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
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)
24 * This file implements VFS file and inode operations for 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
28 * the page is dirty and is used for optimization purposes - dirty pages are
29 * not budgeted so the flag shows that 'ubifs_write_end()' should not release
30 * the budget for this page. The @PG_checked flag is set if full budgeting is
31 * required for the page e.g., when it corresponds to a file hole or it is
32 * beyond the file size. The budgeting is done in 'ubifs_write_begin()', because
33 * it is OK to fail in this function, and the budget is released in
34 * 'ubifs_write_end()'. So the @PG_private and @PG_checked flags carry
35 * information about how the page was budgeted, to make it possible to release
36 * the budget properly.
38 * A thing to keep in mind: inode @i_mutex is locked in most VFS operations we
39 * implement. However, this is not true for 'ubifs_writepage()', which may be
40 * called with @i_mutex unlocked. For example, when flusher thread is doing
41 * background write-back, it calls 'ubifs_writepage()' with unlocked @i_mutex.
42 * At "normal" work-paths the @i_mutex is locked in 'ubifs_writepage()', e.g.
43 * in the "sys_write -> alloc_pages -> direct reclaim path". So, in
44 * 'ubifs_writepage()' we are only guaranteed that the page is locked.
46 * Similarly, @i_mutex is not always locked in 'ubifs_readpage()', e.g., the
47 * read-ahead path does not lock it ("sys_read -> generic_file_aio_read ->
48 * ondemand_readahead -> readpage"). In case of readahead, @I_SYNC flag is not
49 * set as well. However, UBIFS disables readahead.
53 #include <linux/aio.h>
54 #include <linux/mount.h>
55 #include <linux/namei.h>
56 #include <linux/slab.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;
66 data_key_init(c, &key, inode->i_ino, block);
67 err = ubifs_tnc_lookup(c, &key, dn);
70 /* Not found, so it must be a hole */
71 memset(addr, 0, UBIFS_BLOCK_SIZE);
75 ubifs_assert(le64_to_cpu(dn->ch.sqnum) >
76 ubifs_inode(inode)->creat_sqnum);
77 len = le32_to_cpu(dn->size);
78 if (len <= 0 || len > UBIFS_BLOCK_SIZE)
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)
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);
99 ubifs_err("bad data node (block %u, inode %lu)",
100 block, inode->i_ino);
101 ubifs_dump_node(c, dn);
105 static int do_readpage(struct page *page)
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));
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);
130 dn = kmalloc(UBIFS_MAX_DATA_NODE_SZ, GFP_NOFS);
140 if (block >= beyond) {
141 /* Reading beyond inode */
143 memset(addr, 0, UBIFS_BLOCK_SIZE);
145 ret = read_block(inode, addr, block, dn);
150 } else if (block + 1 == beyond) {
151 int dlen = le32_to_cpu(dn->size);
152 int ilen = i_size & (UBIFS_BLOCK_SIZE - 1);
154 if (ilen && ilen < dlen)
155 memset(addr + ilen, 0, dlen - ilen);
158 if (++i >= UBIFS_BLOCKS_PER_PAGE)
161 addr += UBIFS_BLOCK_SIZE;
164 if (err == -ENOENT) {
165 /* Not found, so it must be a hole */
166 SetPageChecked(page);
170 ubifs_err("cannot read page %lu of inode %lu, error %d",
171 page->index, inode->i_ino, err);
178 SetPageUptodate(page);
179 ClearPageError(page);
180 flush_dcache_page(page);
186 ClearPageUptodate(page);
188 flush_dcache_page(page);
194 * release_new_page_budget - release budget of a new page.
195 * @c: UBIFS file-system description object
197 * This is a helper function which releases budget corresponding to the budget
198 * of one new page of data.
200 static void release_new_page_budget(struct ubifs_info *c)
202 struct ubifs_budget_req req = { .recalculate = 1, .new_page = 1 };
204 ubifs_release_budget(c, &req);
208 * release_existing_page_budget - release budget of an existing page.
209 * @c: UBIFS file-system description object
211 * This is a helper function which releases budget corresponding to the budget
212 * of changing one one page of data which already exists on the flash media.
214 static void release_existing_page_budget(struct ubifs_info *c)
216 struct ubifs_budget_req req = { .dd_growth = c->bi.page_budget};
218 ubifs_release_budget(c, &req);
221 static int write_begin_slow(struct address_space *mapping,
222 loff_t pos, unsigned len, struct page **pagep,
225 struct inode *inode = mapping->host;
226 struct ubifs_info *c = inode->i_sb->s_fs_info;
227 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
228 struct ubifs_budget_req req = { .new_page = 1 };
229 int uninitialized_var(err), appending = !!(pos + len > inode->i_size);
232 dbg_gen("ino %lu, pos %llu, len %u, i_size %lld",
233 inode->i_ino, pos, len, inode->i_size);
236 * At the slow path we have to budget before locking the page, because
237 * budgeting may force write-back, which would wait on locked pages and
238 * deadlock if we had the page locked. At this point we do not know
239 * anything about the page, so assume that this is a new page which is
240 * written to a hole. This corresponds to largest budget. Later the
241 * budget will be amended if this is not true.
244 /* We are appending data, budget for inode change */
247 err = ubifs_budget_space(c, &req);
251 page = grab_cache_page_write_begin(mapping, index, flags);
252 if (unlikely(!page)) {
253 ubifs_release_budget(c, &req);
257 if (!PageUptodate(page)) {
258 if (!(pos & ~PAGE_CACHE_MASK) && len == PAGE_CACHE_SIZE)
259 SetPageChecked(page);
261 err = do_readpage(page);
264 page_cache_release(page);
269 SetPageUptodate(page);
270 ClearPageError(page);
273 if (PagePrivate(page))
275 * The page is dirty, which means it was budgeted twice:
276 * o first time the budget was allocated by the task which
277 * made the page dirty and set the PG_private flag;
278 * o and then we budgeted for it for the second time at the
279 * very beginning of this function.
281 * So what we have to do is to release the page budget we
284 release_new_page_budget(c);
285 else if (!PageChecked(page))
287 * We are changing a page which already exists on the media.
288 * This means that changing the page does not make the amount
289 * of indexing information larger, and this part of the budget
290 * which we have already acquired may be released.
292 ubifs_convert_page_budget(c);
295 struct ubifs_inode *ui = ubifs_inode(inode);
298 * 'ubifs_write_end()' is optimized from the fast-path part of
299 * 'ubifs_write_begin()' and expects the @ui_mutex to be locked
300 * if data is appended.
302 mutex_lock(&ui->ui_mutex);
305 * The inode is dirty already, so we may free the
306 * budget we allocated.
308 ubifs_release_dirty_inode_budget(c, ui);
316 * allocate_budget - allocate budget for 'ubifs_write_begin()'.
317 * @c: UBIFS file-system description object
318 * @page: page to allocate budget for
319 * @ui: UBIFS inode object the page belongs to
320 * @appending: non-zero if the page is appended
322 * This is a helper function for 'ubifs_write_begin()' which allocates budget
323 * for the operation. The budget is allocated differently depending on whether
324 * this is appending, whether the page is dirty or not, and so on. This
325 * function leaves the @ui->ui_mutex locked in case of appending. Returns zero
326 * in case of success and %-ENOSPC in case of failure.
328 static int allocate_budget(struct ubifs_info *c, struct page *page,
329 struct ubifs_inode *ui, int appending)
331 struct ubifs_budget_req req = { .fast = 1 };
333 if (PagePrivate(page)) {
336 * The page is dirty and we are not appending, which
337 * means no budget is needed at all.
341 mutex_lock(&ui->ui_mutex);
344 * The page is dirty and we are appending, so the inode
345 * has to be marked as dirty. However, it is already
346 * dirty, so we do not need any budget. We may return,
347 * but @ui->ui_mutex hast to be left locked because we
348 * should prevent write-back from flushing the inode
349 * and freeing the budget. The lock will be released in
350 * 'ubifs_write_end()'.
355 * The page is dirty, we are appending, the inode is clean, so
356 * we need to budget the inode change.
360 if (PageChecked(page))
362 * The page corresponds to a hole and does not
363 * exist on the media. So changing it makes
364 * make the amount of indexing information
365 * larger, and we have to budget for a new
371 * Not a hole, the change will not add any new
372 * indexing information, budget for page
375 req.dirtied_page = 1;
378 mutex_lock(&ui->ui_mutex);
381 * The inode is clean but we will have to mark
382 * it as dirty because we are appending. This
389 return ubifs_budget_space(c, &req);
393 * This function is called when a page of data is going to be written. Since
394 * the page of data will not necessarily go to the flash straight away, UBIFS
395 * has to reserve space on the media for it, which is done by means of
398 * This is the hot-path of the file-system and we are trying to optimize it as
399 * much as possible. For this reasons it is split on 2 parts - slow and fast.
401 * There many budgeting cases:
402 * o a new page is appended - we have to budget for a new page and for
403 * changing the inode; however, if the inode is already dirty, there is
404 * no need to budget for it;
405 * o an existing clean page is changed - we have budget for it; if the page
406 * does not exist on the media (a hole), we have to budget for a new
407 * page; otherwise, we may budget for changing an existing page; the
408 * difference between these cases is that changing an existing page does
409 * not introduce anything new to the FS indexing information, so it does
410 * not grow, and smaller budget is acquired in this case;
411 * o an existing dirty page is changed - no need to budget at all, because
412 * the page budget has been acquired by earlier, when the page has been
415 * UBIFS budgeting sub-system may force write-back if it thinks there is no
416 * space to reserve. This imposes some locking restrictions and makes it
417 * impossible to take into account the above cases, and makes it impossible to
418 * optimize budgeting.
420 * The solution for this is that the fast path of 'ubifs_write_begin()' assumes
421 * there is a plenty of flash space and the budget will be acquired quickly,
422 * without forcing write-back. The slow path does not make this assumption.
424 static int ubifs_write_begin(struct file *file, struct address_space *mapping,
425 loff_t pos, unsigned len, unsigned flags,
426 struct page **pagep, void **fsdata)
428 struct inode *inode = mapping->host;
429 struct ubifs_info *c = inode->i_sb->s_fs_info;
430 struct ubifs_inode *ui = ubifs_inode(inode);
431 pgoff_t index = pos >> PAGE_CACHE_SHIFT;
432 int uninitialized_var(err), appending = !!(pos + len > inode->i_size);
433 int skipped_read = 0;
436 ubifs_assert(ubifs_inode(inode)->ui_size == inode->i_size);
437 ubifs_assert(!c->ro_media && !c->ro_mount);
439 if (unlikely(c->ro_error))
442 /* Try out the fast-path part first */
443 page = grab_cache_page_write_begin(mapping, index, flags);
447 if (!PageUptodate(page)) {
448 /* The page is not loaded from the flash */
449 if (!(pos & ~PAGE_CACHE_MASK) && len == PAGE_CACHE_SIZE) {
451 * We change whole page so no need to load it. But we
452 * do not know whether this page exists on the media or
453 * not, so we assume the latter because it requires
454 * larger budget. The assumption is that it is better
455 * to budget a bit more than to read the page from the
456 * media. Thus, we are setting the @PG_checked flag
459 SetPageChecked(page);
462 err = do_readpage(page);
465 page_cache_release(page);
470 SetPageUptodate(page);
471 ClearPageError(page);
474 err = allocate_budget(c, page, ui, appending);
476 ubifs_assert(err == -ENOSPC);
478 * If we skipped reading the page because we were going to
479 * write all of it, then it is not up to date.
482 ClearPageChecked(page);
483 ClearPageUptodate(page);
486 * Budgeting failed which means it would have to force
487 * write-back but didn't, because we set the @fast flag in the
488 * request. Write-back cannot be done now, while we have the
489 * page locked, because it would deadlock. Unlock and free
490 * everything and fall-back to slow-path.
493 ubifs_assert(mutex_is_locked(&ui->ui_mutex));
494 mutex_unlock(&ui->ui_mutex);
497 page_cache_release(page);
499 return write_begin_slow(mapping, pos, len, pagep, flags);
503 * Whee, we acquired budgeting quickly - without involving
504 * garbage-collection, committing or forcing write-back. We return
505 * with @ui->ui_mutex locked if we are appending pages, and unlocked
506 * otherwise. This is an optimization (slightly hacky though).
514 * cancel_budget - cancel budget.
515 * @c: UBIFS file-system description object
516 * @page: page to cancel budget for
517 * @ui: UBIFS inode object the page belongs to
518 * @appending: non-zero if the page is appended
520 * This is a helper function for a page write operation. It unlocks the
521 * @ui->ui_mutex in case of appending.
523 static void cancel_budget(struct ubifs_info *c, struct page *page,
524 struct ubifs_inode *ui, int appending)
528 ubifs_release_dirty_inode_budget(c, ui);
529 mutex_unlock(&ui->ui_mutex);
531 if (!PagePrivate(page)) {
532 if (PageChecked(page))
533 release_new_page_budget(c);
535 release_existing_page_budget(c);
539 static int ubifs_write_end(struct file *file, struct address_space *mapping,
540 loff_t pos, unsigned len, unsigned copied,
541 struct page *page, void *fsdata)
543 struct inode *inode = mapping->host;
544 struct ubifs_inode *ui = ubifs_inode(inode);
545 struct ubifs_info *c = inode->i_sb->s_fs_info;
546 loff_t end_pos = pos + len;
547 int appending = !!(end_pos > inode->i_size);
549 dbg_gen("ino %lu, pos %llu, pg %lu, len %u, copied %d, i_size %lld",
550 inode->i_ino, pos, page->index, len, copied, inode->i_size);
552 if (unlikely(copied < len && len == PAGE_CACHE_SIZE)) {
554 * VFS copied less data to the page that it intended and
555 * declared in its '->write_begin()' call via the @len
556 * argument. If the page was not up-to-date, and @len was
557 * @PAGE_CACHE_SIZE, the 'ubifs_write_begin()' function did
558 * not load it from the media (for optimization reasons). This
559 * means that part of the page contains garbage. So read the
562 dbg_gen("copied %d instead of %d, read page and repeat",
564 cancel_budget(c, page, ui, appending);
565 ClearPageChecked(page);
568 * Return 0 to force VFS to repeat the whole operation, or the
569 * error code if 'do_readpage()' fails.
571 copied = do_readpage(page);
575 if (!PagePrivate(page)) {
576 SetPagePrivate(page);
577 atomic_long_inc(&c->dirty_pg_cnt);
578 __set_page_dirty_nobuffers(page);
582 i_size_write(inode, end_pos);
583 ui->ui_size = end_pos;
585 * Note, we do not set @I_DIRTY_PAGES (which means that the
586 * inode has dirty pages), this has been done in
587 * '__set_page_dirty_nobuffers()'.
589 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
590 ubifs_assert(mutex_is_locked(&ui->ui_mutex));
591 mutex_unlock(&ui->ui_mutex);
596 page_cache_release(page);
601 * populate_page - copy data nodes into a page for bulk-read.
602 * @c: UBIFS file-system description object
604 * @bu: bulk-read information
605 * @n: next zbranch slot
607 * This function returns %0 on success and a negative error code on failure.
609 static int populate_page(struct ubifs_info *c, struct page *page,
610 struct bu_info *bu, int *n)
612 int i = 0, nn = *n, offs = bu->zbranch[0].offs, hole = 0, read = 0;
613 struct inode *inode = page->mapping->host;
614 loff_t i_size = i_size_read(inode);
615 unsigned int page_block;
619 dbg_gen("ino %lu, pg %lu, i_size %lld, flags %#lx",
620 inode->i_ino, page->index, i_size, page->flags);
622 addr = zaddr = kmap(page);
624 end_index = (i_size - 1) >> PAGE_CACHE_SHIFT;
625 if (!i_size || page->index > end_index) {
627 memset(addr, 0, PAGE_CACHE_SIZE);
631 page_block = page->index << UBIFS_BLOCKS_PER_PAGE_SHIFT;
633 int err, len, out_len, dlen;
637 memset(addr, 0, UBIFS_BLOCK_SIZE);
638 } else if (key_block(c, &bu->zbranch[nn].key) == page_block) {
639 struct ubifs_data_node *dn;
641 dn = bu->buf + (bu->zbranch[nn].offs - offs);
643 ubifs_assert(le64_to_cpu(dn->ch.sqnum) >
644 ubifs_inode(inode)->creat_sqnum);
646 len = le32_to_cpu(dn->size);
647 if (len <= 0 || len > UBIFS_BLOCK_SIZE)
650 dlen = le32_to_cpu(dn->ch.len) - UBIFS_DATA_NODE_SZ;
651 out_len = UBIFS_BLOCK_SIZE;
652 err = ubifs_decompress(&dn->data, dlen, addr, &out_len,
653 le16_to_cpu(dn->compr_type));
654 if (err || len != out_len)
657 if (len < UBIFS_BLOCK_SIZE)
658 memset(addr + len, 0, UBIFS_BLOCK_SIZE - len);
661 read = (i << UBIFS_BLOCK_SHIFT) + len;
662 } else if (key_block(c, &bu->zbranch[nn].key) < page_block) {
667 memset(addr, 0, UBIFS_BLOCK_SIZE);
669 if (++i >= UBIFS_BLOCKS_PER_PAGE)
671 addr += UBIFS_BLOCK_SIZE;
675 if (end_index == page->index) {
676 int len = i_size & (PAGE_CACHE_SIZE - 1);
678 if (len && len < read)
679 memset(zaddr + len, 0, read - len);
684 SetPageChecked(page);
688 SetPageUptodate(page);
689 ClearPageError(page);
690 flush_dcache_page(page);
696 ClearPageUptodate(page);
698 flush_dcache_page(page);
700 ubifs_err("bad data node (block %u, inode %lu)",
701 page_block, inode->i_ino);
706 * ubifs_do_bulk_read - do bulk-read.
707 * @c: UBIFS file-system description object
708 * @bu: bulk-read information
709 * @page1: first page to read
711 * This function returns %1 if the bulk-read is done, otherwise %0 is returned.
713 static int ubifs_do_bulk_read(struct ubifs_info *c, struct bu_info *bu,
716 pgoff_t offset = page1->index, end_index;
717 struct address_space *mapping = page1->mapping;
718 struct inode *inode = mapping->host;
719 struct ubifs_inode *ui = ubifs_inode(inode);
720 int err, page_idx, page_cnt, ret = 0, n = 0;
721 int allocate = bu->buf ? 0 : 1;
724 err = ubifs_tnc_get_bu_keys(c, bu);
729 /* Turn off bulk-read at the end of the file */
730 ui->read_in_a_row = 1;
734 page_cnt = bu->blk_cnt >> UBIFS_BLOCKS_PER_PAGE_SHIFT;
737 * This happens when there are multiple blocks per page and the
738 * blocks for the first page we are looking for, are not
739 * together. If all the pages were like this, bulk-read would
740 * reduce performance, so we turn it off for a while.
748 * Allocate bulk-read buffer depending on how many data
749 * nodes we are going to read.
751 bu->buf_len = bu->zbranch[bu->cnt - 1].offs +
752 bu->zbranch[bu->cnt - 1].len -
754 ubifs_assert(bu->buf_len > 0);
755 ubifs_assert(bu->buf_len <= c->leb_size);
756 bu->buf = kmalloc(bu->buf_len, GFP_NOFS | __GFP_NOWARN);
761 err = ubifs_tnc_bulk_read(c, bu);
766 err = populate_page(c, page1, bu, &n);
773 isize = i_size_read(inode);
776 end_index = ((isize - 1) >> PAGE_CACHE_SHIFT);
778 for (page_idx = 1; page_idx < page_cnt; page_idx++) {
779 pgoff_t page_offset = offset + page_idx;
782 if (page_offset > end_index)
784 page = find_or_create_page(mapping, page_offset,
785 GFP_NOFS | __GFP_COLD);
788 if (!PageUptodate(page))
789 err = populate_page(c, page, bu, &n);
791 page_cache_release(page);
796 ui->last_page_read = offset + page_idx - 1;
804 ubifs_warn("ignoring error %d and skipping bulk-read", err);
808 ui->read_in_a_row = ui->bulk_read = 0;
813 * ubifs_bulk_read - determine whether to bulk-read and, if so, do it.
814 * @page: page from which to start bulk-read.
816 * Some flash media are capable of reading sequentially at faster rates. UBIFS
817 * bulk-read facility is designed to take advantage of that, by reading in one
818 * go consecutive data nodes that are also located consecutively in the same
819 * LEB. This function returns %1 if a bulk-read is done and %0 otherwise.
821 static int ubifs_bulk_read(struct page *page)
823 struct inode *inode = page->mapping->host;
824 struct ubifs_info *c = inode->i_sb->s_fs_info;
825 struct ubifs_inode *ui = ubifs_inode(inode);
826 pgoff_t index = page->index, last_page_read = ui->last_page_read;
828 int err = 0, allocated = 0;
830 ui->last_page_read = index;
835 * Bulk-read is protected by @ui->ui_mutex, but it is an optimization,
836 * so don't bother if we cannot lock the mutex.
838 if (!mutex_trylock(&ui->ui_mutex))
841 if (index != last_page_read + 1) {
842 /* Turn off bulk-read if we stop reading sequentially */
843 ui->read_in_a_row = 1;
849 if (!ui->bulk_read) {
850 ui->read_in_a_row += 1;
851 if (ui->read_in_a_row < 3)
853 /* Three reads in a row, so switch on bulk-read */
858 * If possible, try to use pre-allocated bulk-read information, which
859 * is protected by @c->bu_mutex.
861 if (mutex_trylock(&c->bu_mutex))
864 bu = kmalloc(sizeof(struct bu_info), GFP_NOFS | __GFP_NOWARN);
872 bu->buf_len = c->max_bu_buf_len;
873 data_key_init(c, &bu->key, inode->i_ino,
874 page->index << UBIFS_BLOCKS_PER_PAGE_SHIFT);
875 err = ubifs_do_bulk_read(c, bu, page);
878 mutex_unlock(&c->bu_mutex);
883 mutex_unlock(&ui->ui_mutex);
887 static int ubifs_readpage(struct file *file, struct page *page)
889 if (ubifs_bulk_read(page))
896 static int do_writepage(struct page *page, int len)
898 int err = 0, i, blen;
902 struct inode *inode = page->mapping->host;
903 struct ubifs_info *c = inode->i_sb->s_fs_info;
906 spin_lock(&ui->ui_lock);
907 ubifs_assert(page->index <= ui->synced_i_size << PAGE_CACHE_SIZE);
908 spin_unlock(&ui->ui_lock);
911 /* Update radix tree tags */
912 set_page_writeback(page);
915 block = page->index << UBIFS_BLOCKS_PER_PAGE_SHIFT;
918 blen = min_t(int, len, UBIFS_BLOCK_SIZE);
919 data_key_init(c, &key, inode->i_ino, block);
920 err = ubifs_jnl_write_data(c, inode, &key, addr, blen);
923 if (++i >= UBIFS_BLOCKS_PER_PAGE)
931 ubifs_err("cannot write page %lu of inode %lu, error %d",
932 page->index, inode->i_ino, err);
933 ubifs_ro_mode(c, err);
936 ubifs_assert(PagePrivate(page));
937 if (PageChecked(page))
938 release_new_page_budget(c);
940 release_existing_page_budget(c);
942 atomic_long_dec(&c->dirty_pg_cnt);
943 ClearPagePrivate(page);
944 ClearPageChecked(page);
948 end_page_writeback(page);
953 * When writing-back dirty inodes, VFS first writes-back pages belonging to the
954 * inode, then the inode itself. For UBIFS this may cause a problem. Consider a
955 * situation when a we have an inode with size 0, then a megabyte of data is
956 * appended to the inode, then write-back starts and flushes some amount of the
957 * dirty pages, the journal becomes full, commit happens and finishes, and then
958 * an unclean reboot happens. When the file system is mounted next time, the
959 * inode size would still be 0, but there would be many pages which are beyond
960 * the inode size, they would be indexed and consume flash space. Because the
961 * journal has been committed, the replay would not be able to detect this
962 * situation and correct the inode size. This means UBIFS would have to scan
963 * whole index and correct all inode sizes, which is long an unacceptable.
965 * To prevent situations like this, UBIFS writes pages back only if they are
966 * within the last synchronized inode size, i.e. the size which has been
967 * written to the flash media last time. Otherwise, UBIFS forces inode
968 * write-back, thus making sure the on-flash inode contains current inode size,
969 * and then keeps writing pages back.
971 * Some locking issues explanation. 'ubifs_writepage()' first is called with
972 * the page locked, and it locks @ui_mutex. However, write-back does take inode
973 * @i_mutex, which means other VFS operations may be run on this inode at the
974 * same time. And the problematic one is truncation to smaller size, from where
975 * we have to call 'truncate_setsize()', which first changes @inode->i_size,
976 * then drops the truncated pages. And while dropping the pages, it takes the
977 * page lock. This means that 'do_truncation()' cannot call 'truncate_setsize()'
978 * with @ui_mutex locked, because it would deadlock with 'ubifs_writepage()'.
979 * This means that @inode->i_size is changed while @ui_mutex is unlocked.
981 * XXX(truncate): with the new truncate sequence this is not true anymore,
982 * and the calls to truncate_setsize can be move around freely. They should
983 * be moved to the very end of the truncate sequence.
985 * But in 'ubifs_writepage()' we have to guarantee that we do not write beyond
986 * inode size. How do we do this if @inode->i_size may became smaller while we
987 * are in the middle of 'ubifs_writepage()'? The UBIFS solution is the
988 * @ui->ui_isize "shadow" field which UBIFS uses instead of @inode->i_size
989 * internally and updates it under @ui_mutex.
991 * Q: why we do not worry that if we race with truncation, we may end up with a
992 * situation when the inode is truncated while we are in the middle of
993 * 'do_writepage()', so we do write beyond inode size?
994 * A: If we are in the middle of 'do_writepage()', truncation would be locked
995 * on the page lock and it would not write the truncated inode node to the
996 * journal before we have finished.
998 static int ubifs_writepage(struct page *page, struct writeback_control *wbc)
1000 struct inode *inode = page->mapping->host;
1001 struct ubifs_inode *ui = ubifs_inode(inode);
1002 loff_t i_size = i_size_read(inode), synced_i_size;
1003 pgoff_t end_index = i_size >> PAGE_CACHE_SHIFT;
1004 int err, len = i_size & (PAGE_CACHE_SIZE - 1);
1007 dbg_gen("ino %lu, pg %lu, pg flags %#lx",
1008 inode->i_ino, page->index, page->flags);
1009 ubifs_assert(PagePrivate(page));
1011 /* Is the page fully outside @i_size? (truncate in progress) */
1012 if (page->index > end_index || (page->index == end_index && !len)) {
1017 spin_lock(&ui->ui_lock);
1018 synced_i_size = ui->synced_i_size;
1019 spin_unlock(&ui->ui_lock);
1021 /* Is the page fully inside @i_size? */
1022 if (page->index < end_index) {
1023 if (page->index >= synced_i_size >> PAGE_CACHE_SHIFT) {
1024 err = inode->i_sb->s_op->write_inode(inode, NULL);
1028 * The inode has been written, but the write-buffer has
1029 * not been synchronized, so in case of an unclean
1030 * reboot we may end up with some pages beyond inode
1031 * size, but they would be in the journal (because
1032 * commit flushes write buffers) and recovery would deal
1036 return do_writepage(page, PAGE_CACHE_SIZE);
1040 * The page straddles @i_size. It must be zeroed out on each and every
1041 * writepage invocation because it may be mmapped. "A file is mapped
1042 * in multiples of the page size. For a file that is not a multiple of
1043 * the page size, the remaining memory is zeroed when mapped, and
1044 * writes to that region are not written out to the file."
1046 kaddr = kmap_atomic(page);
1047 memset(kaddr + len, 0, PAGE_CACHE_SIZE - len);
1048 flush_dcache_page(page);
1049 kunmap_atomic(kaddr);
1051 if (i_size > synced_i_size) {
1052 err = inode->i_sb->s_op->write_inode(inode, NULL);
1057 return do_writepage(page, len);
1065 * do_attr_changes - change inode attributes.
1066 * @inode: inode to change attributes for
1067 * @attr: describes attributes to change
1069 static void do_attr_changes(struct inode *inode, const struct iattr *attr)
1071 if (attr->ia_valid & ATTR_UID)
1072 inode->i_uid = attr->ia_uid;
1073 if (attr->ia_valid & ATTR_GID)
1074 inode->i_gid = attr->ia_gid;
1075 if (attr->ia_valid & ATTR_ATIME)
1076 inode->i_atime = timespec_trunc(attr->ia_atime,
1077 inode->i_sb->s_time_gran);
1078 if (attr->ia_valid & ATTR_MTIME)
1079 inode->i_mtime = timespec_trunc(attr->ia_mtime,
1080 inode->i_sb->s_time_gran);
1081 if (attr->ia_valid & ATTR_CTIME)
1082 inode->i_ctime = timespec_trunc(attr->ia_ctime,
1083 inode->i_sb->s_time_gran);
1084 if (attr->ia_valid & ATTR_MODE) {
1085 umode_t mode = attr->ia_mode;
1087 if (!in_group_p(inode->i_gid) && !capable(CAP_FSETID))
1089 inode->i_mode = mode;
1094 * do_truncation - truncate an inode.
1095 * @c: UBIFS file-system description object
1096 * @inode: inode to truncate
1097 * @attr: inode attribute changes description
1099 * This function implements VFS '->setattr()' call when the inode is truncated
1100 * to a smaller size. Returns zero in case of success and a negative error code
1101 * in case of failure.
1103 static int do_truncation(struct ubifs_info *c, struct inode *inode,
1104 const struct iattr *attr)
1107 struct ubifs_budget_req req;
1108 loff_t old_size = inode->i_size, new_size = attr->ia_size;
1109 int offset = new_size & (UBIFS_BLOCK_SIZE - 1), budgeted = 1;
1110 struct ubifs_inode *ui = ubifs_inode(inode);
1112 dbg_gen("ino %lu, size %lld -> %lld", inode->i_ino, old_size, new_size);
1113 memset(&req, 0, sizeof(struct ubifs_budget_req));
1116 * If this is truncation to a smaller size, and we do not truncate on a
1117 * block boundary, budget for changing one data block, because the last
1118 * block will be re-written.
1120 if (new_size & (UBIFS_BLOCK_SIZE - 1))
1121 req.dirtied_page = 1;
1123 req.dirtied_ino = 1;
1124 /* A funny way to budget for truncation node */
1125 req.dirtied_ino_d = UBIFS_TRUN_NODE_SZ;
1126 err = ubifs_budget_space(c, &req);
1129 * Treat truncations to zero as deletion and always allow them,
1130 * just like we do for '->unlink()'.
1132 if (new_size || err != -ENOSPC)
1137 truncate_setsize(inode, new_size);
1140 pgoff_t index = new_size >> PAGE_CACHE_SHIFT;
1143 page = find_lock_page(inode->i_mapping, index);
1145 if (PageDirty(page)) {
1147 * 'ubifs_jnl_truncate()' will try to truncate
1148 * the last data node, but it contains
1149 * out-of-date data because the page is dirty.
1150 * Write the page now, so that
1151 * 'ubifs_jnl_truncate()' will see an already
1152 * truncated (and up to date) data node.
1154 ubifs_assert(PagePrivate(page));
1156 clear_page_dirty_for_io(page);
1157 if (UBIFS_BLOCKS_PER_PAGE_SHIFT)
1159 (PAGE_CACHE_SIZE - 1);
1160 err = do_writepage(page, offset);
1161 page_cache_release(page);
1165 * We could now tell 'ubifs_jnl_truncate()' not
1166 * to read the last block.
1170 * We could 'kmap()' the page and pass the data
1171 * to 'ubifs_jnl_truncate()' to save it from
1172 * having to read it.
1175 page_cache_release(page);
1180 mutex_lock(&ui->ui_mutex);
1181 ui->ui_size = inode->i_size;
1182 /* Truncation changes inode [mc]time */
1183 inode->i_mtime = inode->i_ctime = ubifs_current_time(inode);
1184 /* Other attributes may be changed at the same time as well */
1185 do_attr_changes(inode, attr);
1186 err = ubifs_jnl_truncate(c, inode, old_size, new_size);
1187 mutex_unlock(&ui->ui_mutex);
1191 ubifs_release_budget(c, &req);
1193 c->bi.nospace = c->bi.nospace_rp = 0;
1200 * do_setattr - change inode attributes.
1201 * @c: UBIFS file-system description object
1202 * @inode: inode to change attributes for
1203 * @attr: inode attribute changes description
1205 * This function implements VFS '->setattr()' call for all cases except
1206 * truncations to smaller size. Returns zero in case of success and a negative
1207 * error code in case of failure.
1209 static int do_setattr(struct ubifs_info *c, struct inode *inode,
1210 const struct iattr *attr)
1213 loff_t new_size = attr->ia_size;
1214 struct ubifs_inode *ui = ubifs_inode(inode);
1215 struct ubifs_budget_req req = { .dirtied_ino = 1,
1216 .dirtied_ino_d = ALIGN(ui->data_len, 8) };
1218 err = ubifs_budget_space(c, &req);
1222 if (attr->ia_valid & ATTR_SIZE) {
1223 dbg_gen("size %lld -> %lld", inode->i_size, new_size);
1224 truncate_setsize(inode, new_size);
1227 mutex_lock(&ui->ui_mutex);
1228 if (attr->ia_valid & ATTR_SIZE) {
1229 /* Truncation changes inode [mc]time */
1230 inode->i_mtime = inode->i_ctime = ubifs_current_time(inode);
1231 /* 'truncate_setsize()' changed @i_size, update @ui_size */
1232 ui->ui_size = inode->i_size;
1235 do_attr_changes(inode, attr);
1237 release = ui->dirty;
1238 if (attr->ia_valid & ATTR_SIZE)
1240 * Inode length changed, so we have to make sure
1241 * @I_DIRTY_DATASYNC is set.
1243 __mark_inode_dirty(inode, I_DIRTY_SYNC | I_DIRTY_DATASYNC);
1245 mark_inode_dirty_sync(inode);
1246 mutex_unlock(&ui->ui_mutex);
1249 ubifs_release_budget(c, &req);
1251 err = inode->i_sb->s_op->write_inode(inode, NULL);
1255 int ubifs_setattr(struct dentry *dentry, struct iattr *attr)
1258 struct inode *inode = dentry->d_inode;
1259 struct ubifs_info *c = inode->i_sb->s_fs_info;
1261 dbg_gen("ino %lu, mode %#x, ia_valid %#x",
1262 inode->i_ino, inode->i_mode, attr->ia_valid);
1263 err = inode_change_ok(inode, attr);
1267 err = dbg_check_synced_i_size(c, inode);
1271 if ((attr->ia_valid & ATTR_SIZE) && attr->ia_size < inode->i_size)
1272 /* Truncation to a smaller size */
1273 err = do_truncation(c, inode, attr);
1275 err = do_setattr(c, inode, attr);
1280 static void ubifs_invalidatepage(struct page *page, unsigned long offset)
1282 struct inode *inode = page->mapping->host;
1283 struct ubifs_info *c = inode->i_sb->s_fs_info;
1285 ubifs_assert(PagePrivate(page));
1287 /* Partial page remains dirty */
1290 if (PageChecked(page))
1291 release_new_page_budget(c);
1293 release_existing_page_budget(c);
1295 atomic_long_dec(&c->dirty_pg_cnt);
1296 ClearPagePrivate(page);
1297 ClearPageChecked(page);
1300 static void *ubifs_follow_link(struct dentry *dentry, struct nameidata *nd)
1302 struct ubifs_inode *ui = ubifs_inode(dentry->d_inode);
1304 nd_set_link(nd, ui->data);
1308 int ubifs_fsync(struct file *file, loff_t start, loff_t end, int datasync)
1310 struct inode *inode = file->f_mapping->host;
1311 struct ubifs_info *c = inode->i_sb->s_fs_info;
1314 dbg_gen("syncing inode %lu", inode->i_ino);
1318 * For some really strange reasons VFS does not filter out
1319 * 'fsync()' for R/O mounted file-systems as per 2.6.39.
1323 err = filemap_write_and_wait_range(inode->i_mapping, start, end);
1326 mutex_lock(&inode->i_mutex);
1328 /* Synchronize the inode unless this is a 'datasync()' call. */
1329 if (!datasync || (inode->i_state & I_DIRTY_DATASYNC)) {
1330 err = inode->i_sb->s_op->write_inode(inode, NULL);
1336 * Nodes related to this inode may still sit in a write-buffer. Flush
1339 err = ubifs_sync_wbufs_by_inode(c, inode);
1341 mutex_unlock(&inode->i_mutex);
1346 * mctime_update_needed - check if mtime or ctime update is needed.
1347 * @inode: the inode to do the check for
1348 * @now: current time
1350 * This helper function checks if the inode mtime/ctime should be updated or
1351 * not. If current values of the time-stamps are within the UBIFS inode time
1352 * granularity, they are not updated. This is an optimization.
1354 static inline int mctime_update_needed(const struct inode *inode,
1355 const struct timespec *now)
1357 if (!timespec_equal(&inode->i_mtime, now) ||
1358 !timespec_equal(&inode->i_ctime, now))
1364 * update_ctime - update mtime and ctime of an inode.
1365 * @c: UBIFS file-system description object
1366 * @inode: inode to update
1368 * This function updates mtime and ctime of the inode if it is not equivalent to
1369 * current time. Returns zero in case of success and a negative error code in
1372 static int update_mctime(struct ubifs_info *c, struct inode *inode)
1374 struct timespec now = ubifs_current_time(inode);
1375 struct ubifs_inode *ui = ubifs_inode(inode);
1377 if (mctime_update_needed(inode, &now)) {
1379 struct ubifs_budget_req req = { .dirtied_ino = 1,
1380 .dirtied_ino_d = ALIGN(ui->data_len, 8) };
1382 err = ubifs_budget_space(c, &req);
1386 mutex_lock(&ui->ui_mutex);
1387 inode->i_mtime = inode->i_ctime = ubifs_current_time(inode);
1388 release = ui->dirty;
1389 mark_inode_dirty_sync(inode);
1390 mutex_unlock(&ui->ui_mutex);
1392 ubifs_release_budget(c, &req);
1398 static ssize_t ubifs_aio_write(struct kiocb *iocb, const struct iovec *iov,
1399 unsigned long nr_segs, loff_t pos)
1402 struct inode *inode = iocb->ki_filp->f_mapping->host;
1403 struct ubifs_info *c = inode->i_sb->s_fs_info;
1405 err = update_mctime(c, inode);
1409 return generic_file_aio_write(iocb, iov, nr_segs, pos);
1412 static int ubifs_set_page_dirty(struct page *page)
1416 ret = __set_page_dirty_nobuffers(page);
1418 * An attempt to dirty a page without budgeting for it - should not
1421 ubifs_assert(ret == 0);
1425 static int ubifs_releasepage(struct page *page, gfp_t unused_gfp_flags)
1428 * An attempt to release a dirty page without budgeting for it - should
1431 if (PageWriteback(page))
1433 ubifs_assert(PagePrivate(page));
1435 ClearPagePrivate(page);
1436 ClearPageChecked(page);
1441 * mmap()d file has taken write protection fault and is being made writable.
1442 * UBIFS must ensure page is budgeted for.
1444 static int ubifs_vm_page_mkwrite(struct vm_area_struct *vma,
1445 struct vm_fault *vmf)
1447 struct page *page = vmf->page;
1448 struct inode *inode = file_inode(vma->vm_file);
1449 struct ubifs_info *c = inode->i_sb->s_fs_info;
1450 struct timespec now = ubifs_current_time(inode);
1451 struct ubifs_budget_req req = { .new_page = 1 };
1452 int err, update_time;
1454 dbg_gen("ino %lu, pg %lu, i_size %lld", inode->i_ino, page->index,
1455 i_size_read(inode));
1456 ubifs_assert(!c->ro_media && !c->ro_mount);
1458 if (unlikely(c->ro_error))
1459 return VM_FAULT_SIGBUS; /* -EROFS */
1462 * We have not locked @page so far so we may budget for changing the
1463 * page. Note, we cannot do this after we locked the page, because
1464 * budgeting may cause write-back which would cause deadlock.
1466 * At the moment we do not know whether the page is dirty or not, so we
1467 * assume that it is not and budget for a new page. We could look at
1468 * the @PG_private flag and figure this out, but we may race with write
1469 * back and the page state may change by the time we lock it, so this
1470 * would need additional care. We do not bother with this at the
1471 * moment, although it might be good idea to do. Instead, we allocate
1472 * budget for a new page and amend it later on if the page was in fact
1475 * The budgeting-related logic of this function is similar to what we
1476 * do in 'ubifs_write_begin()' and 'ubifs_write_end()'. Glance there
1477 * for more comments.
1479 update_time = mctime_update_needed(inode, &now);
1482 * We have to change inode time stamp which requires extra
1485 req.dirtied_ino = 1;
1487 err = ubifs_budget_space(c, &req);
1488 if (unlikely(err)) {
1490 ubifs_warn("out of space for mmapped file (inode number %lu)",
1492 return VM_FAULT_SIGBUS;
1496 if (unlikely(page->mapping != inode->i_mapping ||
1497 page_offset(page) > i_size_read(inode))) {
1498 /* Page got truncated out from underneath us */
1503 if (PagePrivate(page))
1504 release_new_page_budget(c);
1506 if (!PageChecked(page))
1507 ubifs_convert_page_budget(c);
1508 SetPagePrivate(page);
1509 atomic_long_inc(&c->dirty_pg_cnt);
1510 __set_page_dirty_nobuffers(page);
1515 struct ubifs_inode *ui = ubifs_inode(inode);
1517 mutex_lock(&ui->ui_mutex);
1518 inode->i_mtime = inode->i_ctime = ubifs_current_time(inode);
1519 release = ui->dirty;
1520 mark_inode_dirty_sync(inode);
1521 mutex_unlock(&ui->ui_mutex);
1523 ubifs_release_dirty_inode_budget(c, ui);
1526 wait_for_stable_page(page);
1532 ubifs_release_budget(c, &req);
1534 err = VM_FAULT_SIGBUS;
1538 static const struct vm_operations_struct ubifs_file_vm_ops = {
1539 .fault = filemap_fault,
1540 .page_mkwrite = ubifs_vm_page_mkwrite,
1541 .remap_pages = generic_file_remap_pages,
1544 static int ubifs_file_mmap(struct file *file, struct vm_area_struct *vma)
1548 err = generic_file_mmap(file, vma);
1551 vma->vm_ops = &ubifs_file_vm_ops;
1555 const struct address_space_operations ubifs_file_address_operations = {
1556 .readpage = ubifs_readpage,
1557 .writepage = ubifs_writepage,
1558 .write_begin = ubifs_write_begin,
1559 .write_end = ubifs_write_end,
1560 .invalidatepage = ubifs_invalidatepage,
1561 .set_page_dirty = ubifs_set_page_dirty,
1562 .releasepage = ubifs_releasepage,
1565 const struct inode_operations ubifs_file_inode_operations = {
1566 .setattr = ubifs_setattr,
1567 .getattr = ubifs_getattr,
1568 .setxattr = ubifs_setxattr,
1569 .getxattr = ubifs_getxattr,
1570 .listxattr = ubifs_listxattr,
1571 .removexattr = ubifs_removexattr,
1574 const struct inode_operations ubifs_symlink_inode_operations = {
1575 .readlink = generic_readlink,
1576 .follow_link = ubifs_follow_link,
1577 .setattr = ubifs_setattr,
1578 .getattr = ubifs_getattr,
1581 const struct file_operations ubifs_file_operations = {
1582 .llseek = generic_file_llseek,
1583 .read = do_sync_read,
1584 .write = do_sync_write,
1585 .aio_read = generic_file_aio_read,
1586 .aio_write = ubifs_aio_write,
1587 .mmap = ubifs_file_mmap,
1588 .fsync = ubifs_fsync,
1589 .unlocked_ioctl = ubifs_ioctl,
1590 .splice_read = generic_file_splice_read,
1591 .splice_write = generic_file_splice_write,
1592 #ifdef CONFIG_COMPAT
1593 .compat_ioctl = ubifs_compat_ioctl,