ACPI: Make Len Brown the ACPI maintainer again
[linux-2.6/next.git] / fs / ubifs / file.c
blob4071d1cae29f7cb0b65221ecd541b3832b37be31
1 /*
2 * This file is part of UBIFS.
4 * Copyright (C) 2006-2008 Nokia Corporation.
6 * This program is free software; you can redistribute it and/or modify it
7 * under the terms of the GNU General Public License version 2 as published by
8 * the Free Software Foundation.
10 * This program is distributed in the hope that it will be useful, but WITHOUT
11 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
12 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
13 * more details.
15 * You should have received a copy of the GNU General Public License along with
16 * this program; if not, write to the Free Software Foundation, Inc., 51
17 * Franklin St, Fifth Floor, Boston, MA 02110-1301 USA
19 * Authors: Artem Bityutskiy (Битюцкий Артём)
20 * Adrian Hunter
24 * This file implements VFS file and inode operations of regular files, device
25 * nodes and symlinks as well as address space operations.
27 * UBIFS uses 2 page flags: PG_private and PG_checked. PG_private is set if the
28 * page is dirty and is used for budgeting purposes - dirty pages should not be
29 * budgeted. The PG_checked flag is set if full budgeting is required for the
30 * page e.g., when it corresponds to a file hole or it is just beyond the file
31 * size. The budgeting is done in 'ubifs_write_begin()', because it is OK to
32 * fail in this function, and the budget is released in 'ubifs_write_end()'. So
33 * the PG_private and PG_checked flags carry the information about how the page
34 * was budgeted, to make it possible to release the budget properly.
36 * A thing to keep in mind: inode's 'i_mutex' is locked in most VFS operations
37 * we implement. However, this is not true for '->writepage()', which might be
38 * called with 'i_mutex' unlocked. For example, when pdflush is performing
39 * write-back, it calls 'writepage()' with unlocked 'i_mutex', although the
40 * inode has 'I_LOCK' flag in this case. At "normal" work-paths 'i_mutex' is
41 * locked in '->writepage', e.g. in "sys_write -> alloc_pages -> direct reclaim
42 * path'. So, in '->writepage()' we are only guaranteed that the page is
43 * locked.
45 * Similarly, 'i_mutex' does not have to be locked in readpage(), e.g.,
46 * readahead path does not have it locked ("sys_read -> generic_file_aio_read
47 * -> ondemand_readahead -> readpage"). In case of readahead, 'I_LOCK' flag is
48 * not set as well. However, UBIFS disables readahead.
50 * This, for example means that there might be 2 concurrent '->writepage()'
51 * calls for the same inode, but different inode dirty pages.
54 #include "ubifs.h"
55 #include <linux/mount.h>
56 #include <linux/namei.h>
58 static int read_block(struct inode *inode, void *addr, unsigned int block,
59 struct ubifs_data_node *dn)
61 struct ubifs_info *c = inode->i_sb->s_fs_info;
62 int err, len, out_len;
63 union ubifs_key key;
64 unsigned int dlen;
66 data_key_init(c, &key, inode->i_ino, block);
67 err = ubifs_tnc_lookup(c, &key, dn);
68 if (err) {
69 if (err == -ENOENT)
70 /* Not found, so it must be a hole */
71 memset(addr, 0, UBIFS_BLOCK_SIZE);
72 return err;
75 ubifs_assert(dn->ch.sqnum > ubifs_inode(inode)->creat_sqnum);
77 len = le32_to_cpu(dn->size);
78 if (len <= 0 || len > UBIFS_BLOCK_SIZE)
79 goto dump;
81 dlen = le32_to_cpu(dn->ch.len) - UBIFS_DATA_NODE_SZ;
82 out_len = UBIFS_BLOCK_SIZE;
83 err = ubifs_decompress(&dn->data, dlen, addr, &out_len,
84 le16_to_cpu(dn->compr_type));
85 if (err || len != out_len)
86 goto dump;
89 * Data length can be less than a full block, even for blocks that are
90 * not the last in the file (e.g., as a result of making a hole and
91 * appending data). Ensure that the remainder is zeroed out.
93 if (len < UBIFS_BLOCK_SIZE)
94 memset(addr + len, 0, UBIFS_BLOCK_SIZE - len);
96 return 0;
98 dump:
99 ubifs_err("bad data node (block %u, inode %lu)",
100 block, inode->i_ino);
101 dbg_dump_node(c, dn);
102 return -EINVAL;
105 static int do_readpage(struct page *page)
107 void *addr;
108 int err = 0, i;
109 unsigned int block, beyond;
110 struct ubifs_data_node *dn;
111 struct inode *inode = page->mapping->host;
112 loff_t i_size = i_size_read(inode);
114 dbg_gen("ino %lu, pg %lu, i_size %lld, flags %#lx",
115 inode->i_ino, page->index, i_size, page->flags);
116 ubifs_assert(!PageChecked(page));
117 ubifs_assert(!PagePrivate(page));
119 addr = kmap(page);
121 block = page->index << UBIFS_BLOCKS_PER_PAGE_SHIFT;
122 beyond = (i_size + UBIFS_BLOCK_SIZE - 1) >> UBIFS_BLOCK_SHIFT;
123 if (block >= beyond) {
124 /* Reading beyond inode */
125 SetPageChecked(page);
126 memset(addr, 0, PAGE_CACHE_SIZE);
127 goto out;
130 dn = kmalloc(UBIFS_MAX_DATA_NODE_SZ, GFP_NOFS);
131 if (!dn) {
132 err = -ENOMEM;
133 goto error;
136 i = 0;
137 while (1) {
138 int ret;
140 if (block >= beyond) {
141 /* Reading beyond inode */
142 err = -ENOENT;
143 memset(addr, 0, UBIFS_BLOCK_SIZE);
144 } else {
145 ret = read_block(inode, addr, block, dn);
146 if (ret) {
147 err = ret;
148 if (err != -ENOENT)
149 break;
152 if (++i >= UBIFS_BLOCKS_PER_PAGE)
153 break;
154 block += 1;
155 addr += UBIFS_BLOCK_SIZE;
157 if (err) {
158 if (err == -ENOENT) {
159 /* Not found, so it must be a hole */
160 SetPageChecked(page);
161 dbg_gen("hole");
162 goto out_free;
164 ubifs_err("cannot read page %lu of inode %lu, error %d",
165 page->index, inode->i_ino, err);
166 goto error;
169 out_free:
170 kfree(dn);
171 out:
172 SetPageUptodate(page);
173 ClearPageError(page);
174 flush_dcache_page(page);
175 kunmap(page);
176 return 0;
178 error:
179 kfree(dn);
180 ClearPageUptodate(page);
181 SetPageError(page);
182 flush_dcache_page(page);
183 kunmap(page);
184 return err;
188 * release_new_page_budget - release budget of a new page.
189 * @c: UBIFS file-system description object
191 * This is a helper function which releases budget corresponding to the budget
192 * of one new page of data.
194 static void release_new_page_budget(struct ubifs_info *c)
196 struct ubifs_budget_req req = { .recalculate = 1, .new_page = 1 };
198 ubifs_release_budget(c, &req);
202 * release_existing_page_budget - release budget of an existing page.
203 * @c: UBIFS file-system description object
205 * This is a helper function which releases budget corresponding to the budget
206 * of changing one one page of data which already exists on the flash media.
208 static void release_existing_page_budget(struct ubifs_info *c)
210 struct ubifs_budget_req req = { .dd_growth = c->page_budget};
212 ubifs_release_budget(c, &req);
215 static int write_begin_slow(struct address_space *mapping,
216 loff_t pos, unsigned len, struct page **pagep)
218 struct inode *inode = mapping->host;
219 struct ubifs_info *c = inode->i_sb->s_fs_info;
220 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
221 struct ubifs_budget_req req = { .new_page = 1 };
222 int uninitialized_var(err), appending = !!(pos + len > inode->i_size);
223 struct page *page;
225 dbg_gen("ino %lu, pos %llu, len %u, i_size %lld",
226 inode->i_ino, pos, len, inode->i_size);
229 * At the slow path we have to budget before locking the page, because
230 * budgeting may force write-back, which would wait on locked pages and
231 * deadlock if we had the page locked. At this point we do not know
232 * anything about the page, so assume that this is a new page which is
233 * written to a hole. This corresponds to largest budget. Later the
234 * budget will be amended if this is not true.
236 if (appending)
237 /* We are appending data, budget for inode change */
238 req.dirtied_ino = 1;
240 err = ubifs_budget_space(c, &req);
241 if (unlikely(err))
242 return err;
244 page = __grab_cache_page(mapping, index);
245 if (unlikely(!page)) {
246 ubifs_release_budget(c, &req);
247 return -ENOMEM;
250 if (!PageUptodate(page)) {
251 if (!(pos & PAGE_CACHE_MASK) && len == PAGE_CACHE_SIZE)
252 SetPageChecked(page);
253 else {
254 err = do_readpage(page);
255 if (err) {
256 unlock_page(page);
257 page_cache_release(page);
258 return err;
262 SetPageUptodate(page);
263 ClearPageError(page);
266 if (PagePrivate(page))
268 * The page is dirty, which means it was budgeted twice:
269 * o first time the budget was allocated by the task which
270 * made the page dirty and set the PG_private flag;
271 * o and then we budgeted for it for the second time at the
272 * very beginning of this function.
274 * So what we have to do is to release the page budget we
275 * allocated.
277 release_new_page_budget(c);
278 else if (!PageChecked(page))
280 * We are changing a page which already exists on the media.
281 * This means that changing the page does not make the amount
282 * of indexing information larger, and this part of the budget
283 * which we have already acquired may be released.
285 ubifs_convert_page_budget(c);
287 if (appending) {
288 struct ubifs_inode *ui = ubifs_inode(inode);
291 * 'ubifs_write_end()' is optimized from the fast-path part of
292 * 'ubifs_write_begin()' and expects the @ui_mutex to be locked
293 * if data is appended.
295 mutex_lock(&ui->ui_mutex);
296 if (ui->dirty)
298 * The inode is dirty already, so we may free the
299 * budget we allocated.
301 ubifs_release_dirty_inode_budget(c, ui);
304 *pagep = page;
305 return 0;
309 * allocate_budget - allocate budget for 'ubifs_write_begin()'.
310 * @c: UBIFS file-system description object
311 * @page: page to allocate budget for
312 * @ui: UBIFS inode object the page belongs to
313 * @appending: non-zero if the page is appended
315 * This is a helper function for 'ubifs_write_begin()' which allocates budget
316 * for the operation. The budget is allocated differently depending on whether
317 * this is appending, whether the page is dirty or not, and so on. This
318 * function leaves the @ui->ui_mutex locked in case of appending. Returns zero
319 * in case of success and %-ENOSPC in case of failure.
321 static int allocate_budget(struct ubifs_info *c, struct page *page,
322 struct ubifs_inode *ui, int appending)
324 struct ubifs_budget_req req = { .fast = 1 };
326 if (PagePrivate(page)) {
327 if (!appending)
329 * The page is dirty and we are not appending, which
330 * means no budget is needed at all.
332 return 0;
334 mutex_lock(&ui->ui_mutex);
335 if (ui->dirty)
337 * The page is dirty and we are appending, so the inode
338 * has to be marked as dirty. However, it is already
339 * dirty, so we do not need any budget. We may return,
340 * but @ui->ui_mutex hast to be left locked because we
341 * should prevent write-back from flushing the inode
342 * and freeing the budget. The lock will be released in
343 * 'ubifs_write_end()'.
345 return 0;
348 * The page is dirty, we are appending, the inode is clean, so
349 * we need to budget the inode change.
351 req.dirtied_ino = 1;
352 } else {
353 if (PageChecked(page))
355 * The page corresponds to a hole and does not
356 * exist on the media. So changing it makes
357 * make the amount of indexing information
358 * larger, and we have to budget for a new
359 * page.
361 req.new_page = 1;
362 else
364 * Not a hole, the change will not add any new
365 * indexing information, budget for page
366 * change.
368 req.dirtied_page = 1;
370 if (appending) {
371 mutex_lock(&ui->ui_mutex);
372 if (!ui->dirty)
374 * The inode is clean but we will have to mark
375 * it as dirty because we are appending. This
376 * needs a budget.
378 req.dirtied_ino = 1;
382 return ubifs_budget_space(c, &req);
386 * This function is called when a page of data is going to be written. Since
387 * the page of data will not necessarily go to the flash straight away, UBIFS
388 * has to reserve space on the media for it, which is done by means of
389 * budgeting.
391 * This is the hot-path of the file-system and we are trying to optimize it as
392 * much as possible. For this reasons it is split on 2 parts - slow and fast.
394 * There many budgeting cases:
395 * o a new page is appended - we have to budget for a new page and for
396 * changing the inode; however, if the inode is already dirty, there is
397 * no need to budget for it;
398 * o an existing clean page is changed - we have budget for it; if the page
399 * does not exist on the media (a hole), we have to budget for a new
400 * page; otherwise, we may budget for changing an existing page; the
401 * difference between these cases is that changing an existing page does
402 * not introduce anything new to the FS indexing information, so it does
403 * not grow, and smaller budget is acquired in this case;
404 * o an existing dirty page is changed - no need to budget at all, because
405 * the page budget has been acquired by earlier, when the page has been
406 * marked dirty.
408 * UBIFS budgeting sub-system may force write-back if it thinks there is no
409 * space to reserve. This imposes some locking restrictions and makes it
410 * impossible to take into account the above cases, and makes it impossible to
411 * optimize budgeting.
413 * The solution for this is that the fast path of 'ubifs_write_begin()' assumes
414 * there is a plenty of flash space and the budget will be acquired quickly,
415 * without forcing write-back. The slow path does not make this assumption.
417 static int ubifs_write_begin(struct file *file, struct address_space *mapping,
418 loff_t pos, unsigned len, unsigned flags,
419 struct page **pagep, void **fsdata)
421 struct inode *inode = mapping->host;
422 struct ubifs_info *c = inode->i_sb->s_fs_info;
423 struct ubifs_inode *ui = ubifs_inode(inode);
424 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
425 int uninitialized_var(err), appending = !!(pos + len > inode->i_size);
426 struct page *page;
429 ubifs_assert(ubifs_inode(inode)->ui_size == inode->i_size);
431 if (unlikely(c->ro_media))
432 return -EROFS;
434 /* Try out the fast-path part first */
435 page = __grab_cache_page(mapping, index);
436 if (unlikely(!page))
437 return -ENOMEM;
439 if (!PageUptodate(page)) {
440 /* The page is not loaded from the flash */
441 if (!(pos & PAGE_CACHE_MASK) && len == PAGE_CACHE_SIZE)
443 * We change whole page so no need to load it. But we
444 * have to set the @PG_checked flag to make the further
445 * code the page is new. This might be not true, but it
446 * is better to budget more that to read the page from
447 * the media.
449 SetPageChecked(page);
450 else {
451 err = do_readpage(page);
452 if (err) {
453 unlock_page(page);
454 page_cache_release(page);
455 return err;
459 SetPageUptodate(page);
460 ClearPageError(page);
463 err = allocate_budget(c, page, ui, appending);
464 if (unlikely(err)) {
465 ubifs_assert(err == -ENOSPC);
467 * Budgeting failed which means it would have to force
468 * write-back but didn't, because we set the @fast flag in the
469 * request. Write-back cannot be done now, while we have the
470 * page locked, because it would deadlock. Unlock and free
471 * everything and fall-back to slow-path.
473 if (appending) {
474 ubifs_assert(mutex_is_locked(&ui->ui_mutex));
475 mutex_unlock(&ui->ui_mutex);
477 unlock_page(page);
478 page_cache_release(page);
480 return write_begin_slow(mapping, pos, len, pagep);
484 * Whee, we aquired budgeting quickly - without involving
485 * garbage-collection, committing or forceing write-back. We return
486 * with @ui->ui_mutex locked if we are appending pages, and unlocked
487 * otherwise. This is an optimization (slightly hacky though).
489 *pagep = page;
490 return 0;
495 * cancel_budget - cancel budget.
496 * @c: UBIFS file-system description object
497 * @page: page to cancel budget for
498 * @ui: UBIFS inode object the page belongs to
499 * @appending: non-zero if the page is appended
501 * This is a helper function for a page write operation. It unlocks the
502 * @ui->ui_mutex in case of appending.
504 static void cancel_budget(struct ubifs_info *c, struct page *page,
505 struct ubifs_inode *ui, int appending)
507 if (appending) {
508 if (!ui->dirty)
509 ubifs_release_dirty_inode_budget(c, ui);
510 mutex_unlock(&ui->ui_mutex);
512 if (!PagePrivate(page)) {
513 if (PageChecked(page))
514 release_new_page_budget(c);
515 else
516 release_existing_page_budget(c);
520 static int ubifs_write_end(struct file *file, struct address_space *mapping,
521 loff_t pos, unsigned len, unsigned copied,
522 struct page *page, void *fsdata)
524 struct inode *inode = mapping->host;
525 struct ubifs_inode *ui = ubifs_inode(inode);
526 struct ubifs_info *c = inode->i_sb->s_fs_info;
527 loff_t end_pos = pos + len;
528 int appending = !!(end_pos > inode->i_size);
530 dbg_gen("ino %lu, pos %llu, pg %lu, len %u, copied %d, i_size %lld",
531 inode->i_ino, pos, page->index, len, copied, inode->i_size);
533 if (unlikely(copied < len && len == PAGE_CACHE_SIZE)) {
535 * VFS copied less data to the page that it intended and
536 * declared in its '->write_begin()' call via the @len
537 * argument. If the page was not up-to-date, and @len was
538 * @PAGE_CACHE_SIZE, the 'ubifs_write_begin()' function did
539 * not load it from the media (for optimization reasons). This
540 * means that part of the page contains garbage. So read the
541 * page now.
543 dbg_gen("copied %d instead of %d, read page and repeat",
544 copied, len);
545 cancel_budget(c, page, ui, appending);
548 * Return 0 to force VFS to repeat the whole operation, or the
549 * error code if 'do_readpage()' failes.
551 copied = do_readpage(page);
552 goto out;
555 if (!PagePrivate(page)) {
556 SetPagePrivate(page);
557 atomic_long_inc(&c->dirty_pg_cnt);
558 __set_page_dirty_nobuffers(page);
561 if (appending) {
562 i_size_write(inode, end_pos);
563 ui->ui_size = end_pos;
565 * Note, we do not set @I_DIRTY_PAGES (which means that the
566 * inode has dirty pages), this has been done in
567 * '__set_page_dirty_nobuffers()'.
569 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
570 ubifs_assert(mutex_is_locked(&ui->ui_mutex));
571 mutex_unlock(&ui->ui_mutex);
574 out:
575 unlock_page(page);
576 page_cache_release(page);
577 return copied;
580 static int ubifs_readpage(struct file *file, struct page *page)
582 do_readpage(page);
583 unlock_page(page);
584 return 0;
587 static int do_writepage(struct page *page, int len)
589 int err = 0, i, blen;
590 unsigned int block;
591 void *addr;
592 union ubifs_key key;
593 struct inode *inode = page->mapping->host;
594 struct ubifs_info *c = inode->i_sb->s_fs_info;
596 #ifdef UBIFS_DEBUG
597 spin_lock(&ui->ui_lock);
598 ubifs_assert(page->index <= ui->synced_i_size << PAGE_CACHE_SIZE);
599 spin_unlock(&ui->ui_lock);
600 #endif
602 /* Update radix tree tags */
603 set_page_writeback(page);
605 addr = kmap(page);
606 block = page->index << UBIFS_BLOCKS_PER_PAGE_SHIFT;
607 i = 0;
608 while (len) {
609 blen = min_t(int, len, UBIFS_BLOCK_SIZE);
610 data_key_init(c, &key, inode->i_ino, block);
611 err = ubifs_jnl_write_data(c, inode, &key, addr, blen);
612 if (err)
613 break;
614 if (++i >= UBIFS_BLOCKS_PER_PAGE)
615 break;
616 block += 1;
617 addr += blen;
618 len -= blen;
620 if (err) {
621 SetPageError(page);
622 ubifs_err("cannot write page %lu of inode %lu, error %d",
623 page->index, inode->i_ino, err);
624 ubifs_ro_mode(c, err);
627 ubifs_assert(PagePrivate(page));
628 if (PageChecked(page))
629 release_new_page_budget(c);
630 else
631 release_existing_page_budget(c);
633 atomic_long_dec(&c->dirty_pg_cnt);
634 ClearPagePrivate(page);
635 ClearPageChecked(page);
637 kunmap(page);
638 unlock_page(page);
639 end_page_writeback(page);
640 return err;
644 * When writing-back dirty inodes, VFS first writes-back pages belonging to the
645 * inode, then the inode itself. For UBIFS this may cause a problem. Consider a
646 * situation when a we have an inode with size 0, then a megabyte of data is
647 * appended to the inode, then write-back starts and flushes some amount of the
648 * dirty pages, the journal becomes full, commit happens and finishes, and then
649 * an unclean reboot happens. When the file system is mounted next time, the
650 * inode size would still be 0, but there would be many pages which are beyond
651 * the inode size, they would be indexed and consume flash space. Because the
652 * journal has been committed, the replay would not be able to detect this
653 * situation and correct the inode size. This means UBIFS would have to scan
654 * whole index and correct all inode sizes, which is long an unacceptable.
656 * To prevent situations like this, UBIFS writes pages back only if they are
657 * within last synchronized inode size, i.e. the the size which has been
658 * written to the flash media last time. Otherwise, UBIFS forces inode
659 * write-back, thus making sure the on-flash inode contains current inode size,
660 * and then keeps writing pages back.
662 * Some locking issues explanation. 'ubifs_writepage()' first is called with
663 * the page locked, and it locks @ui_mutex. However, write-back does take inode
664 * @i_mutex, which means other VFS operations may be run on this inode at the
665 * same time. And the problematic one is truncation to smaller size, from where
666 * we have to call 'vmtruncate()', which first changes @inode->i_size, then
667 * drops the truncated pages. And while dropping the pages, it takes the page
668 * lock. This means that 'do_truncation()' cannot call 'vmtruncate()' with
669 * @ui_mutex locked, because it would deadlock with 'ubifs_writepage()'. This
670 * means that @inode->i_size is changed while @ui_mutex is unlocked.
672 * But in 'ubifs_writepage()' we have to guarantee that we do not write beyond
673 * inode size. How do we do this if @inode->i_size may became smaller while we
674 * are in the middle of 'ubifs_writepage()'? The UBIFS solution is the
675 * @ui->ui_isize "shadow" field which UBIFS uses instead of @inode->i_size
676 * internally and updates it under @ui_mutex.
678 * Q: why we do not worry that if we race with truncation, we may end up with a
679 * situation when the inode is truncated while we are in the middle of
680 * 'do_writepage()', so we do write beyond inode size?
681 * A: If we are in the middle of 'do_writepage()', truncation would be locked
682 * on the page lock and it would not write the truncated inode node to the
683 * journal before we have finished.
685 static int ubifs_writepage(struct page *page, struct writeback_control *wbc)
687 struct inode *inode = page->mapping->host;
688 struct ubifs_inode *ui = ubifs_inode(inode);
689 loff_t i_size = i_size_read(inode), synced_i_size;
690 pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
691 int err, len = i_size & (PAGE_CACHE_SIZE - 1);
692 void *kaddr;
694 dbg_gen("ino %lu, pg %lu, pg flags %#lx",
695 inode->i_ino, page->index, page->flags);
696 ubifs_assert(PagePrivate(page));
698 /* Is the page fully outside @i_size? (truncate in progress) */
699 if (page->index > end_index || (page->index == end_index && !len)) {
700 err = 0;
701 goto out_unlock;
704 spin_lock(&ui->ui_lock);
705 synced_i_size = ui->synced_i_size;
706 spin_unlock(&ui->ui_lock);
708 /* Is the page fully inside @i_size? */
709 if (page->index < end_index) {
710 if (page->index >= synced_i_size >> PAGE_CACHE_SHIFT) {
711 err = inode->i_sb->s_op->write_inode(inode, 1);
712 if (err)
713 goto out_unlock;
715 * The inode has been written, but the write-buffer has
716 * not been synchronized, so in case of an unclean
717 * reboot we may end up with some pages beyond inode
718 * size, but they would be in the journal (because
719 * commit flushes write buffers) and recovery would deal
720 * with this.
723 return do_writepage(page, PAGE_CACHE_SIZE);
727 * The page straddles @i_size. It must be zeroed out on each and every
728 * writepage invocation because it may be mmapped. "A file is mapped
729 * in multiples of the page size. For a file that is not a multiple of
730 * the page size, the remaining memory is zeroed when mapped, and
731 * writes to that region are not written out to the file."
733 kaddr = kmap_atomic(page, KM_USER0);
734 memset(kaddr + len, 0, PAGE_CACHE_SIZE - len);
735 flush_dcache_page(page);
736 kunmap_atomic(kaddr, KM_USER0);
738 if (i_size > synced_i_size) {
739 err = inode->i_sb->s_op->write_inode(inode, 1);
740 if (err)
741 goto out_unlock;
744 return do_writepage(page, len);
746 out_unlock:
747 unlock_page(page);
748 return err;
752 * do_attr_changes - change inode attributes.
753 * @inode: inode to change attributes for
754 * @attr: describes attributes to change
756 static void do_attr_changes(struct inode *inode, const struct iattr *attr)
758 if (attr->ia_valid & ATTR_UID)
759 inode->i_uid = attr->ia_uid;
760 if (attr->ia_valid & ATTR_GID)
761 inode->i_gid = attr->ia_gid;
762 if (attr->ia_valid & ATTR_ATIME)
763 inode->i_atime = timespec_trunc(attr->ia_atime,
764 inode->i_sb->s_time_gran);
765 if (attr->ia_valid & ATTR_MTIME)
766 inode->i_mtime = timespec_trunc(attr->ia_mtime,
767 inode->i_sb->s_time_gran);
768 if (attr->ia_valid & ATTR_CTIME)
769 inode->i_ctime = timespec_trunc(attr->ia_ctime,
770 inode->i_sb->s_time_gran);
771 if (attr->ia_valid & ATTR_MODE) {
772 umode_t mode = attr->ia_mode;
774 if (!in_group_p(inode->i_gid) && !capable(CAP_FSETID))
775 mode &= ~S_ISGID;
776 inode->i_mode = mode;
781 * do_truncation - truncate an inode.
782 * @c: UBIFS file-system description object
783 * @inode: inode to truncate
784 * @attr: inode attribute changes description
786 * This function implements VFS '->setattr()' call when the inode is truncated
787 * to a smaller size. Returns zero in case of success and a negative error code
788 * in case of failure.
790 static int do_truncation(struct ubifs_info *c, struct inode *inode,
791 const struct iattr *attr)
793 int err;
794 struct ubifs_budget_req req;
795 loff_t old_size = inode->i_size, new_size = attr->ia_size;
796 int offset = new_size & (UBIFS_BLOCK_SIZE - 1);
797 struct ubifs_inode *ui = ubifs_inode(inode);
799 dbg_gen("ino %lu, size %lld -> %lld", inode->i_ino, old_size, new_size);
800 memset(&req, 0, sizeof(struct ubifs_budget_req));
803 * If this is truncation to a smaller size, and we do not truncate on a
804 * block boundary, budget for changing one data block, because the last
805 * block will be re-written.
807 if (new_size & (UBIFS_BLOCK_SIZE - 1))
808 req.dirtied_page = 1;
810 req.dirtied_ino = 1;
811 /* A funny way to budget for truncation node */
812 req.dirtied_ino_d = UBIFS_TRUN_NODE_SZ;
813 err = ubifs_budget_space(c, &req);
814 if (err)
815 return err;
817 err = vmtruncate(inode, new_size);
818 if (err)
819 goto out_budg;
821 if (offset) {
822 pgoff_t index = new_size >> PAGE_CACHE_SHIFT;
823 struct page *page;
825 page = find_lock_page(inode->i_mapping, index);
826 if (page) {
827 if (PageDirty(page)) {
829 * 'ubifs_jnl_truncate()' will try to truncate
830 * the last data node, but it contains
831 * out-of-date data because the page is dirty.
832 * Write the page now, so that
833 * 'ubifs_jnl_truncate()' will see an already
834 * truncated (and up to date) data node.
836 ubifs_assert(PagePrivate(page));
838 clear_page_dirty_for_io(page);
839 if (UBIFS_BLOCKS_PER_PAGE_SHIFT)
840 offset = new_size &
841 (PAGE_CACHE_SIZE - 1);
842 err = do_writepage(page, offset);
843 page_cache_release(page);
844 if (err)
845 goto out_budg;
847 * We could now tell 'ubifs_jnl_truncate()' not
848 * to read the last block.
850 } else {
852 * We could 'kmap()' the page and pass the data
853 * to 'ubifs_jnl_truncate()' to save it from
854 * having to read it.
856 unlock_page(page);
857 page_cache_release(page);
862 mutex_lock(&ui->ui_mutex);
863 ui->ui_size = inode->i_size;
864 /* Truncation changes inode [mc]time */
865 inode->i_mtime = inode->i_ctime = ubifs_current_time(inode);
866 /* The other attributes may be changed at the same time as well */
867 do_attr_changes(inode, attr);
869 err = ubifs_jnl_truncate(c, inode, old_size, new_size);
870 mutex_unlock(&ui->ui_mutex);
871 out_budg:
872 ubifs_release_budget(c, &req);
873 return err;
877 * do_setattr - change inode attributes.
878 * @c: UBIFS file-system description object
879 * @inode: inode to change attributes for
880 * @attr: inode attribute changes description
882 * This function implements VFS '->setattr()' call for all cases except
883 * truncations to smaller size. Returns zero in case of success and a negative
884 * error code in case of failure.
886 static int do_setattr(struct ubifs_info *c, struct inode *inode,
887 const struct iattr *attr)
889 int err, release;
890 loff_t new_size = attr->ia_size;
891 struct ubifs_inode *ui = ubifs_inode(inode);
892 struct ubifs_budget_req req = { .dirtied_ino = 1,
893 .dirtied_ino_d = ALIGN(ui->data_len, 8) };
895 err = ubifs_budget_space(c, &req);
896 if (err)
897 return err;
899 if (attr->ia_valid & ATTR_SIZE) {
900 dbg_gen("size %lld -> %lld", inode->i_size, new_size);
901 err = vmtruncate(inode, new_size);
902 if (err)
903 goto out;
906 mutex_lock(&ui->ui_mutex);
907 if (attr->ia_valid & ATTR_SIZE) {
908 /* Truncation changes inode [mc]time */
909 inode->i_mtime = inode->i_ctime = ubifs_current_time(inode);
910 /* 'vmtruncate()' changed @i_size, update @ui_size */
911 ui->ui_size = inode->i_size;
914 do_attr_changes(inode, attr);
916 release = ui->dirty;
917 if (attr->ia_valid & ATTR_SIZE)
919 * Inode length changed, so we have to make sure
920 * @I_DIRTY_DATASYNC is set.
922 __mark_inode_dirty(inode, I_DIRTY_SYNC | I_DIRTY_DATASYNC);
923 else
924 mark_inode_dirty_sync(inode);
925 mutex_unlock(&ui->ui_mutex);
927 if (release)
928 ubifs_release_budget(c, &req);
929 if (IS_SYNC(inode))
930 err = inode->i_sb->s_op->write_inode(inode, 1);
931 return err;
933 out:
934 ubifs_release_budget(c, &req);
935 return err;
938 int ubifs_setattr(struct dentry *dentry, struct iattr *attr)
940 int err;
941 struct inode *inode = dentry->d_inode;
942 struct ubifs_info *c = inode->i_sb->s_fs_info;
944 dbg_gen("ino %lu, mode %#x, ia_valid %#x",
945 inode->i_ino, inode->i_mode, attr->ia_valid);
946 err = inode_change_ok(inode, attr);
947 if (err)
948 return err;
950 err = dbg_check_synced_i_size(inode);
951 if (err)
952 return err;
954 if ((attr->ia_valid & ATTR_SIZE) && attr->ia_size < inode->i_size)
955 /* Truncation to a smaller size */
956 err = do_truncation(c, inode, attr);
957 else
958 err = do_setattr(c, inode, attr);
960 return err;
963 static void ubifs_invalidatepage(struct page *page, unsigned long offset)
965 struct inode *inode = page->mapping->host;
966 struct ubifs_info *c = inode->i_sb->s_fs_info;
968 ubifs_assert(PagePrivate(page));
969 if (offset)
970 /* Partial page remains dirty */
971 return;
973 if (PageChecked(page))
974 release_new_page_budget(c);
975 else
976 release_existing_page_budget(c);
978 atomic_long_dec(&c->dirty_pg_cnt);
979 ClearPagePrivate(page);
980 ClearPageChecked(page);
983 static void *ubifs_follow_link(struct dentry *dentry, struct nameidata *nd)
985 struct ubifs_inode *ui = ubifs_inode(dentry->d_inode);
987 nd_set_link(nd, ui->data);
988 return NULL;
991 int ubifs_fsync(struct file *file, struct dentry *dentry, int datasync)
993 struct inode *inode = dentry->d_inode;
994 struct ubifs_info *c = inode->i_sb->s_fs_info;
995 int err;
997 dbg_gen("syncing inode %lu", inode->i_ino);
1000 * VFS has already synchronized dirty pages for this inode. Synchronize
1001 * the inode unless this is a 'datasync()' call.
1003 if (!datasync || (inode->i_state & I_DIRTY_DATASYNC)) {
1004 err = inode->i_sb->s_op->write_inode(inode, 1);
1005 if (err)
1006 return err;
1010 * Nodes related to this inode may still sit in a write-buffer. Flush
1011 * them.
1013 err = ubifs_sync_wbufs_by_inode(c, inode);
1014 if (err)
1015 return err;
1017 return 0;
1021 * mctime_update_needed - check if mtime or ctime update is needed.
1022 * @inode: the inode to do the check for
1023 * @now: current time
1025 * This helper function checks if the inode mtime/ctime should be updated or
1026 * not. If current values of the time-stamps are within the UBIFS inode time
1027 * granularity, they are not updated. This is an optimization.
1029 static inline int mctime_update_needed(const struct inode *inode,
1030 const struct timespec *now)
1032 if (!timespec_equal(&inode->i_mtime, now) ||
1033 !timespec_equal(&inode->i_ctime, now))
1034 return 1;
1035 return 0;
1039 * update_ctime - update mtime and ctime of an inode.
1040 * @c: UBIFS file-system description object
1041 * @inode: inode to update
1043 * This function updates mtime and ctime of the inode if it is not equivalent to
1044 * current time. Returns zero in case of success and a negative error code in
1045 * case of failure.
1047 static int update_mctime(struct ubifs_info *c, struct inode *inode)
1049 struct timespec now = ubifs_current_time(inode);
1050 struct ubifs_inode *ui = ubifs_inode(inode);
1052 if (mctime_update_needed(inode, &now)) {
1053 int err, release;
1054 struct ubifs_budget_req req = { .dirtied_ino = 1,
1055 .dirtied_ino_d = ALIGN(ui->data_len, 8) };
1057 err = ubifs_budget_space(c, &req);
1058 if (err)
1059 return err;
1061 mutex_lock(&ui->ui_mutex);
1062 inode->i_mtime = inode->i_ctime = ubifs_current_time(inode);
1063 release = ui->dirty;
1064 mark_inode_dirty_sync(inode);
1065 mutex_unlock(&ui->ui_mutex);
1066 if (release)
1067 ubifs_release_budget(c, &req);
1070 return 0;
1073 static ssize_t ubifs_aio_write(struct kiocb *iocb, const struct iovec *iov,
1074 unsigned long nr_segs, loff_t pos)
1076 int err;
1077 ssize_t ret;
1078 struct inode *inode = iocb->ki_filp->f_mapping->host;
1079 struct ubifs_info *c = inode->i_sb->s_fs_info;
1081 err = update_mctime(c, inode);
1082 if (err)
1083 return err;
1085 ret = generic_file_aio_write(iocb, iov, nr_segs, pos);
1086 if (ret < 0)
1087 return ret;
1089 if (ret > 0 && (IS_SYNC(inode) || iocb->ki_filp->f_flags & O_SYNC)) {
1090 err = ubifs_sync_wbufs_by_inode(c, inode);
1091 if (err)
1092 return err;
1095 return ret;
1098 static int ubifs_set_page_dirty(struct page *page)
1100 int ret;
1102 ret = __set_page_dirty_nobuffers(page);
1104 * An attempt to dirty a page without budgeting for it - should not
1105 * happen.
1107 ubifs_assert(ret == 0);
1108 return ret;
1111 static int ubifs_releasepage(struct page *page, gfp_t unused_gfp_flags)
1114 * An attempt to release a dirty page without budgeting for it - should
1115 * not happen.
1117 if (PageWriteback(page))
1118 return 0;
1119 ubifs_assert(PagePrivate(page));
1120 ubifs_assert(0);
1121 ClearPagePrivate(page);
1122 ClearPageChecked(page);
1123 return 1;
1127 * mmap()d file has taken write protection fault and is being made
1128 * writable. UBIFS must ensure page is budgeted for.
1130 static int ubifs_vm_page_mkwrite(struct vm_area_struct *vma, struct page *page)
1132 struct inode *inode = vma->vm_file->f_path.dentry->d_inode;
1133 struct ubifs_info *c = inode->i_sb->s_fs_info;
1134 struct timespec now = ubifs_current_time(inode);
1135 struct ubifs_budget_req req = { .new_page = 1 };
1136 int err, update_time;
1138 dbg_gen("ino %lu, pg %lu, i_size %lld", inode->i_ino, page->index,
1139 i_size_read(inode));
1140 ubifs_assert(!(inode->i_sb->s_flags & MS_RDONLY));
1142 if (unlikely(c->ro_media))
1143 return -EROFS;
1146 * We have not locked @page so far so we may budget for changing the
1147 * page. Note, we cannot do this after we locked the page, because
1148 * budgeting may cause write-back which would cause deadlock.
1150 * At the moment we do not know whether the page is dirty or not, so we
1151 * assume that it is not and budget for a new page. We could look at
1152 * the @PG_private flag and figure this out, but we may race with write
1153 * back and the page state may change by the time we lock it, so this
1154 * would need additional care. We do not bother with this at the
1155 * moment, although it might be good idea to do. Instead, we allocate
1156 * budget for a new page and amend it later on if the page was in fact
1157 * dirty.
1159 * The budgeting-related logic of this function is similar to what we
1160 * do in 'ubifs_write_begin()' and 'ubifs_write_end()'. Glance there
1161 * for more comments.
1163 update_time = mctime_update_needed(inode, &now);
1164 if (update_time)
1166 * We have to change inode time stamp which requires extra
1167 * budgeting.
1169 req.dirtied_ino = 1;
1171 err = ubifs_budget_space(c, &req);
1172 if (unlikely(err)) {
1173 if (err == -ENOSPC)
1174 ubifs_warn("out of space for mmapped file "
1175 "(inode number %lu)", inode->i_ino);
1176 return err;
1179 lock_page(page);
1180 if (unlikely(page->mapping != inode->i_mapping ||
1181 page_offset(page) > i_size_read(inode))) {
1182 /* Page got truncated out from underneath us */
1183 err = -EINVAL;
1184 goto out_unlock;
1187 if (PagePrivate(page))
1188 release_new_page_budget(c);
1189 else {
1190 if (!PageChecked(page))
1191 ubifs_convert_page_budget(c);
1192 SetPagePrivate(page);
1193 atomic_long_inc(&c->dirty_pg_cnt);
1194 __set_page_dirty_nobuffers(page);
1197 if (update_time) {
1198 int release;
1199 struct ubifs_inode *ui = ubifs_inode(inode);
1201 mutex_lock(&ui->ui_mutex);
1202 inode->i_mtime = inode->i_ctime = ubifs_current_time(inode);
1203 release = ui->dirty;
1204 mark_inode_dirty_sync(inode);
1205 mutex_unlock(&ui->ui_mutex);
1206 if (release)
1207 ubifs_release_dirty_inode_budget(c, ui);
1210 unlock_page(page);
1211 return 0;
1213 out_unlock:
1214 unlock_page(page);
1215 ubifs_release_budget(c, &req);
1216 return err;
1219 static struct vm_operations_struct ubifs_file_vm_ops = {
1220 .fault = filemap_fault,
1221 .page_mkwrite = ubifs_vm_page_mkwrite,
1224 static int ubifs_file_mmap(struct file *file, struct vm_area_struct *vma)
1226 int err;
1228 /* 'generic_file_mmap()' takes care of NOMMU case */
1229 err = generic_file_mmap(file, vma);
1230 if (err)
1231 return err;
1232 vma->vm_ops = &ubifs_file_vm_ops;
1233 return 0;
1236 struct address_space_operations ubifs_file_address_operations = {
1237 .readpage = ubifs_readpage,
1238 .writepage = ubifs_writepage,
1239 .write_begin = ubifs_write_begin,
1240 .write_end = ubifs_write_end,
1241 .invalidatepage = ubifs_invalidatepage,
1242 .set_page_dirty = ubifs_set_page_dirty,
1243 .releasepage = ubifs_releasepage,
1246 struct inode_operations ubifs_file_inode_operations = {
1247 .setattr = ubifs_setattr,
1248 .getattr = ubifs_getattr,
1249 #ifdef CONFIG_UBIFS_FS_XATTR
1250 .setxattr = ubifs_setxattr,
1251 .getxattr = ubifs_getxattr,
1252 .listxattr = ubifs_listxattr,
1253 .removexattr = ubifs_removexattr,
1254 #endif
1257 struct inode_operations ubifs_symlink_inode_operations = {
1258 .readlink = generic_readlink,
1259 .follow_link = ubifs_follow_link,
1260 .setattr = ubifs_setattr,
1261 .getattr = ubifs_getattr,
1264 struct file_operations ubifs_file_operations = {
1265 .llseek = generic_file_llseek,
1266 .read = do_sync_read,
1267 .write = do_sync_write,
1268 .aio_read = generic_file_aio_read,
1269 .aio_write = ubifs_aio_write,
1270 .mmap = ubifs_file_mmap,
1271 .fsync = ubifs_fsync,
1272 .unlocked_ioctl = ubifs_ioctl,
1273 .splice_read = generic_file_splice_read,
1274 .splice_write = generic_file_splice_write,
1275 #ifdef CONFIG_COMPAT
1276 .compat_ioctl = ubifs_compat_ioctl,
1277 #endif