate(vi);
1877                 up_write(&vi->i_alloc_sem);
1878                 if (err || NInoTruncateFailed(ni)) {
1879                         if (!err)
1880                                 err = -EIO;
1881                         ntfs_error(vol->sb, "Cannot perform write to inode "
1882                                         "0x%lx, attribute type 0x%x, because "
1883                                         "ntfs_truncate() failed (error code "
1884                                         "%i).", vi->i_ino,
1885                                         (unsigned)le32_to_cpu(ni->type), err);
1886                         return err;
1887                 }
1888         }
1889         /* The first byte after the write. */
1890         end = pos + count;
1891         /*
1892          * If the write goes beyond the allocated size, extend the allocation
1893          * to cover the whole of the write, rounded up to the nearest cluster.
1894          */
1895         read_lock_irqsave(&ni->size_lock, flags);
1896         ll = ni->allocated_size;
1897         read_unlock_irqrestore(&ni->size_lock, flags);
1898         if (end > ll) {
1899                 /* Extend the allocation without changing the data size. */
1900                 ll = ntfs_attr_extend_allocation(ni, end, -1, pos);
1901                 if (likely(ll >= 0)) {
1902                         BUG_ON(pos >= ll);
1903                         /* If the extension was partial truncate the write. */
1904                         if (end > ll) {
1905                                 ntfs_debug("Truncating write to inode 0x%lx, "
1906                                                 "attribute type 0x%x, because "
1907                                                 "the allocation was only "
1908                                                 "partially extended.",
1909                                                 vi->i_ino, (unsigned)
1910                                                 le32_to_cpu(ni->type));
1911                                 end = ll;
1912                                 count = ll - pos;
1913                         }
1914                 } else {
1915                         err = ll;
1916                         read_lock_irqsave(&ni->size_lock, flags);
1917                         ll = ni->allocated_size;
1918                         read_unlock_irqrestore(&ni->size_lock, flags);
1919                         /* Perform a partial write if possible or fail. */
1920                         if (pos < ll) {
1921                                 ntfs_debug("Truncating write to inode 0x%lx, "
1922                                                 "attribute type 0x%x, because "
1923                                                 "extending the allocation "
1924                                                 "failed (error code %i).",
1925                                                 vi->i_ino, (unsigned)
1926                                                 le32_to_cpu(ni->type), err);
1927                                 end = ll;
1928                                 count = ll - pos;
1929                         } else {
1930                                 ntfs_error(vol->sb, "Cannot perform write to "
1931                                                 "inode 0x%lx, attribute type "
1932                                                 "0x%x, because extending the "
1933                                                 "allocation failed (error "
1934                                                 "code %i).", vi->i_ino,
1935                                                 (unsigned)
1936                                                 le32_to_cpu(ni->type), err);
1937                                 return err;
1938                         }
1939                 }
1940         }
1941         pagevec_init(&lru_pvec, 0);
1942         written = 0;
1943         /*
1944          * If the write starts beyond the initialized size, extend it up to the
1945          * beginning of the write and initialize all non-sparse space between
1946          * the old initialized size and the new one.  This automatically also
1947          * increments the vfs inode->i_size to keep it above or equal to the
1948          * initialized_size.
1949          */
1950         read_lock_irqsave(&ni->size_lock, flags);
1951         ll = ni->initialized_size;
1952         read_unlock_irqrestore(&ni->size_lock, flags);
1953         if (pos > ll) {
1954                 err = ntfs_attr_extend_initialized(ni, pos, &cached_page,
1955                                 &lru_pvec);
1956                 if (err < 0) {
1957                         ntfs_error(vol->sb, "Cannot perform write to inode "
1958                                         "0x%lx, attribute type 0x%x, because "
1959                                         "extending the initialized size "
1960                                         "failed (error code %i).", vi->i_ino,
1961                                         (unsigned)le32_to_cpu(ni->type), err);
1962                         status = err;
1963                         goto err_out;
1964                 }
1965         }
1966         /*
1967          * Determine the number of pages per cluster for non-resident
1968          * attributes.
1969          */
1970         nr_pages = 1;
1971         if (vol->cluster_size > PAGE_CACHE_SIZE && NInoNonResident(ni))
1972                 nr_pages = vol->cluster_size >> PAGE_CACHE_SHIFT;
1973         /* Finally, perform the actual write. */
1974         last_vcn = -1;
1975         if (likely(nr_segs == 1))
1976                 buf = iov->iov_base;
1977         do {
1978                 VCN vcn;
1979                 pgoff_t idx, start_idx;
1980                 unsigned ofs, do_pages, u;
1981                 size_t copied;
1982
1983                 start_idx = idx = pos >> PAGE_CACHE_SHIFT;
1984                 ofs = pos & ~PAGE_CACHE_MASK;
1985                 bytes = PAGE_CACHE_SIZE - ofs;
1986                 do_pages = 1;
1987                 if (nr_pages > 1) {
1988                         vcn = pos >> vol->cluster_size_bits;
1989                         if (vcn != last_vcn) {
1990                                 last_vcn = vcn;
1991                                 /*
1992                                  * Get the lcn of the vcn the write is in.  If
1993                                  * it is a hole, need to lock down all pages in
1994                                  * the cluster.
1995                                  */
1996                                 down_read(&ni->runlist.lock);
1997                                 lcn = ntfs_attr_vcn_to_lcn_nolock(ni, pos >>
1998                                                 vol->cluster_size_bits, FALSE);
1999                                 up_read(&ni->runlist.lock);
2000                                 if (unlikely(lcn < LCN_HOLE)) {
2001                                         status = -EIO;
2002                                         if (lcn == LCN_ENOMEM)
2003                                                 status = -ENOMEM;
2004                                         else
2005                                                 ntfs_error(vol->sb, "Cannot "
2006                                                         "perform write to "
2007                                                         "inode 0x%lx, "
2008                                                         "attribute type 0x%x, "
2009                                                         "because the attribute "
2010                                                         "is corrupt.",
2011                                                         vi->i_ino, (unsigned)
2012                                                         le32_to_cpu(ni->type));
2013                                         break;
2014                                 }
2015                                 if (lcn == LCN_HOLE) {
2016                                         start_idx = (pos & ~(s64)
2017                                                         vol->cluster_size_mask)
2018                                                         >> PAGE_CACHE_SHIFT;
2019                                         bytes = vol->cluster_size - (pos &
2020                                                         vol->cluster_size_mask);
2021                                         do_pages = nr_pages;
2022                                 }
2023                         }
2024                 }
2025                 if (bytes > count)
2026                         bytes = count;
2027                 /*
2028                  * Bring in the user page(s) that we will copy from _first_.
2029                  * Otherwise there is a nasty deadlock on copying from the same
2030                  * page(s) as we are writing to, without it/them being marked
2031                  * up-to-date.  Note, at present there is nothing to stop the
2032                  * pages being swapped out between us bringing them into memory
2033                  * and doing the actual copying.
2034                  */
2035                 if (likely(nr_segs == 1))
2036                         ntfs_fault_in_pages_readable(buf, bytes);
2037                 else
2038                         ntfs_fault_in_pages_readable_iovec(iov, iov_ofs, bytes);
2039                 /* Get and lock @do_pages starting at index @start_idx. */
2040                 status = __ntfs_grab_cache_pages(mapping, start_idx, do_pages,
2041                                 pages, &cached_page, &lru_pvec);
2042                 if (unlikely(status))
2043                         break;
2044                 /*
2045                  * For non-resident attributes, we need to fill any holes with
2046                  * actual clusters and ensure all bufferes are mapped.  We also
2047                  * need to bring uptodate any buffers that are only partially
2048                  * being written to.
2049                  */
2050                 if (NInoNonResident(ni)) {
2051                         status = ntfs_prepare_pages_for_non_resident_write(
2052                                         pages, do_pages, pos, bytes);
2053                         if (unlikely(status)) {
2054                                 loff_t i_size;
2055
2056                                 do {
2057                                         unlock_page(pages[--do_pages]);
2058                                         page_cache_release(pages[do_pages]);
2059                                 } while (do_pages);
2060                                 /*
2061                                  * The write preparation may have instantiated
2062                                  * allocated space outside i_size.  Trim this
2063                                  * off again.  We can ignore any errors in this
2064                                  * case as we will just be waisting a bit of
2065                                  * allocated space, which is not a disaster.
2066                                  */
2067                                 i_size = i_size_read(vi);
2068                                 if (pos + bytes > i_size)
2069                                         vmtruncate(vi, i_size);
2070                                 break;
2071                         }
2072                 }
2073                 u = (pos >> PAGE_CACHE_SHIFT) - pages[0]->index;
2074                 if (likely(nr_segs == 1)) {
2075                         copied = ntfs_copy_from_user(pages + u, do_pages - u,
2076                                         ofs, buf, bytes);
2077                         buf += copied;
2078                 } else
2079                         copied = ntfs_copy_from_user_iovec(pages + u,
2080                                         do_pages - u, ofs, &iov, &iov_ofs,
2081                                         bytes);
2082                 ntfs_flush_dcache_pages(pages + u, do_pages - u);
2083                 status = ntfs_commit_pages_after_write(pages, do_pages, pos,
2084                                 bytes);
2085                 if (likely(!status)) {
2086                         written += copied;
2087                         count -= copied;
2088                         pos += copied;
2089                         if (unlikely(copied != bytes))
2090                                 status = -EFAULT;
2091                 }
2092                 do {
2093                         unlock_page(pages[--do_pages]);
2094                         mark_page_accessed(pages[do_pages]);
2095                         page_cache_release(pages[do_pages]);
2096                 } while (do_pages);
2097                 if (unlikely(status))
2098                         break;
2099                 balance_dirty_pages_ratelimited(mapping);
2100                 cond_resched();
2101         } while (count);
2102 err_out:
2103         *ppos = pos;
2104         if (cached_page)
2105                 page_cache_release(cached_page);
2106         /* For now, when the user asks for O_SYNC, we actually give O_DSYNC. */
2107         if (likely(!status)) {
2108                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(vi))) {
2109                         if (!mapping->a_ops->writepage || !is_sync_kiocb(iocb))
2110                                 status = generic_osync_inode(vi, mapping,
2111                                                 OSYNC_METADATA|OSYNC_DATA);
2112                 }
2113         }
2114         pagevec_lru_add(&lru_pvec);
2115         ntfs_debug("Done.  Returning %s (written 0x%lx, status %li).",
2116                         written ? "written" : "status", (unsigned long)written,
2117                         (long)status);
2118         return written ? written : status;
2119 }
2120
2121 /**
2122  * ntfs_file_aio_write_nolock -
2123  */
2124 static ssize_t ntfs_file_aio_write_nolock(struct kiocb *iocb,
2125                 const struct iovec *iov, unsigned long nr_segs, loff_t *ppos)
2126 {
2127         struct file *file = iocb->ki_filp;
2128         struct address_space *mapping = file->f_mapping;
2129         struct inode *inode = mapping->host;
2130         loff_t pos;
2131         unsigned long seg;
2132         size_t count;           /* after file limit checks */
2133         ssize_t written, err;
2134
2135         count = 0;
2136         for (seg = 0; seg < nr_segs; seg++) {
2137                 const struct iovec *iv = &iov[seg];
2138                 /*
2139                  * If any segment has a negative length, or the cumulative
2140                  * length ever wraps negative then return -EINVAL.
2141                  */
2142                 count += iv->iov_len;
2143                 if (unlikely((ssize_t)(count|iv->iov_len) < 0))
2144                         return -EINVAL;
2145                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
2146                         continue;
2147                 if (!seg)
2148                         return -EFAULT;
2149                 nr_segs = seg;
2150                 count -= iv->iov_len;   /* This segment is no good */
2151                 break;
2152         }
2153         pos = *ppos;
2154         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2155         /* We can write back this queue in page reclaim. */
2156         current->backing_dev_info = mapping->backing_dev_info;
2157         written = 0;
2158         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2159         if (err)
2160                 goto out;
2161         if (!count)
2162                 goto out;
2163         err = remove_suid(file->f_dentry);
2164         if (err)
2165                 goto out;
2166         file_update_time(file);
2167         written = ntfs_file_buffered_write(iocb, iov, nr_segs, pos, ppos,
2168                         count);
2169 out:
2170         current->backing_dev_info = NULL;
2171         return written ? written : err;
2172 }
2173
2174 /**
2175  * ntfs_file_aio_write -
2176  */
2177 static ssize_t ntfs_file_aio_write(struct kiocb *iocb, const char __user *buf,
2178                 size_t count, loff_t pos)
2179 {
2180         struct file *file = iocb->ki_filp;
2181         struct address_space *mapping = file->f_mapping;
2182         struct inode *inode = mapping->host;
2183         ssize_t ret;
2184         struct iovec local_iov = { .iov_base = (void __user *)buf,
2185                                    .iov_len = count };
2186
2187         BUG_ON(iocb->ki_pos != pos);
2188
2189         mutex_lock(&inode->i_mutex);
2190         ret = ntfs_file_aio_write_nolock(iocb, &local_iov, 1, &iocb->ki_pos);
2191         mutex_unlock(&inode->i_mutex);
2192         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2193                 int err = sync_page_range(inode, mapping, pos, ret);
2194                 if (err < 0)
2195                         ret = err;
2196         }
2197         return ret;
2198 }
2199
2200 /**
2201  * ntfs_file_writev -
2202  *
2203  * Basically the same as generic_file_writev() except that it ends up calling
2204  * ntfs_file_aio_write_nolock() instead of __generic_file_aio_write_nolock().
2205  */
2206 static ssize_t ntfs_file_writev(struct file *file, const struct iovec *iov,
2207                 unsigned long nr_segs, loff_t *ppos)
2208 {
2209         struct address_space *mapping = file->f_mapping;
2210         struct inode *inode = mapping->host;
2211         struct kiocb kiocb;
2212         ssize_t ret;
2213
2214         mutex_lock(&inode->i_mutex);
2215         init_sync_kiocb(&kiocb, file);
2216         ret = ntfs_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2217         if (ret == -EIOCBQUEUED)
2218                 ret = wait_on_sync_kiocb(&kiocb);
2219         mutex_unlock(&inode->i_mutex);
2220         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2221                 int err = sync_page_range(inode, mapping, *ppos - ret, ret);
2222                 if (err < 0)
2223                         ret = err;
2224         }
2225         return ret;
2226 }
2227
2228 /**
2229  * ntfs_file_write - simple wrapper for ntfs_file_writev()
2230  */
2231 static ssize_t ntfs_file_write(struct file *file, const char __user *buf,
2232                 size_t count, loff_t *ppos)
2233 {
2234         struct iovec local_iov = { .iov_base = (void __user *)buf,
2235                                    .iov_len = count };
2236
2237         return ntfs_file_writev(file, &local_iov, 1, ppos);
2238 }
2239
2240 /**
2241  * ntfs_file_fsync - sync a file to disk
2242  * @filp:       file to be synced
2243  * @dentry:     dentry describing the file to sync
2244  * @datasync:   if non-zero only flush user data and not metadata
2245  *
2246  * Data integrity sync of a file to disk.  Used for fsync, fdatasync, and msync
2247  * system calls.  This function is inspired by fs/buffer.c::file_fsync().
2248  *
2249  * If @datasync is false, write the mft record and all associated extent mft
2250  * records as well as the $DATA attribute and then sync the block device.
2251  *
2252  * If @datasync is true and the attribute is non-resident, we skip the writing
2253  * of the mft record and all associated extent mft records (this might still
2254  * happen due to the write_inode_now() call).
2255  *
2256  * Also, if @datasync is true, we do not wait on the inode to be written out
2257  * but we always wait on the page cache pages to be written out.
2258  *
2259  * Note: In the past @filp could be NULL so we ignore it as we don't need it
2260  * anyway.
2261  *
2262  * Locking: Caller must hold i_mutex on the inode.
2263  *
2264  * TODO: We should probably also write all attribute/index inodes associated
2265  * with this inode but since we have no simple way of getting to them we ignore
2266  * this problem for now.
2267  */
2268 static int ntfs_file_fsync(struct file *filp, struct dentry *dentry,
2269                 int datasync)
2270 {
2271         struct inode *vi = dentry->d_inode;
2272         int err, ret = 0;
2273
2274         ntfs_debug("Entering for inode 0x%lx.", vi->i_ino);
2275         BUG_ON(S_ISDIR(vi->i_mode));
2276         if (!datasync || !NInoNonResident(NTFS_I(vi)))
2277                 ret = ntfs_write_inode(vi, 1);
2278         write_inode_now(vi, !datasync);
2279         /*
2280          * NOTE: If we were to use mapping->private_list (see ext2 and
2281          * fs/buffer.c) for dirty blocks then we could optimize the below to be
2282          * sync_mapping_buffers(vi->i_mapping).
2283          */
2284         err = sync_blockdev(vi->i_sb->s_bdev);
2285         if (unlikely(err && !ret))
2286                 ret = err;
2287         if (likely(!ret))
2288                 ntfs_debug("Done.");
2289         else
2290                 ntfs_warning(vi->i_sb, "Failed to f%ssync inode 0x%lx.  Error "
2291                                 "%u.", datasync ? "data" : "", vi->i_ino, -ret);
2292         return ret;
2293 }
2294
2295 #endif /* NTFS_RW */
2296
2297 struct file_operations ntfs_file_ops = {
2298         .llseek         = generic_file_llseek,   /* Seek inside file. */
2299         .read           = generic_file_read,     /* Read from file. */
2300         .aio_read       = generic_file_aio_read, /* Async read from file. */
2301         .readv          = generic_file_readv,    /* Read from file. */
2302 #ifdef NTFS_RW
2303         .write          = ntfs_file_write,       /* Write to file. */
2304         .aio_write      = ntfs_file_aio_write,   /* Async write to file. */
2305         .writev         = ntfs_file_writev,      /* Write to file. */
2306         /*.release      = ,*/                    /* Last file is closed.  See
2307                                                     fs/ext2/file.c::
2308                                                     ext2_release_file() for
2309                                                     how to use this to discard
2310                                                     preallocated space for
2311                                                     write opened files. */
2312         .fsync          = ntfs_file_fsync,       /* Sync a file to disk. */
2313         /*.aio_fsync    = ,*/                    /* Sync all outstanding async
2314                                                     i/o operations on a
2315                                                     kiocb. */
2316 #endif /* NTFS_RW */
2317         /*.ioctl        = ,*/                    /* Perform function on the
2318                                                     mounted filesystem. */
2319         .mmap           = generic_file_mmap,     /* Mmap file. */
2320         .open           = ntfs_file_open,        /* Open file. */
2321         .sendfile       = generic_file_sendfile, /* Zero-copy data send with
2322                                                     the data source being on
2323                                                     the ntfs partition.  We do
2324                                                     not need to care about the
2325                                                     data destination. */
2326         /*.sendpage     = ,*/                    /* Zero-copy data send with
2327                                                     the data destination being
2328                                                     on the ntfs partition.  We
2329                                                     do not need to care about
2330                                                     the data source. */
2331 };
2332
2333 struct inode_operations ntfs_file_inode_ops = {
2334 #ifdef NTFS_RW
2335         .truncate       = ntfs_truncate_vfs,
2336         .setattr        = ntfs_setattr,
2337 #endif /* NTFS_RW */
2338 };
2339
2340 struct file_operations ntfs_empty_file_ops = {};
2341
2342 struct inode_operations ntfs_empty_inode_ops = {};