1 /**
2 * aops.c - NTFS kernel address space operations and page cache handling.
3 * Part of the Linux-NTFS project.
4 *
5 * Copyright (c) 2001-2007 Anton Altaparmakov
6 * Copyright (c) 2002 Richard Russon
7 *
8 * This program/include file is free software; you can redistribute it and/or
9 * modify it under the terms of the GNU General Public License as published
10 * by the Free Software Foundation; either version 2 of the License, or
11 * (at your option) any later version.
12 *
13 * This program/include file is distributed in the hope that it will be
14 * useful, but WITHOUT ANY WARRANTY; without even the implied warranty
15 * of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
16 * GNU General Public License for more details.
17 *
18 * You should have received a copy of the GNU General Public License
19 * along with this program (in the main directory of the Linux-NTFS
20 * distribution in the file COPYING); if not, write to the Free Software
21 * Foundation,Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
22 */
23
24 #include <linux/errno.h>
25 #include <linux/fs.h>
26 #include <linux/mm.h>
27 #include <linux/pagemap.h>
28 #include <linux/swap.h>
29 #include <linux/buffer_head.h>
30 #include <linux/writeback.h>
31 #include <linux/bit_spinlock.h>
32
33 #include "aops.h"
34 #include "attrib.h"
35 #include "debug.h"
36 #include "inode.h"
37 #include "mft.h"
38 #include "runlist.h"
39 #include "types.h"
40 #include "ntfs.h"
41
42 /**
43 * ntfs_end_buffer_async_read - async io completion for reading attributes
44 * @bh: buffer head on which io is completed
45 * @uptodate: whether @bh is now uptodate or not
46 *
47 * Asynchronous I/O completion handler for reading pages belonging to the
48 * attribute address space of an inode. The inodes can either be files or
49 * directories or they can be fake inodes describing some attribute.
50 *
51 * If NInoMstProtected(), perform the post read mst fixups when all IO on the
52 * page has been completed and mark the page uptodate or set the error bit on
53 * the page. To determine the size of the records that need fixing up, we
54 * cheat a little bit by setting the index_block_size in ntfs_inode to the ntfs
55 * record size, and index_block_size_bits, to the log(base 2) of the ntfs
56 * record size.
57 */
58 static void ntfs_end_buffer_async_read(struct buffer_head *bh, int uptodate)
59 {
60 unsigned long flags;
61 struct buffer_head *first, *tmp;
62 struct page *page;
63 struct inode *vi;
64 ntfs_inode *ni;
65 int page_uptodate = 1;
66
67 page = bh->b_page;
68 vi = page->mapping->host;
69 ni = NTFS_I(vi);
70
71 if (likely(uptodate)) {
72 loff_t i_size;
73 s64 file_ofs, init_size;
74
75 set_buffer_uptodate(bh);
76
77 file_ofs = ((s64)page->index << PAGE_CACHE_SHIFT) +
78 bh_offset(bh);
79 read_lock_irqsave(&ni->size_lock, flags);
80 init_size = ni->initialized_size;
81 i_size = i_size_read(vi);
82 read_unlock_irqrestore(&ni->size_lock, flags);
83 if (unlikely(init_size > i_size)) {
84 /* Race with shrinking truncate. */
85 init_size = i_size;
86 }
87 /* Check for the current buffer head overflowing. */
88 if (unlikely(file_ofs + bh->b_size > init_size)) {
89 int ofs;
90 void *kaddr;
91
92 ofs = 0;
93 if (file_ofs < init_size)
94 ofs = init_size - file_ofs;
95 local_irq_save(flags);
96 kaddr = kmap_atomic(page, KM_BIO_SRC_IRQ);
97 memset(kaddr + bh_offset(bh) + ofs, 0,
98 bh->b_size - ofs);
99 flush_dcache_page(page);
100 kunmap_atomic(kaddr, KM_BIO_SRC_IRQ);
101 local_irq_restore(flags);
102 }
103 } else {
104 clear_buffer_uptodate(bh);
105 SetPageError(page);
106 ntfs_error(ni->vol->sb, "Buffer I/O error, logical block "
107 "0x%llx.", (unsigned long long)bh->b_blocknr);
108 }
109 first = page_buffers(page);
110 spin_lock_irqsave(&first->b_uptodate_lock, flags);
111 clear_buffer_async_read(bh);
112 unlock_buffer(bh);
113 tmp = bh;
114 do {
115 if (!buffer_uptodate(tmp))
116 page_uptodate = 0;
117 if (buffer_async_read(tmp)) {
118 if (likely(buffer_locked(tmp)))
119 goto still_busy;
120 /* Async buffers must be locked. */
121 BUG();
122 }
123 tmp = tmp->b_this_page;
124 } while (tmp != bh);
125 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
126 /*
127 * If none of the buffers had errors then we can set the page uptodate,
128 * but we first have to perform the post read mst fixups, if the
129 * attribute is mst protected, i.e. if NInoMstProteced(ni) is true.
130 * Note we ignore fixup errors as those are detected when
131 * map_mft_record() is called which gives us per record granularity
132 * rather than per page granularity.
133 */
134 if (!NInoMstProtected(ni)) {
135 if (likely(page_uptodate && !PageError(page)))
136 SetPageUptodate(page);
137 } else {
138 u8 *kaddr;
139 unsigned int i, recs;
140 u32 rec_size;
141
142 rec_size = ni->itype.index.block_size;
143 recs = PAGE_CACHE_SIZE / rec_size;
144 /* Should have been verified before we got here... */
145 BUG_ON(!recs);
146 local_irq_save_nort(flags);
147 kaddr = kmap_atomic(page, KM_BIO_SRC_IRQ);
148 for (i = 0; i < recs; i++)
149 post_read_mst_fixup((NTFS_RECORD*)(kaddr +
150 i * rec_size), rec_size);
151 kunmap_atomic(kaddr, KM_BIO_SRC_IRQ);
152 local_irq_restore_nort(flags);
153 flush_dcache_page(page);
154 if (likely(page_uptodate && !PageError(page)))
155 SetPageUptodate(page);
156 }
157 unlock_page(page);
158 return;
159 still_busy:
160 spin_unlock_irqrestore(&first->b_uptodate_lock, flags);
161 return;
162 }
163
164 /**
165 * ntfs_read_block - fill a @page of an address space with data
166 * @page: page cache page to fill with data
167 *
168 * Fill the page @page of the address space belonging to the @page->host inode.
169 * We read each buffer asynchronously and when all buffers are read in, our io
170 * completion handler ntfs_end_buffer_read_async(), if required, automatically
171 * applies the mst fixups to the page before finally marking it uptodate and
172 * unlocking it.
173 *
174 * We only enforce allocated_size limit because i_size is checked for in
175 * generic_file_read().
176 *
177 * Return 0 on success and -errno on error.
178 *
179 * Contains an adapted version of fs/buffer.c::block_read_full_page().
180 */
181 static int ntfs_read_block(struct page *page)
182 {
183 loff_t i_size;
184 VCN vcn;
185 LCN lcn;
186 s64 init_size;
187 struct inode *vi;
188 ntfs_inode *ni;
189 ntfs_volume *vol;
190 runlist_element *rl;
191 struct buffer_head *bh, *head, *arr[MAX_BUF_PER_PAGE];
192 sector_t iblock, lblock, zblock;
193 unsigned long flags;
194 unsigned int blocksize, vcn_ofs;
195 int i, nr;
196 unsigned char blocksize_bits;
197
198 vi = page->mapping->host;
199 ni = NTFS_I(vi);
200 vol = ni->vol;
201
202 /* $MFT/$DATA must have its complete runlist in memory at all times. */
203 BUG_ON(!ni->runlist.rl && !ni->mft_no && !NInoAttr(ni));
204
205 blocksize = vol->sb->s_blocksize;
206 blocksize_bits = vol->sb->s_blocksize_bits;
207
208 if (!page_has_buffers(page)) {
209 create_empty_buffers(page, blocksize, 0);
210 if (unlikely(!page_has_buffers(page))) {
211 unlock_page(page);
212 return -ENOMEM;
213 }
214 }
215 bh = head = page_buffers(page);
216 BUG_ON(!bh);
217
218 /*
219 * We may be racing with truncate. To avoid some of the problems we
220 * now take a snapshot of the various sizes and use those for the whole
221 * of the function. In case of an extending truncate it just means we
222 * may leave some buffers unmapped which are now allocated. This is
223 * not a problem since these buffers will just get mapped when a write
224 * occurs. In case of a shrinking truncate, we will detect this later
225 * on due to the runlist being incomplete and if the page is being
226 * fully truncated, truncate will throw it away as soon as we unlock
227 * it so no need to worry what we do with it.
228 */
229 iblock = (s64)page->index << (PAGE_CACHE_SHIFT - blocksize_bits);
230 read_lock_irqsave(&ni->size_lock, flags);
231 lblock = (ni->allocated_size + blocksize - 1) >> blocksize_bits;
232 init_size = ni->initialized_size;
233 i_size = i_size_read(vi);
234 read_unlock_irqrestore(&ni->size_lock, flags);
235 if (unlikely(init_size > i_size)) {
236 /* Race with shrinking truncate. */
237 init_size = i_size;
238 }
239 zblock = (init_size + blocksize - 1) >> blocksize_bits;
240
241 /* Loop through all the buffers in the page. */
242 rl = NULL;
243 nr = i = 0;
244 do {
245 int err = 0;
246
247 if (unlikely(buffer_uptodate(bh)))
248 continue;
249 if (unlikely(buffer_mapped(bh))) {
250 arr[nr++] = bh;
251 continue;
252 }
253 bh->b_bdev = vol->sb->s_bdev;
254 /* Is the block within the allowed limits? */
255 if (iblock < lblock) {
256 bool is_retry = false;
257
258 /* Convert iblock into corresponding vcn and offset. */
259 vcn = (VCN)iblock << blocksize_bits >>
260 vol->cluster_size_bits;
261 vcn_ofs = ((VCN)iblock << blocksize_bits) &
262 vol->cluster_size_mask;
263 if (!rl) {
264 lock_retry_remap:
265 down_read(&ni->runlist.lock);
266 rl = ni->runlist.rl;
267 }
268 if (likely(rl != NULL)) {
269 /* Seek to element containing target vcn. */
270 while (rl->length && rl[1].vcn <= vcn)
271 rl++;
272 lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
273 } else
274 lcn = LCN_RL_NOT_MAPPED;
275 /* Successful remap. */
276 if (lcn >= 0) {
277 /* Setup buffer head to correct block. */
278 bh->b_blocknr = ((lcn << vol->cluster_size_bits)
279 + vcn_ofs) >> blocksize_bits;
280 set_buffer_mapped(bh);
281 /* Only read initialized data blocks. */
282 if (iblock < zblock) {
283 arr[nr++] = bh;
284 continue;
285 }
286 /* Fully non-initialized data block, zero it. */
287 goto handle_zblock;
288 }
289 /* It is a hole, need to zero it. */
290 if (lcn == LCN_HOLE)
291 goto handle_hole;
292 /* If first try and runlist unmapped, map and retry. */
293 if (!is_retry && lcn == LCN_RL_NOT_MAPPED) {
294 is_retry = true;
295 /*
296 * Attempt to map runlist, dropping lock for
297 * the duration.
298 */
299 up_read(&ni->runlist.lock);
300 err = ntfs_map_runlist(ni, vcn);
301 if (likely(!err))
302 goto lock_retry_remap;
303 rl = NULL;
304 } else if (!rl)
305 up_read(&ni->runlist.lock);
306 /*
307 * If buffer is outside the runlist, treat it as a
308 * hole. This can happen due to concurrent truncate
309 * for example.
310 */
311 if (err == -ENOENT || lcn == LCN_ENOENT) {
312 err = 0;
313 goto handle_hole;
314 }
315 /* Hard error, zero out region. */
316 if (!err)
317 err = -EIO;
318 bh->b_blocknr = -1;
319 SetPageError(page);
320 ntfs_error(vol->sb, "Failed to read from inode 0x%lx, "
321 "attribute type 0x%x, vcn 0x%llx, "
322 "offset 0x%x because its location on "
323 "disk could not be determined%s "
324 "(error code %i).", ni->mft_no,
325 ni->type, (unsigned long long)vcn,
326 vcn_ofs, is_retry ? " even after "
327 "retrying" : "", err);
328 }
329 /*
330 * Either iblock was outside lblock limits or
331 * ntfs_rl_vcn_to_lcn() returned error. Just zero that portion
332 * of the page and set the buffer uptodate.
333 */
334 handle_hole:
335 bh->b_blocknr = -1UL;
336 clear_buffer_mapped(bh);
337 handle_zblock:
338 zero_user(page, i * blocksize, blocksize);
339 if (likely(!err))
340 set_buffer_uptodate(bh);
341 } while (i++, iblock++, (bh = bh->b_this_page) != head);
342
343 /* Release the lock if we took it. */
344 if (rl)
345 up_read(&ni->runlist.lock);
346
347 /* Check we have at least one buffer ready for i/o. */
348 if (nr) {
349 struct buffer_head *tbh;
350
351 /* Lock the buffers. */
352 for (i = 0; i < nr; i++) {
353 tbh = arr[i];
354 lock_buffer(tbh);
355 tbh->b_end_io = ntfs_end_buffer_async_read;
356 set_buffer_async_read(tbh);
357 }
358 /* Finally, start i/o on the buffers. */
359 for (i = 0; i < nr; i++) {
360 tbh = arr[i];
361 if (likely(!buffer_uptodate(tbh)))
362 submit_bh(READ, tbh);
363 else
364 ntfs_end_buffer_async_read(tbh, 1);
365 }
366 return 0;
367 }
368 /* No i/o was scheduled on any of the buffers. */
369 if (likely(!PageError(page)))
370 SetPageUptodate(page);
371 else /* Signal synchronous i/o error. */
372 nr = -EIO;
373 unlock_page(page);
374 return nr;
375 }
376
377 /**
378 * ntfs_readpage - fill a @page of a @file with data from the device
379 * @file: open file to which the page @page belongs or NULL
380 * @page: page cache page to fill with data
381 *
382 * For non-resident attributes, ntfs_readpage() fills the @page of the open
383 * file @file by calling the ntfs version of the generic block_read_full_page()
384 * function, ntfs_read_block(), which in turn creates and reads in the buffers
385 * associated with the page asynchronously.
386 *
387 * For resident attributes, OTOH, ntfs_readpage() fills @page by copying the
388 * data from the mft record (which at this stage is most likely in memory) and
389 * fills the remainder with zeroes. Thus, in this case, I/O is synchronous, as
390 * even if the mft record is not cached at this point in time, we need to wait
391 * for it to be read in before we can do the copy.
392 *
393 * Return 0 on success and -errno on error.
394 */
395 static int ntfs_readpage(struct file *file, struct page *page)
396 {
397 loff_t i_size;
398 struct inode *vi;
399 ntfs_inode *ni, *base_ni;
400 u8 *addr;
401 ntfs_attr_search_ctx *ctx;
402 MFT_RECORD *mrec;
403 unsigned long flags;
404 u32 attr_len;
405 int err = 0;
406
407 retry_readpage:
408 BUG_ON(!PageLocked(page));
409 vi = page->mapping->host;
410 i_size = i_size_read(vi);
411 /* Is the page fully outside i_size? (truncate in progress) */
412 if (unlikely(page->index >= (i_size + PAGE_CACHE_SIZE - 1) >>
413 PAGE_CACHE_SHIFT)) {
414 zero_user(page, 0, PAGE_CACHE_SIZE);
415 ntfs_debug("Read outside i_size - truncated?");
416 goto done;
417 }
418 /*
419 * This can potentially happen because we clear PageUptodate() during
420 * ntfs_writepage() of MstProtected() attributes.
421 */
422 if (PageUptodate(page)) {
423 unlock_page(page);
424 return 0;
425 }
426 ni = NTFS_I(vi);
427 /*
428 * Only $DATA attributes can be encrypted and only unnamed $DATA
429 * attributes can be compressed. Index root can have the flags set but
430 * this means to create compressed/encrypted files, not that the
431 * attribute is compressed/encrypted. Note we need to check for
432 * AT_INDEX_ALLOCATION since this is the type of both directory and
433 * index inodes.
434 */
435 if (ni->type != AT_INDEX_ALLOCATION) {
436 /* If attribute is encrypted, deny access, just like NT4. */
437 if (NInoEncrypted(ni)) {
438 BUG_ON(ni->type != AT_DATA);
439 err = -EACCES;
440 goto err_out;
441 }
442 /* Compressed data streams are handled in compress.c. */
443 if (NInoNonResident(ni) && NInoCompressed(ni)) {
444 BUG_ON(ni->type != AT_DATA);
445 BUG_ON(ni->name_len);
446 return ntfs_read_compressed_block(page);
447 }
448 }
449 /* NInoNonResident() == NInoIndexAllocPresent() */
450 if (NInoNonResident(ni)) {
451 /* Normal, non-resident data stream. */
452 return ntfs_read_block(page);
453 }
454 /*
455 * Attribute is resident, implying it is not compressed or encrypted.
456 * This also means the attribute is smaller than an mft record and
457 * hence smaller than a page, so can simply zero out any pages with
458 * index above 0. Note the attribute can actually be marked compressed
459 * but if it is resident the actual data is not compressed so we are
460 * ok to ignore the compressed flag here.
461 */
462 if (unlikely(page->index > 0)) {
463 zero_user(page, 0, PAGE_CACHE_SIZE);
464 goto done;
465 }
466 if (!NInoAttr(ni))
467 base_ni = ni;
468 else
469 base_ni = ni->ext.base_ntfs_ino;
470 /* Map, pin, and lock the mft record. */
471 mrec = map_mft_record(base_ni);
472 if (IS_ERR(mrec)) {
473 err = PTR_ERR(mrec);
474 goto err_out;
475 }
476 /*
477 * If a parallel write made the attribute non-resident, drop the mft
478 * record and retry the readpage.
479 */
480 if (unlikely(NInoNonResident(ni))) {
481 unmap_mft_record(base_ni);
482 goto retry_readpage;
483 }
484 ctx = ntfs_attr_get_search_ctx(base_ni, mrec);
485 if (unlikely(!ctx)) {
486 err = -ENOMEM;
487 goto unm_err_out;
488 }
489 err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
490 CASE_SENSITIVE, 0, NULL, 0, ctx);
491 if (unlikely(err))
492 goto put_unm_err_out;
493 attr_len = le32_to_cpu(ctx->attr->data.resident.value_length);
494 read_lock_irqsave(&ni->size_lock, flags);
495 if (unlikely(attr_len > ni->initialized_size))
496 attr_len = ni->initialized_size;
497 i_size = i_size_read(vi);
498 read_unlock_irqrestore(&ni->size_lock, flags);
499 if (unlikely(attr_len > i_size)) {
500 /* Race with shrinking truncate. */
501 attr_len = i_size;
502 }
503 addr = kmap_atomic(page, KM_USER0);
504 /* Copy the data to the page. */
505 memcpy(addr, (u8*)ctx->attr +
506 le16_to_cpu(ctx->attr->data.resident.value_offset),
507 attr_len);
508 /* Zero the remainder of the page. */
509 memset(addr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
510 flush_dcache_page(page);
511 kunmap_atomic(addr, KM_USER0);
512 put_unm_err_out:
513 ntfs_attr_put_search_ctx(ctx);
514 unm_err_out:
515 unmap_mft_record(base_ni);
516 done:
517 SetPageUptodate(page);
518 err_out:
519 unlock_page(page);
520 return err;
521 }
522
523 #ifdef NTFS_RW
524
525 /**
526 * ntfs_write_block - write a @page to the backing store
527 * @page: page cache page to write out
528 * @wbc: writeback control structure
529 *
530 * This function is for writing pages belonging to non-resident, non-mst
531 * protected attributes to their backing store.
532 *
533 * For a page with buffers, map and write the dirty buffers asynchronously
534 * under page writeback. For a page without buffers, create buffers for the
535 * page, then proceed as above.
536 *
537 * If a page doesn't have buffers the page dirty state is definitive. If a page
538 * does have buffers, the page dirty state is just a hint, and the buffer dirty
539 * state is definitive. (A hint which has rules: dirty buffers against a clean
540 * page is illegal. Other combinations are legal and need to be handled. In
541 * particular a dirty page containing clean buffers for example.)
542 *
543 * Return 0 on success and -errno on error.
544 *
545 * Based on ntfs_read_block() and __block_write_full_page().
546 */
547 static int ntfs_write_block(struct page *page, struct writeback_control *wbc)
548 {
549 VCN vcn;
550 LCN lcn;
551 s64 initialized_size;
552 loff_t i_size;
553 sector_t block, dblock, iblock;
554 struct inode *vi;
555 ntfs_inode *ni;
556 ntfs_volume *vol;
557 runlist_element *rl;
558 struct buffer_head *bh, *head;
559 unsigned long flags;
560 unsigned int blocksize, vcn_ofs;
561 int err;
562 bool need_end_writeback;
563 unsigned char blocksize_bits;
564
565 vi = page->mapping->host;
566 ni = NTFS_I(vi);
567 vol = ni->vol;
568
569 ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
570 "0x%lx.", ni->mft_no, ni->type, page->index);
571
572 BUG_ON(!NInoNonResident(ni));
573 BUG_ON(NInoMstProtected(ni));
574 blocksize = vol->sb->s_blocksize;
575 blocksize_bits = vol->sb->s_blocksize_bits;
576 if (!page_has_buffers(page)) {
577 BUG_ON(!PageUptodate(page));
578 create_empty_buffers(page, blocksize,
579 (1 << BH_Uptodate) | (1 << BH_Dirty));
580 if (unlikely(!page_has_buffers(page))) {
581 ntfs_warning(vol->sb, "Error allocating page "
582 "buffers. Redirtying page so we try "
583 "again later.");
584 /*
585 * Put the page back on mapping->dirty_pages, but leave
586 * its buffers' dirty state as-is.
587 */
588 redirty_page_for_writepage(wbc, page);
589 unlock_page(page);
590 return 0;
591 }
592 }
593 bh = head = page_buffers(page);
594 BUG_ON(!bh);
595
596 /* NOTE: Different naming scheme to ntfs_read_block()! */
597
598 /* The first block in the page. */
599 block = (s64)page->index << (PAGE_CACHE_SHIFT - blocksize_bits);
600
601 read_lock_irqsave(&ni->size_lock, flags);
602 i_size = i_size_read(vi);
603 initialized_size = ni->initialized_size;
604 read_unlock_irqrestore(&ni->size_lock, flags);
605
606 /* The first out of bounds block for the data size. */
607 dblock = (i_size + blocksize - 1) >> blocksize_bits;
608
609 /* The last (fully or partially) initialized block. */
610 iblock = initialized_size >> blocksize_bits;
611
612 /*
613 * Be very careful. We have no exclusion from __set_page_dirty_buffers
614 * here, and the (potentially unmapped) buffers may become dirty at
615 * any time. If a buffer becomes dirty here after we've inspected it
616 * then we just miss that fact, and the page stays dirty.
617 *
618 * Buffers outside i_size may be dirtied by __set_page_dirty_buffers;
619 * handle that here by just cleaning them.
620 */
621
622 /*
623 * Loop through all the buffers in the page, mapping all the dirty
624 * buffers to disk addresses and handling any aliases from the
625 * underlying block device's mapping.
626 */
627 rl = NULL;
628 err = 0;
629 do {
630 bool is_retry = false;
631
632 if (unlikely(block >= dblock)) {
633 /*
634 * Mapped buffers outside i_size will occur, because
635 * this page can be outside i_size when there is a
636 * truncate in progress. The contents of such buffers
637 * were zeroed by ntfs_writepage().
638 *
639 * FIXME: What about the small race window where
640 * ntfs_writepage() has not done any clearing because
641 * the page was within i_size but before we get here,
642 * vmtruncate() modifies i_size?
643 */
644 clear_buffer_dirty(bh);
645 set_buffer_uptodate(bh);
646 continue;
647 }
648
649 /* Clean buffers are not written out, so no need to map them. */
650 if (!buffer_dirty(bh))
651 continue;
652
653 /* Make sure we have enough initialized size. */
654 if (unlikely((block >= iblock) &&
655 (initialized_size < i_size))) {
656 /*
657 * If this page is fully outside initialized size, zero
658 * out all pages between the current initialized size
659 * and the current page. Just use ntfs_readpage() to do
660 * the zeroing transparently.
661 */
662 if (block > iblock) {
663 // TODO:
664 // For each page do:
665 // - read_cache_page()
666 // Again for each page do:
667 // - wait_on_page_locked()
668 // - Check (PageUptodate(page) &&
669 // !PageError(page))
670 // Update initialized size in the attribute and
671 // in the inode.
672 // Again, for each page do:
673 // __set_page_dirty_buffers();
674 // page_cache_release()
675 // We don't need to wait on the writes.
676 // Update iblock.
677 }
678 /*
679 * The current page straddles initialized size. Zero
680 * all non-uptodate buffers and set them uptodate (and
681 * dirty?). Note, there aren't any non-uptodate buffers
682 * if the page is uptodate.
683 * FIXME: For an uptodate page, the buffers may need to
684 * be written out because they were not initialized on
685 * disk before.
686 */
687 if (!PageUptodate(page)) {
688 // TODO:
689 // Zero any non-uptodate buffers up to i_size.
690 // Set them uptodate and dirty.
691 }
692 // TODO:
693 // Update initialized size in the attribute and in the
694 // inode (up to i_size).
695 // Update iblock.
696 // FIXME: This is inefficient. Try to batch the two
697 // size changes to happen in one go.
698 ntfs_error(vol->sb, "Writing beyond initialized size "
699 "is not supported yet. Sorry.");
700 err = -EOPNOTSUPP;
701 break;
702 // Do NOT set_buffer_new() BUT DO clear buffer range
703 // outside write request range.
704 // set_buffer_uptodate() on complete buffers as well as
705 // set_buffer_dirty().
706 }
707
708 /* No need to map buffers that are already mapped. */
709 if (buffer_mapped(bh))
710 continue;
711
712 /* Unmapped, dirty buffer. Need to map it. */
713 bh->b_bdev = vol->sb->s_bdev;
714
715 /* Convert block into corresponding vcn and offset. */
716 vcn = (VCN)block << blocksize_bits;
717 vcn_ofs = vcn & vol->cluster_size_mask;
718 vcn >>= vol->cluster_size_bits;
719 if (!rl) {
720 lock_retry_remap:
721 down_read(&ni->runlist.lock);
722 rl = ni->runlist.rl;
723 }
724 if (likely(rl != NULL)) {
725 /* Seek to element containing target vcn. */
726 while (rl->length && rl[1].vcn <= vcn)
727 rl++;
728 lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
729 } else
730 lcn = LCN_RL_NOT_MAPPED;
731 /* Successful remap. */
732 if (lcn >= 0) {
733 /* Setup buffer head to point to correct block. */
734 bh->b_blocknr = ((lcn << vol->cluster_size_bits) +
735 vcn_ofs) >> blocksize_bits;
736 set_buffer_mapped(bh);
737 continue;
738 }
739 /* It is a hole, need to instantiate it. */
740 if (lcn == LCN_HOLE) {
741 u8 *kaddr;
742 unsigned long *bpos, *bend;
743
744 /* Check if the buffer is zero. */
745 kaddr = kmap_atomic(page, KM_USER0);
746 bpos = (unsigned long *)(kaddr + bh_offset(bh));
747 bend = (unsigned long *)((u8*)bpos + blocksize);
748 do {
749 if (unlikely(*bpos))
750 break;
751 } while (likely(++bpos < bend));
752 kunmap_atomic(kaddr, KM_USER0);
753 if (bpos == bend) {
754 /*
755 * Buffer is zero and sparse, no need to write
756 * it.
757 */
758 bh->b_blocknr = -1;
759 clear_buffer_dirty(bh);
760 continue;
761 }
762 // TODO: Instantiate the hole.
763 // clear_buffer_new(bh);
764 // unmap_underlying_metadata(bh->b_bdev, bh->b_blocknr);
765 ntfs_error(vol->sb, "Writing into sparse regions is "
766 "not supported yet. Sorry.");
767 err = -EOPNOTSUPP;
768 break;
769 }
770 /* If first try and runlist unmapped, map and retry. */
771 if (!is_retry && lcn == LCN_RL_NOT_MAPPED) {
772 is_retry = true;
773 /*
774 * Attempt to map runlist, dropping lock for
775 * the duration.
776 */
777 up_read(&ni->runlist.lock);
778 err = ntfs_map_runlist(ni, vcn);
779 if (likely(!err))
780 goto lock_retry_remap;
781 rl = NULL;
782 } else if (!rl)
783 up_read(&ni->runlist.lock);
784 /*
785 * If buffer is outside the runlist, truncate has cut it out
786 * of the runlist. Just clean and clear the buffer and set it
787 * uptodate so it can get discarded by the VM.
788 */
789 if (err == -ENOENT || lcn == LCN_ENOENT) {
790 bh->b_blocknr = -1;
791 clear_buffer_dirty(bh);
792 zero_user(page, bh_offset(bh), blocksize);
793 set_buffer_uptodate(bh);
794 err = 0;
795 continue;
796 }
797 /* Failed to map the buffer, even after retrying. */
798 if (!err)
799 err = -EIO;
800 bh->b_blocknr = -1;
801 ntfs_error(vol->sb, "Failed to write to inode 0x%lx, "
802 "attribute type 0x%x, vcn 0x%llx, offset 0x%x "
803 "because its location on disk could not be "
804 "determined%s (error code %i).", ni->mft_no,
805 ni->type, (unsigned long long)vcn,
806 vcn_ofs, is_retry ? " even after "
807 "retrying" : "", err);
808 break;
809 } while (block++, (bh = bh->b_this_page) != head);
810
811 /* Release the lock if we took it. */
812 if (rl)
813 up_read(&ni->runlist.lock);
814
815 /* For the error case, need to reset bh to the beginning. */
816 bh = head;
817
818 /* Just an optimization, so ->readpage() is not called later. */
819 if (unlikely(!PageUptodate(page))) {
820 int uptodate = 1;
821 do {
822 if (!buffer_uptodate(bh)) {
823 uptodate = 0;
824 bh = head;
825 break;
826 }
827 } while ((bh = bh->b_this_page) != head);
828 if (uptodate)
829 SetPageUptodate(page);
830 }
831
832 /* Setup all mapped, dirty buffers for async write i/o. */
833 do {
834 if (buffer_mapped(bh) && buffer_dirty(bh)) {
835 lock_buffer(bh);
836 if (test_clear_buffer_dirty(bh)) {
837 BUG_ON(!buffer_uptodate(bh));
838 mark_buffer_async_write(bh);
839 } else
840 unlock_buffer(bh);
841 } else if (unlikely(err)) {
842 /*
843 * For the error case. The buffer may have been set
844 * dirty during attachment to a dirty page.
845 */
846 if (err != -ENOMEM)
847 clear_buffer_dirty(bh);
848 }
849 } while ((bh = bh->b_this_page) != head);
850
851 if (unlikely(err)) {
852 // TODO: Remove the -EOPNOTSUPP check later on...
853 if (unlikely(err == -EOPNOTSUPP))
854 err = 0;
855 else if (err == -ENOMEM) {
856 ntfs_warning(vol->sb, "Error allocating memory. "
857 "Redirtying page so we try again "
858 "later.");
859 /*
860 * Put the page back on mapping->dirty_pages, but
861 * leave its buffer's dirty state as-is.
862 */
863 redirty_page_for_writepage(wbc, page);
864 err = 0;
865 } else
866 SetPageError(page);
867 }
868
869 BUG_ON(PageWriteback(page));
870 set_page_writeback(page); /* Keeps try_to_free_buffers() away. */
871
872 /* Submit the prepared buffers for i/o. */
873 need_end_writeback = true;
874 do {
875 struct buffer_head *next = bh->b_this_page;
876 if (buffer_async_write(bh)) {
877 submit_bh(WRITE, bh);
878 need_end_writeback = false;
879 }
880 bh = next;
881 } while (bh != head);
882 unlock_page(page);
883
884 /* If no i/o was started, need to end_page_writeback(). */
885 if (unlikely(need_end_writeback))
886 end_page_writeback(page);
887
888 ntfs_debug("Done.");
889 return err;
890 }
891
892 /**
893 * ntfs_write_mst_block - write a @page to the backing store
894 * @page: page cache page to write out
895 * @wbc: writeback control structure
896 *
897 * This function is for writing pages belonging to non-resident, mst protected
898 * attributes to their backing store. The only supported attributes are index
899 * allocation and $MFT/$DATA. Both directory inodes and index inodes are
900 * supported for the index allocation case.
901 *
902 * The page must remain locked for the duration of the write because we apply
903 * the mst fixups, write, and then undo the fixups, so if we were to unlock the
904 * page before undoing the fixups, any other user of the page will see the
905 * page contents as corrupt.
906 *
907 * We clear the page uptodate flag for the duration of the function to ensure
908 * exclusion for the $MFT/$DATA case against someone mapping an mft record we
909 * are about to apply the mst fixups to.
910 *
911 * Return 0 on success and -errno on error.
912 *
913 * Based on ntfs_write_block(), ntfs_mft_writepage(), and
914 * write_mft_record_nolock().
915 */
916 static int ntfs_write_mst_block(struct page *page,
917 struct writeback_control *wbc)
918 {
919 sector_t block, dblock, rec_block;
920 struct inode *vi = page->mapping->host;
921 ntfs_inode *ni = NTFS_I(vi);
922 ntfs_volume *vol = ni->vol;
923 u8 *kaddr;
924 unsigned int rec_size = ni->itype.index.block_size;
925 ntfs_inode *locked_nis[PAGE_CACHE_SIZE / rec_size];
926 struct buffer_head *bh, *head, *tbh, *rec_start_bh;
927 struct buffer_head *bhs[MAX_BUF_PER_PAGE];
928 runlist_element *rl;
929 int i, nr_locked_nis, nr_recs, nr_bhs, max_bhs, bhs_per_rec, err, err2;
930 unsigned bh_size, rec_size_bits;
931 bool sync, is_mft, page_is_dirty, rec_is_dirty;
932 unsigned char bh_size_bits;
933
934 ntfs_debug("Entering for inode 0x%lx, attribute type 0x%x, page index "
935 "0x%lx.", vi->i_ino, ni->type, page->index);
936 BUG_ON(!NInoNonResident(ni));
937 BUG_ON(!NInoMstProtected(ni));
938 is_mft = (S_ISREG(vi->i_mode) && !vi->i_ino);
939 /*
940 * NOTE: ntfs_write_mst_block() would be called for $MFTMirr if a page
941 * in its page cache were to be marked dirty. However this should
942 * never happen with the current driver and considering we do not
943 * handle this case here we do want to BUG(), at least for now.
944 */
945 BUG_ON(!(is_mft || S_ISDIR(vi->i_mode) ||
946 (NInoAttr(ni) && ni->type == AT_INDEX_ALLOCATION)));
947 bh_size = vol->sb->s_blocksize;
948 bh_size_bits = vol->sb->s_blocksize_bits;
949 max_bhs = PAGE_CACHE_SIZE / bh_size;
950 BUG_ON(!max_bhs);
951 BUG_ON(max_bhs > MAX_BUF_PER_PAGE);
952
953 /* Were we called for sync purposes? */
954 sync = (wbc->sync_mode == WB_SYNC_ALL);
955
956 /* Make sure we have mapped buffers. */
957 bh = head = page_buffers(page);
958 BUG_ON(!bh);
959
960 rec_size_bits = ni->itype.index.block_size_bits;
961 BUG_ON(!(PAGE_CACHE_SIZE >> rec_size_bits));
962 bhs_per_rec = rec_size >> bh_size_bits;
963 BUG_ON(!bhs_per_rec);
964
965 /* The first block in the page. */
966 rec_block = block = (sector_t)page->index <<
967 (PAGE_CACHE_SHIFT - bh_size_bits);
968
969 /* The first out of bounds block for the data size. */
970 dblock = (i_size_read(vi) + bh_size - 1) >> bh_size_bits;
971
972 rl = NULL;
973 err = err2 = nr_bhs = nr_recs = nr_locked_nis = 0;
974 page_is_dirty = rec_is_dirty = false;
975 rec_start_bh = NULL;
976 do {
977 bool is_retry = false;
978
979 if (likely(block < rec_block)) {
980 if (unlikely(block >= dblock)) {
981 clear_buffer_dirty(bh);
982 set_buffer_uptodate(bh);
983 continue;
984 }
985 /*
986 * This block is not the first one in the record. We
987 * ignore the buffer's dirty state because we could
988 * have raced with a parallel mark_ntfs_record_dirty().
989 */
990 if (!rec_is_dirty)
991 continue;
992 if (unlikely(err2)) {
993 if (err2 != -ENOMEM)
994 clear_buffer_dirty(bh);
995 continue;
996 }
997 } else /* if (block == rec_block) */ {
998 BUG_ON(block > rec_block);
999 /* This block is the first one in the record. */
1000 rec_block += bhs_per_rec;
1001 err2 = 0;
1002 if (unlikely(block >= dblock)) {
1003 clear_buffer_dirty(bh);
1004 continue;
1005 }
1006 if (!buffer_dirty(bh)) {
1007 /* Clean records are not written out. */
1008 rec_is_dirty = false;
1009 continue;
1010 }
1011 rec_is_dirty = true;
1012 rec_start_bh = bh;
1013 }
1014 /* Need to map the buffer if it is not mapped already. */
1015 if (unlikely(!buffer_mapped(bh))) {
1016 VCN vcn;
1017 LCN lcn;
1018 unsigned int vcn_ofs;
1019
1020 bh->b_bdev = vol->sb->s_bdev;
1021 /* Obtain the vcn and offset of the current block. */
1022 vcn = (VCN)block << bh_size_bits;
1023 vcn_ofs = vcn & vol->cluster_size_mask;
1024 vcn >>= vol->cluster_size_bits;
1025 if (!rl) {
1026 lock_retry_remap:
1027 down_read(&ni->runlist.lock);
1028 rl = ni->runlist.rl;
1029 }
1030 if (likely(rl != NULL)) {
1031 /* Seek to element containing target vcn. */
1032 while (rl->length && rl[1].vcn <= vcn)
1033 rl++;
1034 lcn = ntfs_rl_vcn_to_lcn(rl, vcn);
1035 } else
1036 lcn = LCN_RL_NOT_MAPPED;
1037 /* Successful remap. */
1038 if (likely(lcn >= 0)) {
1039 /* Setup buffer head to correct block. */
1040 bh->b_blocknr = ((lcn <<
1041 vol->cluster_size_bits) +
1042 vcn_ofs) >> bh_size_bits;
1043 set_buffer_mapped(bh);
1044 } else {
1045 /*
1046 * Remap failed. Retry to map the runlist once
1047 * unless we are working on $MFT which always
1048 * has the whole of its runlist in memory.
1049 */
1050 if (!is_mft && !is_retry &&
1051 lcn == LCN_RL_NOT_MAPPED) {
1052 is_retry = true;
1053 /*
1054 * Attempt to map runlist, dropping
1055 * lock for the duration.
1056 */
1057 up_read(&ni->runlist.lock);
1058 err2 = ntfs_map_runlist(ni, vcn);
1059 if (likely(!err2))
1060 goto lock_retry_remap;
1061 if (err2 == -ENOMEM)
1062 page_is_dirty = true;
1063 lcn = err2;
1064 } else {
1065 err2 = -EIO;
1066 if (!rl)
1067 up_read(&ni->runlist.lock);
1068 }
1069 /* Hard error. Abort writing this record. */
1070 if (!err || err == -ENOMEM)
1071 err = err2;
1072 bh->b_blocknr = -1;
1073 ntfs_error(vol->sb, "Cannot write ntfs record "
1074 "0x%llx (inode 0x%lx, "
1075 "attribute type 0x%x) because "
1076 "its location on disk could "
1077 "not be determined (error "
1078 "code %lli).",
1079 (long long)block <<
1080 bh_size_bits >>
1081 vol->mft_record_size_bits,
1082 ni->mft_no, ni->type,
1083 (long long)lcn);
1084 /*
1085 * If this is not the first buffer, remove the
1086 * buffers in this record from the list of
1087 * buffers to write and clear their dirty bit
1088 * if not error -ENOMEM.
1089 */
1090 if (rec_start_bh != bh) {
1091 while (bhs[--nr_bhs] != rec_start_bh)
1092 ;
1093 if (err2 != -ENOMEM) {
1094 do {
1095 clear_buffer_dirty(
1096 rec_start_bh);
1097 } while ((rec_start_bh =
1098 rec_start_bh->
1099 b_this_page) !=
1100 bh);
1101 }
1102 }
1103 continue;
1104 }
1105 }
1106 BUG_ON(!buffer_uptodate(bh));
1107 BUG_ON(nr_bhs >= max_bhs);
1108 bhs[nr_bhs++] = bh;
1109 } while (block++, (bh = bh->b_this_page) != head);
1110 if (unlikely(rl))
1111 up_read(&ni->runlist.lock);
1112 /* If there were no dirty buffers, we are done. */
1113 if (!nr_bhs)
1114 goto done;
1115 /* Map the page so we can access its contents. */
1116 kaddr = kmap(page);
1117 /* Clear the page uptodate flag whilst the mst fixups are applied. */
1118 BUG_ON(!PageUptodate(page));
1119 ClearPageUptodate(page);
1120 for (i = 0; i < nr_bhs; i++) {
1121 unsigned int ofs;
1122
1123 /* Skip buffers which are not at the beginning of records. */
1124 if (i % bhs_per_rec)
1125 continue;
1126 tbh = bhs[i];
1127 ofs = bh_offset(tbh);
1128 if (is_mft) {
1129 ntfs_inode *tni;
1130 unsigned long mft_no;
1131
1132 /* Get the mft record number. */
1133 mft_no = (((s64)page->index << PAGE_CACHE_SHIFT) + ofs)
1134 >> rec_size_bits;
1135 /* Check whether to write this mft record. */
1136 tni = NULL;
1137 if (!ntfs_may_write_mft_record(vol, mft_no,
1138 (MFT_RECORD*)(kaddr + ofs), &tni)) {
1139 /*
1140 * The record should not be written. This
1141 * means we need to redirty the page before
1142 * returning.
1143 */
1144 page_is_dirty = true;
1145 /*
1146 * Remove the buffers in this mft record from
1147 * the list of buffers to write.
1148 */
1149 do {
1150 bhs[i] = NULL;
1151 } while (++i % bhs_per_rec);
1152 continue;
1153 }
1154 /*
1155 * The record should be written. If a locked ntfs
1156 * inode was returned, add it to the array of locked
1157 * ntfs inodes.
1158 */
1159 if (tni)
1160 locked_nis[nr_locked_nis++] = tni;
1161 }
1162 /* Apply the mst protection fixups. */
1163 err2 = pre_write_mst_fixup((NTFS_RECORD*)(kaddr + ofs),
1164 rec_size);
1165 if (unlikely(err2)) {
1166 if (!err || err == -ENOMEM)
1167 err = -EIO;
1168 ntfs_error(vol->sb, "Failed to apply mst fixups "
1169 "(inode 0x%lx, attribute type 0x%x, "
1170 "page index 0x%lx, page offset 0x%x)!"
1171 " Unmount and run chkdsk.", vi->i_ino,
1172 ni->type, page->index, ofs);
1173 /*
1174 * Mark all the buffers in this record clean as we do
1175 * not want to write corrupt data to disk.
1176 */
1177 do {
1178 clear_buffer_dirty(bhs[i]);
1179 bhs[i] = NULL;
1180 } while (++i % bhs_per_rec);
1181 continue;
1182 }
1183 nr_recs++;
1184 }
1185 /* If no records are to be written out, we are done. */
1186 if (!nr_recs)
1187 goto unm_done;
1188 flush_dcache_page(page);
1189 /* Lock buffers and start synchronous write i/o on them. */
1190 for (i = 0; i < nr_bhs; i++) {
1191 tbh = bhs[i];
1192 if (!tbh)
1193 continue;
1194 if (unlikely(test_set_buffer_locked(tbh)))
1195 BUG();
1196 /* The buffer dirty state is now irrelevant, just clean it. */
1197 clear_buffer_dirty(tbh);
1198 BUG_ON(!buffer_uptodate(tbh));
1199 BUG_ON(!buffer_mapped(tbh));
1200 get_bh(tbh);
1201 tbh->b_end_io = end_buffer_write_sync;
1202 submit_bh(WRITE, tbh);
1203 }
1204 /* Synchronize the mft mirror now if not @sync. */
1205 if (is_mft && !sync)
1206 goto do_mirror;
1207 do_wait:
1208 /* Wait on i/o completion of buffers. */
1209 for (i = 0; i < nr_bhs; i++) {
1210 tbh = bhs[i];
1211 if (!tbh)
1212 continue;
1213 wait_on_buffer(tbh);
1214 if (unlikely(!buffer_uptodate(tbh))) {
1215 ntfs_error(vol->sb, "I/O error while writing ntfs "
1216 "record buffer (inode 0x%lx, "
1217 "attribute type 0x%x, page index "
1218 "0x%lx, page offset 0x%lx)! Unmount "
1219 "and run chkdsk.", vi->i_ino, ni->type,
1220 page->index, bh_offset(tbh));
1221 if (!err || err == -ENOMEM)
1222 err = -EIO;
1223 /*
1224 * Set the buffer uptodate so the page and buffer
1225 * states do not become out of sync.
1226 */
1227 set_buffer_uptodate(tbh);
1228 }
1229 }
1230 /* If @sync, now synchronize the mft mirror. */
1231 if (is_mft && sync) {
1232 do_mirror:
1233 for (i = 0; i < nr_bhs; i++) {
1234 unsigned long mft_no;
1235 unsigned int ofs;
1236
1237 /*
1238 * Skip buffers which are not at the beginning of
1239 * records.
1240 */
1241 if (i % bhs_per_rec)
1242 continue;
1243 tbh = bhs[i];
1244 /* Skip removed buffers (and hence records). */
1245 if (!tbh)
1246 continue;
1247 ofs = bh_offset(tbh);
1248 /* Get the mft record number. */
1249 mft_no = (((s64)page->index << PAGE_CACHE_SHIFT) + ofs)
1250 >> rec_size_bits;
1251 if (mft_no < vol->mftmirr_size)
1252 ntfs_sync_mft_mirror(vol, mft_no,
1253 (MFT_RECORD*)(kaddr + ofs),
1254 sync);
1255 }
1256 if (!sync)
1257 goto do_wait;
1258 }
1259 /* Remove the mst protection fixups again. */
1260 for (i = 0; i < nr_bhs; i++) {
1261 if (!(i % bhs_per_rec)) {
1262 tbh = bhs[i];
1263 if (!tbh)
1264 continue;
1265 post_write_mst_fixup((NTFS_RECORD*)(kaddr +
1266 bh_offset(tbh)));
1267 }
1268 }
1269 flush_dcache_page(page);
1270 unm_done:
1271 /* Unlock any locked inodes. */
1272 while (nr_locked_nis-- > 0) {
1273 ntfs_inode *tni, *base_tni;
1274
1275 tni = locked_nis[nr_locked_nis];
1276 /* Get the base inode. */
1277 mutex_lock(&tni->extent_lock);
1278 if (tni->nr_extents >= 0)
1279 base_tni = tni;
1280 else {
1281 base_tni = tni->ext.base_ntfs_ino;
1282 BUG_ON(!base_tni);
1283 }
1284 mutex_unlock(&tni->extent_lock);
1285 ntfs_debug("Unlocking %s inode 0x%lx.",
1286 tni == base_tni ? "base" : "extent",
1287 tni->mft_no);
1288 mutex_unlock(&tni->mrec_lock);
1289 atomic_dec(&tni->count);
1290 iput(VFS_I(base_tni));
1291 }
1292 SetPageUptodate(page);
1293 kunmap(page);
1294 done:
1295 if (unlikely(err && err != -ENOMEM)) {
1296 /*
1297 * Set page error if there is only one ntfs record in the page.
1298 * Otherwise we would loose per-record granularity.
1299 */
1300 if (ni->itype.index.block_size == PAGE_CACHE_SIZE)
1301 SetPageError(page);
1302 NVolSetErrors(vol);
1303 }
1304 if (page_is_dirty) {
1305 ntfs_debug("Page still contains one or more dirty ntfs "
1306 "records. Redirtying the page starting at "
1307 "record 0x%lx.", page->index <<
1308 (PAGE_CACHE_SHIFT - rec_size_bits));
1309 redirty_page_for_writepage(wbc, page);
1310 unlock_page(page);
1311 } else {
1312 /*
1313 * Keep the VM happy. This must be done otherwise the
1314 * radix-tree tag PAGECACHE_TAG_DIRTY remains set even though
1315 * the page is clean.
1316 */
1317 BUG_ON(PageWriteback(page));
1318 set_page_writeback(page);
1319 unlock_page(page);
1320 end_page_writeback(page);
1321 }
1322 if (likely(!err))
1323 ntfs_debug("Done.");
1324 return err;
1325 }
1326
1327 /**
1328 * ntfs_writepage - write a @page to the backing store
1329 * @page: page cache page to write out
1330 * @wbc: writeback control structure
1331 *
1332 * This is called from the VM when it wants to have a dirty ntfs page cache
1333 * page cleaned. The VM has already locked the page and marked it clean.
1334 *
1335 * For non-resident attributes, ntfs_writepage() writes the @page by calling
1336 * the ntfs version of the generic block_write_full_page() function,
1337 * ntfs_write_block(), which in turn if necessary creates and writes the
1338 * buffers associated with the page asynchronously.
1339 *
1340 * For resident attributes, OTOH, ntfs_writepage() writes the @page by copying
1341 * the data to the mft record (which at this stage is most likely in memory).
1342 * The mft record is then marked dirty and written out asynchronously via the
1343 * vfs inode dirty code path for the inode the mft record belongs to or via the
1344 * vm page dirty code path for the page the mft record is in.
1345 *
1346 * Based on ntfs_readpage() and fs/buffer.c::block_write_full_page().
1347 *
1348 * Return 0 on success and -errno on error.
1349 */
1350 static int ntfs_writepage(struct page *page, struct writeback_control *wbc)
1351 {
1352 loff_t i_size;
1353 struct inode *vi = page->mapping->host;
1354 ntfs_inode *base_ni = NULL, *ni = NTFS_I(vi);
1355 char *addr;
1356 ntfs_attr_search_ctx *ctx = NULL;
1357 MFT_RECORD *m = NULL;
1358 u32 attr_len;
1359 int err;
1360
1361 retry_writepage:
1362 BUG_ON(!PageLocked(page));
1363 i_size = i_size_read(vi);
1364 /* Is the page fully outside i_size? (truncate in progress) */
1365 if (unlikely(page->index >= (i_size + PAGE_CACHE_SIZE - 1) >>
1366 PAGE_CACHE_SHIFT)) {
1367 /*
1368 * The page may have dirty, unmapped buffers. Make them
1369 * freeable here, so the page does not leak.
1370 */
1371 block_invalidatepage(page, 0);
1372 unlock_page(page);
1373 ntfs_debug("Write outside i_size - truncated?");
1374 return 0;
1375 }
1376 /*
1377 * Only $DATA attributes can be encrypted and only unnamed $DATA
1378 * attributes can be compressed. Index root can have the flags set but
1379 * this means to create compressed/encrypted files, not that the
1380 * attribute is compressed/encrypted. Note we need to check for
1381 * AT_INDEX_ALLOCATION since this is the type of both directory and
1382 * index inodes.
1383 */
1384 if (ni->type != AT_INDEX_ALLOCATION) {
1385 /* If file is encrypted, deny access, just like NT4. */
1386 if (NInoEncrypted(ni)) {
1387 unlock_page(page);
1388 BUG_ON(ni->type != AT_DATA);
1389 ntfs_debug("Denying write access to encrypted file.");
1390 return -EACCES;
1391 }
1392 /* Compressed data streams are handled in compress.c. */
1393 if (NInoNonResident(ni) && NInoCompressed(ni)) {
1394 BUG_ON(ni->type != AT_DATA);
1395 BUG_ON(ni->name_len);
1396 // TODO: Implement and replace this with
1397 // return ntfs_write_compressed_block(page);
1398 unlock_page(page);
1399 ntfs_error(vi->i_sb, "Writing to compressed files is "
1400 "not supported yet. Sorry.");
1401 return -EOPNOTSUPP;
1402 }
1403 // TODO: Implement and remove this check.
1404 if (NInoNonResident(ni) && NInoSparse(ni)) {
1405 unlock_page(page);
1406 ntfs_error(vi->i_sb, "Writing to sparse files is not "
1407 "supported yet. Sorry.");
1408 return -EOPNOTSUPP;
1409 }
1410 }
1411 /* NInoNonResident() == NInoIndexAllocPresent() */
1412 if (NInoNonResident(ni)) {
1413 /* We have to zero every time due to mmap-at-end-of-file. */
1414 if (page->index >= (i_size >> PAGE_CACHE_SHIFT)) {
1415 /* The page straddles i_size. */
1416 unsigned int ofs = i_size & ~PAGE_CACHE_MASK;
1417 zero_user_segment(page, ofs, PAGE_CACHE_SIZE);
1418 }
1419 /* Handle mst protected attributes. */
1420 if (NInoMstProtected(ni))
1421 return ntfs_write_mst_block(page, wbc);
1422 /* Normal, non-resident data stream. */
1423 return ntfs_write_block(page, wbc);
1424 }
1425 /*
1426 * Attribute is resident, implying it is not compressed, encrypted, or
1427 * mst protected. This also means the attribute is smaller than an mft
1428 * record and hence smaller than a page, so can simply return error on
1429 * any pages with index above 0. Note the attribute can actually be
1430 * marked compressed but if it is resident the actual data is not
1431 * compressed so we are ok to ignore the compressed flag here.
1432 */
1433 BUG_ON(page_has_buffers(page));
1434 BUG_ON(!PageUptodate(page));
1435 if (unlikely(page->index > 0)) {
1436 ntfs_error(vi->i_sb, "BUG()! page->index (0x%lx) > 0. "
1437 "Aborting write.", page->index);
1438 BUG_ON(PageWriteback(page));
1439 set_page_writeback(page);
1440 unlock_page(page);
1441 end_page_writeback(page);
1442 return -EIO;
1443 }
1444 if (!NInoAttr(ni))
1445 base_ni = ni;
1446 else
1447 base_ni = ni->ext.base_ntfs_ino;
1448 /* Map, pin, and lock the mft record. */
1449 m = map_mft_record(base_ni);
1450 if (IS_ERR(m)) {
1451 err = PTR_ERR(m);
1452 m = NULL;
1453 ctx = NULL;
1454 goto err_out;
1455 }
1456 /*
1457 * If a parallel write made the attribute non-resident, drop the mft
1458 * record and retry the writepage.
1459 */
1460 if (unlikely(NInoNonResident(ni))) {
1461 unmap_mft_record(base_ni);
1462 goto retry_writepage;
1463 }
1464 ctx = ntfs_attr_get_search_ctx(base_ni, m);
1465 if (unlikely(!ctx)) {
1466 err = -ENOMEM;
1467 goto err_out;
1468 }
1469 err = ntfs_attr_lookup(ni->type, ni->name, ni->name_len,
1470 CASE_SENSITIVE, 0, NULL, 0, ctx);
1471 if (unlikely(err))
1472 goto err_out;
1473 /*
1474 * Keep the VM happy. This must be done otherwise the radix-tree tag
1475 * PAGECACHE_TAG_DIRTY remains set even though the page is clean.
1476 */
1477 BUG_ON(PageWriteback(page));
1478 set_page_writeback(page);
1479 unlock_page(page);
1480 attr_len = le32_to_cpu(ctx->attr->data.resident.value_length);
1481 i_size = i_size_read(vi);
1482 if (unlikely(attr_len > i_size)) {
1483 /* Race with shrinking truncate or a failed truncate. */
1484 attr_len = i_size;
1485 /*
1486 * If the truncate failed, fix it up now. If a concurrent
1487 * truncate, we do its job, so it does not have to do anything.
1488 */
1489 err = ntfs_resident_attr_value_resize(ctx->mrec, ctx->attr,
1490 attr_len);
1491 /* Shrinking cannot fail. */
1492 BUG_ON(err);
1493 }
1494 addr = kmap_atomic(page, KM_USER0);
1495 /* Copy the data from the page to the mft record. */
1496 memcpy((u8*)ctx->attr +
1497 le16_to_cpu(ctx->attr->data.resident.value_offset),
1498 addr, attr_len);
1499 /* Zero out of bounds area in the page cache page. */
1500 memset(addr + attr_len, 0, PAGE_CACHE_SIZE - attr_len);
1501 kunmap_atomic(addr, KM_USER0);
1502 flush_dcache_page(page);
1503 flush_dcache_mft_record_page(ctx->ntfs_ino);
1504 /* We are done with the page. */
1505 end_page_writeback(page);
1506 /* Finally, mark the mft record dirty, so it gets written back. */
1507 mark_mft_record_dirty(ctx->ntfs_ino);
1508 ntfs_attr_put_search_ctx(ctx);
1509 unmap_mft_record(base_ni);
1510 return 0;
1511 err_out:
1512 if (err == -ENOMEM) {
1513 ntfs_warning(vi->i_sb, "Error allocating memory. Redirtying "
1514 "page so we try again later.");
1515 /*
1516 * Put the page back on mapping->dirty_pages, but leave its
1517 * buffers' dirty state as-is.
1518 */
1519 redirty_page_for_writepage(wbc, page);
1520 err = 0;
1521 } else {
1522 ntfs_error(vi->i_sb, "Resident attribute write failed with "
1523 "error %i.", err);
1524 SetPageError(page);
1525 NVolSetErrors(ni->vol);
1526 }
1527 unlock_page(page);
1528 if (ctx)
1529 ntfs_attr_put_search_ctx(ctx);
1530 if (m)
1531 unmap_mft_record(base_ni);
1532 return err;
1533 }
1534
1535 #endif /* NTFS_RW */
1536
1537 /**
1538 * ntfs_aops - general address space operations for inodes and attributes
1539 */
1540 const struct address_space_operations ntfs_aops = {
1541 .readpage = ntfs_readpage, /* Fill page with data. */
1542 .sync_page = block_sync_page, /* Currently, just unplugs the
1543 disk request queue. */
1544 #ifdef NTFS_RW
1545 .writepage = ntfs_writepage, /* Write dirty page to disk. */
1546 #endif /* NTFS_RW */
1547 .migratepage = buffer_migrate_page, /* Move a page cache page from
1548 one physical page to an
1549 other. */
1550 };
1551
1552 /**
1553 * ntfs_mst_aops - general address space operations for mst protecteed inodes
1554 * and attributes
1555 */
1556 const struct address_space_operations ntfs_mst_aops = {
1557 .readpage = ntfs_readpage, /* Fill page with data. */
1558 .sync_page = block_sync_page, /* Currently, just unplugs the
1559 disk request queue. */
1560 #ifdef NTFS_RW
1561 .writepage = ntfs_writepage, /* Write dirty page to disk. */
1562 .set_page_dirty = __set_page_dirty_nobuffers, /* Set the page dirty
1563 without touching the buffers
1564 belonging to the page. */
1565 #endif /* NTFS_RW */
1566 .migratepage = buffer_migrate_page, /* Move a page cache page from
1567 one physical page to an
1568 other. */
1569 };
1570
1571 #ifdef NTFS_RW
1572
1573 /**
1574 * mark_ntfs_record_dirty - mark an ntfs record dirty
1575 * @page: page containing the ntfs record to mark dirty
1576 * @ofs: byte offset within @page at which the ntfs record begins
1577 *
1578 * Set the buffers and the page in which the ntfs record is located dirty.
1579 *
1580 * The latter also marks the vfs inode the ntfs record belongs to dirty
1581 * (I_DIRTY_PAGES only).
1582 *
1583 * If the page does not have buffers, we create them and set them uptodate.
1584 * The page may not be locked which is why we need to handle the buffers under
1585 * the mapping->private_lock. Once the buffers are marked dirty we no longer
1586 * need the lock since try_to_free_buffers() does not free dirty buffers.
1587 */
1588 void mark_ntfs_record_dirty(struct page *page, const unsigned int ofs) {
1589 struct address_space *mapping = page->mapping;
1590 ntfs_inode *ni = NTFS_I(mapping->host);
1591 struct buffer_head *bh, *head, *buffers_to_free = NULL;
1592 unsigned int end, bh_size, bh_ofs;
1593
1594 BUG_ON(!PageUptodate(page));
1595 end = ofs + ni->itype.index.block_size;
1596 bh_size = VFS_I(ni)->i_sb->s_blocksize;
1597 spin_lock(&mapping->private_lock);
1598 if (unlikely(!page_has_buffers(page))) {
1599 spin_unlock(&mapping->private_lock);
1600 bh = head = alloc_page_buffers(page, bh_size, 1);
1601 spin_lock(&mapping->private_lock);
1602 if (likely(!page_has_buffers(page))) {
1603 struct buffer_head *tail;
1604
1605 do {
1606 set_buffer_uptodate(bh);
1607 tail = bh;
1608 bh = bh->b_this_page;
1609 } while (bh);
1610 tail->b_this_page = head;
1611 attach_page_buffers(page, head);
1612 } else
1613 buffers_to_free = bh;
1614 }
1615 bh = head = page_buffers(page);
1616 BUG_ON(!bh);
1617 do {
1618 bh_ofs = bh_offset(bh);
1619 if (bh_ofs + bh_size <= ofs)
1620 continue;
1621 if (unlikely(bh_ofs >= end))
1622 break;
1623 set_buffer_dirty(bh);
1624 } while ((bh = bh->b_this_page) != head);
1625 spin_unlock(&mapping->private_lock);
1626 __set_page_dirty_nobuffers(page);
1627 if (unlikely(buffers_to_free)) {
1628 do {
1629 bh = buffers_to_free->b_this_page;
1630 free_buffer_head(buffers_to_free);
1631 buffers_to_free = bh;
1632 } while (buffers_to_free);
1633 }
1634 }
1635
1636 #endif /* NTFS_RW */
1637
|
This page was automatically generated by the
LXR engine.
|