[PATCH] Vectorize aio_read/aio_write fileop methods
[linux-2.6.git] / fs / ntfs / file.c
1 /*
2  * file.c - NTFS kernel file operations.  Part of the Linux-NTFS project.
3  *
4  * Copyright (c) 2001-2006 Anton Altaparmakov
5  *
6  * This program/include file is free software; you can redistribute it and/or
7  * modify it under the terms of the GNU General Public License as published
8  * by the Free Software Foundation; either version 2 of the License, or
9  * (at your option) any later version.
10  *
11  * This program/include file is distributed in the hope that it will be
12  * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
13  * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
14  * GNU General Public License for more details.
15  *
16  * You should have received a copy of the GNU General Public License
17  * along with this program (in the main directory of the Linux-NTFS
18  * distribution in the file COPYING); if not, write to the Free Software
19  * Foundation,Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
20  */
21
22 #include <linux/buffer_head.h>
23 #include <linux/pagemap.h>
24 #include <linux/pagevec.h>
25 #include <linux/sched.h>
26 #include <linux/swap.h>
27 #include <linux/uio.h>
28 #include <linux/writeback.h>
29
30 #include <asm/page.h>
31 #include <asm/uaccess.h>
32
33 #include "attrib.h"
34 #include "bitmap.h"
35 #include "inode.h"
36 #include "debug.h"
37 #include "lcnalloc.h"
38 #include "malloc.h"
39 #include "mft.h"
40 #include "ntfs.h"
41
42 /**
43  * ntfs_file_open - called when an inode is about to be opened
44  * @vi:         inode to be opened
45  * @filp:       file structure describing the inode
46  *
47  * Limit file size to the page cache limit on architectures where unsigned long
48  * is 32-bits. This is the most we can do for now without overflowing the page
49  * cache page index. Doing it this way means we don't run into problems because
50  * of existing too large files. It would be better to allow the user to read
51  * the beginning of the file but I doubt very much anyone is going to hit this
52  * check on a 32-bit architecture, so there is no point in adding the extra
53  * complexity required to support this.
54  *
55  * On 64-bit architectures, the check is hopefully optimized away by the
56  * compiler.
57  *
58  * After the check passes, just call generic_file_open() to do its work.
59  */
60 static int ntfs_file_open(struct inode *vi, struct file *filp)
61 {
62         if (sizeof(unsigned long) < 8) {
63                 if (i_size_read(vi) > MAX_LFS_FILESIZE)
64                         return -EFBIG;
65         }
66         return generic_file_open(vi, filp);
67 }
68
69 #ifdef NTFS_RW
70
71 /**
72  * ntfs_attr_extend_initialized - extend the initialized size of an attribute
73  * @ni:                 ntfs inode of the attribute to extend
74  * @new_init_size:      requested new initialized size in bytes
75  * @cached_page:        store any allocated but unused page here
76  * @lru_pvec:           lru-buffering pagevec of the caller
77  *
78  * Extend the initialized size of an attribute described by the ntfs inode @ni
79  * to @new_init_size bytes.  This involves zeroing any non-sparse space between
80  * the old initialized size and @new_init_size both in the page cache and on
81  * disk (if relevant complete pages are already uptodate in the page cache then
82  * these are simply marked dirty).
83  *
84  * As a side-effect, the file size (vfs inode->i_size) may be incremented as,
85  * in the resident attribute case, it is tied to the initialized size and, in
86  * the non-resident attribute case, it may not fall below the initialized size.
87  *
88  * Note that if the attribute is resident, we do not need to touch the page
89  * cache at all.  This is because if the page cache page is not uptodate we
90  * bring it uptodate later, when doing the write to the mft record since we
91  * then already have the page mapped.  And if the page is uptodate, the
92  * non-initialized region will already have been zeroed when the page was
93  * brought uptodate and the region may in fact already have been overwritten
94  * with new data via mmap() based writes, so we cannot just zero it.  And since
95  * POSIX specifies that the behaviour of resizing a file whilst it is mmap()ped
96  * is unspecified, we choose not to do zeroing and thus we do not need to touch
97  * the page at all.  For a more detailed explanation see ntfs_truncate() in
98  * fs/ntfs/inode.c.
99  *
100  * @cached_page and @lru_pvec are just optimizations for dealing with multiple
101  * pages.
102  *
103  * Return 0 on success and -errno on error.  In the case that an error is
104  * encountered it is possible that the initialized size will already have been
105  * incremented some way towards @new_init_size but it is guaranteed that if
106  * this is the case, the necessary zeroing will also have happened and that all
107  * metadata is self-consistent.
108  *
109  * Locking: i_mutex on the vfs inode corrseponsind to the ntfs inode @ni must be
110  *          held by the caller.
111  */
112 static int ntfs_attr_extend_initialized(ntfs_inode *ni, const s64 new_init_size,
113                 struct page **cached_page, struct pagevec *lru_pvec)
114 {
115         s64 old_init_size;
116         loff_t old_i_size;
117         pgoff_t index, end_index;
118         unsigned long flags;
119         struct inode *vi = VFS_I(ni);
120         ntfs_inode *base_ni;
121         MFT_RECORD *m = NULL;
122         ATTR_RECORD *a;
123         ntfs_attr_search_ctx *ctx = NULL;
124         struct address_space *mapping;
125         struct page *page = NULL;
126         u8 *kattr;
127         int err;
128         u32 attr_len;
129
130         read_lock_irqsave(&ni->size_lock, flags);
131         old_init_size = ni->initialized_size;
132         old_i_size = i_size_read(vi);
133         BUG_ON(new_init_size > ni->allocated_size);
134         read_unlock_irqrestore(&ni->size_lock, flags);
135         ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, "
136                         "old_initialized_size 0x%llx, "
137                         "new_initialized_size 0x%llx, i_size 0x%llx.",
138                         vi->i_ino, (unsigned)le32_to_cpu(ni->type),
139                         (unsigned long long)old_init_size,
140                         (unsigned long long)new_init_size, old_i_size);
141         if (!NInoAttr(ni))
142                 base_ni = ni;
143         else
144                 base_ni = ni->ext.base_ntfs_ino;
145         /* Use goto to reduce indentation and we need the label below anyway. */
146         if (NInoNonResident(ni))
147                 goto do_non_resident_extend;
148         BUG_ON(old_init_size != old_i_size);
149         m = map_mft_record(base_ni);
150         if (IS_ERR(m)) {
151                 err = PTR_ERR(m);
152                 m = NULL;
153                 goto err_out;
154         }
155         ctx = ntfs_attr_get_search_ctx(base_ni, m);
156         if (unlikely(!ctx)) {
157                 err = -ENOMEM;
158                 goto err_out;
159         }
160         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
161                         CASE_SENSITIVE, 0, NULL, 0, ctx);
162         if (unlikely(err)) {
163                 if (err == -ENOENT)
164                         err = -EIO;
165                 goto err_out;
166         }
167         m = ctx->mrec;
168         a = ctx->attr;
169         BUG_ON(a->non_resident);
170         /* The total length of the attribute value. */
171         attr_len = le32_to_cpu(a->data.resident.value_length);
172         BUG_ON(old_i_size != (loff_t)attr_len);
173         /*
174          * Do the zeroing in the mft record and update the attribute size in
175          * the mft record.
176          */
177         kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
178         memset(kattr + attr_len, 0, new_init_size - attr_len);
179         a->data.resident.value_length = cpu_to_le32((u32)new_init_size);
180         /* Finally, update the sizes in the vfs and ntfs inodes. */
181         write_lock_irqsave(&ni->size_lock, flags);
182         i_size_write(vi, new_init_size);
183         ni->initialized_size = new_init_size;
184         write_unlock_irqrestore(&ni->size_lock, flags);
185         goto done;
186 do_non_resident_extend:
187         /*
188          * If the new initialized size @new_init_size exceeds the current file
189          * size (vfs inode->i_size), we need to extend the file size to the
190          * new initialized size.
191          */
192         if (new_init_size > old_i_size) {
193                 m = map_mft_record(base_ni);
194                 if (IS_ERR(m)) {
195                         err = PTR_ERR(m);
196                         m = NULL;
197                         goto err_out;
198                 }
199                 ctx = ntfs_attr_get_search_ctx(base_ni, m);
200                 if (unlikely(!ctx)) {
201                         err = -ENOMEM;
202                         goto err_out;
203                 }
204                 err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
205                                 CASE_SENSITIVE, 0, NULL, 0, ctx);
206                 if (unlikely(err)) {
207                         if (err == -ENOENT)
208                                 err = -EIO;
209                         goto err_out;
210                 }
211                 m = ctx->mrec;
212                 a = ctx->attr;
213                 BUG_ON(!a->non_resident);
214                 BUG_ON(old_i_size != (loff_t)
215                                 sle64_to_cpu(a->data.non_resident.data_size));
216                 a->data.non_resident.data_size = cpu_to_sle64(new_init_size);
217                 flush_dcache_mft_record_page(ctx->ntfs_ino);
218                 mark_mft_record_dirty(ctx->ntfs_ino);
219                 /* Update the file size in the vfs inode. */
220                 i_size_write(vi, new_init_size);
221                 ntfs_attr_put_search_ctx(ctx);
222                 ctx = NULL;
223                 unmap_mft_record(base_ni);
224                 m = NULL;
225         }
226         mapping = vi->i_mapping;
227         index = old_init_size >> PAGE_CACHE_SHIFT;
228         end_index = (new_init_size + PAGE_CACHE_SIZE - 1) >> PAGE_CACHE_SHIFT;
229         do {
230                 /*
231                  * Read the page.  If the page is not present, this will zero
232                  * the uninitialized regions for us.
233                  */
234                 page = read_mapping_page(mapping, index, NULL);
235                 if (IS_ERR(page)) {
236                         err = PTR_ERR(page);
237                         goto init_err_out;
238                 }
239                 wait_on_page_locked(page);
240                 if (unlikely(!PageUptodate(page) || PageError(page))) {
241                         page_cache_release(page);
242                         err = -EIO;
243                         goto init_err_out;
244                 }
245                 /*
246                  * Update the initialized size in the ntfs inode.  This is
247                  * enough to make ntfs_writepage() work.
248                  */
249                 write_lock_irqsave(&ni->size_lock, flags);
250                 ni->initialized_size = (s64)(index + 1) << PAGE_CACHE_SHIFT;
251                 if (ni->initialized_size > new_init_size)
252                         ni->initialized_size = new_init_size;
253                 write_unlock_irqrestore(&ni->size_lock, flags);
254                 /* Set the page dirty so it gets written out. */
255                 set_page_dirty(page);
256                 page_cache_release(page);
257                 /*
258                  * Play nice with the vm and the rest of the system.  This is
259                  * very much needed as we can potentially be modifying the
260                  * initialised size from a very small value to a really huge
261                  * value, e.g.
262                  *      f = open(somefile, O_TRUNC);
263                  *      truncate(f, 10GiB);
264                  *      seek(f, 10GiB);
265                  *      write(f, 1);
266                  * And this would mean we would be marking dirty hundreds of
267                  * thousands of pages or as in the above example more than
268                  * two and a half million pages!
269                  *
270                  * TODO: For sparse pages could optimize this workload by using
271                  * the FsMisc / MiscFs page bit as a "PageIsSparse" bit.  This
272                  * would be set in readpage for sparse pages and here we would
273                  * not need to mark dirty any pages which have this bit set.
274                  * The only caveat is that we have to clear the bit everywhere
275                  * where we allocate any clusters that lie in the page or that
276                  * contain the page.
277                  *
278                  * TODO: An even greater optimization would be for us to only
279                  * call readpage() on pages which are not in sparse regions as
280                  * determined from the runlist.  This would greatly reduce the
281                  * number of pages we read and make dirty in the case of sparse
282                  * files.
283                  */
284                 balance_dirty_pages_ratelimited(mapping);
285                 cond_resched();
286         } while (++index < end_index);
287         read_lock_irqsave(&ni->size_lock, flags);
288         BUG_ON(ni->initialized_size != new_init_size);
289         read_unlock_irqrestore(&ni->size_lock, flags);
290         /* Now bring in sync the initialized_size in the mft record. */
291         m = map_mft_record(base_ni);
292         if (IS_ERR(m)) {
293                 err = PTR_ERR(m);
294                 m = NULL;
295                 goto init_err_out;
296         }
297         ctx = ntfs_attr_get_search_ctx(base_ni, m);
298         if (unlikely(!ctx)) {
299                 err = -ENOMEM;
300                 goto init_err_out;
301         }
302         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
303                         CASE_SENSITIVE, 0, NULL, 0, ctx);
304         if (unlikely(err)) {
305                 if (err == -ENOENT)
306                         err = -EIO;
307                 goto init_err_out;
308         }
309         m = ctx->mrec;
310         a = ctx->attr;
311         BUG_ON(!a->non_resident);
312         a->data.non_resident.initialized_size = cpu_to_sle64(new_init_size);
313 done:
314         flush_dcache_mft_record_page(ctx->ntfs_ino);
315         mark_mft_record_dirty(ctx->ntfs_ino);
316         if (ctx)
317                 ntfs_attr_put_search_ctx(ctx);
318         if (m)
319                 unmap_mft_record(base_ni);
320         ntfs_debug("Done, initialized_size 0x%llx, i_size 0x%llx.",
321                         (unsigned long long)new_init_size, i_size_read(vi));
322         return 0;
323 init_err_out:
324         write_lock_irqsave(&ni->size_lock, flags);
325         ni->initialized_size = old_init_size;
326         write_unlock_irqrestore(&ni->size_lock, flags);
327 err_out:
328         if (ctx)
329                 ntfs_attr_put_search_ctx(ctx);
330         if (m)
331                 unmap_mft_record(base_ni);
332         ntfs_debug("Failed.  Returning error code %i.", err);
333         return err;
334 }
335
336 /**
337  * ntfs_fault_in_pages_readable -
338  *
339  * Fault a number of userspace pages into pagetables.
340  *
341  * Unlike include/linux/pagemap.h::fault_in_pages_readable(), this one copes
342  * with more than two userspace pages as well as handling the single page case
343  * elegantly.
344  *
345  * If you find this difficult to understand, then think of the while loop being
346  * the following code, except that we do without the integer variable ret:
347  *
348  *      do {
349  *              ret = __get_user(c, uaddr);
350  *              uaddr += PAGE_SIZE;
351  *      } while (!ret && uaddr < end);
352  *
353  * Note, the final __get_user() may well run out-of-bounds of the user buffer,
354  * but _not_ out-of-bounds of the page the user buffer belongs to, and since
355  * this is only a read and not a write, and since it is still in the same page,
356  * it should not matter and this makes the code much simpler.
357  */
358 static inline void ntfs_fault_in_pages_readable(const char __user *uaddr,
359                 int bytes)
360 {
361         const char __user *end;
362         volatile char c;
363
364         /* Set @end to the first byte outside the last page we care about. */
365         end = (const char __user*)PAGE_ALIGN((ptrdiff_t __user)uaddr + bytes);
366
367         while (!__get_user(c, uaddr) && (uaddr += PAGE_SIZE, uaddr < end))
368                 ;
369 }
370
371 /**
372  * ntfs_fault_in_pages_readable_iovec -
373  *
374  * Same as ntfs_fault_in_pages_readable() but operates on an array of iovecs.
375  */
376 static inline void ntfs_fault_in_pages_readable_iovec(const struct iovec *iov,
377                 size_t iov_ofs, int bytes)
378 {
379         do {
380                 const char __user *buf;
381                 unsigned len;
382
383                 buf = iov->iov_base + iov_ofs;
384                 len = iov->iov_len - iov_ofs;
385                 if (len > bytes)
386                         len = bytes;
387                 ntfs_fault_in_pages_readable(buf, len);
388                 bytes -= len;
389                 iov++;
390                 iov_ofs = 0;
391         } while (bytes);
392 }
393
394 /**
395  * __ntfs_grab_cache_pages - obtain a number of locked pages
396  * @mapping:    address space mapping from which to obtain page cache pages
397  * @index:      starting index in @mapping at which to begin obtaining pages
398  * @nr_pages:   number of page cache pages to obtain
399  * @pages:      array of pages in which to return the obtained page cache pages
400  * @cached_page: allocated but as yet unused page
401  * @lru_pvec:   lru-buffering pagevec of caller
402  *
403  * Obtain @nr_pages locked page cache pages from the mapping @maping and
404  * starting at index @index.
405  *
406  * If a page is newly created, increment its refcount and add it to the
407  * caller's lru-buffering pagevec @lru_pvec.
408  *
409  * This is the same as mm/filemap.c::__grab_cache_page(), except that @nr_pages
410  * are obtained at once instead of just one page and that 0 is returned on
411  * success and -errno on error.
412  *
413  * Note, the page locks are obtained in ascending page index order.
414  */
415 static inline int __ntfs_grab_cache_pages(struct address_space *mapping,
416                 pgoff_t index, const unsigned nr_pages, struct page **pages,
417                 struct page **cached_page, struct pagevec *lru_pvec)
418 {
419         int err, nr;
420
421         BUG_ON(!nr_pages);
422         err = nr = 0;
423         do {
424                 pages[nr] = find_lock_page(mapping, index);
425                 if (!pages[nr]) {
426                         if (!*cached_page) {
427                                 *cached_page = page_cache_alloc(mapping);
428                                 if (unlikely(!*cached_page)) {
429                                         err = -ENOMEM;
430                                         goto err_out;
431                                 }
432                         }
433                         err = add_to_page_cache(*cached_page, mapping, index,
434                                         GFP_KERNEL);
435                         if (unlikely(err)) {
436                                 if (err == -EEXIST)
437                                         continue;
438                                 goto err_out;
439                         }
440                         pages[nr] = *cached_page;
441                         page_cache_get(*cached_page);
442                         if (unlikely(!pagevec_add(lru_pvec, *cached_page)))
443                                 __pagevec_lru_add(lru_pvec);
444                         *cached_page = NULL;
445                 }
446                 index++;
447                 nr++;
448         } while (nr < nr_pages);
449 out:
450         return err;
451 err_out:
452         while (nr > 0) {
453                 unlock_page(pages[--nr]);
454                 page_cache_release(pages[nr]);
455         }
456         goto out;
457 }
458
459 static inline int ntfs_submit_bh_for_read(struct buffer_head *bh)
460 {
461         lock_buffer(bh);
462         get_bh(bh);
463         bh->b_end_io = end_buffer_read_sync;
464         return submit_bh(READ, bh);
465 }
466
467 /**
468  * ntfs_prepare_pages_for_non_resident_write - prepare pages for receiving data
469  * @pages:      array of destination pages
470  * @nr_pages:   number of pages in @pages
471  * @pos:        byte position in file at which the write begins
472  * @bytes:      number of bytes to be written
473  *
474  * This is called for non-resident attributes from ntfs_file_buffered_write()
475  * with i_mutex held on the inode (@pages[0]->mapping->host).  There are
476  * @nr_pages pages in @pages which are locked but not kmap()ped.  The source
477  * data has not yet been copied into the @pages.
478  * 
479  * Need to fill any holes with actual clusters, allocate buffers if necessary,
480  * ensure all the buffers are mapped, and bring uptodate any buffers that are
481  * only partially being written to.
482  *
483  * If @nr_pages is greater than one, we are guaranteed that the cluster size is
484  * greater than PAGE_CACHE_SIZE, that all pages in @pages are entirely inside
485  * the same cluster and that they are the entirety of that cluster, and that
486  * the cluster is sparse, i.e. we need to allocate a cluster to fill the hole.
487  *
488  * i_size is not to be modified yet.
489  *
490  * Return 0 on success or -errno on error.
491  */
492 static int ntfs_prepare_pages_for_non_resident_write(struct page **pages,
493                 unsigned nr_pages, s64 pos, size_t bytes)
494 {
495         VCN vcn, highest_vcn = 0, cpos, cend, bh_cpos, bh_cend;
496         LCN lcn;
497         s64 bh_pos, vcn_len, end, initialized_size;
498         sector_t lcn_block;
499         struct page *page;
500         struct inode *vi;
501         ntfs_inode *ni, *base_ni = NULL;
502         ntfs_volume *vol;
503         runlist_element *rl, *rl2;
504         struct buffer_head *bh, *head, *wait[2], **wait_bh = wait;
505         ntfs_attr_search_ctx *ctx = NULL;
506         MFT_RECORD *m = NULL;
507         ATTR_RECORD *a = NULL;
508         unsigned long flags;
509         u32 attr_rec_len = 0;
510         unsigned blocksize, u;
511         int err, mp_size;
512         bool rl_write_locked, was_hole, is_retry;
513         unsigned char blocksize_bits;
514         struct {
515                 u8 runlist_merged:1;
516                 u8 mft_attr_mapped:1;
517                 u8 mp_rebuilt:1;
518                 u8 attr_switched:1;
519         } status = { 0, 0, 0, 0 };
520
521         BUG_ON(!nr_pages);
522         BUG_ON(!pages);
523         BUG_ON(!*pages);
524         vi = pages[0]->mapping->host;
525         ni = NTFS_I(vi);
526         vol = ni->vol;
527         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, start page "
528                         "index 0x%lx, nr_pages 0x%x, pos 0x%llx, bytes 0x%zx.",
529                         vi->i_ino, ni->type, pages[0]->index, nr_pages,
530                         (long long)pos, bytes);
531         blocksize = vol->sb->s_blocksize;
532         blocksize_bits = vol->sb->s_blocksize_bits;
533         u = 0;
534         do {
535                 struct page *page = pages[u];
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                                         u8 *kaddr = kmap_atomic(page, KM_USER0);
611                                         memset(kaddr + bh_offset(bh), 0,
612                                                         blocksize);
613                                         kunmap_atomic(kaddr, KM_USER0);
614                                         flush_dcache_page(page);
615                                         set_buffer_uptodate(bh);
616                                 }
617                         }
618                         continue;
619                 }
620                 /* Unmapped buffer.  Need to map it. */
621                 bh->b_bdev = vol->sb->s_bdev;
622                 /*
623                  * If the current buffer is in the same clusters as the map
624                  * cache, there is no need to check the runlist again.  The
625                  * map cache is made up of @vcn, which is the first cached file
626                  * cluster, @vcn_len which is the number of cached file
627                  * clusters, @lcn is the device cluster corresponding to @vcn,
628                  * and @lcn_block is the block number corresponding to @lcn.
629                  */
630                 cdelta = bh_cpos - vcn;
631                 if (likely(!cdelta || (cdelta > 0 && cdelta < vcn_len))) {
632 map_buffer_cached:
633                         BUG_ON(lcn < 0);
634                         bh->b_blocknr = lcn_block +
635                                         (cdelta << (vol->cluster_size_bits -
636                                         blocksize_bits)) +
637                                         (bh_cofs >> blocksize_bits);
638                         set_buffer_mapped(bh);
639                         /*
640                          * If the page is uptodate so is the buffer.  If the
641                          * buffer is fully outside the write, we ignore it if
642                          * it was already allocated and we mark it dirty so it
643                          * gets written out if we allocated it.  On the other
644                          * hand, if we allocated the buffer but we are not
645                          * marking it dirty we set buffer_new so we can do
646                          * error recovery.
647                          */
648                         if (PageUptodate(page)) {
649                                 if (!buffer_uptodate(bh))
650                                         set_buffer_uptodate(bh);
651                                 if (unlikely(was_hole)) {
652                                         /* We allocated the buffer. */
653                                         unmap_underlying_metadata(bh->b_bdev,
654                                                         bh->b_blocknr);
655                                         if (bh_end <= pos || bh_pos >= end)
656                                                 mark_buffer_dirty(bh);
657                                         else
658                                                 set_buffer_new(bh);
659                                 }
660                                 continue;
661                         }
662                         /* Page is _not_ uptodate. */
663                         if (likely(!was_hole)) {
664                                 /*
665                                  * Buffer was already allocated.  If it is not
666                                  * uptodate and is only partially being written
667                                  * to, we need to read it in before the write,
668                                  * i.e. now.
669                                  */
670                                 if (!buffer_uptodate(bh) && bh_pos < end &&
671                                                 bh_end > pos &&
672                                                 (bh_pos < pos ||
673                                                 bh_end > end)) {
674                                         /*
675                                          * If the buffer is fully or partially
676                                          * within the initialized size, do an
677                                          * actual read.  Otherwise, simply zero
678                                          * the buffer.
679                                          */
680                                         read_lock_irqsave(&ni->size_lock,
681                                                         flags);
682                                         initialized_size = ni->initialized_size;
683                                         read_unlock_irqrestore(&ni->size_lock,
684                                                         flags);
685                                         if (bh_pos < initialized_size) {
686                                                 ntfs_submit_bh_for_read(bh);
687                                                 *wait_bh++ = bh;
688                                         } else {
689                                                 u8 *kaddr = kmap_atomic(page,
690                                                                 KM_USER0);
691                                                 memset(kaddr + bh_offset(bh),
692                                                                 0, blocksize);
693                                                 kunmap_atomic(kaddr, KM_USER0);
694                                                 flush_dcache_page(page);
695                                                 set_buffer_uptodate(bh);
696                                         }
697                                 }
698                                 continue;
699                         }
700                         /* We allocated the buffer. */
701                         unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
702                         /*
703                          * If the buffer is fully outside the write, zero it,
704                          * set it uptodate, and mark it dirty so it gets
705                          * written out.  If it is partially being written to,
706                          * zero region surrounding the write but leave it to
707                          * commit write to do anything else.  Finally, if the
708                          * buffer is fully being overwritten, do nothing.
709                          */
710                         if (bh_end <= pos || bh_pos >= end) {
711                                 if (!buffer_uptodate(bh)) {
712                                         u8 *kaddr = kmap_atomic(page, KM_USER0);
713                                         memset(kaddr + bh_offset(bh), 0,
714                                                         blocksize);
715                                         kunmap_atomic(kaddr, KM_USER0);
716                                         flush_dcache_page(page);
717                                         set_buffer_uptodate(bh);
718                                 }
719                                 mark_buffer_dirty(bh);
720                                 continue;
721                         }
722                         set_buffer_new(bh);
723                         if (!buffer_uptodate(bh) &&
724                                         (bh_pos < pos || bh_end > end)) {
725                                 u8 *kaddr;
726                                 unsigned pofs;
727                                         
728                                 kaddr = kmap_atomic(page, KM_USER0);
729                                 if (bh_pos < pos) {
730                                         pofs = bh_pos & ~PAGE_CACHE_MASK;
731                                         memset(kaddr + pofs, 0, pos - bh_pos);
732                                 }
733                                 if (bh_end > end) {
734                                         pofs = end & ~PAGE_CACHE_MASK;
735                                         memset(kaddr + pofs, 0, bh_end - end);
736                                 }
737                                 kunmap_atomic(kaddr, KM_USER0);
738                                 flush_dcache_page(page);
739                         }
740                         continue;
741                 }
742                 /*
743                  * Slow path: this is the first buffer in the cluster.  If it
744                  * is outside allocated size and is not uptodate, zero it and
745                  * set it uptodate.
746                  */
747                 read_lock_irqsave(&ni->size_lock, flags);
748                 initialized_size = ni->allocated_size;
749                 read_unlock_irqrestore(&ni->size_lock, flags);
750                 if (bh_pos > initialized_size) {
751                         if (PageUptodate(page)) {
752                                 if (!buffer_uptodate(bh))
753                                         set_buffer_uptodate(bh);
754                         } else if (!buffer_uptodate(bh)) {
755                                 u8 *kaddr = kmap_atomic(page, KM_USER0);
756                                 memset(kaddr + bh_offset(bh), 0, blocksize);
757                                 kunmap_atomic(kaddr, KM_USER0);
758                                 flush_dcache_page(page);
759                                 set_buffer_uptodate(bh);
760                         }
761                         continue;
762                 }
763                 is_retry = false;
764                 if (!rl) {
765                         down_read(&ni->runlist.lock);
766 retry_remap:
767                         rl = ni->runlist.rl;
768                 }
769                 if (likely(rl != NULL)) {
770                         /* Seek to element containing target cluster. */
771                         while (rl->length && rl[1].vcn <= bh_cpos)
772                                 rl++;
773                         lcn = ntfs_rl_vcn_to_lcn(rl, bh_cpos);
774                         if (likely(lcn >= 0)) {
775                                 /*
776                                  * Successful remap, setup the map cache and
777                                  * use that to deal with the buffer.
778                                  */
779                                 was_hole = false;
780                                 vcn = bh_cpos;
781                                 vcn_len = rl[1].vcn - vcn;
782                                 lcn_block = lcn << (vol->cluster_size_bits -
783                                                 blocksize_bits);
784                                 cdelta = 0;
785                                 /*
786                                  * If the number of remaining clusters touched
787                                  * by the write is smaller or equal to the
788                                  * number of cached clusters, unlock the
789                                  * runlist as the map cache will be used from
790                                  * now on.
791                                  */
792                                 if (likely(vcn + vcn_len >= cend)) {
793                                         if (rl_write_locked) {
794                                                 up_write(&ni->runlist.lock);
795                                                 rl_write_locked = false;
796                                         } else
797                                                 up_read(&ni->runlist.lock);
798                                         rl = NULL;
799                                 }
800                                 goto map_buffer_cached;
801                         }
802                 } else
803                         lcn = LCN_RL_NOT_MAPPED;
804                 /*
805                  * If it is not a hole and not out of bounds, the runlist is
806                  * probably unmapped so try to map it now.
807                  */
808                 if (unlikely(lcn != LCN_HOLE && lcn != LCN_ENOENT)) {
809                         if (likely(!is_retry && lcn == LCN_RL_NOT_MAPPED)) {
810                                 /* Attempt to map runlist. */
811                                 if (!rl_write_locked) {
812                                         /*
813                                          * We need the runlist locked for
814                                          * writing, so if it is locked for
815                                          * reading relock it now and retry in
816                                          * case it changed whilst we dropped
817                                          * the lock.
818                                          */
819                                         up_read(&ni->runlist.lock);
820                                         down_write(&ni->runlist.lock);
821                                         rl_write_locked = true;
822                                         goto retry_remap;
823                                 }
824                                 err = ntfs_map_runlist_nolock(ni, bh_cpos,
825                                                 NULL);
826                                 if (likely(!err)) {
827                                         is_retry = true;
828                                         goto retry_remap;
829                                 }
830                                 /*
831                                  * If @vcn is out of bounds, pretend @lcn is
832                                  * LCN_ENOENT.  As long as the buffer is out
833                                  * of bounds this will work fine.
834                                  */
835                                 if (err == -ENOENT) {
836                                         lcn = LCN_ENOENT;
837                                         err = 0;
838                                         goto rl_not_mapped_enoent;
839                                 }
840                         } else
841                                 err = -EIO;
842                         /* Failed to map the buffer, even after retrying. */
843                         bh->b_blocknr = -1;
844                         ntfs_error(vol->sb, "Failed to write to inode 0x%lx, "
845                                         "attribute type 0x%x, vcn 0x%llx, "
846                                         "vcn offset 0x%x, because its "
847                                         "location on disk could not be "
848                                         "determined%s (error code %i).",
849                                         ni->mft_no, ni->type,
850                                         (unsigned long long)bh_cpos,
851                                         (unsigned)bh_pos &
852                                         vol->cluster_size_mask,
853                                         is_retry ? " even after retrying" : "",
854                                         err);
855                         break;
856                 }
857 rl_not_mapped_enoent:
858                 /*
859                  * The buffer is in a hole or out of bounds.  We need to fill
860                  * the hole, unless the buffer is in a cluster which is not
861                  * touched by the write, in which case we just leave the buffer
862                  * unmapped.  This can only happen when the cluster size is
863                  * less than the page cache size.
864                  */
865                 if (unlikely(vol->cluster_size < PAGE_CACHE_SIZE)) {
866                         bh_cend = (bh_end + vol->cluster_size - 1) >>
867                                         vol->cluster_size_bits;
868                         if ((bh_cend <= cpos || bh_cpos >= cend)) {
869                                 bh->b_blocknr = -1;
870                                 /*
871                                  * If the buffer is uptodate we skip it.  If it
872                                  * is not but the page is uptodate, we can set
873                                  * the buffer uptodate.  If the page is not
874                                  * uptodate, we can clear the buffer and set it
875                                  * uptodate.  Whether this is worthwhile is
876                                  * debatable and this could be removed.
877                                  */
878                                 if (PageUptodate(page)) {
879                                         if (!buffer_uptodate(bh))
880                                                 set_buffer_uptodate(bh);
881                                 } else if (!buffer_uptodate(bh)) {
882                                         u8 *kaddr = kmap_atomic(page, KM_USER0);
883                                         memset(kaddr + bh_offset(bh), 0,
884                                                         blocksize);
885                                         kunmap_atomic(kaddr, KM_USER0);
886                                         flush_dcache_page(page);
887                                         set_buffer_uptodate(bh);
888                                 }
889                                 continue;
890                         }
891                 }
892                 /*
893                  * Out of bounds buffer is invalid if it was not really out of
894                  * bounds.
895                  */
896                 BUG_ON(lcn != LCN_HOLE);
897                 /*
898                  * We need the runlist locked for writing, so if it is locked
899                  * for reading relock it now and retry in case it changed
900                  * whilst we dropped the lock.
901                  */
902                 BUG_ON(!rl);
903                 if (!rl_write_locked) {
904                         up_read(&ni->runlist.lock);
905                         down_write(&ni->runlist.lock);
906                         rl_write_locked = true;
907                         goto retry_remap;
908                 }
909                 /* Find the previous last allocated cluster. */
910                 BUG_ON(rl->lcn != LCN_HOLE);
911                 lcn = -1;
912                 rl2 = rl;
913                 while (--rl2 >= ni->runlist.rl) {
914                         if (rl2->lcn >= 0) {
915                                 lcn = rl2->lcn + rl2->length;
916                                 break;
917                         }
918                 }
919                 rl2 = ntfs_cluster_alloc(vol, bh_cpos, 1, lcn, DATA_ZONE,
920                                 false);
921                 if (IS_ERR(rl2)) {
922                         err = PTR_ERR(rl2);
923                         ntfs_debug("Failed to allocate cluster, error code %i.",
924                                         err);
925                         break;
926                 }
927                 lcn = rl2->lcn;
928                 rl = ntfs_runlists_merge(ni->runlist.rl, rl2);
929                 if (IS_ERR(rl)) {
930                         err = PTR_ERR(rl);
931                         if (err != -ENOMEM)
932                                 err = -EIO;
933                         if (ntfs_cluster_free_from_rl(vol, rl2)) {
934                                 ntfs_error(vol->sb, "Failed to release "
935                                                 "allocated cluster in error "
936                                                 "code path.  Run chkdsk to "
937                                                 "recover the lost cluster.");
938                                 NVolSetErrors(vol);
939                         }
940                         ntfs_free(rl2);
941                         break;
942                 }
943                 ni->runlist.rl = rl;
944                 status.runlist_merged = 1;
945                 ntfs_debug("Allocated cluster, lcn 0x%llx.",
946                                 (unsigned long long)lcn);
947                 /* Map and lock the mft record and get the attribute record. */
948                 if (!NInoAttr(ni))
949                         base_ni = ni;
950                 else
951                         base_ni = ni->ext.base_ntfs_ino;
952                 m = map_mft_record(base_ni);
953                 if (IS_ERR(m)) {
954                         err = PTR_ERR(m);
955                         break;
956                 }
957                 ctx = ntfs_attr_get_search_ctx(base_ni, m);
958                 if (unlikely(!ctx)) {
959                         err = -ENOMEM;
960                         unmap_mft_record(base_ni);
961                         break;
962                 }
963                 status.mft_attr_mapped = 1;
964                 err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
965                                 CASE_SENSITIVE, bh_cpos, NULL, 0, ctx);
966                 if (unlikely(err)) {
967                         if (err == -ENOENT)
968                                 err = -EIO;
969                         break;
970                 }
971                 m = ctx->mrec;
972                 a = ctx->attr;
973                 /*
974                  * Find the runlist element with which the attribute extent
975                  * starts.  Note, we cannot use the _attr_ version because we
976                  * have mapped the mft record.  That is ok because we know the
977                  * runlist fragment must be mapped already to have ever gotten
978                  * here, so we can just use the _rl_ version.
979                  */
980                 vcn = sle64_to_cpu(a->data.non_resident.lowest_vcn);
981                 rl2 = ntfs_rl_find_vcn_nolock(rl, vcn);
982                 BUG_ON(!rl2);
983                 BUG_ON(!rl2->length);
984                 BUG_ON(rl2->lcn < LCN_HOLE);
985                 highest_vcn = sle64_to_cpu(a->data.non_resident.highest_vcn);
986                 /*
987                  * If @highest_vcn is zero, calculate the real highest_vcn
988                  * (which can really be zero).
989                  */
990                 if (!highest_vcn)
991                         highest_vcn = (sle64_to_cpu(
992                                         a->data.non_resident.allocated_size) >>
993                                         vol->cluster_size_bits) - 1;
994                 /*
995                  * Determine the size of the mapping pairs array for the new
996                  * extent, i.e. the old extent with the hole filled.
997                  */
998                 mp_size = ntfs_get_size_for_mapping_pairs(vol, rl2, vcn,
999                                 highest_vcn);
1000                 if (unlikely(mp_size <= 0)) {
1001                         if (!(err = mp_size))
1002                                 err = -EIO;
1003                         ntfs_debug("Failed to get size for mapping pairs "
1004                                         "array, error code %i.", err);
1005                         break;
1006                 }
1007                 /*
1008                  * Resize the attribute record to fit the new mapping pairs
1009                  * array.
1010                  */
1011                 attr_rec_len = le32_to_cpu(a->length);
1012                 err = ntfs_attr_record_resize(m, a, mp_size + le16_to_cpu(
1013                                 a->data.non_resident.mapping_pairs_offset));
1014                 if (unlikely(err)) {
1015                         BUG_ON(err != -ENOSPC);
1016                         // TODO: Deal with this by using the current attribute
1017                         // and fill it with as much of the mapping pairs
1018                         // array as possible.  Then loop over each attribute
1019                         // extent rewriting the mapping pairs arrays as we go
1020                         // along and if when we reach the end we have not
1021                         // enough space, try to resize the last attribute
1022                         // extent and if even that fails, add a new attribute
1023                         // extent.
1024                         // We could also try to resize at each step in the hope
1025                         // that we will not need to rewrite every single extent.
1026                         // Note, we may need to decompress some extents to fill
1027                         // the runlist as we are walking the extents...
1028                         ntfs_error(vol->sb, "Not enough space in the mft "
1029                                         "record for the extended attribute "
1030                                         "record.  This case is not "
1031                                         "implemented yet.");
1032                         err = -EOPNOTSUPP;
1033                         break ;
1034                 }
1035                 status.mp_rebuilt = 1;
1036                 /*
1037                  * Generate the mapping pairs array directly into the attribute
1038                  * record.
1039                  */
1040                 err = ntfs_mapping_pairs_build(vol, (u8*)a + le16_to_cpu(
1041                                 a->data.non_resident.mapping_pairs_offset),
1042                                 mp_size, rl2, vcn, highest_vcn, NULL);
1043                 if (unlikely(err)) {
1044                         ntfs_error(vol->sb, "Cannot fill hole in inode 0x%lx, "
1045                                         "attribute type 0x%x, because building "
1046                                         "the mapping pairs failed with error "
1047                                         "code %i.", vi->i_ino,
1048                                         (unsigned)le32_to_cpu(ni->type), err);
1049                         err = -EIO;
1050                         break;
1051                 }
1052                 /* Update the highest_vcn but only if it was not set. */
1053                 if (unlikely(!a->data.non_resident.highest_vcn))
1054                         a->data.non_resident.highest_vcn =
1055                                         cpu_to_sle64(highest_vcn);
1056                 /*
1057                  * If the attribute is sparse/compressed, update the compressed
1058                  * size in the ntfs_inode structure and the attribute record.
1059                  */
1060                 if (likely(NInoSparse(ni) || NInoCompressed(ni))) {
1061                         /*
1062                          * If we are not in the first attribute extent, switch
1063                          * to it, but first ensure the changes will make it to
1064                          * disk later.
1065                          */
1066                         if (a->data.non_resident.lowest_vcn) {
1067                                 flush_dcache_mft_record_page(ctx->ntfs_ino);
1068                                 mark_mft_record_dirty(ctx->ntfs_ino);
1069                                 ntfs_attr_reinit_search_ctx(ctx);
1070                                 err = ntfs_attr_lookup(ni->type, ni->name,
1071                                                 ni->name_len, CASE_SENSITIVE,
1072                                                 0, NULL, 0, ctx);
1073                                 if (unlikely(err)) {
1074                                         status.attr_switched = 1;
1075                                         break;
1076                                 }
1077                                 /* @m is not used any more so do not set it. */
1078                                 a = ctx->attr;
1079                         }
1080                         write_lock_irqsave(&ni->size_lock, flags);
1081                         ni->itype.compressed.size += vol->cluster_size;
1082                         a->data.non_resident.compressed_size =
1083                                         cpu_to_sle64(ni->itype.compressed.size);
1084                         write_unlock_irqrestore(&ni->size_lock, flags);
1085                 }
1086                 /* Ensure the changes make it to disk. */
1087                 flush_dcache_mft_record_page(ctx->ntfs_ino);
1088                 mark_mft_record_dirty(ctx->ntfs_ino);
1089                 ntfs_attr_put_search_ctx(ctx);
1090                 unmap_mft_record(base_ni);
1091                 /* Successfully filled the hole. */
1092                 status.runlist_merged = 0;
1093                 status.mft_attr_mapped = 0;
1094                 status.mp_rebuilt = 0;
1095                 /* Setup the map cache and use that to deal with the buffer. */
1096                 was_hole = true;
1097                 vcn = bh_cpos;
1098                 vcn_len = 1;
1099                 lcn_block = lcn << (vol->cluster_size_bits - blocksize_bits);
1100                 cdelta = 0;
1101                 /*
1102                  * If the number of remaining clusters in the @pages is smaller
1103                  * or equal to the number of cached clusters, unlock the
1104                  * runlist as the map cache will be used from now on.
1105                  */
1106                 if (likely(vcn + vcn_len >= cend)) {
1107                         up_write(&ni->runlist.lock);
1108                         rl_write_locked = false;
1109                         rl = NULL;
1110                 }
1111                 goto map_buffer_cached;
1112         } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
1113         /* If there are no errors, do the next page. */
1114         if (likely(!err && ++u < nr_pages))
1115                 goto do_next_page;
1116         /* If there are no errors, release the runlist lock if we took it. */
1117         if (likely(!err)) {
1118                 if (unlikely(rl_write_locked)) {
1119                         up_write(&ni->runlist.lock);
1120                         rl_write_locked = false;
1121                 } else if (unlikely(rl))
1122                         up_read(&ni->runlist.lock);
1123                 rl = NULL;
1124         }
1125         /* If we issued read requests, let them complete. */
1126         read_lock_irqsave(&ni->size_lock, flags);
1127         initialized_size = ni->initialized_size;
1128         read_unlock_irqrestore(&ni->size_lock, flags);
1129         while (wait_bh > wait) {
1130                 bh = *--wait_bh;
1131                 wait_on_buffer(bh);
1132                 if (likely(buffer_uptodate(bh))) {
1133                         page = bh->b_page;
1134                         bh_pos = ((s64)page->index << PAGE_CACHE_SHIFT) +
1135                                         bh_offset(bh);
1136                         /*
1137                          * If the buffer overflows the initialized size, need
1138                          * to zero the overflowing region.
1139                          */
1140                         if (unlikely(bh_pos + blocksize > initialized_size)) {
1141                                 u8 *kaddr;
1142                                 int ofs = 0;
1143
1144                                 if (likely(bh_pos < initialized_size))
1145                                         ofs = initialized_size - bh_pos;
1146                                 kaddr = kmap_atomic(page, KM_USER0);
1147                                 memset(kaddr + bh_offset(bh) + ofs, 0,
1148                                                 blocksize - ofs);
1149                                 kunmap_atomic(kaddr, KM_USER0);
1150                                 flush_dcache_page(page);
1151                         }
1152                 } else /* if (unlikely(!buffer_uptodate(bh))) */
1153                         err = -EIO;
1154         }
1155         if (likely(!err)) {
1156                 /* Clear buffer_new on all buffers. */
1157                 u = 0;
1158                 do {
1159                         bh = head = page_buffers(pages[u]);
1160                         do {
1161                                 if (buffer_new(bh))
1162                                         clear_buffer_new(bh);
1163                         } while ((bh = bh->b_this_page) != head);
1164                 } while (++u < nr_pages);
1165                 ntfs_debug("Done.");
1166                 return err;
1167         }
1168         if (status.attr_switched) {
1169                 /* Get back to the attribute extent we modified. */
1170                 ntfs_attr_reinit_search_ctx(ctx);
1171                 if (ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1172                                 CASE_SENSITIVE, bh_cpos, NULL, 0, ctx)) {
1173                         ntfs_error(vol->sb, "Failed to find required "
1174                                         "attribute extent of attribute in "
1175                                         "error code path.  Run chkdsk to "
1176                                         "recover.");
1177                         write_lock_irqsave(&ni->size_lock, flags);
1178                         ni->itype.compressed.size += vol->cluster_size;
1179                         write_unlock_irqrestore(&ni->size_lock, flags);
1180                         flush_dcache_mft_record_page(ctx->ntfs_ino);
1181                         mark_mft_record_dirty(ctx->ntfs_ino);
1182                         /*
1183                          * The only thing that is now wrong is the compressed
1184                          * size of the base attribute extent which chkdsk
1185                          * should be able to fix.
1186                          */
1187                         NVolSetErrors(vol);
1188                 } else {
1189                         m = ctx->mrec;
1190                         a = ctx->attr;
1191                         status.attr_switched = 0;
1192                 }
1193         }
1194         /*
1195          * If the runlist has been modified, need to restore it by punching a
1196          * hole into it and we then need to deallocate the on-disk cluster as
1197          * well.  Note, we only modify the runlist if we are able to generate a
1198          * new mapping pairs array, i.e. only when the mapped attribute extent
1199          * is not switched.
1200          */
1201         if (status.runlist_merged && !status.attr_switched) {
1202                 BUG_ON(!rl_write_locked);
1203                 /* Make the file cluster we allocated sparse in the runlist. */
1204                 if (ntfs_rl_punch_nolock(vol, &ni->runlist, bh_cpos, 1)) {
1205                         ntfs_error(vol->sb, "Failed to punch hole into "
1206                                         "attribute runlist in error code "
1207                                         "path.  Run chkdsk to recover the "
1208                                         "lost cluster.");
1209                         NVolSetErrors(vol);
1210                 } else /* if (success) */ {
1211                         status.runlist_merged = 0;
1212                         /*
1213                          * Deallocate the on-disk cluster we allocated but only
1214                          * if we succeeded in punching its vcn out of the
1215                          * runlist.
1216                          */
1217                         down_write(&vol->lcnbmp_lock);
1218                         if (ntfs_bitmap_clear_bit(vol->lcnbmp_ino, lcn)) {
1219                                 ntfs_error(vol->sb, "Failed to release "
1220                                                 "allocated cluster in error "
1221                                                 "code path.  Run chkdsk to "
1222                                                 "recover the lost cluster.");
1223                                 NVolSetErrors(vol);
1224                         }
1225                         up_write(&vol->lcnbmp_lock);
1226                 }
1227         }
1228         /*
1229          * Resize the attribute record to its old size and rebuild the mapping
1230          * pairs array.  Note, we only can do this if the runlist has been
1231          * restored to its old state which also implies that the mapped
1232          * attribute extent is not switched.
1233          */
1234         if (status.mp_rebuilt && !status.runlist_merged) {
1235                 if (ntfs_attr_record_resize(m, a, attr_rec_len)) {
1236                         ntfs_error(vol->sb, "Failed to restore attribute "
1237                                         "record in error code path.  Run "
1238                                         "chkdsk to recover.");
1239                         NVolSetErrors(vol);
1240                 } else /* if (success) */ {
1241                         if (ntfs_mapping_pairs_build(vol, (u8*)a +
1242                                         le16_to_cpu(a->data.non_resident.
1243                                         mapping_pairs_offset), attr_rec_len -
1244                                         le16_to_cpu(a->data.non_resident.
1245                                         mapping_pairs_offset), ni->runlist.rl,
1246                                         vcn, highest_vcn, NULL)) {
1247                                 ntfs_error(vol->sb, "Failed to restore "
1248                                                 "mapping pairs array in error "
1249                                                 "code path.  Run chkdsk to "
1250                                                 "recover.");
1251                                 NVolSetErrors(vol);
1252                         }
1253                         flush_dcache_mft_record_page(ctx->ntfs_ino);
1254                         mark_mft_record_dirty(ctx->ntfs_ino);
1255                 }
1256         }
1257         /* Release the mft record and the attribute. */
1258         if (status.mft_attr_mapped) {
1259                 ntfs_attr_put_search_ctx(ctx);
1260                 unmap_mft_record(base_ni);
1261         }
1262         /* Release the runlist lock. */
1263         if (rl_write_locked)
1264                 up_write(&ni->runlist.lock);
1265         else if (rl)
1266                 up_read(&ni->runlist.lock);
1267         /*
1268          * Zero out any newly allocated blocks to avoid exposing stale data.
1269          * If BH_New is set, we know that the block was newly allocated above
1270          * and that it has not been fully zeroed and marked dirty yet.
1271          */
1272         nr_pages = u;
1273         u = 0;
1274         end = bh_cpos << vol->cluster_size_bits;
1275         do {
1276                 page = pages[u];
1277                 bh = head = page_buffers(page);
1278                 do {
1279                         if (u == nr_pages &&
1280                                         ((s64)page->index << PAGE_CACHE_SHIFT) +
1281                                         bh_offset(bh) >= end)
1282                                 break;
1283                         if (!buffer_new(bh))
1284                                 continue;
1285                         clear_buffer_new(bh);
1286                         if (!buffer_uptodate(bh)) {
1287                                 if (PageUptodate(page))
1288                                         set_buffer_uptodate(bh);
1289                                 else {
1290                                         u8 *kaddr = kmap_atomic(page, KM_USER0);
1291                                         memset(kaddr + bh_offset(bh), 0,
1292                                                         blocksize);
1293                                         kunmap_atomic(kaddr, KM_USER0);
1294                                         flush_dcache_page(page);
1295                                         set_buffer_uptodate(bh);
1296                                 }
1297                         }
1298                         mark_buffer_dirty(bh);
1299                 } while ((bh = bh->b_this_page) != head);
1300         } while (++u <= nr_pages);
1301         ntfs_error(vol->sb, "Failed.  Returning error code %i.", err);
1302         return err;
1303 }
1304
1305 /*
1306  * Copy as much as we can into the pages and return the number of bytes which
1307  * were sucessfully copied.  If a fault is encountered then clear the pages
1308  * out to (ofs + bytes) and return the number of bytes which were copied.
1309  */
1310 static inline size_t ntfs_copy_from_user(struct page **pages,
1311                 unsigned nr_pages, unsigned ofs, const char __user *buf,
1312                 size_t bytes)
1313 {
1314         struct page **last_page = pages + nr_pages;
1315         char *kaddr;
1316         size_t total = 0;
1317         unsigned len;
1318         int left;
1319
1320         do {
1321                 len = PAGE_CACHE_SIZE - ofs;
1322                 if (len > bytes)
1323                         len = bytes;
1324                 kaddr = kmap_atomic(*pages, KM_USER0);
1325                 left = __copy_from_user_inatomic(kaddr + ofs, buf, len);
1326                 kunmap_atomic(kaddr, KM_USER0);
1327                 if (unlikely(left)) {
1328                         /* Do it the slow way. */
1329                         kaddr = kmap(*pages);
1330                         left = __copy_from_user(kaddr + ofs, buf, len);
1331                         kunmap(*pages);
1332                         if (unlikely(left))
1333                                 goto err_out;
1334                 }
1335                 total += len;
1336                 bytes -= len;
1337                 if (!bytes)
1338                         break;
1339                 buf += len;
1340                 ofs = 0;
1341         } while (++pages < last_page);
1342 out:
1343         return total;
1344 err_out:
1345         total += len - left;
1346         /* Zero the rest of the target like __copy_from_user(). */
1347         while (++pages < last_page) {
1348                 bytes -= len;
1349                 if (!bytes)
1350                         break;
1351                 len = PAGE_CACHE_SIZE;
1352                 if (len > bytes)
1353                         len = bytes;
1354                 kaddr = kmap_atomic(*pages, KM_USER0);
1355                 memset(kaddr, 0, len);
1356                 kunmap_atomic(kaddr, KM_USER0);
1357         }
1358         goto out;
1359 }
1360
1361 static size_t __ntfs_copy_from_user_iovec_inatomic(char *vaddr,
1362                 const struct iovec *iov, size_t iov_ofs, size_t bytes)
1363 {
1364         size_t total = 0;
1365
1366         while (1) {
1367                 const char __user *buf = iov->iov_base + iov_ofs;
1368                 unsigned len;
1369                 size_t left;
1370
1371                 len = iov->iov_len - iov_ofs;
1372                 if (len > bytes)
1373                         len = bytes;
1374                 left = __copy_from_user_inatomic(vaddr, buf, len);
1375                 total += len;
1376                 bytes -= len;
1377                 vaddr += len;
1378                 if (unlikely(left)) {
1379                         total -= left;
1380                         break;
1381                 }
1382                 if (!bytes)
1383                         break;
1384                 iov++;
1385                 iov_ofs = 0;
1386         }
1387         return total;
1388 }
1389
1390 static inline void ntfs_set_next_iovec(const struct iovec **iovp,
1391                 size_t *iov_ofsp, size_t bytes)
1392 {
1393         const struct iovec *iov = *iovp;
1394         size_t iov_ofs = *iov_ofsp;
1395
1396         while (bytes) {
1397                 unsigned len;
1398
1399                 len = iov->iov_len - iov_ofs;
1400                 if (len > bytes)
1401                         len = bytes;
1402                 bytes -= len;
1403                 iov_ofs += len;
1404                 if (iov->iov_len == iov_ofs) {
1405                         iov++;
1406                         iov_ofs = 0;
1407                 }
1408         }
1409         *iovp = iov;
1410         *iov_ofsp = iov_ofs;
1411 }
1412
1413 /*
1414  * This has the same side-effects and return value as ntfs_copy_from_user().
1415  * The difference is that on a fault we need to memset the remainder of the
1416  * pages (out to offset + bytes), to emulate ntfs_copy_from_user()'s
1417  * single-segment behaviour.
1418  *
1419  * We call the same helper (__ntfs_copy_from_user_iovec_inatomic()) both
1420  * when atomic and when not atomic.  This is ok because
1421  * __ntfs_copy_from_user_iovec_inatomic() calls __copy_from_user_inatomic()
1422  * and it is ok to call this when non-atomic.
1423  * Infact, the only difference between __copy_from_user_inatomic() and
1424  * __copy_from_user() is that the latter calls might_sleep() and the former
1425  * should not zero the tail of the buffer on error.  And on many
1426  * architectures __copy_from_user_inatomic() is just defined to
1427  * __copy_from_user() so it makes no difference at all on those architectures.
1428  */
1429 static inline size_t ntfs_copy_from_user_iovec(struct page **pages,
1430                 unsigned nr_pages, unsigned ofs, const struct iovec **iov,
1431                 size_t *iov_ofs, size_t bytes)
1432 {
1433         struct page **last_page = pages + nr_pages;
1434         char *kaddr;
1435         size_t copied, len, total = 0;
1436
1437         do {
1438                 len = PAGE_CACHE_SIZE - ofs;
1439                 if (len > bytes)
1440                         len = bytes;
1441                 kaddr = kmap_atomic(*pages, KM_USER0);
1442                 copied = __ntfs_copy_from_user_iovec_inatomic(kaddr + ofs,
1443                                 *iov, *iov_ofs, len);
1444                 kunmap_atomic(kaddr, KM_USER0);
1445                 if (unlikely(copied != len)) {
1446                         /* Do it the slow way. */
1447                         kaddr = kmap(*pages);
1448                         copied = __ntfs_copy_from_user_iovec_inatomic(kaddr + ofs,
1449                                         *iov, *iov_ofs, len);
1450                         /*
1451                          * Zero the rest of the target like __copy_from_user().
1452                          */
1453                         memset(kaddr + ofs + copied, 0, len - copied);
1454                         kunmap(*pages);
1455                         if (unlikely(copied != len))
1456                                 goto err_out;
1457                 }
1458                 total += len;
1459                 bytes -= len;
1460                 if (!bytes)
1461                         break;
1462                 ntfs_set_next_iovec(iov, iov_ofs, len);
1463                 ofs = 0;
1464         } while (++pages < last_page);
1465 out:
1466         return total;
1467 err_out:
1468         total += copied;
1469         /* Zero the rest of the target like __copy_from_user(). */
1470         while (++pages < last_page) {
1471                 bytes -= len;
1472                 if (!bytes)
1473                         break;
1474                 len = PAGE_CACHE_SIZE;
1475                 if (len > bytes)
1476                         len = bytes;
1477                 kaddr = kmap_atomic(*pages, KM_USER0);
1478                 memset(kaddr, 0, len);
1479                 kunmap_atomic(kaddr, KM_USER0);
1480         }
1481         goto out;
1482 }
1483
1484 static inline void ntfs_flush_dcache_pages(struct page **pages,
1485                 unsigned nr_pages)
1486 {
1487         BUG_ON(!nr_pages);
1488         /*
1489          * Warning: Do not do the decrement at the same time as the call to
1490          * flush_dcache_page() because it is a NULL macro on i386 and hence the
1491          * decrement never happens so the loop never terminates.
1492          */
1493         do {
1494                 --nr_pages;
1495                 flush_dcache_page(pages[nr_pages]);
1496         } while (nr_pages > 0);
1497 }
1498
1499 /**
1500  * ntfs_commit_pages_after_non_resident_write - commit the received data
1501  * @pages:      array of destination pages
1502  * @nr_pages:   number of pages in @pages
1503  * @pos:        byte position in file at which the write begins
1504  * @bytes:      number of bytes to be written
1505  *
1506  * See description of ntfs_commit_pages_after_write(), below.
1507  */
1508 static inline int ntfs_commit_pages_after_non_resident_write(
1509                 struct page **pages, const unsigned nr_pages,
1510                 s64 pos, size_t bytes)
1511 {
1512         s64 end, initialized_size;
1513         struct inode *vi;
1514         ntfs_inode *ni, *base_ni;
1515         struct buffer_head *bh, *head;
1516         ntfs_attr_search_ctx *ctx;
1517         MFT_RECORD *m;
1518         ATTR_RECORD *a;
1519         unsigned long flags;
1520         unsigned blocksize, u;
1521         int err;
1522
1523         vi = pages[0]->mapping->host;
1524         ni = NTFS_I(vi);
1525         blocksize = vi->i_sb->s_blocksize;
1526         end = pos + bytes;
1527         u = 0;
1528         do {
1529                 s64 bh_pos;
1530                 struct page *page;
1531                 bool partial;
1532
1533                 page = pages[u];
1534                 bh_pos = (s64)page->index << PAGE_CACHE_SHIFT;
1535                 bh = head = page_buffers(page);
1536                 partial = false;
1537                 do {
1538                         s64 bh_end;
1539
1540                         bh_end = bh_pos + blocksize;
1541                         if (bh_end <= pos || bh_pos >= end) {
1542                                 if (!buffer_uptodate(bh))
1543                                         partial = true;
1544                         } else {
1545                                 set_buffer_uptodate(bh);
1546                                 mark_buffer_dirty(bh);
1547                         }
1548                 } while (bh_pos += blocksize, (bh = bh->b_this_page) != head);
1549                 /*
1550                  * If all buffers are now uptodate but the page is not, set the
1551                  * page uptodate.
1552                  */
1553                 if (!partial && !PageUptodate(page))
1554                         SetPageUptodate(page);
1555         } while (++u < nr_pages);
1556         /*
1557          * Finally, if we do not need to update initialized_size or i_size we
1558          * are finished.
1559          */
1560         read_lock_irqsave(&ni->size_lock, flags);
1561         initialized_size = ni->initialized_size;
1562         read_unlock_irqrestore(&ni->size_lock, flags);
1563         if (end <= initialized_size) {
1564                 ntfs_debug("Done.");
1565                 return 0;
1566         }
1567         /*
1568          * Update initialized_size/i_size as appropriate, both in the inode and
1569          * the mft record.
1570          */
1571         if (!NInoAttr(ni))
1572                 base_ni = ni;
1573         else
1574                 base_ni = ni->ext.base_ntfs_ino;
1575         /* Map, pin, and lock the mft record. */
1576         m = map_mft_record(base_ni);
1577         if (IS_ERR(m)) {
1578                 err = PTR_ERR(m);
1579                 m = NULL;
1580                 ctx = NULL;
1581                 goto err_out;
1582         }
1583         BUG_ON(!NInoNonResident(ni));
1584         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1585         if (unlikely(!ctx)) {
1586                 err = -ENOMEM;
1587                 goto err_out;
1588         }
1589         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1590                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1591         if (unlikely(err)) {
1592                 if (err == -ENOENT)
1593                         err = -EIO;
1594                 goto err_out;
1595         }
1596         a = ctx->attr;
1597         BUG_ON(!a->non_resident);
1598         write_lock_irqsave(&ni->size_lock, flags);
1599         BUG_ON(end > ni->allocated_size);
1600         ni->initialized_size = end;
1601         a->data.non_resident.initialized_size = cpu_to_sle64(end);
1602         if (end > i_size_read(vi)) {
1603                 i_size_write(vi, end);
1604                 a->data.non_resident.data_size =
1605                                 a->data.non_resident.initialized_size;
1606         }
1607         write_unlock_irqrestore(&ni->size_lock, flags);
1608         /* Mark the mft record dirty, so it gets written back. */
1609         flush_dcache_mft_record_page(ctx->ntfs_ino);
1610         mark_mft_record_dirty(ctx->ntfs_ino);
1611         ntfs_attr_put_search_ctx(ctx);
1612         unmap_mft_record(base_ni);
1613         ntfs_debug("Done.");
1614         return 0;
1615 err_out:
1616         if (ctx)
1617                 ntfs_attr_put_search_ctx(ctx);
1618         if (m)
1619                 unmap_mft_record(base_ni);
1620         ntfs_error(vi->i_sb, "Failed to update initialized_size/i_size (error "
1621                         "code %i).", err);
1622         if (err != -ENOMEM)
1623                 NVolSetErrors(ni->vol);
1624         return err;
1625 }
1626
1627 /**
1628  * ntfs_commit_pages_after_write - commit the received data
1629  * @pages:      array of destination pages
1630  * @nr_pages:   number of pages in @pages
1631  * @pos:        byte position in file at which the write begins
1632  * @bytes:      number of bytes to be written
1633  *
1634  * This is called from ntfs_file_buffered_write() with i_mutex held on the inode
1635  * (@pages[0]->mapping->host).  There are @nr_pages pages in @pages which are
1636  * locked but not kmap()ped.  The source data has already been copied into the
1637  * @page.  ntfs_prepare_pages_for_non_resident_write() has been called before
1638  * the data was copied (for non-resident attributes only) and it returned
1639  * success.
1640  *
1641  * Need to set uptodate and mark dirty all buffers within the boundary of the
1642  * write.  If all buffers in a page are uptodate we set the page uptodate, too.
1643  *
1644  * Setting the buffers dirty ensures that they get written out later when
1645  * ntfs_writepage() is invoked by the VM.
1646  *
1647  * Finally, we need to update i_size and initialized_size as appropriate both
1648  * in the inode and the mft record.
1649  *
1650  * This is modelled after fs/buffer.c::generic_commit_write(), which marks
1651  * buffers uptodate and dirty, sets the page uptodate if all buffers in the
1652  * page are uptodate, and updates i_size if the end of io is beyond i_size.  In
1653  * that case, it also marks the inode dirty.
1654  *
1655  * If things have gone as outlined in
1656  * ntfs_prepare_pages_for_non_resident_write(), we do not need to do any page
1657  * content modifications here for non-resident attributes.  For resident
1658  * attributes we need to do the uptodate bringing here which we combine with
1659  * the copying into the mft record which means we save one atomic kmap.
1660  *
1661  * Return 0 on success or -errno on error.
1662  */
1663 static int ntfs_commit_pages_after_write(struct page **pages,
1664                 const unsigned nr_pages, s64 pos, size_t bytes)
1665 {
1666         s64 end, initialized_size;
1667         loff_t i_size;
1668         struct inode *vi;
1669         ntfs_inode *ni, *base_ni;
1670         struct page *page;
1671         ntfs_attr_search_ctx *ctx;
1672         MFT_RECORD *m;
1673         ATTR_RECORD *a;
1674         char *kattr, *kaddr;
1675         unsigned long flags;
1676         u32 attr_len;
1677         int err;
1678
1679         BUG_ON(!nr_pages);
1680         BUG_ON(!pages);
1681         page = pages[0];
1682         BUG_ON(!page);
1683         vi = page->mapping->host;
1684         ni = NTFS_I(vi);
1685         ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, start page "
1686                         "index 0x%lx, nr_pages 0x%x, pos 0x%llx, bytes 0x%zx.",
1687                         vi->i_ino, ni->type, page->index, nr_pages,
1688                         (long long)pos, bytes);
1689         if (NInoNonResident(ni))
1690                 return ntfs_commit_pages_after_non_resident_write(pages,
1691                                 nr_pages, pos, bytes);
1692         BUG_ON(nr_pages > 1);
1693         /*
1694          * Attribute is resident, implying it is not compressed, encrypted, or
1695          * sparse.
1696          */
1697         if (!NInoAttr(ni))
1698                 base_ni = ni;
1699         else
1700                 base_ni = ni->ext.base_ntfs_ino;
1701         BUG_ON(NInoNonResident(ni));
1702         /* Map, pin, and lock the mft record. */
1703         m = map_mft_record(base_ni);
1704         if (IS_ERR(m)) {
1705                 err = PTR_ERR(m);
1706                 m = NULL;
1707                 ctx = NULL;
1708                 goto err_out;
1709         }
1710         ctx = ntfs_attr_get_search_ctx(base_ni, m);
1711         if (unlikely(!ctx)) {
1712                 err = -ENOMEM;
1713                 goto err_out;
1714         }
1715         err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1716                         CASE_SENSITIVE, 0, NULL, 0, ctx);
1717         if (unlikely(err)) {
1718                 if (err == -ENOENT)
1719                         err = -EIO;
1720                 goto err_out;
1721         }
1722         a = ctx->attr;
1723         BUG_ON(a->non_resident);
1724         /* The total length of the attribute value. */
1725         attr_len = le32_to_cpu(a->data.resident.value_length);
1726         i_size = i_size_read(vi);
1727         BUG_ON(attr_len != i_size);
1728         BUG_ON(pos > attr_len);
1729         end = pos + bytes;
1730         BUG_ON(end > le32_to_cpu(a->length) -
1731                         le16_to_cpu(a->data.resident.value_offset));
1732         kattr = (u8*)a + le16_to_cpu(a->data.resident.value_offset);
1733         kaddr = kmap_atomic(page, KM_USER0);
1734         /* Copy the received data from the page to the mft record. */
1735         memcpy(kattr + pos, kaddr + pos, bytes);
1736         /* Update the attribute length if necessary. */
1737         if (end > attr_len) {
1738                 attr_len = end;
1739                 a->data.resident.value_length = cpu_to_le32(attr_len);
1740         }
1741         /*
1742          * If the page is not uptodate, bring the out of bounds area(s)
1743          * uptodate by copying data from the mft record to the page.
1744          */
1745         if (!PageUptodate(page)) {
1746                 if (pos > 0)
1747                         memcpy(kaddr, kattr, pos);
1748                 if (end < attr_len)
1749                         memcpy(kaddr + end, kattr + end, attr_len - end);
1750                 /* Zero the region outside the end of the attribute value. */
1751                 memset(kaddr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
1752                 flush_dcache_page(page);
1753                 SetPageUptodate(page);
1754         }
1755         kunmap_atomic(kaddr, KM_USER0);
1756         /* Update initialized_size/i_size if necessary. */
1757         read_lock_irqsave(&ni->size_lock, flags);
1758         initialized_size = ni->initialized_size;
1759         BUG_ON(end > ni->allocated_size);
1760         read_unlock_irqrestore(&ni->size_lock, flags);
1761         BUG_ON(initialized_size != i_size);
1762         if (end > initialized_size) {
1763                 unsigned long flags;
1764
1765                 write_lock_irqsave(&ni->size_lock, flags);
1766                 ni->initialized_size = end;
1767                 i_size_write(vi, end);
1768                 write_unlock_irqrestore(&ni->size_lock, flags);
1769         }
1770         /* Mark the mft record dirty, so it gets written back. */
1771         flush_dcache_mft_record_page(ctx->ntfs_ino);
1772         mark_mft_record_dirty(ctx->ntfs_ino);
1773         ntfs_attr_put_search_ctx(ctx);
1774         unmap_mft_record(base_ni);
1775         ntfs_debug("Done.");
1776         return 0;
1777 err_out:
1778         if (err == -ENOMEM) {
1779                 ntfs_warning(vi->i_sb, "Error allocating memory required to "
1780                                 "commit the write.");
1781                 if (PageUptodate(page)) {
1782                         ntfs_warning(vi->i_sb, "Page is uptodate, setting "
1783                                         "dirty so the write will be retried "
1784                                         "later on by the VM.");
1785                         /*
1786                          * Put the page on mapping->dirty_pages, but leave its
1787                          * buffers' dirty state as-is.
1788                          */
1789                         __set_page_dirty_nobuffers(page);
1790                         err = 0;
1791                 } else
1792                         ntfs_error(vi->i_sb, "Page is not uptodate.  Written "
1793                                         "data has been lost.");
1794         } else {
1795                 ntfs_error(vi->i_sb, "Resident attribute commit write failed "
1796                                 "with error %i.", err);
1797                 NVolSetErrors(ni->vol);
1798         }
1799         if (ctx)
1800                 ntfs_attr_put_search_ctx(ctx);
1801         if (m)
1802                 unmap_mft_record(base_ni);
1803         return err;
1804 }
1805
1806 /**
1807  * ntfs_file_buffered_write -
1808  *
1809  * Locking: The vfs is holding ->i_mutex on the inode.
1810  */
1811 static ssize_t ntfs_file_buffered_write(struct kiocb *iocb,
1812                 const struct iovec *iov, unsigned long nr_segs,
1813                 loff_t pos, loff_t *ppos, size_t count)
1814 {
1815         struct file *file = iocb->ki_filp;
1816         struct address_space *mapping = file->f_mapping;
1817         struct inode *vi = mapping->host;
1818         ntfs_inode *ni = NTFS_I(vi);
1819         ntfs_volume *vol = ni->vol;
1820         struct page *pages[NTFS_MAX_PAGES_PER_CLUSTER];
1821         struct page *cached_page = NULL;
1822         char __user *buf = NULL;
1823         s64 end, ll;
1824         VCN last_vcn;
1825         LCN lcn;
1826         unsigned long flags;
1827         size_t bytes, iov_ofs = 0;      /* Offset in the current iovec. */
1828         ssize_t status, written;
1829         unsigned nr_pages;
1830         int err;
1831         struct pagevec lru_pvec;
1832
1833         ntfs_debug("Entering for i_ino 0x%lx, attribute type 0x%x, "
1834                         "pos 0x%llx, count 0x%lx.",
1835                         vi->i_ino, (unsigned)le32_to_cpu(ni->type),
1836                         (unsigned long long)pos, (unsigned long)count);
1837         if (unlikely(!count))
1838                 return 0;
1839         BUG_ON(NInoMstProtected(ni));
1840         /*
1841          * If the attribute is not an index root and it is encrypted or
1842          * compressed, we cannot write to it yet.  Note we need to check for
1843          * AT_INDEX_ALLOCATION since this is the type of both directory and
1844          * index inodes.
1845          */
1846         if (ni->type != AT_INDEX_ALLOCATION) {
1847                 /* If file is encrypted, deny access, just like NT4. */
1848                 if (NInoEncrypted(ni)) {
1849                         /*
1850                          * Reminder for later: Encrypted files are _always_
1851                          * non-resident so that the content can always be
1852                          * encrypted.
1853                          */
1854                         ntfs_debug("Denying write access to encrypted file.");
1855                         return -EACCES;
1856                 }
1857                 if (NInoCompressed(ni)) {
1858                         /* Only unnamed $DATA attribute can be compressed. */
1859                         BUG_ON(ni->type != AT_DATA);
1860                         BUG_ON(ni->name_len);
1861                         /*
1862                          * Reminder for later: If resident, the data is not
1863                          * actually compressed.  Only on the switch to non-
1864                          * resident does compression kick in.  This is in
1865                          * contrast to encrypted files (see above).
1866                          */
1867                         ntfs_error(vi->i_sb, "Writing to compressed files is "
1868                                         "not implemented yet.  Sorry.");
1869                         return -EOPNOTSUPP;
1870                 }
1871         }
1872         /*
1873          * If a previous ntfs_truncate() failed, repeat it and abort if it
1874          * fails again.
1875          */
1876         if (unlikely(NInoTruncateFailed(ni))) {
1877                 down_write(&vi->i_alloc_sem);
1878                 err = ntfs_truncate(vi);
1879                 up_write(&vi->i_alloc_sem);
1880                 if (err || NInoTruncateFailed(ni)) {
1881                         if (!err)
1882                                 err = -EIO;
1883                         ntfs_error(vol->sb, "Cannot perform write to inode "
1884                                         "0x%lx, attribute type 0x%x, because "
1885                                         "ntfs_truncate() failed (error code "
1886                                         "%i).", vi->i_ino,
1887                                         (unsigned)le32_to_cpu(ni->type), err);
1888                         return err;
1889                 }
1890         }
1891         /* The first byte after the write. */
1892         end = pos + count;
1893         /*
1894          * If the write goes beyond the allocated size, extend the allocation
1895          * to cover the whole of the write, rounded up to the nearest cluster.
1896          */
1897         read_lock_irqsave(&ni->size_lock, flags);
1898         ll = ni->allocated_size;
1899         read_unlock_irqrestore(&ni->size_lock, flags);
1900         if (end > ll) {
1901                 /* Extend the allocation without changing the data size. */
1902                 ll = ntfs_attr_extend_allocation(ni, end, -1, pos);
1903                 if (likely(ll >= 0)) {
1904                         BUG_ON(pos >= ll);
1905                         /* If the extension was partial truncate the write. */
1906                         if (end > ll) {
1907                                 ntfs_debug("Truncating write to inode 0x%lx, "
1908                                                 "attribute type 0x%x, because "
1909                                                 "the allocation was only "
1910                                                 "partially extended.",
1911                                                 vi->i_ino, (unsigned)
1912                                                 le32_to_cpu(ni->type));
1913                                 end = ll;
1914                                 count = ll - pos;
1915                         }
1916                 } else {
1917                         err = ll;
1918                         read_lock_irqsave(&ni->size_lock, flags);
1919                         ll = ni->allocated_size;
1920                         read_unlock_irqrestore(&ni->size_lock, flags);
1921                         /* Perform a partial write if possible or fail. */
1922                         if (pos < ll) {
1923                                 ntfs_debug("Truncating write to inode 0x%lx, "
1924                                                 "attribute type 0x%x, because "
1925                                                 "extending the allocation "
1926                                                 "failed (error code %i).",
1927                                                 vi->i_ino, (unsigned)
1928                                                 le32_to_cpu(ni->type), err);
1929                                 end = ll;
1930                                 count = ll - pos;
1931                         } else {
1932                                 ntfs_error(vol->sb, "Cannot perform write to "
1933                                                 "inode 0x%lx, attribute type "
1934                                                 "0x%x, because extending the "
1935                                                 "allocation failed (error "
1936                                                 "code %i).", vi->i_ino,
1937                                                 (unsigned)
1938                                                 le32_to_cpu(ni->type), err);
1939                                 return err;
1940                         }
1941                 }
1942         }
1943         pagevec_init(&lru_pvec, 0);
1944         written = 0;
1945         /*
1946          * If the write starts beyond the initialized size, extend it up to the
1947          * beginning of the write and initialize all non-sparse space between
1948          * the old initialized size and the new one.  This automatically also
1949          * increments the vfs inode->i_size to keep it above or equal to the
1950          * initialized_size.
1951          */
1952         read_lock_irqsave(&ni->size_lock, flags);
1953         ll = ni->initialized_size;
1954         read_unlock_irqrestore(&ni->size_lock, flags);
1955         if (pos > ll) {
1956                 err = ntfs_attr_extend_initialized(ni, pos, &cached_page,
1957                                 &lru_pvec);
1958                 if (err < 0) {
1959                         ntfs_error(vol->sb, "Cannot perform write to inode "
1960                                         "0x%lx, attribute type 0x%x, because "
1961                                         "extending the initialized size "
1962                                         "failed (error code %i).", vi->i_ino,
1963                                         (unsigned)le32_to_cpu(ni->type), err);
1964                         status = err;
1965                         goto err_out;
1966                 }
1967         }
1968         /*
1969          * Determine the number of pages per cluster for non-resident
1970          * attributes.
1971          */
1972         nr_pages = 1;
1973         if (vol->cluster_size > PAGE_CACHE_SIZE && NInoNonResident(ni))
1974                 nr_pages = vol->cluster_size >> PAGE_CACHE_SHIFT;
1975         /* Finally, perform the actual write. */
1976         last_vcn = -1;
1977         if (likely(nr_segs == 1))
1978                 buf = iov->iov_base;
1979         do {
1980                 VCN vcn;
1981                 pgoff_t idx, start_idx;
1982                 unsigned ofs, do_pages, u;
1983                 size_t copied;
1984
1985                 start_idx = idx = pos >> PAGE_CACHE_SHIFT;
1986                 ofs = pos & ~PAGE_CACHE_MASK;
1987                 bytes = PAGE_CACHE_SIZE - ofs;
1988                 do_pages = 1;
1989                 if (nr_pages > 1) {
1990                         vcn = pos >> vol->cluster_size_bits;
1991                         if (vcn != last_vcn) {
1992                                 last_vcn = vcn;
1993                                 /*
1994                                  * Get the lcn of the vcn the write is in.  If
1995                                  * it is a hole, need to lock down all pages in
1996                                  * the cluster.
1997                                  */
1998                                 down_read(&ni->runlist.lock);
1999                                 lcn = ntfs_attr_vcn_to_lcn_nolock(ni, pos >>
2000                                                 vol->cluster_size_bits, false);
2001                                 up_read(&ni->runlist.lock);
2002                                 if (unlikely(lcn < LCN_HOLE)) {
2003                                         status = -EIO;
2004                                         if (lcn == LCN_ENOMEM)
2005                                                 status = -ENOMEM;
2006                                         else
2007                                                 ntfs_error(vol->sb, "Cannot "
2008                                                         "perform write to "
2009                                                         "inode 0x%lx, "
2010                                                         "attribute type 0x%x, "
2011                                                         "because the attribute "
2012                                                         "is corrupt.",
2013                                                         vi->i_ino, (unsigned)
2014                                                         le32_to_cpu(ni->type));
2015                                         break;
2016                                 }
2017                                 if (lcn == LCN_HOLE) {
2018                                         start_idx = (pos & ~(s64)
2019                                                         vol->cluster_size_mask)
2020                                                         >> PAGE_CACHE_SHIFT;
2021                                         bytes = vol->cluster_size - (pos &
2022                                                         vol->cluster_size_mask);
2023                                         do_pages = nr_pages;
2024                                 }
2025                         }
2026                 }
2027                 if (bytes > count)
2028                         bytes = count;
2029                 /*
2030                  * Bring in the user page(s) that we will copy from _first_.
2031                  * Otherwise there is a nasty deadlock on copying from the same
2032                  * page(s) as we are writing to, without it/them being marked
2033                  * up-to-date.  Note, at present there is nothing to stop the
2034                  * pages being swapped out between us bringing them into memory
2035                  * and doing the actual copying.
2036                  */
2037                 if (likely(nr_segs == 1))
2038                         ntfs_fault_in_pages_readable(buf, bytes);
2039                 else
2040                         ntfs_fault_in_pages_readable_iovec(iov, iov_ofs, bytes);
2041                 /* Get and lock @do_pages starting at index @start_idx. */
2042                 status = __ntfs_grab_cache_pages(mapping, start_idx, do_pages,
2043                                 pages, &cached_page, &lru_pvec);
2044                 if (unlikely(status))
2045                         break;
2046                 /*
2047                  * For non-resident attributes, we need to fill any holes with
2048                  * actual clusters and ensure all bufferes are mapped.  We also
2049                  * need to bring uptodate any buffers that are only partially
2050                  * being written to.
2051                  */
2052                 if (NInoNonResident(ni)) {
2053                         status = ntfs_prepare_pages_for_non_resident_write(
2054                                         pages, do_pages, pos, bytes);
2055                         if (unlikely(status)) {
2056                                 loff_t i_size;
2057
2058                                 do {
2059                                         unlock_page(pages[--do_pages]);
2060                                         page_cache_release(pages[do_pages]);
2061                                 } while (do_pages);
2062                                 /*
2063                                  * The write preparation may have instantiated
2064                                  * allocated space outside i_size.  Trim this
2065                                  * off again.  We can ignore any errors in this
2066                                  * case as we will just be waisting a bit of
2067                                  * allocated space, which is not a disaster.
2068                                  */
2069                                 i_size = i_size_read(vi);
2070                                 if (pos + bytes > i_size)
2071                                         vmtruncate(vi, i_size);
2072                                 break;
2073                         }
2074                 }
2075                 u = (pos >> PAGE_CACHE_SHIFT) - pages[0]->index;
2076                 if (likely(nr_segs == 1)) {
2077                         copied = ntfs_copy_from_user(pages + u, do_pages - u,
2078                                         ofs, buf, bytes);
2079                         buf += copied;
2080                 } else
2081                         copied = ntfs_copy_from_user_iovec(pages + u,
2082                                         do_pages - u, ofs, &iov, &iov_ofs,
2083                                         bytes);
2084                 ntfs_flush_dcache_pages(pages + u, do_pages - u);
2085                 status = ntfs_commit_pages_after_write(pages, do_pages, pos,
2086                                 bytes);
2087                 if (likely(!status)) {
2088                         written += copied;
2089                         count -= copied;
2090                         pos += copied;
2091                         if (unlikely(copied != bytes))
2092                                 status = -EFAULT;
2093                 }
2094                 do {
2095                         unlock_page(pages[--do_pages]);
2096                         mark_page_accessed(pages[do_pages]);
2097                         page_cache_release(pages[do_pages]);
2098                 } while (do_pages);
2099                 if (unlikely(status))
2100                         break;
2101                 balance_dirty_pages_ratelimited(mapping);
2102                 cond_resched();
2103         } while (count);
2104 err_out:
2105         *ppos = pos;
2106         if (cached_page)
2107                 page_cache_release(cached_page);
2108         /* For now, when the user asks for O_SYNC, we actually give O_DSYNC. */
2109         if (likely(!status)) {
2110                 if (unlikely((file->f_flags & O_SYNC) || IS_SYNC(vi))) {
2111                         if (!mapping->a_ops->writepage || !is_sync_kiocb(iocb))
2112                                 status = generic_osync_inode(vi, mapping,
2113                                                 OSYNC_METADATA|OSYNC_DATA);
2114                 }
2115         }
2116         pagevec_lru_add(&lru_pvec);
2117         ntfs_debug("Done.  Returning %s (written 0x%lx, status %li).",
2118                         written ? "written" : "status", (unsigned long)written,
2119                         (long)status);
2120         return written ? written : status;
2121 }
2122
2123 /**
2124  * ntfs_file_aio_write_nolock -
2125  */
2126 static ssize_t ntfs_file_aio_write_nolock(struct kiocb *iocb,
2127                 const struct iovec *iov, unsigned long nr_segs, loff_t *ppos)
2128 {
2129         struct file *file = iocb->ki_filp;
2130         struct address_space *mapping = file->f_mapping;
2131         struct inode *inode = mapping->host;
2132         loff_t pos;
2133         unsigned long seg;
2134         size_t count;           /* after file limit checks */
2135         ssize_t written, err;
2136
2137         count = 0;
2138         for (seg = 0; seg < nr_segs; seg++) {
2139                 const struct iovec *iv = &iov[seg];
2140                 /*
2141                  * If any segment has a negative length, or the cumulative
2142                  * length ever wraps negative then return -EINVAL.
2143                  */
2144                 count += iv->iov_len;
2145                 if (unlikely((ssize_t)(count|iv->iov_len) < 0))
2146                         return -EINVAL;
2147                 if (access_ok(VERIFY_READ, iv->iov_base, iv->iov_len))
2148                         continue;
2149                 if (!seg)
2150                         return -EFAULT;
2151                 nr_segs = seg;
2152                 count -= iv->iov_len;   /* This segment is no good */
2153                 break;
2154         }
2155         pos = *ppos;
2156         vfs_check_frozen(inode->i_sb, SB_FREEZE_WRITE);
2157         /* We can write back this queue in page reclaim. */
2158         current->backing_dev_info = mapping->backing_dev_info;
2159         written = 0;
2160         err = generic_write_checks(file, &pos, &count, S_ISBLK(inode->i_mode));
2161         if (err)
2162                 goto out;
2163         if (!count)
2164                 goto out;
2165         err = remove_suid(file->f_dentry);
2166         if (err)
2167                 goto out;
2168         file_update_time(file);
2169         written = ntfs_file_buffered_write(iocb, iov, nr_segs, pos, ppos,
2170                         count);
2171 out:
2172         current->backing_dev_info = NULL;
2173         return written ? written : err;
2174 }
2175
2176 /**
2177  * ntfs_file_aio_write -
2178  */
2179 static ssize_t ntfs_file_aio_write(struct kiocb *iocb, const struct iovec *iov,
2180                 unsigned long nr_segs, loff_t pos)
2181 {
2182         struct file *file = iocb->ki_filp;
2183         struct address_space *mapping = file->f_mapping;
2184         struct inode *inode = mapping->host;
2185         ssize_t ret;
2186
2187         BUG_ON(iocb->ki_pos != pos);
2188
2189         mutex_lock(&inode->i_mutex);
2190         ret = ntfs_file_aio_write_nolock(iocb, iov, nr_segs, &iocb->ki_pos);
2191         mutex_unlock(&inode->i_mutex);
2192         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2193                 int err = sync_page_range(inode, mapping, pos, ret);
2194                 if (err < 0)
2195                         ret = err;
2196         }
2197         return ret;
2198 }
2199
2200 /**
2201  * ntfs_file_writev -
2202  *
2203  * Basically the same as generic_file_writev() except that it ends up calling
2204  * ntfs_file_aio_write_nolock() instead of __generic_file_aio_write_nolock().
2205  */
2206 static ssize_t ntfs_file_writev(struct file *file, const struct iovec *iov,
2207                 unsigned long nr_segs, loff_t *ppos)
2208 {
2209         struct address_space *mapping = file->f_mapping;
2210         struct inode *inode = mapping->host;
2211         struct kiocb kiocb;
2212         ssize_t ret;
2213
2214         mutex_lock(&inode->i_mutex);
2215         init_sync_kiocb(&kiocb, file);
2216         ret = ntfs_file_aio_write_nolock(&kiocb, iov, nr_segs, ppos);
2217         if (ret == -EIOCBQUEUED)
2218                 ret = wait_on_sync_kiocb(&kiocb);
2219         mutex_unlock(&inode->i_mutex);
2220         if (ret > 0 && ((file->f_flags & O_SYNC) || IS_SYNC(inode))) {
2221                 int err = sync_page_range(inode, mapping, *ppos - ret, ret);
2222                 if (err < 0)
2223                         ret = err;
2224         }
2225         return ret;
2226 }
2227
2228 /**
2229  * ntfs_file_write - simple wrapper for ntfs_file_writev()
2230  */
2231 static ssize_t ntfs_file_write(struct file *file, const char __user *buf,
2232                 size_t count, loff_t *ppos)
2233 {
2234         struct iovec local_iov = { .iov_base = (void __user *)buf,
2235                                    .iov_len = count };
2236
2237         return ntfs_file_writev(file, &local_iov, 1, ppos);
2238 }
2239
2240 /**
2241  * ntfs_file_fsync - sync a file to disk
2242  * @filp:       file to be synced
2243  * @dentry:     dentry describing the file to sync
2244  * @datasync:   if non-zero only flush user data and not metadata
2245  *
2246  * Data integrity sync of a file to disk.  Used for fsync, fdatasync, and msync
2247  * system calls.  This function is inspired by fs/buffer.c::file_fsync().
2248  *
2249  * If @datasync is false, write the mft record and all associated extent mft
2250  * records as well as the $DATA attribute and then sync the block device.
2251  *
2252  * If @datasync is true and the attribute is non-resident, we skip the writing
2253  * of the mft record and all associated extent mft records (this might still
2254  * happen due to the write_inode_now() call).
2255  *
2256  * Also, if @datasync is true, we do not wait on the inode to be written out
2257  * but we always wait on the page cache pages to be written out.
2258  *
2259  * Note: In the past @filp could be NULL so we ignore it as we don't need it
2260  * anyway.
2261  *
2262  * Locking: Caller must hold i_mutex on the inode.
2263  *
2264  * TODO: We should probably also write all attribute/index inodes associated
2265  * with this inode but since we have no simple way of getting to them we ignore
2266  * this problem for now.
2267  */
2268 static int ntfs_file_fsync(struct file *filp, struct dentry *dentry,
2269                 int datasync)
2270 {
2271         struct inode *vi = dentry->d_inode;
2272         int err, ret = 0;
2273
2274         ntfs_debug("Entering for inode 0x%lx.", vi->i_ino);
2275         BUG_ON(S_ISDIR(vi->i_mode));
2276         if (!datasync || !NInoNonResident(NTFS_I(vi)))
2277                 ret = ntfs_write_inode(vi, 1);
2278         write_inode_now(vi, !datasync);
2279         /*
2280          * NOTE: If we were to use mapping->private_list (see ext2 and
2281          * fs/buffer.c) for dirty blocks then we could optimize the below to be
2282          * sync_mapping_buffers(vi->i_mapping).
2283          */
2284         err = sync_blockdev(vi->i_sb->s_bdev);
2285         if (unlikely(err && !ret))
2286                 ret = err;
2287         if (likely(!ret))
2288                 ntfs_debug("Done.");
2289         else
2290                 ntfs_warning(vi->i_sb, "Failed to f%ssync inode 0x%lx.  Error "
2291                                 "%u.", datasync ? "data" : "", vi->i_ino, -ret);
2292         return ret;
2293 }
2294
2295 #endif /* NTFS_RW */
2296
2297 const struct file_operations ntfs_file_ops = {
2298         .llseek         = generic_file_llseek,   /* Seek inside file. */
2299         .read           = generic_file_read,     /* Read from file. */
2300         .aio_read       = generic_file_aio_read, /* Async read from file. */
2301         .readv          = generic_file_readv,    /* Read from file. */
2302 #ifdef NTFS_RW
2303         .write          = ntfs_file_write,       /* Write to file. */
2304         .aio_write      = ntfs_file_aio_write,   /* Async write to file. */
2305         .writev         = ntfs_file_writev,      /* Write to file. */
2306         /*.release      = ,*/                    /* Last file is closed.  See
2307                                                     fs/ext2/file.c::
2308                                                     ext2_release_file() for
2309                                                     how to use this to discard
2310                                                     preallocated space for
2311                                                     write opened files. */
2312         .fsync          = ntfs_file_fsync,       /* Sync a file to disk. */
2313         /*.aio_fsync    = ,*/                    /* Sync all outstanding async
2314                                                     i/o operations on a
2315                                                     kiocb. */
2316 #endif /* NTFS_RW */
2317         /*.ioctl        = ,*/                    /* Perform function on the
2318                                                     mounted filesystem. */
2319         .mmap           = generic_file_mmap,     /* Mmap file. */
2320         .open           = ntfs_file_open,        /* Open file. */
2321         .sendfile       = generic_file_sendfile, /* Zero-copy data send with
2322                                                     the data source being on
2323                                                     the ntfs partition.  We do
2324                                                     not need to care about the
2325                                                     data destination. */
2326         /*.sendpage     = ,*/                    /* Zero-copy data send with
2327                                                     the data destination being
2328                                                     on the ntfs partition.  We
2329                                                     do not need to care about
2330                                                     the data source. */
2331 };
2332
2333 struct inode_operations ntfs_file_inode_ops = {
2334 #ifdef NTFS_RW
2335         .truncate       = ntfs_truncate_vfs,
2336         .setattr        = ntfs_setattr,
2337 #endif /* NTFS_RW */
2338 };
2339
2340 const struct file_operations ntfs_empty_file_ops = {};
2341
2342 struct inode_operations ntfs_empty_inode_ops = {};