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