1 // SPDX-License-Identifier: GPL-2.0-only
3 * This file is part of UBIFS.
5 * Copyright (C) 2006-2008 Nokia Corporation.
7 * Authors: Artem Bityutskiy (Битюцкий Артём)
12 * This file implements VFS file and inode operations for regular files, device
13 * nodes and symlinks as well as address space operations.
15 * UBIFS uses 2 page flags: @PG_private and @PG_checked. @PG_private is set if
16 * the page is dirty and is used for optimization purposes - dirty pages are
17 * not budgeted so the flag shows that 'ubifs_write_end()' should not release
18 * the budget for this page. The @PG_checked flag is set if full budgeting is
19 * required for the page e.g., when it corresponds to a file hole or it is
20 * beyond the file size. The budgeting is done in 'ubifs_write_begin()', because
21 * it is OK to fail in this function, and the budget is released in
22 * 'ubifs_write_end()'. So the @PG_private and @PG_checked flags carry
23 * information about how the page was budgeted, to make it possible to release
24 * the budget properly.
26 * A thing to keep in mind: inode @i_mutex is locked in most VFS operations we
27 * implement. However, this is not true for 'ubifs_writepage()', which may be
28 * called with @i_mutex unlocked. For example, when flusher thread is doing
29 * background write-back, it calls 'ubifs_writepage()' with unlocked @i_mutex.
30 * At "normal" work-paths the @i_mutex is locked in 'ubifs_writepage()', e.g.
31 * in the "sys_write -> alloc_pages -> direct reclaim path". So, in
32 * 'ubifs_writepage()' we are only guaranteed that the page is locked.
34 * Similarly, @i_mutex is not always locked in 'ubifs_readpage()', e.g., the
35 * read-ahead path does not lock it ("sys_read -> generic_file_aio_read ->
36 * ondemand_readahead -> readpage"). In case of readahead, @I_SYNC flag is not
37 * set as well. However, UBIFS disables readahead.
41 #include <linux/mount.h>
42 #include <linux/slab.h>
43 #include <linux/migrate.h>
45 static int read_block(struct inode *inode, void *addr, unsigned int block,
46 struct ubifs_data_node *dn)
48 struct ubifs_info *c = inode->i_sb->s_fs_info;
49 int err, len, out_len;
53 data_key_init(c, &key, inode->i_ino, block);
54 err = ubifs_tnc_lookup(c, &key, dn);
57 /* Not found, so it must be a hole */
58 memset(addr, 0, UBIFS_BLOCK_SIZE);
62 ubifs_assert(c, le64_to_cpu(dn->ch.sqnum) >
63 ubifs_inode(inode)->creat_sqnum);
64 len = le32_to_cpu(dn->size);
65 if (len <= 0 || len > UBIFS_BLOCK_SIZE)
68 dlen = le32_to_cpu(dn->ch.len) - UBIFS_DATA_NODE_SZ;
70 if (IS_ENCRYPTED(inode)) {
71 err = ubifs_decrypt(inode, dn, &dlen, block);
76 out_len = UBIFS_BLOCK_SIZE;
77 err = ubifs_decompress(c, &dn->data, dlen, addr, &out_len,
78 le16_to_cpu(dn->compr_type));
79 if (err || len != out_len)
83 * Data length can be less than a full block, even for blocks that are
84 * not the last in the file (e.g., as a result of making a hole and
85 * appending data). Ensure that the remainder is zeroed out.
87 if (len < UBIFS_BLOCK_SIZE)
88 memset(addr + len, 0, UBIFS_BLOCK_SIZE - len);
93 ubifs_err(c, "bad data node (block %u, inode %lu)",
95 ubifs_dump_node(c, dn);
99 static int do_readpage(struct page *page)
103 unsigned int block, beyond;
104 struct ubifs_data_node *dn;
105 struct inode *inode = page->mapping->host;
106 struct ubifs_info *c = inode->i_sb->s_fs_info;
107 loff_t i_size = i_size_read(inode);
109 dbg_gen("ino %lu, pg %lu, i_size %lld, flags %#lx",
110 inode->i_ino, page->index, i_size, page->flags);
111 ubifs_assert(c, !PageChecked(page));
112 ubifs_assert(c, !PagePrivate(page));
116 block = page->index << UBIFS_BLOCKS_PER_PAGE_SHIFT;
117 beyond = (i_size + UBIFS_BLOCK_SIZE - 1) >> UBIFS_BLOCK_SHIFT;
118 if (block >= beyond) {
119 /* Reading beyond inode */
120 SetPageChecked(page);
121 memset(addr, 0, PAGE_SIZE);
125 dn = kmalloc(UBIFS_MAX_DATA_NODE_SZ, GFP_NOFS);
135 if (block >= beyond) {
136 /* Reading beyond inode */
138 memset(addr, 0, UBIFS_BLOCK_SIZE);
140 ret = read_block(inode, addr, block, dn);
145 } else if (block + 1 == beyond) {
146 int dlen = le32_to_cpu(dn->size);
147 int ilen = i_size & (UBIFS_BLOCK_SIZE - 1);
149 if (ilen && ilen < dlen)
150 memset(addr + ilen, 0, dlen - ilen);
153 if (++i >= UBIFS_BLOCKS_PER_PAGE)
156 addr += UBIFS_BLOCK_SIZE;
159 struct ubifs_info *c = inode->i_sb->s_fs_info;
160 if (err == -ENOENT) {
161 /* Not found, so it must be a hole */
162 SetPageChecked(page);
166 ubifs_err(c, "cannot read page %lu of inode %lu, error %d",
167 page->index, inode->i_ino, err);
174 SetPageUptodate(page);
175 ClearPageError(page);
176 flush_dcache_page(page);
182 ClearPageUptodate(page);
184 flush_dcache_page(page);
190 * release_new_page_budget - release budget of a new page.
191 * @c: UBIFS file-system description object
193 * This is a helper function which releases budget corresponding to the budget
194 * of one new page of data.
196 static void release_new_page_budget(struct ubifs_info *c)
198 struct ubifs_budget_req req = { .recalculate = 1, .new_page = 1 };
200 ubifs_release_budget(c, &req);
204 * release_existing_page_budget - release budget of an existing page.
205 * @c: UBIFS file-system description object
207 * This is a helper function which releases budget corresponding to the budget
208 * of changing one one page of data which already exists on the flash media.
210 static void release_existing_page_budget(struct ubifs_info *c)
212 struct ubifs_budget_req req = { .dd_growth = c->bi.page_budget};
214 ubifs_release_budget(c, &req);
217 static int write_begin_slow(struct address_space *mapping,
218 loff_t pos, unsigned len, struct page **pagep,
221 struct inode *inode = mapping->host;
222 struct ubifs_info *c = inode->i_sb->s_fs_info;
223 pgoff_t index = pos >> PAGE_SHIFT;
224 struct ubifs_budget_req req = { .new_page = 1 };
225 int err, appending = !!(pos + len > inode->i_size);
228 dbg_gen("ino %lu, pos %llu, len %u, i_size %lld",
229 inode->i_ino, pos, len, inode->i_size);
232 * At the slow path we have to budget before locking the page, because
233 * budgeting may force write-back, which would wait on locked pages and
234 * deadlock if we had the page locked. At this point we do not know
235 * anything about the page, so assume that this is a new page which is
236 * written to a hole. This corresponds to largest budget. Later the
237 * budget will be amended if this is not true.
240 /* We are appending data, budget for inode change */
243 err = ubifs_budget_space(c, &req);
247 page = grab_cache_page_write_begin(mapping, index, flags);
248 if (unlikely(!page)) {
249 ubifs_release_budget(c, &req);
253 if (!PageUptodate(page)) {
254 if (!(pos & ~PAGE_MASK) && len == PAGE_SIZE)
255 SetPageChecked(page);
257 err = do_readpage(page);
261 ubifs_release_budget(c, &req);
267 if (PagePrivate(page))
269 * The page is dirty, which means it was budgeted twice:
270 * o first time the budget was allocated by the task which
271 * made the page dirty and set the PG_private flag;
272 * o and then we budgeted for it for the second time at the
273 * very beginning of this function.
275 * So what we have to do is to release the page budget we
278 release_new_page_budget(c);
279 else if (!PageChecked(page))
281 * We are changing a page which already exists on the media.
282 * This means that changing the page does not make the amount
283 * of indexing information larger, and this part of the budget
284 * which we have already acquired may be released.
286 ubifs_convert_page_budget(c);
289 struct ubifs_inode *ui = ubifs_inode(inode);
292 * 'ubifs_write_end()' is optimized from the fast-path part of
293 * 'ubifs_write_begin()' and expects the @ui_mutex to be locked
294 * if data is appended.
296 mutex_lock(&ui->ui_mutex);
299 * The inode is dirty already, so we may free the
300 * budget we allocated.
302 ubifs_release_dirty_inode_budget(c, ui);
310 * allocate_budget - allocate budget for 'ubifs_write_begin()'.
311 * @c: UBIFS file-system description object
312 * @page: page to allocate budget for
313 * @ui: UBIFS inode object the page belongs to
314 * @appending: non-zero if the page is appended
316 * This is a helper function for 'ubifs_write_begin()' which allocates budget
317 * for the operation. The budget is allocated differently depending on whether
318 * this is appending, whether the page is dirty or not, and so on. This
319 * function leaves the @ui->ui_mutex locked in case of appending. Returns zero
320 * in case of success and %-ENOSPC in case of failure.
322 static int allocate_budget(struct ubifs_info *c, struct page *page,
323 struct ubifs_inode *ui, int appending)
325 struct ubifs_budget_req req = { .fast = 1 };
327 if (PagePrivate(page)) {
330 * The page is dirty and we are not appending, which
331 * means no budget is needed at all.
335 mutex_lock(&ui->ui_mutex);
338 * The page is dirty and we are appending, so the inode
339 * has to be marked as dirty. However, it is already
340 * dirty, so we do not need any budget. We may return,
341 * but @ui->ui_mutex hast to be left locked because we
342 * should prevent write-back from flushing the inode
343 * and freeing the budget. The lock will be released in
344 * 'ubifs_write_end()'.
349 * The page is dirty, we are appending, the inode is clean, so
350 * we need to budget the inode change.
354 if (PageChecked(page))
356 * The page corresponds to a hole and does not
357 * exist on the media. So changing it makes
358 * make the amount of indexing information
359 * larger, and we have to budget for a new
365 * Not a hole, the change will not add any new
366 * indexing information, budget for page
369 req.dirtied_page = 1;
372 mutex_lock(&ui->ui_mutex);
375 * The inode is clean but we will have to mark
376 * it as dirty because we are appending. This
383 return ubifs_budget_space(c, &req);
387 * This function is called when a page of data is going to be written. Since
388 * the page of data will not necessarily go to the flash straight away, UBIFS
389 * has to reserve space on the media for it, which is done by means of
392 * This is the hot-path of the file-system and we are trying to optimize it as
393 * much as possible. For this reasons it is split on 2 parts - slow and fast.
395 * There many budgeting cases:
396 * o a new page is appended - we have to budget for a new page and for
397 * changing the inode; however, if the inode is already dirty, there is
398 * no need to budget for it;
399 * o an existing clean page is changed - we have budget for it; if the page
400 * does not exist on the media (a hole), we have to budget for a new
401 * page; otherwise, we may budget for changing an existing page; the
402 * difference between these cases is that changing an existing page does
403 * not introduce anything new to the FS indexing information, so it does
404 * not grow, and smaller budget is acquired in this case;
405 * o an existing dirty page is changed - no need to budget at all, because
406 * the page budget has been acquired by earlier, when the page has been
409 * UBIFS budgeting sub-system may force write-back if it thinks there is no
410 * space to reserve. This imposes some locking restrictions and makes it
411 * impossible to take into account the above cases, and makes it impossible to
412 * optimize budgeting.
414 * The solution for this is that the fast path of 'ubifs_write_begin()' assumes
415 * there is a plenty of flash space and the budget will be acquired quickly,
416 * without forcing write-back. The slow path does not make this assumption.
418 static int ubifs_write_begin(struct file *file, struct address_space *mapping,
419 loff_t pos, unsigned len, unsigned flags,
420 struct page **pagep, void **fsdata)
422 struct inode *inode = mapping->host;
423 struct ubifs_info *c = inode->i_sb->s_fs_info;
424 struct ubifs_inode *ui = ubifs_inode(inode);
425 pgoff_t index = pos >> PAGE_SHIFT;
426 int err, appending = !!(pos + len > inode->i_size);
427 int skipped_read = 0;
430 ubifs_assert(c, ubifs_inode(inode)->ui_size == inode->i_size);
431 ubifs_assert(c, !c->ro_media && !c->ro_mount);
433 if (unlikely(c->ro_error))
436 /* Try out the fast-path part first */
437 page = grab_cache_page_write_begin(mapping, index, flags);
441 if (!PageUptodate(page)) {
442 /* The page is not loaded from the flash */
443 if (!(pos & ~PAGE_MASK) && len == PAGE_SIZE) {
445 * We change whole page so no need to load it. But we
446 * do not know whether this page exists on the media or
447 * not, so we assume the latter because it requires
448 * larger budget. The assumption is that it is better
449 * to budget a bit more than to read the page from the
450 * media. Thus, we are setting the @PG_checked flag
453 SetPageChecked(page);
456 err = do_readpage(page);
465 err = allocate_budget(c, page, ui, appending);
467 ubifs_assert(c, err == -ENOSPC);
469 * If we skipped reading the page because we were going to
470 * write all of it, then it is not up to date.
473 ClearPageChecked(page);
475 * Budgeting failed which means it would have to force
476 * write-back but didn't, because we set the @fast flag in the
477 * request. Write-back cannot be done now, while we have the
478 * page locked, because it would deadlock. Unlock and free
479 * everything and fall-back to slow-path.
482 ubifs_assert(c, mutex_is_locked(&ui->ui_mutex));
483 mutex_unlock(&ui->ui_mutex);
488 return write_begin_slow(mapping, pos, len, pagep, flags);
492 * Whee, we acquired budgeting quickly - without involving
493 * garbage-collection, committing or forcing write-back. We return
494 * with @ui->ui_mutex locked if we are appending pages, and unlocked
495 * otherwise. This is an optimization (slightly hacky though).
503 * cancel_budget - cancel budget.
504 * @c: UBIFS file-system description object
505 * @page: page to cancel budget for
506 * @ui: UBIFS inode object the page belongs to
507 * @appending: non-zero if the page is appended
509 * This is a helper function for a page write operation. It unlocks the
510 * @ui->ui_mutex in case of appending.
512 static void cancel_budget(struct ubifs_info *c, struct page *page,
513 struct ubifs_inode *ui, int appending)
517 ubifs_release_dirty_inode_budget(c, ui);
518 mutex_unlock(&ui->ui_mutex);
520 if (!PagePrivate(page)) {
521 if (PageChecked(page))
522 release_new_page_budget(c);
524 release_existing_page_budget(c);
528 static int ubifs_write_end(struct file *file, struct address_space *mapping,
529 loff_t pos, unsigned len, unsigned copied,
530 struct page *page, void *fsdata)
532 struct inode *inode = mapping->host;
533 struct ubifs_inode *ui = ubifs_inode(inode);
534 struct ubifs_info *c = inode->i_sb->s_fs_info;
535 loff_t end_pos = pos + len;
536 int appending = !!(end_pos > inode->i_size);
538 dbg_gen("ino %lu, pos %llu, pg %lu, len %u, copied %d, i_size %lld",
539 inode->i_ino, pos, page->index, len, copied, inode->i_size);
541 if (unlikely(copied < len && len == PAGE_SIZE)) {
543 * VFS copied less data to the page that it intended and
544 * declared in its '->write_begin()' call via the @len
545 * argument. If the page was not up-to-date, and @len was
546 * @PAGE_SIZE, the 'ubifs_write_begin()' function did
547 * not load it from the media (for optimization reasons). This
548 * means that part of the page contains garbage. So read the
551 dbg_gen("copied %d instead of %d, read page and repeat",
553 cancel_budget(c, page, ui, appending);
554 ClearPageChecked(page);
557 * Return 0 to force VFS to repeat the whole operation, or the
558 * error code if 'do_readpage()' fails.
560 copied = do_readpage(page);
564 if (len == PAGE_SIZE)
565 SetPageUptodate(page);
567 if (!PagePrivate(page)) {
568 attach_page_private(page, (void *)1);
569 atomic_long_inc(&c->dirty_pg_cnt);
570 __set_page_dirty_nobuffers(page);
574 i_size_write(inode, end_pos);
575 ui->ui_size = end_pos;
577 * Note, we do not set @I_DIRTY_PAGES (which means that the
578 * inode has dirty pages), this has been done in
579 * '__set_page_dirty_nobuffers()'.
581 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
582 ubifs_assert(c, mutex_is_locked(&ui->ui_mutex));
583 mutex_unlock(&ui->ui_mutex);
593 * populate_page - copy data nodes into a page for bulk-read.
594 * @c: UBIFS file-system description object
596 * @bu: bulk-read information
597 * @n: next zbranch slot
599 * This function returns %0 on success and a negative error code on failure.
601 static int populate_page(struct ubifs_info *c, struct page *page,
602 struct bu_info *bu, int *n)
604 int i = 0, nn = *n, offs = bu->zbranch[0].offs, hole = 0, read = 0;
605 struct inode *inode = page->mapping->host;
606 loff_t i_size = i_size_read(inode);
607 unsigned int page_block;
611 dbg_gen("ino %lu, pg %lu, i_size %lld, flags %#lx",
612 inode->i_ino, page->index, i_size, page->flags);
614 addr = zaddr = kmap(page);
616 end_index = (i_size - 1) >> PAGE_SHIFT;
617 if (!i_size || page->index > end_index) {
619 memset(addr, 0, PAGE_SIZE);
623 page_block = page->index << UBIFS_BLOCKS_PER_PAGE_SHIFT;
625 int err, len, out_len, dlen;
629 memset(addr, 0, UBIFS_BLOCK_SIZE);
630 } else if (key_block(c, &bu->zbranch[nn].key) == page_block) {
631 struct ubifs_data_node *dn;
633 dn = bu->buf + (bu->zbranch[nn].offs - offs);
635 ubifs_assert(c, le64_to_cpu(dn->ch.sqnum) >
636 ubifs_inode(inode)->creat_sqnum);
638 len = le32_to_cpu(dn->size);
639 if (len <= 0 || len > UBIFS_BLOCK_SIZE)
642 dlen = le32_to_cpu(dn->ch.len) - UBIFS_DATA_NODE_SZ;
643 out_len = UBIFS_BLOCK_SIZE;
645 if (IS_ENCRYPTED(inode)) {
646 err = ubifs_decrypt(inode, dn, &dlen, page_block);
651 err = ubifs_decompress(c, &dn->data, dlen, addr, &out_len,
652 le16_to_cpu(dn->compr_type));
653 if (err || len != out_len)
656 if (len < UBIFS_BLOCK_SIZE)
657 memset(addr + len, 0, UBIFS_BLOCK_SIZE - len);
660 read = (i << UBIFS_BLOCK_SHIFT) + len;
661 } else if (key_block(c, &bu->zbranch[nn].key) < page_block) {
666 memset(addr, 0, UBIFS_BLOCK_SIZE);
668 if (++i >= UBIFS_BLOCKS_PER_PAGE)
670 addr += UBIFS_BLOCK_SIZE;
674 if (end_index == page->index) {
675 int len = i_size & (PAGE_SIZE - 1);
677 if (len && len < read)
678 memset(zaddr + len, 0, read - len);
683 SetPageChecked(page);
687 SetPageUptodate(page);
688 ClearPageError(page);
689 flush_dcache_page(page);
695 ClearPageUptodate(page);
697 flush_dcache_page(page);
699 ubifs_err(c, "bad data node (block %u, inode %lu)",
700 page_block, inode->i_ino);
705 * ubifs_do_bulk_read - do bulk-read.
706 * @c: UBIFS file-system description object
707 * @bu: bulk-read information
708 * @page1: first page to read
710 * This function returns %1 if the bulk-read is done, otherwise %0 is returned.
712 static int ubifs_do_bulk_read(struct ubifs_info *c, struct bu_info *bu,
715 pgoff_t offset = page1->index, end_index;
716 struct address_space *mapping = page1->mapping;
717 struct inode *inode = mapping->host;
718 struct ubifs_inode *ui = ubifs_inode(inode);
719 int err, page_idx, page_cnt, ret = 0, n = 0;
720 int allocate = bu->buf ? 0 : 1;
722 gfp_t ra_gfp_mask = readahead_gfp_mask(mapping) & ~__GFP_FS;
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(c, bu->buf_len > 0);
755 ubifs_assert(c, 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_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 = pagecache_get_page(mapping, page_offset,
785 FGP_LOCK|FGP_ACCESSED|FGP_CREAT|FGP_NOWAIT,
789 if (!PageUptodate(page))
790 err = populate_page(c, page, bu, &n);
797 ui->last_page_read = offset + page_idx - 1;
805 ubifs_warn(c, "ignoring error %d and skipping bulk-read", err);
809 ui->read_in_a_row = ui->bulk_read = 0;
814 * ubifs_bulk_read - determine whether to bulk-read and, if so, do it.
815 * @page: page from which to start bulk-read.
817 * Some flash media are capable of reading sequentially at faster rates. UBIFS
818 * bulk-read facility is designed to take advantage of that, by reading in one
819 * go consecutive data nodes that are also located consecutively in the same
820 * LEB. This function returns %1 if a bulk-read is done and %0 otherwise.
822 static int ubifs_bulk_read(struct page *page)
824 struct inode *inode = page->mapping->host;
825 struct ubifs_info *c = inode->i_sb->s_fs_info;
826 struct ubifs_inode *ui = ubifs_inode(inode);
827 pgoff_t index = page->index, last_page_read = ui->last_page_read;
829 int err = 0, allocated = 0;
831 ui->last_page_read = index;
836 * Bulk-read is protected by @ui->ui_mutex, but it is an optimization,
837 * so don't bother if we cannot lock the mutex.
839 if (!mutex_trylock(&ui->ui_mutex))
842 if (index != last_page_read + 1) {
843 /* Turn off bulk-read if we stop reading sequentially */
844 ui->read_in_a_row = 1;
850 if (!ui->bulk_read) {
851 ui->read_in_a_row += 1;
852 if (ui->read_in_a_row < 3)
854 /* Three reads in a row, so switch on bulk-read */
859 * If possible, try to use pre-allocated bulk-read information, which
860 * is protected by @c->bu_mutex.
862 if (mutex_trylock(&c->bu_mutex))
865 bu = kmalloc(sizeof(struct bu_info), GFP_NOFS | __GFP_NOWARN);
873 bu->buf_len = c->max_bu_buf_len;
874 data_key_init(c, &bu->key, inode->i_ino,
875 page->index << UBIFS_BLOCKS_PER_PAGE_SHIFT);
876 err = ubifs_do_bulk_read(c, bu, page);
879 mutex_unlock(&c->bu_mutex);
884 mutex_unlock(&ui->ui_mutex);
888 static int ubifs_readpage(struct file *file, struct page *page)
890 if (ubifs_bulk_read(page))
897 static int do_writepage(struct page *page, int len)
899 int err = 0, i, blen;
903 struct inode *inode = page->mapping->host;
904 struct ubifs_info *c = inode->i_sb->s_fs_info;
907 struct ubifs_inode *ui = ubifs_inode(inode);
908 spin_lock(&ui->ui_lock);
909 ubifs_assert(c, page->index <= ui->synced_i_size >> PAGE_SHIFT);
910 spin_unlock(&ui->ui_lock);
913 /* Update radix tree tags */
914 set_page_writeback(page);
917 block = page->index << UBIFS_BLOCKS_PER_PAGE_SHIFT;
920 blen = min_t(int, len, UBIFS_BLOCK_SIZE);
921 data_key_init(c, &key, inode->i_ino, block);
922 err = ubifs_jnl_write_data(c, inode, &key, addr, blen);
925 if (++i >= UBIFS_BLOCKS_PER_PAGE)
933 ubifs_err(c, "cannot write page %lu of inode %lu, error %d",
934 page->index, inode->i_ino, err);
935 ubifs_ro_mode(c, err);
938 ubifs_assert(c, PagePrivate(page));
939 if (PageChecked(page))
940 release_new_page_budget(c);
942 release_existing_page_budget(c);
944 atomic_long_dec(&c->dirty_pg_cnt);
945 detach_page_private(page);
946 ClearPageChecked(page);
950 end_page_writeback(page);
955 * When writing-back dirty inodes, VFS first writes-back pages belonging to the
956 * inode, then the inode itself. For UBIFS this may cause a problem. Consider a
957 * situation when a we have an inode with size 0, then a megabyte of data is
958 * appended to the inode, then write-back starts and flushes some amount of the
959 * dirty pages, the journal becomes full, commit happens and finishes, and then
960 * an unclean reboot happens. When the file system is mounted next time, the
961 * inode size would still be 0, but there would be many pages which are beyond
962 * the inode size, they would be indexed and consume flash space. Because the
963 * journal has been committed, the replay would not be able to detect this
964 * situation and correct the inode size. This means UBIFS would have to scan
965 * whole index and correct all inode sizes, which is long an unacceptable.
967 * To prevent situations like this, UBIFS writes pages back only if they are
968 * within the last synchronized inode size, i.e. the size which has been
969 * written to the flash media last time. Otherwise, UBIFS forces inode
970 * write-back, thus making sure the on-flash inode contains current inode size,
971 * and then keeps writing pages back.
973 * Some locking issues explanation. 'ubifs_writepage()' first is called with
974 * the page locked, and it locks @ui_mutex. However, write-back does take inode
975 * @i_mutex, which means other VFS operations may be run on this inode at the
976 * same time. And the problematic one is truncation to smaller size, from where
977 * we have to call 'truncate_setsize()', which first changes @inode->i_size,
978 * then drops the truncated pages. And while dropping the pages, it takes the
979 * page lock. This means that 'do_truncation()' cannot call 'truncate_setsize()'
980 * with @ui_mutex locked, because it would deadlock with 'ubifs_writepage()'.
981 * This means that @inode->i_size is changed while @ui_mutex is unlocked.
983 * XXX(truncate): with the new truncate sequence this is not true anymore,
984 * and the calls to truncate_setsize can be move around freely. They should
985 * be moved to the very end of the truncate sequence.
987 * But in 'ubifs_writepage()' we have to guarantee that we do not write beyond
988 * inode size. How do we do this if @inode->i_size may became smaller while we
989 * are in the middle of 'ubifs_writepage()'? The UBIFS solution is the
990 * @ui->ui_isize "shadow" field which UBIFS uses instead of @inode->i_size
991 * internally and updates it under @ui_mutex.
993 * Q: why we do not worry that if we race with truncation, we may end up with a
994 * situation when the inode is truncated while we are in the middle of
995 * 'do_writepage()', so we do write beyond inode size?
996 * A: If we are in the middle of 'do_writepage()', truncation would be locked
997 * on the page lock and it would not write the truncated inode node to the
998 * journal before we have finished.
1000 static int ubifs_writepage(struct page *page, struct writeback_control *wbc)
1002 struct inode *inode = page->mapping->host;
1003 struct ubifs_info *c = inode->i_sb->s_fs_info;
1004 struct ubifs_inode *ui = ubifs_inode(inode);
1005 loff_t i_size = i_size_read(inode), synced_i_size;
1006 pgoff_t end_index = i_size >> PAGE_SHIFT;
1007 int err, len = i_size & (PAGE_SIZE - 1);
1010 dbg_gen("ino %lu, pg %lu, pg flags %#lx",
1011 inode->i_ino, page->index, page->flags);
1012 ubifs_assert(c, PagePrivate(page));
1014 /* Is the page fully outside @i_size? (truncate in progress) */
1015 if (page->index > end_index || (page->index == end_index && !len)) {
1020 spin_lock(&ui->ui_lock);
1021 synced_i_size = ui->synced_i_size;
1022 spin_unlock(&ui->ui_lock);
1024 /* Is the page fully inside @i_size? */
1025 if (page->index < end_index) {
1026 if (page->index >= synced_i_size >> PAGE_SHIFT) {
1027 err = inode->i_sb->s_op->write_inode(inode, NULL);
1031 * The inode has been written, but the write-buffer has
1032 * not been synchronized, so in case of an unclean
1033 * reboot we may end up with some pages beyond inode
1034 * size, but they would be in the journal (because
1035 * commit flushes write buffers) and recovery would deal
1039 return do_writepage(page, PAGE_SIZE);
1043 * The page straddles @i_size. It must be zeroed out on each and every
1044 * writepage invocation because it may be mmapped. "A file is mapped
1045 * in multiples of the page size. For a file that is not a multiple of
1046 * the page size, the remaining memory is zeroed when mapped, and
1047 * writes to that region are not written out to the file."
1049 kaddr = kmap_atomic(page);
1050 memset(kaddr + len, 0, PAGE_SIZE - len);
1051 flush_dcache_page(page);
1052 kunmap_atomic(kaddr);
1054 if (i_size > synced_i_size) {
1055 err = inode->i_sb->s_op->write_inode(inode, NULL);
1060 return do_writepage(page, len);
1063 * redirty_page_for_writepage() won't call ubifs_dirty_inode() because
1064 * it passes I_DIRTY_PAGES flag while calling __mark_inode_dirty(), so
1065 * there is no need to do space budget for dirty inode.
1067 redirty_page_for_writepage(wbc, page);
1074 * do_attr_changes - change inode attributes.
1075 * @inode: inode to change attributes for
1076 * @attr: describes attributes to change
1078 static void do_attr_changes(struct inode *inode, const struct iattr *attr)
1080 if (attr->ia_valid & ATTR_UID)
1081 inode->i_uid = attr->ia_uid;
1082 if (attr->ia_valid & ATTR_GID)
1083 inode->i_gid = attr->ia_gid;
1084 if (attr->ia_valid & ATTR_ATIME)
1085 inode->i_atime = attr->ia_atime;
1086 if (attr->ia_valid & ATTR_MTIME)
1087 inode->i_mtime = attr->ia_mtime;
1088 if (attr->ia_valid & ATTR_CTIME)
1089 inode->i_ctime = attr->ia_ctime;
1090 if (attr->ia_valid & ATTR_MODE) {
1091 umode_t mode = attr->ia_mode;
1093 if (!in_group_p(inode->i_gid) && !capable(CAP_FSETID))
1095 inode->i_mode = mode;
1100 * do_truncation - truncate an inode.
1101 * @c: UBIFS file-system description object
1102 * @inode: inode to truncate
1103 * @attr: inode attribute changes description
1105 * This function implements VFS '->setattr()' call when the inode is truncated
1106 * to a smaller size. Returns zero in case of success and a negative error code
1107 * in case of failure.
1109 static int do_truncation(struct ubifs_info *c, struct inode *inode,
1110 const struct iattr *attr)
1113 struct ubifs_budget_req req;
1114 loff_t old_size = inode->i_size, new_size = attr->ia_size;
1115 int offset = new_size & (UBIFS_BLOCK_SIZE - 1), budgeted = 1;
1116 struct ubifs_inode *ui = ubifs_inode(inode);
1118 dbg_gen("ino %lu, size %lld -> %lld", inode->i_ino, old_size, new_size);
1119 memset(&req, 0, sizeof(struct ubifs_budget_req));
1122 * If this is truncation to a smaller size, and we do not truncate on a
1123 * block boundary, budget for changing one data block, because the last
1124 * block will be re-written.
1126 if (new_size & (UBIFS_BLOCK_SIZE - 1))
1127 req.dirtied_page = 1;
1129 req.dirtied_ino = 1;
1130 /* A funny way to budget for truncation node */
1131 req.dirtied_ino_d = UBIFS_TRUN_NODE_SZ;
1132 err = ubifs_budget_space(c, &req);
1135 * Treat truncations to zero as deletion and always allow them,
1136 * just like we do for '->unlink()'.
1138 if (new_size || err != -ENOSPC)
1143 truncate_setsize(inode, new_size);
1146 pgoff_t index = new_size >> PAGE_SHIFT;
1149 page = find_lock_page(inode->i_mapping, index);
1151 if (PageDirty(page)) {
1153 * 'ubifs_jnl_truncate()' will try to truncate
1154 * the last data node, but it contains
1155 * out-of-date data because the page is dirty.
1156 * Write the page now, so that
1157 * 'ubifs_jnl_truncate()' will see an already
1158 * truncated (and up to date) data node.
1160 ubifs_assert(c, PagePrivate(page));
1162 clear_page_dirty_for_io(page);
1163 if (UBIFS_BLOCKS_PER_PAGE_SHIFT)
1166 err = do_writepage(page, offset);
1171 * We could now tell 'ubifs_jnl_truncate()' not
1172 * to read the last block.
1176 * We could 'kmap()' the page and pass the data
1177 * to 'ubifs_jnl_truncate()' to save it from
1178 * having to read it.
1186 mutex_lock(&ui->ui_mutex);
1187 ui->ui_size = inode->i_size;
1188 /* Truncation changes inode [mc]time */
1189 inode->i_mtime = inode->i_ctime = current_time(inode);
1190 /* Other attributes may be changed at the same time as well */
1191 do_attr_changes(inode, attr);
1192 err = ubifs_jnl_truncate(c, inode, old_size, new_size);
1193 mutex_unlock(&ui->ui_mutex);
1197 ubifs_release_budget(c, &req);
1199 c->bi.nospace = c->bi.nospace_rp = 0;
1206 * do_setattr - change inode attributes.
1207 * @c: UBIFS file-system description object
1208 * @inode: inode to change attributes for
1209 * @attr: inode attribute changes description
1211 * This function implements VFS '->setattr()' call for all cases except
1212 * truncations to smaller size. Returns zero in case of success and a negative
1213 * error code in case of failure.
1215 static int do_setattr(struct ubifs_info *c, struct inode *inode,
1216 const struct iattr *attr)
1219 loff_t new_size = attr->ia_size;
1220 struct ubifs_inode *ui = ubifs_inode(inode);
1221 struct ubifs_budget_req req = { .dirtied_ino = 1,
1222 .dirtied_ino_d = ALIGN(ui->data_len, 8) };
1224 err = ubifs_budget_space(c, &req);
1228 if (attr->ia_valid & ATTR_SIZE) {
1229 dbg_gen("size %lld -> %lld", inode->i_size, new_size);
1230 truncate_setsize(inode, new_size);
1233 mutex_lock(&ui->ui_mutex);
1234 if (attr->ia_valid & ATTR_SIZE) {
1235 /* Truncation changes inode [mc]time */
1236 inode->i_mtime = inode->i_ctime = current_time(inode);
1237 /* 'truncate_setsize()' changed @i_size, update @ui_size */
1238 ui->ui_size = inode->i_size;
1241 do_attr_changes(inode, attr);
1243 release = ui->dirty;
1244 if (attr->ia_valid & ATTR_SIZE)
1246 * Inode length changed, so we have to make sure
1247 * @I_DIRTY_DATASYNC is set.
1249 __mark_inode_dirty(inode, I_DIRTY_DATASYNC);
1251 mark_inode_dirty_sync(inode);
1252 mutex_unlock(&ui->ui_mutex);
1255 ubifs_release_budget(c, &req);
1257 err = inode->i_sb->s_op->write_inode(inode, NULL);
1261 int ubifs_setattr(struct dentry *dentry, struct iattr *attr)
1264 struct inode *inode = d_inode(dentry);
1265 struct ubifs_info *c = inode->i_sb->s_fs_info;
1267 dbg_gen("ino %lu, mode %#x, ia_valid %#x",
1268 inode->i_ino, inode->i_mode, attr->ia_valid);
1269 err = setattr_prepare(dentry, attr);
1273 err = dbg_check_synced_i_size(c, inode);
1277 err = fscrypt_prepare_setattr(dentry, attr);
1281 if ((attr->ia_valid & ATTR_SIZE) && attr->ia_size < inode->i_size)
1282 /* Truncation to a smaller size */
1283 err = do_truncation(c, inode, attr);
1285 err = do_setattr(c, inode, attr);
1290 static void ubifs_invalidatepage(struct page *page, unsigned int offset,
1291 unsigned int length)
1293 struct inode *inode = page->mapping->host;
1294 struct ubifs_info *c = inode->i_sb->s_fs_info;
1296 ubifs_assert(c, PagePrivate(page));
1297 if (offset || length < PAGE_SIZE)
1298 /* Partial page remains dirty */
1301 if (PageChecked(page))
1302 release_new_page_budget(c);
1304 release_existing_page_budget(c);
1306 atomic_long_dec(&c->dirty_pg_cnt);
1307 detach_page_private(page);
1308 ClearPageChecked(page);
1311 int ubifs_fsync(struct file *file, loff_t start, loff_t end, int datasync)
1313 struct inode *inode = file->f_mapping->host;
1314 struct ubifs_info *c = inode->i_sb->s_fs_info;
1317 dbg_gen("syncing inode %lu", inode->i_ino);
1321 * For some really strange reasons VFS does not filter out
1322 * 'fsync()' for R/O mounted file-systems as per 2.6.39.
1326 err = file_write_and_wait_range(file, start, end);
1331 /* Synchronize the inode unless this is a 'datasync()' call. */
1332 if (!datasync || (inode->i_state & I_DIRTY_DATASYNC)) {
1333 err = inode->i_sb->s_op->write_inode(inode, NULL);
1339 * Nodes related to this inode may still sit in a write-buffer. Flush
1342 err = ubifs_sync_wbufs_by_inode(c, inode);
1344 inode_unlock(inode);
1349 * mctime_update_needed - check if mtime or ctime update is needed.
1350 * @inode: the inode to do the check for
1351 * @now: current time
1353 * This helper function checks if the inode mtime/ctime should be updated or
1354 * not. If current values of the time-stamps are within the UBIFS inode time
1355 * granularity, they are not updated. This is an optimization.
1357 static inline int mctime_update_needed(const struct inode *inode,
1358 const struct timespec64 *now)
1360 if (!timespec64_equal(&inode->i_mtime, now) ||
1361 !timespec64_equal(&inode->i_ctime, now))
1367 * ubifs_update_time - update time of inode.
1368 * @inode: inode to update
1370 * This function updates time of the inode.
1372 int ubifs_update_time(struct inode *inode, struct timespec64 *time,
1375 struct ubifs_inode *ui = ubifs_inode(inode);
1376 struct ubifs_info *c = inode->i_sb->s_fs_info;
1377 struct ubifs_budget_req req = { .dirtied_ino = 1,
1378 .dirtied_ino_d = ALIGN(ui->data_len, 8) };
1381 if (!IS_ENABLED(CONFIG_UBIFS_ATIME_SUPPORT))
1382 return generic_update_time(inode, time, flags);
1384 err = ubifs_budget_space(c, &req);
1388 mutex_lock(&ui->ui_mutex);
1389 if (flags & S_ATIME)
1390 inode->i_atime = *time;
1391 if (flags & S_CTIME)
1392 inode->i_ctime = *time;
1393 if (flags & S_MTIME)
1394 inode->i_mtime = *time;
1396 release = ui->dirty;
1397 __mark_inode_dirty(inode, I_DIRTY_SYNC);
1398 mutex_unlock(&ui->ui_mutex);
1400 ubifs_release_budget(c, &req);
1405 * update_mctime - update mtime and ctime of an inode.
1406 * @inode: inode to update
1408 * This function updates mtime and ctime of the inode if it is not equivalent to
1409 * current time. Returns zero in case of success and a negative error code in
1412 static int update_mctime(struct inode *inode)
1414 struct timespec64 now = current_time(inode);
1415 struct ubifs_inode *ui = ubifs_inode(inode);
1416 struct ubifs_info *c = inode->i_sb->s_fs_info;
1418 if (mctime_update_needed(inode, &now)) {
1420 struct ubifs_budget_req req = { .dirtied_ino = 1,
1421 .dirtied_ino_d = ALIGN(ui->data_len, 8) };
1423 err = ubifs_budget_space(c, &req);
1427 mutex_lock(&ui->ui_mutex);
1428 inode->i_mtime = inode->i_ctime = current_time(inode);
1429 release = ui->dirty;
1430 mark_inode_dirty_sync(inode);
1431 mutex_unlock(&ui->ui_mutex);
1433 ubifs_release_budget(c, &req);
1439 static ssize_t ubifs_write_iter(struct kiocb *iocb, struct iov_iter *from)
1441 int err = update_mctime(file_inode(iocb->ki_filp));
1445 return generic_file_write_iter(iocb, from);
1448 static int ubifs_set_page_dirty(struct page *page)
1451 struct inode *inode = page->mapping->host;
1452 struct ubifs_info *c = inode->i_sb->s_fs_info;
1454 ret = __set_page_dirty_nobuffers(page);
1456 * An attempt to dirty a page without budgeting for it - should not
1459 ubifs_assert(c, ret == 0);
1463 #ifdef CONFIG_MIGRATION
1464 static int ubifs_migrate_page(struct address_space *mapping,
1465 struct page *newpage, struct page *page, enum migrate_mode mode)
1469 rc = migrate_page_move_mapping(mapping, newpage, page, 0);
1470 if (rc != MIGRATEPAGE_SUCCESS)
1473 if (PagePrivate(page)) {
1474 detach_page_private(page);
1475 attach_page_private(newpage, (void *)1);
1478 if (mode != MIGRATE_SYNC_NO_COPY)
1479 migrate_page_copy(newpage, page);
1481 migrate_page_states(newpage, page);
1482 return MIGRATEPAGE_SUCCESS;
1486 static int ubifs_releasepage(struct page *page, gfp_t unused_gfp_flags)
1488 struct inode *inode = page->mapping->host;
1489 struct ubifs_info *c = inode->i_sb->s_fs_info;
1492 * An attempt to release a dirty page without budgeting for it - should
1495 if (PageWriteback(page))
1497 ubifs_assert(c, PagePrivate(page));
1499 detach_page_private(page);
1500 ClearPageChecked(page);
1505 * mmap()d file has taken write protection fault and is being made writable.
1506 * UBIFS must ensure page is budgeted for.
1508 static vm_fault_t ubifs_vm_page_mkwrite(struct vm_fault *vmf)
1510 struct page *page = vmf->page;
1511 struct inode *inode = file_inode(vmf->vma->vm_file);
1512 struct ubifs_info *c = inode->i_sb->s_fs_info;
1513 struct timespec64 now = current_time(inode);
1514 struct ubifs_budget_req req = { .new_page = 1 };
1515 int err, update_time;
1517 dbg_gen("ino %lu, pg %lu, i_size %lld", inode->i_ino, page->index,
1518 i_size_read(inode));
1519 ubifs_assert(c, !c->ro_media && !c->ro_mount);
1521 if (unlikely(c->ro_error))
1522 return VM_FAULT_SIGBUS; /* -EROFS */
1525 * We have not locked @page so far so we may budget for changing the
1526 * page. Note, we cannot do this after we locked the page, because
1527 * budgeting may cause write-back which would cause deadlock.
1529 * At the moment we do not know whether the page is dirty or not, so we
1530 * assume that it is not and budget for a new page. We could look at
1531 * the @PG_private flag and figure this out, but we may race with write
1532 * back and the page state may change by the time we lock it, so this
1533 * would need additional care. We do not bother with this at the
1534 * moment, although it might be good idea to do. Instead, we allocate
1535 * budget for a new page and amend it later on if the page was in fact
1538 * The budgeting-related logic of this function is similar to what we
1539 * do in 'ubifs_write_begin()' and 'ubifs_write_end()'. Glance there
1540 * for more comments.
1542 update_time = mctime_update_needed(inode, &now);
1545 * We have to change inode time stamp which requires extra
1548 req.dirtied_ino = 1;
1550 err = ubifs_budget_space(c, &req);
1551 if (unlikely(err)) {
1553 ubifs_warn(c, "out of space for mmapped file (inode number %lu)",
1555 return VM_FAULT_SIGBUS;
1559 if (unlikely(page->mapping != inode->i_mapping ||
1560 page_offset(page) > i_size_read(inode))) {
1561 /* Page got truncated out from underneath us */
1565 if (PagePrivate(page))
1566 release_new_page_budget(c);
1568 if (!PageChecked(page))
1569 ubifs_convert_page_budget(c);
1570 attach_page_private(page, (void *)1);
1571 atomic_long_inc(&c->dirty_pg_cnt);
1572 __set_page_dirty_nobuffers(page);
1577 struct ubifs_inode *ui = ubifs_inode(inode);
1579 mutex_lock(&ui->ui_mutex);
1580 inode->i_mtime = inode->i_ctime = current_time(inode);
1581 release = ui->dirty;
1582 mark_inode_dirty_sync(inode);
1583 mutex_unlock(&ui->ui_mutex);
1585 ubifs_release_dirty_inode_budget(c, ui);
1588 wait_for_stable_page(page);
1589 return VM_FAULT_LOCKED;
1593 ubifs_release_budget(c, &req);
1594 return VM_FAULT_SIGBUS;
1597 static const struct vm_operations_struct ubifs_file_vm_ops = {
1598 .fault = filemap_fault,
1599 .map_pages = filemap_map_pages,
1600 .page_mkwrite = ubifs_vm_page_mkwrite,
1603 static int ubifs_file_mmap(struct file *file, struct vm_area_struct *vma)
1607 err = generic_file_mmap(file, vma);
1610 vma->vm_ops = &ubifs_file_vm_ops;
1612 if (IS_ENABLED(CONFIG_UBIFS_ATIME_SUPPORT))
1613 file_accessed(file);
1618 static const char *ubifs_get_link(struct dentry *dentry,
1619 struct inode *inode,
1620 struct delayed_call *done)
1622 struct ubifs_inode *ui = ubifs_inode(inode);
1624 if (!IS_ENCRYPTED(inode))
1628 return ERR_PTR(-ECHILD);
1630 return fscrypt_get_symlink(inode, ui->data, ui->data_len, done);
1633 static int ubifs_symlink_getattr(const struct path *path, struct kstat *stat,
1634 u32 request_mask, unsigned int query_flags)
1636 ubifs_getattr(path, stat, request_mask, query_flags);
1638 if (IS_ENCRYPTED(d_inode(path->dentry)))
1639 return fscrypt_symlink_getattr(path, stat);
1643 const struct address_space_operations ubifs_file_address_operations = {
1644 .readpage = ubifs_readpage,
1645 .writepage = ubifs_writepage,
1646 .write_begin = ubifs_write_begin,
1647 .write_end = ubifs_write_end,
1648 .invalidatepage = ubifs_invalidatepage,
1649 .set_page_dirty = ubifs_set_page_dirty,
1650 #ifdef CONFIG_MIGRATION
1651 .migratepage = ubifs_migrate_page,
1653 .releasepage = ubifs_releasepage,
1656 const struct inode_operations ubifs_file_inode_operations = {
1657 .setattr = ubifs_setattr,
1658 .getattr = ubifs_getattr,
1659 #ifdef CONFIG_UBIFS_FS_XATTR
1660 .listxattr = ubifs_listxattr,
1662 .update_time = ubifs_update_time,
1665 const struct inode_operations ubifs_symlink_inode_operations = {
1666 .get_link = ubifs_get_link,
1667 .setattr = ubifs_setattr,
1668 .getattr = ubifs_symlink_getattr,
1669 #ifdef CONFIG_UBIFS_FS_XATTR
1670 .listxattr = ubifs_listxattr,
1672 .update_time = ubifs_update_time,
1675 const struct file_operations ubifs_file_operations = {
1676 .llseek = generic_file_llseek,
1677 .read_iter = generic_file_read_iter,
1678 .write_iter = ubifs_write_iter,
1679 .mmap = ubifs_file_mmap,
1680 .fsync = ubifs_fsync,
1681 .unlocked_ioctl = ubifs_ioctl,
1682 .splice_read = generic_file_splice_read,
1683 .splice_write = iter_file_splice_write,
1684 .open = fscrypt_file_open,
1685 #ifdef CONFIG_COMPAT
1686 .compat_ioctl = ubifs_compat_ioctl,