4 * As should be obvious for Linux kernel code, license is GPLv2
6 * Copyright (c) 2005-2008 Joern Engel <joern@logfs.org>
9 * Actually contains five sets of very similar functions:
10 * read read blocks from a file
11 * seek_hole find next hole
12 * seek_data find next data block
13 * valid check whether a block still belongs to a file
14 * write write blocks to a file
15 * delete delete a block (for directories and ifile)
16 * rewrite move existing blocks of a file to a new location (gc helper)
17 * truncate truncate a file
20 #include <linux/sched.h>
21 #include <linux/slab.h>
23 static u64 adjust_bix(u64 bix, level_t level)
29 return max_t(u64, bix, I0_BLOCKS);
31 return max_t(u64, bix, I1_BLOCKS);
33 return max_t(u64, bix, I2_BLOCKS);
35 return max_t(u64, bix, I3_BLOCKS);
37 return max_t(u64, bix, I4_BLOCKS);
44 static inline u64 maxbix(u8 height)
46 return 1ULL << (LOGFS_BLOCK_BITS * height);
50 * The inode address space is cut in two halves. Lower half belongs to data
51 * pages, upper half to indirect blocks. If the high bit (INDIRECT_BIT) is
52 * set, the actual block index (bix) and level can be derived from the page
55 * The lowest three bits of the block index are set to 0 after packing and
56 * unpacking. Since the lowest n bits (9 for 4KiB blocksize) are ignored
57 * anyway this is harmless.
59 #define ARCH_SHIFT (BITS_PER_LONG - 32)
60 #define INDIRECT_BIT (0x80000000UL << ARCH_SHIFT)
61 #define LEVEL_SHIFT (28 + ARCH_SHIFT)
62 static inline pgoff_t first_indirect_block(void)
64 return INDIRECT_BIT | (1ULL << LEVEL_SHIFT);
67 pgoff_t logfs_pack_index(u64 bix, level_t level)
71 BUG_ON(bix >= INDIRECT_BIT);
76 index |= (__force long)level << LEVEL_SHIFT;
77 index |= bix >> ((__force u8)level * LOGFS_BLOCK_BITS);
81 void logfs_unpack_index(pgoff_t index, u64 *bix, level_t *level)
85 if (!(index & INDIRECT_BIT)) {
91 __level = (index & ~INDIRECT_BIT) >> LEVEL_SHIFT;
92 *level = LEVEL(__level);
93 *bix = (index << (__level * LOGFS_BLOCK_BITS)) & ~INDIRECT_BIT;
94 *bix = adjust_bix(*bix, *level);
102 * Time is stored as nanoseconds since the epoch.
104 static struct timespec be64_to_timespec(__be64 betime)
106 return ns_to_timespec(be64_to_cpu(betime));
109 static __be64 timespec_to_be64(struct timespec tsp)
111 return cpu_to_be64((u64)tsp.tv_sec * NSEC_PER_SEC + tsp.tv_nsec);
114 static void logfs_disk_to_inode(struct logfs_disk_inode *di, struct inode*inode)
116 struct logfs_inode *li = logfs_inode(inode);
119 inode->i_mode = be16_to_cpu(di->di_mode);
120 li->li_height = di->di_height;
121 li->li_flags = be32_to_cpu(di->di_flags);
122 inode->i_uid = be32_to_cpu(di->di_uid);
123 inode->i_gid = be32_to_cpu(di->di_gid);
124 inode->i_size = be64_to_cpu(di->di_size);
125 logfs_set_blocks(inode, be64_to_cpu(di->di_used_bytes));
126 inode->i_atime = be64_to_timespec(di->di_atime);
127 inode->i_ctime = be64_to_timespec(di->di_ctime);
128 inode->i_mtime = be64_to_timespec(di->di_mtime);
129 inode->i_nlink = be32_to_cpu(di->di_refcount);
130 inode->i_generation = be32_to_cpu(di->di_generation);
132 switch (inode->i_mode & S_IFMT) {
133 case S_IFSOCK: /* fall through */
134 case S_IFBLK: /* fall through */
135 case S_IFCHR: /* fall through */
137 inode->i_rdev = be64_to_cpu(di->di_data[0]);
139 case S_IFDIR: /* fall through */
140 case S_IFREG: /* fall through */
142 for (i = 0; i < LOGFS_EMBEDDED_FIELDS; i++)
143 li->li_data[i] = be64_to_cpu(di->di_data[i]);
150 static void logfs_inode_to_disk(struct inode *inode, struct logfs_disk_inode*di)
152 struct logfs_inode *li = logfs_inode(inode);
155 di->di_mode = cpu_to_be16(inode->i_mode);
156 di->di_height = li->li_height;
158 di->di_flags = cpu_to_be32(li->li_flags);
159 di->di_uid = cpu_to_be32(inode->i_uid);
160 di->di_gid = cpu_to_be32(inode->i_gid);
161 di->di_size = cpu_to_be64(i_size_read(inode));
162 di->di_used_bytes = cpu_to_be64(li->li_used_bytes);
163 di->di_atime = timespec_to_be64(inode->i_atime);
164 di->di_ctime = timespec_to_be64(inode->i_ctime);
165 di->di_mtime = timespec_to_be64(inode->i_mtime);
166 di->di_refcount = cpu_to_be32(inode->i_nlink);
167 di->di_generation = cpu_to_be32(inode->i_generation);
169 switch (inode->i_mode & S_IFMT) {
170 case S_IFSOCK: /* fall through */
171 case S_IFBLK: /* fall through */
172 case S_IFCHR: /* fall through */
174 di->di_data[0] = cpu_to_be64(inode->i_rdev);
176 case S_IFDIR: /* fall through */
177 case S_IFREG: /* fall through */
179 for (i = 0; i < LOGFS_EMBEDDED_FIELDS; i++)
180 di->di_data[i] = cpu_to_be64(li->li_data[i]);
187 static void __logfs_set_blocks(struct inode *inode)
189 struct super_block *sb = inode->i_sb;
190 struct logfs_inode *li = logfs_inode(inode);
192 inode->i_blocks = ULONG_MAX;
193 if (li->li_used_bytes >> sb->s_blocksize_bits < ULONG_MAX)
194 inode->i_blocks = ALIGN(li->li_used_bytes, 512) >> 9;
197 void logfs_set_blocks(struct inode *inode, u64 bytes)
199 struct logfs_inode *li = logfs_inode(inode);
201 li->li_used_bytes = bytes;
202 __logfs_set_blocks(inode);
205 static void prelock_page(struct super_block *sb, struct page *page, int lock)
207 struct logfs_super *super = logfs_super(sb);
209 BUG_ON(!PageLocked(page));
211 BUG_ON(PagePreLocked(page));
212 SetPagePreLocked(page);
214 /* We are in GC path. */
215 if (PagePreLocked(page))
216 super->s_lock_count++;
218 SetPagePreLocked(page);
222 static void preunlock_page(struct super_block *sb, struct page *page, int lock)
224 struct logfs_super *super = logfs_super(sb);
226 BUG_ON(!PageLocked(page));
228 ClearPagePreLocked(page);
230 /* We are in GC path. */
231 BUG_ON(!PagePreLocked(page));
232 if (super->s_lock_count)
233 super->s_lock_count--;
235 ClearPagePreLocked(page);
240 * Logfs is prone to an AB-BA deadlock where one task tries to acquire
241 * s_write_mutex with a locked page and GC tries to get that page while holding
243 * To solve this issue logfs will ignore the page lock iff the page in question
244 * is waiting for s_write_mutex. We annotate this fact by setting PG_pre_locked
245 * in addition to PG_locked.
247 static void logfs_get_wblocks(struct super_block *sb, struct page *page,
250 struct logfs_super *super = logfs_super(sb);
253 prelock_page(sb, page, lock);
256 mutex_lock(&super->s_write_mutex);
258 /* FIXME: We also have to check for shadowed space
259 * and mempool fill grade */
263 static void logfs_put_wblocks(struct super_block *sb, struct page *page,
266 struct logfs_super *super = logfs_super(sb);
269 preunlock_page(sb, page, lock);
270 /* Order matters - we must clear PG_pre_locked before releasing
271 * s_write_mutex or we could race against another task. */
273 mutex_unlock(&super->s_write_mutex);
276 static struct page *logfs_get_read_page(struct inode *inode, u64 bix,
279 return find_or_create_page(inode->i_mapping,
280 logfs_pack_index(bix, level), GFP_NOFS);
283 static void logfs_put_read_page(struct page *page)
286 page_cache_release(page);
289 static void logfs_lock_write_page(struct page *page)
293 while (unlikely(!trylock_page(page))) {
294 if (loop++ > 0x1000) {
295 /* Has been observed once so far... */
296 printk(KERN_ERR "stack at %p\n", &loop);
299 if (PagePreLocked(page)) {
300 /* Holder of page lock is waiting for us, it
301 * is safe to use this page. */
304 /* Some other process has this page locked and has
305 * nothing to do with us. Wait for it to finish.
309 BUG_ON(!PageLocked(page));
312 static struct page *logfs_get_write_page(struct inode *inode, u64 bix,
315 struct address_space *mapping = inode->i_mapping;
316 pgoff_t index = logfs_pack_index(bix, level);
321 page = find_get_page(mapping, index);
323 page = __page_cache_alloc(GFP_NOFS);
326 err = add_to_page_cache_lru(page, mapping, index, GFP_NOFS);
328 page_cache_release(page);
333 } else logfs_lock_write_page(page);
334 BUG_ON(!PageLocked(page));
338 static void logfs_unlock_write_page(struct page *page)
340 if (!PagePreLocked(page))
344 static void logfs_put_write_page(struct page *page)
346 logfs_unlock_write_page(page);
347 page_cache_release(page);
350 static struct page *logfs_get_page(struct inode *inode, u64 bix, level_t level,
354 return logfs_get_read_page(inode, bix, level);
356 return logfs_get_write_page(inode, bix, level);
359 static void logfs_put_page(struct page *page, int rw)
362 logfs_put_read_page(page);
364 logfs_put_write_page(page);
367 static unsigned long __get_bits(u64 val, int skip, int no)
377 static unsigned long get_bits(u64 val, level_t skip)
379 return __get_bits(val, (__force int)skip, LOGFS_BLOCK_BITS);
382 static inline void init_shadow_tree(struct super_block *sb,
383 struct shadow_tree *tree)
385 struct logfs_super *super = logfs_super(sb);
387 btree_init_mempool64(&tree->new, super->s_btree_pool);
388 btree_init_mempool64(&tree->old, super->s_btree_pool);
391 static void indirect_write_block(struct logfs_block *block)
398 inode = page->mapping->host;
399 logfs_lock_write_page(page);
400 ret = logfs_write_buf(inode, page, 0);
401 logfs_unlock_write_page(page);
403 * This needs some rework. Unless you want your filesystem to run
404 * completely synchronously (you don't), the filesystem will always
405 * report writes as 'successful' before the actual work has been
406 * done. The actual work gets done here and this is where any errors
407 * will show up. And there isn't much we can do about it, really.
409 * Some attempts to fix the errors (move from bad blocks, retry io,...)
410 * have already been done, so anything left should be either a broken
411 * device or a bug somewhere in logfs itself. Being relatively new,
412 * the odds currently favor a bug, so for now the line below isn't
418 static void inode_write_block(struct logfs_block *block)
423 inode = block->inode;
424 if (inode->i_ino == LOGFS_INO_MASTER)
425 logfs_write_anchor(inode->i_sb);
427 ret = __logfs_write_inode(inode, 0);
428 /* see indirect_write_block comment */
434 * This silences a false, yet annoying gcc warning. I hate it when my editor
435 * jumps into bitops.h each time I recompile this file.
436 * TODO: Complain to gcc folks about this and upgrade compiler.
438 static unsigned long fnb(const unsigned long *addr,
439 unsigned long size, unsigned long offset)
441 return find_next_bit(addr, size, offset);
444 static __be64 inode_val0(struct inode *inode)
446 struct logfs_inode *li = logfs_inode(inode);
450 * Explicit shifting generates good code, but must match the format
451 * of the structure. Add some paranoia just in case.
453 BUILD_BUG_ON(offsetof(struct logfs_disk_inode, di_mode) != 0);
454 BUILD_BUG_ON(offsetof(struct logfs_disk_inode, di_height) != 2);
455 BUILD_BUG_ON(offsetof(struct logfs_disk_inode, di_flags) != 4);
457 val = (u64)inode->i_mode << 48 |
458 (u64)li->li_height << 40 |
460 return cpu_to_be64(val);
463 static int inode_write_alias(struct super_block *sb,
464 struct logfs_block *block, write_alias_t *write_one_alias)
466 struct inode *inode = block->inode;
467 struct logfs_inode *li = logfs_inode(inode);
474 for (pos = 0; ; pos++) {
475 pos = fnb(block->alias_map, LOGFS_BLOCK_FACTOR, pos);
476 if (pos >= LOGFS_EMBEDDED_FIELDS + INODE_POINTER_OFS)
480 case INODE_HEIGHT_OFS:
481 val = inode_val0(inode);
484 val = cpu_to_be64(li->li_used_bytes);;
487 val = cpu_to_be64(i_size_read(inode));
489 case INODE_POINTER_OFS ... INODE_POINTER_OFS + LOGFS_EMBEDDED_FIELDS - 1:
490 val = cpu_to_be64(li->li_data[pos - INODE_POINTER_OFS]);
496 ino = LOGFS_INO_MASTER;
499 err = write_one_alias(sb, ino, bix, level, pos, val);
505 static int indirect_write_alias(struct super_block *sb,
506 struct logfs_block *block, write_alias_t *write_one_alias)
509 struct page *page = block->page;
515 for (pos = 0; ; pos++) {
516 pos = fnb(block->alias_map, LOGFS_BLOCK_FACTOR, pos);
517 if (pos >= LOGFS_BLOCK_FACTOR)
520 ino = page->mapping->host->i_ino;
521 logfs_unpack_index(page->index, &bix, &level);
522 child = kmap_atomic(page, KM_USER0);
524 kunmap_atomic(child, KM_USER0);
525 err = write_one_alias(sb, ino, bix, level, pos, val);
531 int logfs_write_obj_aliases_pagecache(struct super_block *sb)
533 struct logfs_super *super = logfs_super(sb);
534 struct logfs_block *block;
537 list_for_each_entry(block, &super->s_object_alias, alias_list) {
538 err = block->ops->write_alias(sb, block, write_alias_journal);
545 void __free_block(struct super_block *sb, struct logfs_block *block)
547 BUG_ON(!list_empty(&block->item_list));
548 list_del(&block->alias_list);
549 mempool_free(block, logfs_super(sb)->s_block_pool);
552 static void inode_free_block(struct super_block *sb, struct logfs_block *block)
554 struct inode *inode = block->inode;
556 logfs_inode(inode)->li_block = NULL;
557 __free_block(sb, block);
560 static void indirect_free_block(struct super_block *sb,
561 struct logfs_block *block)
563 ClearPagePrivate(block->page);
564 block->page->private = 0;
565 __free_block(sb, block);
569 static struct logfs_block_ops inode_block_ops = {
570 .write_block = inode_write_block,
571 .free_block = inode_free_block,
572 .write_alias = inode_write_alias,
575 struct logfs_block_ops indirect_block_ops = {
576 .write_block = indirect_write_block,
577 .free_block = indirect_free_block,
578 .write_alias = indirect_write_alias,
581 struct logfs_block *__alloc_block(struct super_block *sb,
582 u64 ino, u64 bix, level_t level)
584 struct logfs_super *super = logfs_super(sb);
585 struct logfs_block *block;
587 block = mempool_alloc(super->s_block_pool, GFP_NOFS);
588 memset(block, 0, sizeof(*block));
589 INIT_LIST_HEAD(&block->alias_list);
590 INIT_LIST_HEAD(&block->item_list);
594 block->level = level;
598 static void alloc_inode_block(struct inode *inode)
600 struct logfs_inode *li = logfs_inode(inode);
601 struct logfs_block *block;
606 block = __alloc_block(inode->i_sb, LOGFS_INO_MASTER, inode->i_ino, 0);
607 block->inode = inode;
608 li->li_block = block;
609 block->ops = &inode_block_ops;
612 void initialize_block_counters(struct page *page, struct logfs_block *block,
613 __be64 *array, int page_is_empty)
621 if (page->index < first_indirect_block()) {
622 /* Counters are pointless on level 0 */
625 if (page->index == first_indirect_block()) {
626 /* Skip unused pointers */
628 block->full = I0_BLOCKS;
630 if (!page_is_empty) {
631 for (i = start; i < LOGFS_BLOCK_FACTOR; i++) {
632 ptr = be64_to_cpu(array[i]);
635 if (ptr & LOGFS_FULLY_POPULATED)
641 static void alloc_data_block(struct inode *inode, struct page *page)
643 struct logfs_block *block;
647 if (PagePrivate(page))
650 logfs_unpack_index(page->index, &bix, &level);
651 block = __alloc_block(inode->i_sb, inode->i_ino, bix, level);
653 SetPagePrivate(page);
654 page->private = (unsigned long)block;
655 block->ops = &indirect_block_ops;
658 static void alloc_indirect_block(struct inode *inode, struct page *page,
661 struct logfs_block *block;
664 if (PagePrivate(page))
667 alloc_data_block(inode, page);
669 block = logfs_block(page);
670 array = kmap_atomic(page, KM_USER0);
671 initialize_block_counters(page, block, array, page_is_empty);
672 kunmap_atomic(array, KM_USER0);
675 static void block_set_pointer(struct page *page, int index, u64 ptr)
677 struct logfs_block *block = logfs_block(page);
682 array = kmap_atomic(page, KM_USER0);
683 oldptr = be64_to_cpu(array[index]);
684 array[index] = cpu_to_be64(ptr);
685 kunmap_atomic(array, KM_USER0);
686 SetPageUptodate(page);
688 block->full += !!(ptr & LOGFS_FULLY_POPULATED)
689 - !!(oldptr & LOGFS_FULLY_POPULATED);
690 block->partial += !!ptr - !!oldptr;
693 static u64 block_get_pointer(struct page *page, int index)
698 block = kmap_atomic(page, KM_USER0);
699 ptr = be64_to_cpu(block[index]);
700 kunmap_atomic(block, KM_USER0);
704 static int logfs_read_empty(struct page *page)
706 zero_user_segment(page, 0, PAGE_CACHE_SIZE);
710 static int logfs_read_direct(struct inode *inode, struct page *page)
712 struct logfs_inode *li = logfs_inode(inode);
713 pgoff_t index = page->index;
716 block = li->li_data[index];
718 return logfs_read_empty(page);
720 return logfs_segment_read(inode, page, block, index, 0);
723 static int logfs_read_loop(struct inode *inode, struct page *page,
726 struct logfs_inode *li = logfs_inode(inode);
727 u64 bix, bofs = li->li_data[INDIRECT_INDEX];
728 level_t level, target_level;
732 logfs_unpack_index(page->index, &bix, &target_level);
734 return logfs_read_empty(page);
736 if (bix >= maxbix(li->li_height))
737 return logfs_read_empty(page);
739 for (level = LEVEL(li->li_height);
740 (__force u8)level > (__force u8)target_level;
741 level = SUBLEVEL(level)){
742 ipage = logfs_get_page(inode, bix, level, rw_context);
746 ret = logfs_segment_read(inode, ipage, bofs, bix, level);
748 logfs_put_read_page(ipage);
752 bofs = block_get_pointer(ipage, get_bits(bix, SUBLEVEL(level)));
753 logfs_put_page(ipage, rw_context);
755 return logfs_read_empty(page);
758 return logfs_segment_read(inode, page, bofs, bix, 0);
761 static int logfs_read_block(struct inode *inode, struct page *page,
764 pgoff_t index = page->index;
766 if (index < I0_BLOCKS)
767 return logfs_read_direct(inode, page);
768 return logfs_read_loop(inode, page, rw_context);
771 static int logfs_exist_loop(struct inode *inode, u64 bix)
773 struct logfs_inode *li = logfs_inode(inode);
774 u64 bofs = li->li_data[INDIRECT_INDEX];
781 if (bix >= maxbix(li->li_height))
784 for (level = LEVEL(li->li_height); level != 0; level = SUBLEVEL(level)) {
785 ipage = logfs_get_read_page(inode, bix, level);
789 ret = logfs_segment_read(inode, ipage, bofs, bix, level);
791 logfs_put_read_page(ipage);
795 bofs = block_get_pointer(ipage, get_bits(bix, SUBLEVEL(level)));
796 logfs_put_read_page(ipage);
804 int logfs_exist_block(struct inode *inode, u64 bix)
806 struct logfs_inode *li = logfs_inode(inode);
809 return !!li->li_data[bix];
810 return logfs_exist_loop(inode, bix);
813 static u64 seek_holedata_direct(struct inode *inode, u64 bix, int data)
815 struct logfs_inode *li = logfs_inode(inode);
817 for (; bix < I0_BLOCKS; bix++)
818 if (data ^ (li->li_data[bix] == 0))
823 static u64 seek_holedata_loop(struct inode *inode, u64 bix, int data)
825 struct logfs_inode *li = logfs_inode(inode);
827 u64 increment, bofs = li->li_data[INDIRECT_INDEX];
834 for (level = LEVEL(li->li_height); level != 0; level = SUBLEVEL(level)) {
835 increment = 1 << (LOGFS_BLOCK_BITS * ((__force u8)level-1));
836 page = logfs_get_read_page(inode, bix, level);
840 ret = logfs_segment_read(inode, page, bofs, bix, level);
842 logfs_put_read_page(page);
846 slot = get_bits(bix, SUBLEVEL(level));
847 rblock = kmap_atomic(page, KM_USER0);
848 while (slot < LOGFS_BLOCK_FACTOR) {
849 if (data && (rblock[slot] != 0))
851 if (!data && !(be64_to_cpu(rblock[slot]) & LOGFS_FULLY_POPULATED))
855 bix &= ~(increment - 1);
857 if (slot >= LOGFS_BLOCK_FACTOR) {
858 kunmap_atomic(rblock, KM_USER0);
859 logfs_put_read_page(page);
862 bofs = be64_to_cpu(rblock[slot]);
863 kunmap_atomic(rblock, KM_USER0);
864 logfs_put_read_page(page);
874 * logfs_seek_hole - find next hole starting at a given block index
875 * @inode: inode to search in
876 * @bix: block index to start searching
878 * Returns next hole. If the file doesn't contain any further holes, the
879 * block address next to eof is returned instead.
881 u64 logfs_seek_hole(struct inode *inode, u64 bix)
883 struct logfs_inode *li = logfs_inode(inode);
885 if (bix < I0_BLOCKS) {
886 bix = seek_holedata_direct(inode, bix, 0);
891 if (!li->li_data[INDIRECT_INDEX])
893 else if (li->li_data[INDIRECT_INDEX] & LOGFS_FULLY_POPULATED)
894 bix = maxbix(li->li_height);
896 bix = seek_holedata_loop(inode, bix, 0);
897 if (bix < maxbix(li->li_height))
899 /* Should not happen anymore. But if some port writes semi-
900 * corrupt images (as this one used to) we might run into it.
902 WARN_ON_ONCE(bix == maxbix(li->li_height));
908 static u64 __logfs_seek_data(struct inode *inode, u64 bix)
910 struct logfs_inode *li = logfs_inode(inode);
912 if (bix < I0_BLOCKS) {
913 bix = seek_holedata_direct(inode, bix, 1);
918 if (bix < maxbix(li->li_height)) {
919 if (!li->li_data[INDIRECT_INDEX])
920 bix = maxbix(li->li_height);
922 return seek_holedata_loop(inode, bix, 1);
929 * logfs_seek_data - find next data block after a given block index
930 * @inode: inode to search in
931 * @bix: block index to start searching
933 * Returns next data block. If the file doesn't contain any further data
934 * blocks, the last block in the file is returned instead.
936 u64 logfs_seek_data(struct inode *inode, u64 bix)
938 struct super_block *sb = inode->i_sb;
941 ret = __logfs_seek_data(inode, bix);
942 end = i_size_read(inode) >> sb->s_blocksize_bits;
948 static int logfs_is_valid_direct(struct logfs_inode *li, u64 bix, u64 ofs)
950 return pure_ofs(li->li_data[bix]) == ofs;
953 static int __logfs_is_valid_loop(struct inode *inode, u64 bix,
956 struct logfs_inode *li = logfs_inode(inode);
961 for (level = LEVEL(li->li_height); level != 0; level = SUBLEVEL(level)){
962 page = logfs_get_write_page(inode, bix, level);
965 ret = logfs_segment_read(inode, page, bofs, bix, level);
967 logfs_put_write_page(page);
971 bofs = block_get_pointer(page, get_bits(bix, SUBLEVEL(level)));
972 logfs_put_write_page(page);
976 if (pure_ofs(bofs) == ofs)
982 static int logfs_is_valid_loop(struct inode *inode, u64 bix, u64 ofs)
984 struct logfs_inode *li = logfs_inode(inode);
985 u64 bofs = li->li_data[INDIRECT_INDEX];
990 if (bix >= maxbix(li->li_height))
993 if (pure_ofs(bofs) == ofs)
996 return __logfs_is_valid_loop(inode, bix, ofs, bofs);
999 static int __logfs_is_valid_block(struct inode *inode, u64 bix, u64 ofs)
1001 struct logfs_inode *li = logfs_inode(inode);
1003 if ((inode->i_nlink == 0) && atomic_read(&inode->i_count) == 1)
1006 if (bix < I0_BLOCKS)
1007 return logfs_is_valid_direct(li, bix, ofs);
1008 return logfs_is_valid_loop(inode, bix, ofs);
1012 * logfs_is_valid_block - check whether this block is still valid
1015 * @ofs - block physical offset
1016 * @ino - block inode number
1017 * @bix - block index
1018 * @level - block level
1020 * Returns 0 if the block is invalid, 1 if it is valid and 2 if it will
1021 * become invalid once the journal is written.
1023 int logfs_is_valid_block(struct super_block *sb, u64 ofs, u64 ino, u64 bix,
1024 gc_level_t gc_level)
1026 struct logfs_super *super = logfs_super(sb);
1027 struct inode *inode;
1030 /* Umount closes a segment with free blocks remaining. Those
1031 * blocks are by definition invalid. */
1035 LOGFS_BUG_ON((u64)(u_long)ino != ino, sb);
1037 inode = logfs_safe_iget(sb, ino, &cookie);
1041 ret = __logfs_is_valid_block(inode, bix, ofs);
1042 logfs_safe_iput(inode, cookie);
1047 /* Block is nominally invalid, but may still sit in the shadow tree,
1048 * waiting for a journal commit.
1050 if (btree_lookup64(&super->s_shadow_tree.old, ofs))
1055 int logfs_readpage_nolock(struct page *page)
1057 struct inode *inode = page->mapping->host;
1060 ret = logfs_read_block(inode, page, READ);
1063 ClearPageUptodate(page);
1066 SetPageUptodate(page);
1067 ClearPageError(page);
1069 flush_dcache_page(page);
1074 static int logfs_reserve_bytes(struct inode *inode, int bytes)
1076 struct logfs_super *super = logfs_super(inode->i_sb);
1077 u64 available = super->s_free_bytes + super->s_dirty_free_bytes
1078 - super->s_dirty_used_bytes - super->s_dirty_pages;
1083 if (available < bytes)
1086 if (available < bytes + super->s_root_reserve &&
1087 !capable(CAP_SYS_RESOURCE))
1093 int get_page_reserve(struct inode *inode, struct page *page)
1095 struct logfs_super *super = logfs_super(inode->i_sb);
1098 if (logfs_block(page) && logfs_block(page)->reserved_bytes)
1101 logfs_get_wblocks(inode->i_sb, page, WF_LOCK);
1102 ret = logfs_reserve_bytes(inode, 6 * LOGFS_MAX_OBJECTSIZE);
1104 alloc_data_block(inode, page);
1105 logfs_block(page)->reserved_bytes += 6 * LOGFS_MAX_OBJECTSIZE;
1106 super->s_dirty_pages += 6 * LOGFS_MAX_OBJECTSIZE;
1108 logfs_put_wblocks(inode->i_sb, page, WF_LOCK);
1113 * We are protected by write lock. Push victims up to superblock level
1114 * and release transaction when appropriate.
1116 /* FIXME: This is currently called from the wrong spots. */
1117 static void logfs_handle_transaction(struct inode *inode,
1118 struct logfs_transaction *ta)
1120 struct logfs_super *super = logfs_super(inode->i_sb);
1124 logfs_inode(inode)->li_block->ta = NULL;
1126 if (inode->i_ino != LOGFS_INO_MASTER) {
1127 BUG(); /* FIXME: Yes, this needs more thought */
1128 /* just remember the transaction until inode is written */
1129 //BUG_ON(logfs_inode(inode)->li_transaction);
1130 //logfs_inode(inode)->li_transaction = ta;
1134 switch (ta->state) {
1135 case CREATE_1: /* fall through */
1137 BUG_ON(super->s_victim_ino);
1138 super->s_victim_ino = ta->ino;
1140 case CREATE_2: /* fall through */
1142 BUG_ON(super->s_victim_ino != ta->ino);
1143 super->s_victim_ino = 0;
1144 /* transaction ends here - free it */
1147 case CROSS_RENAME_1:
1148 BUG_ON(super->s_rename_dir);
1149 BUG_ON(super->s_rename_pos);
1150 super->s_rename_dir = ta->dir;
1151 super->s_rename_pos = ta->pos;
1153 case CROSS_RENAME_2:
1154 BUG_ON(super->s_rename_dir != ta->dir);
1155 BUG_ON(super->s_rename_pos != ta->pos);
1156 super->s_rename_dir = 0;
1157 super->s_rename_pos = 0;
1160 case TARGET_RENAME_1:
1161 BUG_ON(super->s_rename_dir);
1162 BUG_ON(super->s_rename_pos);
1163 BUG_ON(super->s_victim_ino);
1164 super->s_rename_dir = ta->dir;
1165 super->s_rename_pos = ta->pos;
1166 super->s_victim_ino = ta->ino;
1168 case TARGET_RENAME_2:
1169 BUG_ON(super->s_rename_dir != ta->dir);
1170 BUG_ON(super->s_rename_pos != ta->pos);
1171 BUG_ON(super->s_victim_ino != ta->ino);
1172 super->s_rename_dir = 0;
1173 super->s_rename_pos = 0;
1175 case TARGET_RENAME_3:
1176 BUG_ON(super->s_rename_dir);
1177 BUG_ON(super->s_rename_pos);
1178 BUG_ON(super->s_victim_ino != ta->ino);
1179 super->s_victim_ino = 0;
1188 * Not strictly a reservation, but rather a check that we still have enough
1189 * space to satisfy the write.
1191 static int logfs_reserve_blocks(struct inode *inode, int blocks)
1193 return logfs_reserve_bytes(inode, blocks * LOGFS_MAX_OBJECTSIZE);
1196 struct write_control {
1201 static struct logfs_shadow *alloc_shadow(struct inode *inode, u64 bix,
1202 level_t level, u64 old_ofs)
1204 struct logfs_super *super = logfs_super(inode->i_sb);
1205 struct logfs_shadow *shadow;
1207 shadow = mempool_alloc(super->s_shadow_pool, GFP_NOFS);
1208 memset(shadow, 0, sizeof(*shadow));
1209 shadow->ino = inode->i_ino;
1211 shadow->gc_level = expand_level(inode->i_ino, level);
1212 shadow->old_ofs = old_ofs & ~LOGFS_FULLY_POPULATED;
1216 static void free_shadow(struct inode *inode, struct logfs_shadow *shadow)
1218 struct logfs_super *super = logfs_super(inode->i_sb);
1220 mempool_free(shadow, super->s_shadow_pool);
1223 static void mark_segment(struct shadow_tree *tree, u32 segno)
1227 if (!btree_lookup32(&tree->segment_map, segno)) {
1228 err = btree_insert32(&tree->segment_map, segno, (void *)1,
1231 tree->no_shadowed_segments++;
1236 * fill_shadow_tree - Propagate shadow tree changes due to a write
1237 * @inode: Inode owning the page
1238 * @page: Struct page that was written
1239 * @shadow: Shadow for the current write
1241 * Writes in logfs can result in two semi-valid objects. The old object
1242 * is still valid as long as it can be reached by following pointers on
1243 * the medium. Only when writes propagate all the way up to the journal
1244 * has the new object safely replaced the old one.
1246 * To handle this problem, a struct logfs_shadow is used to represent
1247 * every single write. It is attached to the indirect block, which is
1248 * marked dirty. When the indirect block is written, its shadows are
1249 * handed up to the next indirect block (or inode). Untimately they
1250 * will reach the master inode and be freed upon journal commit.
1252 * This function handles a single step in the propagation. It adds the
1253 * shadow for the current write to the tree, along with any shadows in
1254 * the page's tree, in case it was an indirect block. If a page is
1255 * written, the inode parameter is left NULL, if an inode is written,
1256 * the page parameter is left NULL.
1258 static void fill_shadow_tree(struct inode *inode, struct page *page,
1259 struct logfs_shadow *shadow)
1261 struct logfs_super *super = logfs_super(inode->i_sb);
1262 struct logfs_block *block = logfs_block(page);
1263 struct shadow_tree *tree = &super->s_shadow_tree;
1265 if (PagePrivate(page)) {
1266 if (block->alias_map)
1267 super->s_no_object_aliases -= bitmap_weight(
1268 block->alias_map, LOGFS_BLOCK_FACTOR);
1269 logfs_handle_transaction(inode, block->ta);
1270 block->ops->free_block(inode->i_sb, block);
1273 if (shadow->old_ofs)
1274 btree_insert64(&tree->old, shadow->old_ofs, shadow,
1277 btree_insert64(&tree->new, shadow->new_ofs, shadow,
1280 super->s_dirty_used_bytes += shadow->new_len;
1281 super->s_dirty_free_bytes += shadow->old_len;
1282 mark_segment(tree, shadow->old_ofs >> super->s_segshift);
1283 mark_segment(tree, shadow->new_ofs >> super->s_segshift);
1287 static void logfs_set_alias(struct super_block *sb, struct logfs_block *block,
1290 struct logfs_super *super = logfs_super(sb);
1292 if (block->inode && block->inode->i_ino == LOGFS_INO_MASTER) {
1293 /* Aliases in the master inode are pointless. */
1297 if (!test_bit(child_no, block->alias_map)) {
1298 set_bit(child_no, block->alias_map);
1299 super->s_no_object_aliases++;
1301 list_move_tail(&block->alias_list, &super->s_object_alias);
1305 * Object aliases can and often do change the size and occupied space of a
1306 * file. So not only do we have to change the pointers, we also have to
1307 * change inode->i_size and li->li_used_bytes. Which is done by setting
1308 * another two object aliases for the inode itself.
1310 static void set_iused(struct inode *inode, struct logfs_shadow *shadow)
1312 struct logfs_inode *li = logfs_inode(inode);
1314 if (shadow->new_len == shadow->old_len)
1317 alloc_inode_block(inode);
1318 li->li_used_bytes += shadow->new_len - shadow->old_len;
1319 __logfs_set_blocks(inode);
1320 logfs_set_alias(inode->i_sb, li->li_block, INODE_USED_OFS);
1321 logfs_set_alias(inode->i_sb, li->li_block, INODE_SIZE_OFS);
1324 static int logfs_write_i0(struct inode *inode, struct page *page,
1325 struct write_control *wc)
1327 struct logfs_shadow *shadow;
1332 logfs_unpack_index(page->index, &bix, &level);
1334 if (logfs_reserve_blocks(inode, 1))
1337 shadow = alloc_shadow(inode, bix, level, wc->ofs);
1338 if (wc->flags & WF_WRITE)
1339 err = logfs_segment_write(inode, page, shadow);
1340 if (wc->flags & WF_DELETE)
1341 logfs_segment_delete(inode, shadow);
1343 free_shadow(inode, shadow);
1347 set_iused(inode, shadow);
1350 alloc_indirect_block(inode, page, 0);
1351 full = logfs_block(page)->full == LOGFS_BLOCK_FACTOR;
1353 fill_shadow_tree(inode, page, shadow);
1354 wc->ofs = shadow->new_ofs;
1355 if (wc->ofs && full)
1356 wc->ofs |= LOGFS_FULLY_POPULATED;
1360 static int logfs_write_direct(struct inode *inode, struct page *page,
1363 struct logfs_inode *li = logfs_inode(inode);
1364 struct write_control wc = {
1365 .ofs = li->li_data[page->index],
1370 alloc_inode_block(inode);
1372 err = logfs_write_i0(inode, page, &wc);
1376 li->li_data[page->index] = wc.ofs;
1377 logfs_set_alias(inode->i_sb, li->li_block,
1378 page->index + INODE_POINTER_OFS);
1382 static int ptr_change(u64 ofs, struct page *page)
1384 struct logfs_block *block = logfs_block(page);
1385 int empty0, empty1, full0, full1;
1388 empty1 = block->partial == 0;
1389 if (empty0 != empty1)
1392 /* The !! is necessary to shrink result to int */
1393 full0 = !!(ofs & LOGFS_FULLY_POPULATED);
1394 full1 = block->full == LOGFS_BLOCK_FACTOR;
1400 static int __logfs_write_rec(struct inode *inode, struct page *page,
1401 struct write_control *this_wc,
1402 pgoff_t bix, level_t target_level, level_t level)
1404 int ret, page_empty = 0;
1405 int child_no = get_bits(bix, SUBLEVEL(level));
1407 struct write_control child_wc = {
1408 .flags = this_wc->flags,
1411 ipage = logfs_get_write_page(inode, bix, level);
1416 ret = logfs_segment_read(inode, ipage, this_wc->ofs, bix, level);
1419 } else if (!PageUptodate(ipage)) {
1421 logfs_read_empty(ipage);
1424 child_wc.ofs = block_get_pointer(ipage, child_no);
1426 if ((__force u8)level-1 > (__force u8)target_level)
1427 ret = __logfs_write_rec(inode, page, &child_wc, bix,
1428 target_level, SUBLEVEL(level));
1430 ret = logfs_write_i0(inode, page, &child_wc);
1435 alloc_indirect_block(inode, ipage, page_empty);
1436 block_set_pointer(ipage, child_no, child_wc.ofs);
1437 /* FIXME: first condition seems superfluous */
1438 if (child_wc.ofs || logfs_block(ipage)->partial)
1439 this_wc->flags |= WF_WRITE;
1440 /* the condition on this_wc->ofs ensures that we won't consume extra
1441 * space for indirect blocks in the future, which we cannot reserve */
1442 if (!this_wc->ofs || ptr_change(this_wc->ofs, ipage))
1443 ret = logfs_write_i0(inode, ipage, this_wc);
1445 logfs_set_alias(inode->i_sb, logfs_block(ipage), child_no);
1447 logfs_put_write_page(ipage);
1451 static int logfs_write_rec(struct inode *inode, struct page *page,
1452 pgoff_t bix, level_t target_level, long flags)
1454 struct logfs_inode *li = logfs_inode(inode);
1455 struct write_control wc = {
1456 .ofs = li->li_data[INDIRECT_INDEX],
1461 alloc_inode_block(inode);
1463 if (li->li_height > (__force u8)target_level)
1464 ret = __logfs_write_rec(inode, page, &wc, bix, target_level,
1465 LEVEL(li->li_height));
1467 ret = logfs_write_i0(inode, page, &wc);
1471 if (li->li_data[INDIRECT_INDEX] != wc.ofs) {
1472 li->li_data[INDIRECT_INDEX] = wc.ofs;
1473 logfs_set_alias(inode->i_sb, li->li_block,
1474 INDIRECT_INDEX + INODE_POINTER_OFS);
1479 void logfs_add_transaction(struct inode *inode, struct logfs_transaction *ta)
1481 alloc_inode_block(inode);
1482 logfs_inode(inode)->li_block->ta = ta;
1485 void logfs_del_transaction(struct inode *inode, struct logfs_transaction *ta)
1487 struct logfs_block *block = logfs_inode(inode)->li_block;
1489 if (block && block->ta)
1493 static int grow_inode(struct inode *inode, u64 bix, level_t level)
1495 struct logfs_inode *li = logfs_inode(inode);
1496 u8 height = (__force u8)level;
1498 struct write_control wc = {
1503 BUG_ON(height > 5 || li->li_height > 5);
1504 while (height > li->li_height || bix >= maxbix(li->li_height)) {
1505 page = logfs_get_write_page(inode, I0_BLOCKS + 1,
1506 LEVEL(li->li_height + 1));
1509 logfs_read_empty(page);
1510 alloc_indirect_block(inode, page, 1);
1511 block_set_pointer(page, 0, li->li_data[INDIRECT_INDEX]);
1512 err = logfs_write_i0(inode, page, &wc);
1513 logfs_put_write_page(page);
1516 li->li_data[INDIRECT_INDEX] = wc.ofs;
1519 logfs_set_alias(inode->i_sb, li->li_block, INODE_HEIGHT_OFS);
1524 static int __logfs_write_buf(struct inode *inode, struct page *page, long flags)
1526 struct logfs_super *super = logfs_super(inode->i_sb);
1527 pgoff_t index = page->index;
1532 flags |= WF_WRITE | WF_DELETE;
1533 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1535 logfs_unpack_index(index, &bix, &level);
1536 if (logfs_block(page) && logfs_block(page)->reserved_bytes)
1537 super->s_dirty_pages -= logfs_block(page)->reserved_bytes;
1539 if (index < I0_BLOCKS)
1540 return logfs_write_direct(inode, page, flags);
1542 bix = adjust_bix(bix, level);
1543 err = grow_inode(inode, bix, level);
1546 return logfs_write_rec(inode, page, bix, level, flags);
1549 int logfs_write_buf(struct inode *inode, struct page *page, long flags)
1551 struct super_block *sb = inode->i_sb;
1554 logfs_get_wblocks(sb, page, flags & WF_LOCK);
1555 ret = __logfs_write_buf(inode, page, flags);
1556 logfs_put_wblocks(sb, page, flags & WF_LOCK);
1560 static int __logfs_delete(struct inode *inode, struct page *page)
1562 long flags = WF_DELETE;
1564 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1566 if (page->index < I0_BLOCKS)
1567 return logfs_write_direct(inode, page, flags);
1568 return logfs_write_rec(inode, page, page->index, 0, flags);
1571 int logfs_delete(struct inode *inode, pgoff_t index,
1572 struct shadow_tree *shadow_tree)
1574 struct super_block *sb = inode->i_sb;
1578 page = logfs_get_read_page(inode, index, 0);
1582 logfs_get_wblocks(sb, page, 1);
1583 ret = __logfs_delete(inode, page);
1584 logfs_put_wblocks(sb, page, 1);
1586 logfs_put_read_page(page);
1591 int logfs_rewrite_block(struct inode *inode, u64 bix, u64 ofs,
1592 gc_level_t gc_level, long flags)
1594 level_t level = shrink_level(gc_level);
1598 page = logfs_get_write_page(inode, bix, level);
1602 err = logfs_segment_read(inode, page, ofs, bix, level);
1605 alloc_indirect_block(inode, page, 0);
1606 err = logfs_write_buf(inode, page, flags);
1607 if (!err && shrink_level(gc_level) == 0) {
1608 /* Rewrite cannot mark the inode dirty but has to
1609 * write it immediatly.
1610 * Q: Can't we just create an alias for the inode
1611 * instead? And if not, why not?
1613 if (inode->i_ino == LOGFS_INO_MASTER)
1614 logfs_write_anchor(inode->i_sb);
1616 err = __logfs_write_inode(inode, flags);
1620 logfs_put_write_page(page);
1624 static int truncate_data_block(struct inode *inode, struct page *page,
1625 u64 ofs, struct logfs_shadow *shadow, u64 size)
1627 loff_t pageofs = page->index << inode->i_sb->s_blocksize_bits;
1632 /* Does truncation happen within this page? */
1633 if (size <= pageofs || size - pageofs >= PAGE_SIZE)
1636 logfs_unpack_index(page->index, &bix, &level);
1639 err = logfs_segment_read(inode, page, ofs, bix, level);
1643 zero_user_segment(page, size - pageofs, PAGE_CACHE_SIZE);
1644 return logfs_segment_write(inode, page, shadow);
1647 static int logfs_truncate_i0(struct inode *inode, struct page *page,
1648 struct write_control *wc, u64 size)
1650 struct logfs_shadow *shadow;
1655 logfs_unpack_index(page->index, &bix, &level);
1657 shadow = alloc_shadow(inode, bix, level, wc->ofs);
1659 err = truncate_data_block(inode, page, wc->ofs, shadow, size);
1661 free_shadow(inode, shadow);
1665 logfs_segment_delete(inode, shadow);
1666 set_iused(inode, shadow);
1667 fill_shadow_tree(inode, page, shadow);
1668 wc->ofs = shadow->new_ofs;
1672 static int logfs_truncate_direct(struct inode *inode, u64 size)
1674 struct logfs_inode *li = logfs_inode(inode);
1675 struct write_control wc;
1680 alloc_inode_block(inode);
1682 for (e = I0_BLOCKS - 1; e >= 0; e--) {
1683 if (size > (e+1) * LOGFS_BLOCKSIZE)
1686 wc.ofs = li->li_data[e];
1690 page = logfs_get_write_page(inode, e, 0);
1693 err = logfs_segment_read(inode, page, wc.ofs, e, 0);
1695 logfs_put_write_page(page);
1698 err = logfs_truncate_i0(inode, page, &wc, size);
1699 logfs_put_write_page(page);
1703 li->li_data[e] = wc.ofs;
1708 /* FIXME: these need to become per-sb once we support different blocksizes */
1709 static u64 __logfs_step[] = {
1716 static u64 __logfs_start_index[] = {
1723 static inline u64 logfs_step(level_t level)
1725 return __logfs_step[(__force u8)level];
1728 static inline u64 logfs_factor(u8 level)
1730 return __logfs_step[level] * LOGFS_BLOCKSIZE;
1733 static inline u64 logfs_start_index(level_t level)
1735 return __logfs_start_index[(__force u8)level];
1738 static void logfs_unpack_raw_index(pgoff_t index, u64 *bix, level_t *level)
1740 logfs_unpack_index(index, bix, level);
1741 if (*bix <= logfs_start_index(SUBLEVEL(*level)))
1745 static int __logfs_truncate_rec(struct inode *inode, struct page *ipage,
1746 struct write_control *this_wc, u64 size)
1748 int truncate_happened = 0;
1750 u64 bix, child_bix, next_bix;
1753 struct write_control child_wc = { /* FIXME: flags */ };
1755 logfs_unpack_raw_index(ipage->index, &bix, &level);
1756 err = logfs_segment_read(inode, ipage, this_wc->ofs, bix, level);
1760 for (e = LOGFS_BLOCK_FACTOR - 1; e >= 0; e--) {
1761 child_bix = bix + e * logfs_step(SUBLEVEL(level));
1762 next_bix = child_bix + logfs_step(SUBLEVEL(level));
1763 if (size > next_bix * LOGFS_BLOCKSIZE)
1766 child_wc.ofs = pure_ofs(block_get_pointer(ipage, e));
1770 page = logfs_get_write_page(inode, child_bix, SUBLEVEL(level));
1774 if ((__force u8)level > 1)
1775 err = __logfs_truncate_rec(inode, page, &child_wc, size);
1777 err = logfs_truncate_i0(inode, page, &child_wc, size);
1778 logfs_put_write_page(page);
1782 truncate_happened = 1;
1783 alloc_indirect_block(inode, ipage, 0);
1784 block_set_pointer(ipage, e, child_wc.ofs);
1787 if (!truncate_happened) {
1788 printk("ineffectual truncate (%lx, %lx, %llx)\n", inode->i_ino, ipage->index, size);
1792 this_wc->flags = WF_DELETE;
1793 if (logfs_block(ipage)->partial)
1794 this_wc->flags |= WF_WRITE;
1796 return logfs_write_i0(inode, ipage, this_wc);
1799 static int logfs_truncate_rec(struct inode *inode, u64 size)
1801 struct logfs_inode *li = logfs_inode(inode);
1802 struct write_control wc = {
1803 .ofs = li->li_data[INDIRECT_INDEX],
1808 alloc_inode_block(inode);
1813 page = logfs_get_write_page(inode, 0, LEVEL(li->li_height));
1817 err = __logfs_truncate_rec(inode, page, &wc, size);
1818 logfs_put_write_page(page);
1822 if (li->li_data[INDIRECT_INDEX] != wc.ofs)
1823 li->li_data[INDIRECT_INDEX] = wc.ofs;
1827 static int __logfs_truncate(struct inode *inode, u64 size)
1831 if (size >= logfs_factor(logfs_inode(inode)->li_height))
1834 ret = logfs_truncate_rec(inode, size);
1838 return logfs_truncate_direct(inode, size);
1842 * Truncate, by changing the segment file, can consume a fair amount
1843 * of resources. So back off from time to time and do some GC.
1844 * 8 or 2048 blocks should be well within safety limits even if
1845 * every single block resided in a different segment.
1847 #define TRUNCATE_STEP (8 * 1024 * 1024)
1848 int logfs_truncate(struct inode *inode, u64 target)
1850 struct super_block *sb = inode->i_sb;
1851 u64 size = i_size_read(inode);
1854 size = ALIGN(size, TRUNCATE_STEP);
1855 while (size > target) {
1856 if (size > TRUNCATE_STEP)
1857 size -= TRUNCATE_STEP;
1863 logfs_get_wblocks(sb, NULL, 1);
1864 err = __logfs_truncate(inode, target);
1866 err = __logfs_write_inode(inode, 0);
1867 logfs_put_wblocks(sb, NULL, 1);
1871 err = vmtruncate(inode, target);
1873 /* I don't trust error recovery yet. */
1878 static void move_page_to_inode(struct inode *inode, struct page *page)
1880 struct logfs_inode *li = logfs_inode(inode);
1881 struct logfs_block *block = logfs_block(page);
1886 log_blockmove("move_page_to_inode(%llx, %llx, %x)\n",
1887 block->ino, block->bix, block->level);
1888 BUG_ON(li->li_block);
1889 block->ops = &inode_block_ops;
1890 block->inode = inode;
1891 li->li_block = block;
1895 ClearPagePrivate(page);
1898 static void move_inode_to_page(struct page *page, struct inode *inode)
1900 struct logfs_inode *li = logfs_inode(inode);
1901 struct logfs_block *block = li->li_block;
1906 log_blockmove("move_inode_to_page(%llx, %llx, %x)\n",
1907 block->ino, block->bix, block->level);
1908 BUG_ON(PagePrivate(page));
1909 block->ops = &indirect_block_ops;
1911 page->private = (unsigned long)block;
1912 SetPagePrivate(page);
1914 block->inode = NULL;
1915 li->li_block = NULL;
1918 int logfs_read_inode(struct inode *inode)
1920 struct super_block *sb = inode->i_sb;
1921 struct logfs_super *super = logfs_super(sb);
1922 struct inode *master_inode = super->s_master_inode;
1924 struct logfs_disk_inode *di;
1925 u64 ino = inode->i_ino;
1927 if (ino << sb->s_blocksize_bits > i_size_read(master_inode))
1929 if (!logfs_exist_block(master_inode, ino))
1932 page = read_cache_page(master_inode->i_mapping, ino,
1933 (filler_t *)logfs_readpage, NULL);
1935 return PTR_ERR(page);
1937 di = kmap_atomic(page, KM_USER0);
1938 logfs_disk_to_inode(di, inode);
1939 kunmap_atomic(di, KM_USER0);
1940 move_page_to_inode(inode, page);
1941 page_cache_release(page);
1945 /* Caller must logfs_put_write_page(page); */
1946 static struct page *inode_to_page(struct inode *inode)
1948 struct inode *master_inode = logfs_super(inode->i_sb)->s_master_inode;
1949 struct logfs_disk_inode *di;
1952 BUG_ON(inode->i_ino == LOGFS_INO_MASTER);
1954 page = logfs_get_write_page(master_inode, inode->i_ino, 0);
1958 di = kmap_atomic(page, KM_USER0);
1959 logfs_inode_to_disk(inode, di);
1960 kunmap_atomic(di, KM_USER0);
1961 move_inode_to_page(page, inode);
1965 /* Cheaper version of write_inode. All changes are concealed in
1966 * aliases, which are moved back. No write to the medium happens.
1968 void logfs_clear_inode(struct inode *inode)
1970 struct super_block *sb = inode->i_sb;
1971 struct logfs_inode *li = logfs_inode(inode);
1972 struct logfs_block *block = li->li_block;
1975 /* Only deleted files may be dirty at this point */
1976 BUG_ON(inode->i_state & I_DIRTY && inode->i_nlink);
1979 if ((logfs_super(sb)->s_flags & LOGFS_SB_FLAG_SHUTDOWN)) {
1980 block->ops->free_block(inode->i_sb, block);
1984 BUG_ON(inode->i_ino < LOGFS_RESERVED_INOS);
1985 page = inode_to_page(inode);
1986 BUG_ON(!page); /* FIXME: Use emergency page */
1987 logfs_put_write_page(page);
1990 static int do_write_inode(struct inode *inode)
1992 struct super_block *sb = inode->i_sb;
1993 struct inode *master_inode = logfs_super(sb)->s_master_inode;
1994 loff_t size = (inode->i_ino + 1) << inode->i_sb->s_blocksize_bits;
1998 BUG_ON(inode->i_ino == LOGFS_INO_MASTER);
1999 /* FIXME: lock inode */
2001 if (i_size_read(master_inode) < size)
2002 i_size_write(master_inode, size);
2004 /* TODO: Tell vfs this inode is clean now */
2006 page = inode_to_page(inode);
2010 /* FIXME: transaction is part of logfs_block now. Is that enough? */
2011 err = logfs_write_buf(master_inode, page, 0);
2012 logfs_put_write_page(page);
2016 static void logfs_mod_segment_entry(struct super_block *sb, u32 segno,
2018 void (*change_se)(struct logfs_segment_entry *, long),
2021 struct logfs_super *super = logfs_super(sb);
2022 struct inode *inode;
2024 struct logfs_segment_entry *se;
2028 page_no = segno >> (sb->s_blocksize_bits - 3);
2029 child_no = segno & ((sb->s_blocksize >> 3) - 1);
2031 inode = super->s_segfile_inode;
2032 page = logfs_get_write_page(inode, page_no, 0);
2033 BUG_ON(!page); /* FIXME: We need some reserve page for this case */
2034 if (!PageUptodate(page))
2035 logfs_read_block(inode, page, WRITE);
2038 alloc_indirect_block(inode, page, 0);
2039 se = kmap_atomic(page, KM_USER0);
2040 change_se(se + child_no, arg);
2042 logfs_set_alias(sb, logfs_block(page), child_no);
2043 BUG_ON((int)be32_to_cpu(se[child_no].valid) > super->s_segsize);
2045 kunmap_atomic(se, KM_USER0);
2047 logfs_put_write_page(page);
2050 static void __get_segment_entry(struct logfs_segment_entry *se, long _target)
2052 struct logfs_segment_entry *target = (void *)_target;
2057 void logfs_get_segment_entry(struct super_block *sb, u32 segno,
2058 struct logfs_segment_entry *se)
2060 logfs_mod_segment_entry(sb, segno, 0, __get_segment_entry, (long)se);
2063 static void __set_segment_used(struct logfs_segment_entry *se, long increment)
2067 valid = be32_to_cpu(se->valid);
2069 se->valid = cpu_to_be32(valid);
2072 void logfs_set_segment_used(struct super_block *sb, u64 ofs, int increment)
2074 struct logfs_super *super = logfs_super(sb);
2075 u32 segno = ofs >> super->s_segshift;
2080 logfs_mod_segment_entry(sb, segno, 1, __set_segment_used, increment);
2083 static void __set_segment_erased(struct logfs_segment_entry *se, long ec_level)
2085 se->ec_level = cpu_to_be32(ec_level);
2088 void logfs_set_segment_erased(struct super_block *sb, u32 segno, u32 ec,
2089 gc_level_t gc_level)
2091 u32 ec_level = ec << 4 | (__force u8)gc_level;
2093 logfs_mod_segment_entry(sb, segno, 1, __set_segment_erased, ec_level);
2096 static void __set_segment_reserved(struct logfs_segment_entry *se, long ignore)
2098 se->valid = cpu_to_be32(RESERVED);
2101 void logfs_set_segment_reserved(struct super_block *sb, u32 segno)
2103 logfs_mod_segment_entry(sb, segno, 1, __set_segment_reserved, 0);
2106 static void __set_segment_unreserved(struct logfs_segment_entry *se,
2110 se->ec_level = cpu_to_be32(ec_level);
2113 void logfs_set_segment_unreserved(struct super_block *sb, u32 segno, u32 ec)
2115 u32 ec_level = ec << 4;
2117 logfs_mod_segment_entry(sb, segno, 1, __set_segment_unreserved,
2121 int __logfs_write_inode(struct inode *inode, long flags)
2123 struct super_block *sb = inode->i_sb;
2126 logfs_get_wblocks(sb, NULL, flags & WF_LOCK);
2127 ret = do_write_inode(inode);
2128 logfs_put_wblocks(sb, NULL, flags & WF_LOCK);
2132 static int do_delete_inode(struct inode *inode)
2134 struct super_block *sb = inode->i_sb;
2135 struct inode *master_inode = logfs_super(sb)->s_master_inode;
2139 page = logfs_get_write_page(master_inode, inode->i_ino, 0);
2143 move_inode_to_page(page, inode);
2145 logfs_get_wblocks(sb, page, 1);
2146 ret = __logfs_delete(master_inode, page);
2147 logfs_put_wblocks(sb, page, 1);
2149 logfs_put_write_page(page);
2154 * ZOMBIE inodes have already been deleted before and should remain dead,
2155 * if it weren't for valid checking. No need to kill them again here.
2157 void logfs_delete_inode(struct inode *inode)
2159 struct logfs_inode *li = logfs_inode(inode);
2161 if (!(li->li_flags & LOGFS_IF_ZOMBIE)) {
2162 li->li_flags |= LOGFS_IF_ZOMBIE;
2163 if (i_size_read(inode) > 0)
2164 logfs_truncate(inode, 0);
2165 do_delete_inode(inode);
2167 truncate_inode_pages(&inode->i_data, 0);
2171 void btree_write_block(struct logfs_block *block)
2173 struct inode *inode;
2177 inode = logfs_safe_iget(block->sb, block->ino, &cookie);
2178 page = logfs_get_write_page(inode, block->bix, block->level);
2180 err = logfs_readpage_nolock(page);
2182 BUG_ON(!PagePrivate(page));
2183 BUG_ON(logfs_block(page) != block);
2184 err = __logfs_write_buf(inode, page, 0);
2186 BUG_ON(PagePrivate(page) || page->private);
2188 logfs_put_write_page(page);
2189 logfs_safe_iput(inode, cookie);
2193 * logfs_inode_write - write inode or dentry objects
2195 * @inode: parent inode (ifile or directory)
2196 * @buf: object to write (inode or dentry)
2198 * @_pos: object number (file position in blocks/objects)
2199 * @flags: write flags
2200 * @lock: 0 if write lock is already taken, 1 otherwise
2201 * @shadow_tree: shadow below this inode
2203 * FIXME: All caller of this put a 200-300 byte variable on the stack,
2204 * only to call here and do a memcpy from that stack variable. A good
2205 * example of wasted performance and stack space.
2207 int logfs_inode_write(struct inode *inode, const void *buf, size_t count,
2208 loff_t bix, long flags, struct shadow_tree *shadow_tree)
2210 loff_t pos = bix << inode->i_sb->s_blocksize_bits;
2215 BUG_ON(pos & (LOGFS_BLOCKSIZE-1));
2216 BUG_ON(count > LOGFS_BLOCKSIZE);
2217 page = logfs_get_write_page(inode, bix, 0);
2221 pagebuf = kmap_atomic(page, KM_USER0);
2222 memcpy(pagebuf, buf, count);
2223 flush_dcache_page(page);
2224 kunmap_atomic(pagebuf, KM_USER0);
2226 if (i_size_read(inode) < pos + LOGFS_BLOCKSIZE)
2227 i_size_write(inode, pos + LOGFS_BLOCKSIZE);
2229 err = logfs_write_buf(inode, page, flags);
2230 logfs_put_write_page(page);
2234 int logfs_open_segfile(struct super_block *sb)
2236 struct logfs_super *super = logfs_super(sb);
2237 struct inode *inode;
2239 inode = logfs_read_meta_inode(sb, LOGFS_INO_SEGFILE);
2241 return PTR_ERR(inode);
2242 super->s_segfile_inode = inode;
2246 int logfs_init_rw(struct super_block *sb)
2248 struct logfs_super *super = logfs_super(sb);
2249 int min_fill = 3 * super->s_no_blocks;
2251 INIT_LIST_HEAD(&super->s_object_alias);
2252 mutex_init(&super->s_write_mutex);
2253 super->s_block_pool = mempool_create_kmalloc_pool(min_fill,
2254 sizeof(struct logfs_block));
2255 super->s_shadow_pool = mempool_create_kmalloc_pool(min_fill,
2256 sizeof(struct logfs_shadow));
2260 void logfs_cleanup_rw(struct super_block *sb)
2262 struct logfs_super *super = logfs_super(sb);
2264 destroy_meta_inode(super->s_segfile_inode);
2265 logfs_mempool_destroy(super->s_block_pool);
2266 logfs_mempool_destroy(super->s_shadow_pool);