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 set_nlink(inode, 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 void logfs_get_wblocks(struct super_block *sb, struct page *page, int lock)
249 struct logfs_super *super = logfs_super(sb);
252 prelock_page(sb, page, lock);
255 mutex_lock(&super->s_write_mutex);
257 /* FIXME: We also have to check for shadowed space
258 * and mempool fill grade */
262 void logfs_put_wblocks(struct super_block *sb, struct page *page, int lock)
264 struct logfs_super *super = logfs_super(sb);
267 preunlock_page(sb, page, lock);
268 /* Order matters - we must clear PG_pre_locked before releasing
269 * s_write_mutex or we could race against another task. */
271 mutex_unlock(&super->s_write_mutex);
274 static struct page *logfs_get_read_page(struct inode *inode, u64 bix,
277 return find_or_create_page(inode->i_mapping,
278 logfs_pack_index(bix, level), GFP_NOFS);
281 static void logfs_put_read_page(struct page *page)
284 page_cache_release(page);
287 static void logfs_lock_write_page(struct page *page)
291 while (unlikely(!trylock_page(page))) {
292 if (loop++ > 0x1000) {
293 /* Has been observed once so far... */
294 printk(KERN_ERR "stack at %p\n", &loop);
297 if (PagePreLocked(page)) {
298 /* Holder of page lock is waiting for us, it
299 * is safe to use this page. */
302 /* Some other process has this page locked and has
303 * nothing to do with us. Wait for it to finish.
307 BUG_ON(!PageLocked(page));
310 static struct page *logfs_get_write_page(struct inode *inode, u64 bix,
313 struct address_space *mapping = inode->i_mapping;
314 pgoff_t index = logfs_pack_index(bix, level);
319 page = find_get_page(mapping, index);
321 page = __page_cache_alloc(GFP_NOFS);
324 err = add_to_page_cache_lru(page, mapping, index, GFP_NOFS);
326 page_cache_release(page);
331 } else logfs_lock_write_page(page);
332 BUG_ON(!PageLocked(page));
336 static void logfs_unlock_write_page(struct page *page)
338 if (!PagePreLocked(page))
342 static void logfs_put_write_page(struct page *page)
344 logfs_unlock_write_page(page);
345 page_cache_release(page);
348 static struct page *logfs_get_page(struct inode *inode, u64 bix, level_t level,
352 return logfs_get_read_page(inode, bix, level);
354 return logfs_get_write_page(inode, bix, level);
357 static void logfs_put_page(struct page *page, int rw)
360 logfs_put_read_page(page);
362 logfs_put_write_page(page);
365 static unsigned long __get_bits(u64 val, int skip, int no)
375 static unsigned long get_bits(u64 val, level_t skip)
377 return __get_bits(val, (__force int)skip, LOGFS_BLOCK_BITS);
380 static inline void init_shadow_tree(struct super_block *sb,
381 struct shadow_tree *tree)
383 struct logfs_super *super = logfs_super(sb);
385 btree_init_mempool64(&tree->new, super->s_btree_pool);
386 btree_init_mempool64(&tree->old, super->s_btree_pool);
389 static void indirect_write_block(struct logfs_block *block)
396 inode = page->mapping->host;
397 logfs_lock_write_page(page);
398 ret = logfs_write_buf(inode, page, 0);
399 logfs_unlock_write_page(page);
401 * This needs some rework. Unless you want your filesystem to run
402 * completely synchronously (you don't), the filesystem will always
403 * report writes as 'successful' before the actual work has been
404 * done. The actual work gets done here and this is where any errors
405 * will show up. And there isn't much we can do about it, really.
407 * Some attempts to fix the errors (move from bad blocks, retry io,...)
408 * have already been done, so anything left should be either a broken
409 * device or a bug somewhere in logfs itself. Being relatively new,
410 * the odds currently favor a bug, so for now the line below isn't
416 static void inode_write_block(struct logfs_block *block)
421 inode = block->inode;
422 if (inode->i_ino == LOGFS_INO_MASTER)
423 logfs_write_anchor(inode->i_sb);
425 ret = __logfs_write_inode(inode, 0);
426 /* see indirect_write_block comment */
432 * This silences a false, yet annoying gcc warning. I hate it when my editor
433 * jumps into bitops.h each time I recompile this file.
434 * TODO: Complain to gcc folks about this and upgrade compiler.
436 static unsigned long fnb(const unsigned long *addr,
437 unsigned long size, unsigned long offset)
439 return find_next_bit(addr, size, offset);
442 static __be64 inode_val0(struct inode *inode)
444 struct logfs_inode *li = logfs_inode(inode);
448 * Explicit shifting generates good code, but must match the format
449 * of the structure. Add some paranoia just in case.
451 BUILD_BUG_ON(offsetof(struct logfs_disk_inode, di_mode) != 0);
452 BUILD_BUG_ON(offsetof(struct logfs_disk_inode, di_height) != 2);
453 BUILD_BUG_ON(offsetof(struct logfs_disk_inode, di_flags) != 4);
455 val = (u64)inode->i_mode << 48 |
456 (u64)li->li_height << 40 |
458 return cpu_to_be64(val);
461 static int inode_write_alias(struct super_block *sb,
462 struct logfs_block *block, write_alias_t *write_one_alias)
464 struct inode *inode = block->inode;
465 struct logfs_inode *li = logfs_inode(inode);
472 for (pos = 0; ; pos++) {
473 pos = fnb(block->alias_map, LOGFS_BLOCK_FACTOR, pos);
474 if (pos >= LOGFS_EMBEDDED_FIELDS + INODE_POINTER_OFS)
478 case INODE_HEIGHT_OFS:
479 val = inode_val0(inode);
482 val = cpu_to_be64(li->li_used_bytes);
485 val = cpu_to_be64(i_size_read(inode));
487 case INODE_POINTER_OFS ... INODE_POINTER_OFS + LOGFS_EMBEDDED_FIELDS - 1:
488 val = cpu_to_be64(li->li_data[pos - INODE_POINTER_OFS]);
494 ino = LOGFS_INO_MASTER;
497 err = write_one_alias(sb, ino, bix, level, pos, val);
503 static int indirect_write_alias(struct super_block *sb,
504 struct logfs_block *block, write_alias_t *write_one_alias)
507 struct page *page = block->page;
513 for (pos = 0; ; pos++) {
514 pos = fnb(block->alias_map, LOGFS_BLOCK_FACTOR, pos);
515 if (pos >= LOGFS_BLOCK_FACTOR)
518 ino = page->mapping->host->i_ino;
519 logfs_unpack_index(page->index, &bix, &level);
520 child = kmap_atomic(page, KM_USER0);
522 kunmap_atomic(child, KM_USER0);
523 err = write_one_alias(sb, ino, bix, level, pos, val);
529 int logfs_write_obj_aliases_pagecache(struct super_block *sb)
531 struct logfs_super *super = logfs_super(sb);
532 struct logfs_block *block;
535 list_for_each_entry(block, &super->s_object_alias, alias_list) {
536 err = block->ops->write_alias(sb, block, write_alias_journal);
543 void __free_block(struct super_block *sb, struct logfs_block *block)
545 BUG_ON(!list_empty(&block->item_list));
546 list_del(&block->alias_list);
547 mempool_free(block, logfs_super(sb)->s_block_pool);
550 static void inode_free_block(struct super_block *sb, struct logfs_block *block)
552 struct inode *inode = block->inode;
554 logfs_inode(inode)->li_block = NULL;
555 __free_block(sb, block);
558 static void indirect_free_block(struct super_block *sb,
559 struct logfs_block *block)
561 struct page *page = block->page;
563 if (PagePrivate(page)) {
564 ClearPagePrivate(page);
565 page_cache_release(page);
566 set_page_private(page, 0);
568 __free_block(sb, block);
572 static struct logfs_block_ops inode_block_ops = {
573 .write_block = inode_write_block,
574 .free_block = inode_free_block,
575 .write_alias = inode_write_alias,
578 struct logfs_block_ops indirect_block_ops = {
579 .write_block = indirect_write_block,
580 .free_block = indirect_free_block,
581 .write_alias = indirect_write_alias,
584 struct logfs_block *__alloc_block(struct super_block *sb,
585 u64 ino, u64 bix, level_t level)
587 struct logfs_super *super = logfs_super(sb);
588 struct logfs_block *block;
590 block = mempool_alloc(super->s_block_pool, GFP_NOFS);
591 memset(block, 0, sizeof(*block));
592 INIT_LIST_HEAD(&block->alias_list);
593 INIT_LIST_HEAD(&block->item_list);
597 block->level = level;
601 static void alloc_inode_block(struct inode *inode)
603 struct logfs_inode *li = logfs_inode(inode);
604 struct logfs_block *block;
609 block = __alloc_block(inode->i_sb, LOGFS_INO_MASTER, inode->i_ino, 0);
610 block->inode = inode;
611 li->li_block = block;
612 block->ops = &inode_block_ops;
615 void initialize_block_counters(struct page *page, struct logfs_block *block,
616 __be64 *array, int page_is_empty)
624 if (page->index < first_indirect_block()) {
625 /* Counters are pointless on level 0 */
628 if (page->index == first_indirect_block()) {
629 /* Skip unused pointers */
631 block->full = I0_BLOCKS;
633 if (!page_is_empty) {
634 for (i = start; i < LOGFS_BLOCK_FACTOR; i++) {
635 ptr = be64_to_cpu(array[i]);
638 if (ptr & LOGFS_FULLY_POPULATED)
644 static void alloc_data_block(struct inode *inode, struct page *page)
646 struct logfs_block *block;
650 if (PagePrivate(page))
653 logfs_unpack_index(page->index, &bix, &level);
654 block = __alloc_block(inode->i_sb, inode->i_ino, bix, level);
657 SetPagePrivate(page);
658 page_cache_get(page);
659 set_page_private(page, (unsigned long) block);
661 block->ops = &indirect_block_ops;
664 static void alloc_indirect_block(struct inode *inode, struct page *page,
667 struct logfs_block *block;
670 if (PagePrivate(page))
673 alloc_data_block(inode, page);
675 block = logfs_block(page);
676 array = kmap_atomic(page, KM_USER0);
677 initialize_block_counters(page, block, array, page_is_empty);
678 kunmap_atomic(array, KM_USER0);
681 static void block_set_pointer(struct page *page, int index, u64 ptr)
683 struct logfs_block *block = logfs_block(page);
688 array = kmap_atomic(page, KM_USER0);
689 oldptr = be64_to_cpu(array[index]);
690 array[index] = cpu_to_be64(ptr);
691 kunmap_atomic(array, KM_USER0);
692 SetPageUptodate(page);
694 block->full += !!(ptr & LOGFS_FULLY_POPULATED)
695 - !!(oldptr & LOGFS_FULLY_POPULATED);
696 block->partial += !!ptr - !!oldptr;
699 static u64 block_get_pointer(struct page *page, int index)
704 block = kmap_atomic(page, KM_USER0);
705 ptr = be64_to_cpu(block[index]);
706 kunmap_atomic(block, KM_USER0);
710 static int logfs_read_empty(struct page *page)
712 zero_user_segment(page, 0, PAGE_CACHE_SIZE);
716 static int logfs_read_direct(struct inode *inode, struct page *page)
718 struct logfs_inode *li = logfs_inode(inode);
719 pgoff_t index = page->index;
722 block = li->li_data[index];
724 return logfs_read_empty(page);
726 return logfs_segment_read(inode, page, block, index, 0);
729 static int logfs_read_loop(struct inode *inode, struct page *page,
732 struct logfs_inode *li = logfs_inode(inode);
733 u64 bix, bofs = li->li_data[INDIRECT_INDEX];
734 level_t level, target_level;
738 logfs_unpack_index(page->index, &bix, &target_level);
740 return logfs_read_empty(page);
742 if (bix >= maxbix(li->li_height))
743 return logfs_read_empty(page);
745 for (level = LEVEL(li->li_height);
746 (__force u8)level > (__force u8)target_level;
747 level = SUBLEVEL(level)){
748 ipage = logfs_get_page(inode, bix, level, rw_context);
752 ret = logfs_segment_read(inode, ipage, bofs, bix, level);
754 logfs_put_read_page(ipage);
758 bofs = block_get_pointer(ipage, get_bits(bix, SUBLEVEL(level)));
759 logfs_put_page(ipage, rw_context);
761 return logfs_read_empty(page);
764 return logfs_segment_read(inode, page, bofs, bix, 0);
767 static int logfs_read_block(struct inode *inode, struct page *page,
770 pgoff_t index = page->index;
772 if (index < I0_BLOCKS)
773 return logfs_read_direct(inode, page);
774 return logfs_read_loop(inode, page, rw_context);
777 static int logfs_exist_loop(struct inode *inode, u64 bix)
779 struct logfs_inode *li = logfs_inode(inode);
780 u64 bofs = li->li_data[INDIRECT_INDEX];
787 if (bix >= maxbix(li->li_height))
790 for (level = LEVEL(li->li_height); level != 0; level = SUBLEVEL(level)) {
791 ipage = logfs_get_read_page(inode, bix, level);
795 ret = logfs_segment_read(inode, ipage, bofs, bix, level);
797 logfs_put_read_page(ipage);
801 bofs = block_get_pointer(ipage, get_bits(bix, SUBLEVEL(level)));
802 logfs_put_read_page(ipage);
810 int logfs_exist_block(struct inode *inode, u64 bix)
812 struct logfs_inode *li = logfs_inode(inode);
815 return !!li->li_data[bix];
816 return logfs_exist_loop(inode, bix);
819 static u64 seek_holedata_direct(struct inode *inode, u64 bix, int data)
821 struct logfs_inode *li = logfs_inode(inode);
823 for (; bix < I0_BLOCKS; bix++)
824 if (data ^ (li->li_data[bix] == 0))
829 static u64 seek_holedata_loop(struct inode *inode, u64 bix, int data)
831 struct logfs_inode *li = logfs_inode(inode);
833 u64 increment, bofs = li->li_data[INDIRECT_INDEX];
840 for (level = LEVEL(li->li_height); level != 0; level = SUBLEVEL(level)) {
841 increment = 1 << (LOGFS_BLOCK_BITS * ((__force u8)level-1));
842 page = logfs_get_read_page(inode, bix, level);
846 ret = logfs_segment_read(inode, page, bofs, bix, level);
848 logfs_put_read_page(page);
852 slot = get_bits(bix, SUBLEVEL(level));
853 rblock = kmap_atomic(page, KM_USER0);
854 while (slot < LOGFS_BLOCK_FACTOR) {
855 if (data && (rblock[slot] != 0))
857 if (!data && !(be64_to_cpu(rblock[slot]) & LOGFS_FULLY_POPULATED))
861 bix &= ~(increment - 1);
863 if (slot >= LOGFS_BLOCK_FACTOR) {
864 kunmap_atomic(rblock, KM_USER0);
865 logfs_put_read_page(page);
868 bofs = be64_to_cpu(rblock[slot]);
869 kunmap_atomic(rblock, KM_USER0);
870 logfs_put_read_page(page);
880 * logfs_seek_hole - find next hole starting at a given block index
881 * @inode: inode to search in
882 * @bix: block index to start searching
884 * Returns next hole. If the file doesn't contain any further holes, the
885 * block address next to eof is returned instead.
887 u64 logfs_seek_hole(struct inode *inode, u64 bix)
889 struct logfs_inode *li = logfs_inode(inode);
891 if (bix < I0_BLOCKS) {
892 bix = seek_holedata_direct(inode, bix, 0);
897 if (!li->li_data[INDIRECT_INDEX])
899 else if (li->li_data[INDIRECT_INDEX] & LOGFS_FULLY_POPULATED)
900 bix = maxbix(li->li_height);
901 else if (bix >= maxbix(li->li_height))
904 bix = seek_holedata_loop(inode, bix, 0);
905 if (bix < maxbix(li->li_height))
907 /* Should not happen anymore. But if some port writes semi-
908 * corrupt images (as this one used to) we might run into it.
910 WARN_ON_ONCE(bix == maxbix(li->li_height));
916 static u64 __logfs_seek_data(struct inode *inode, u64 bix)
918 struct logfs_inode *li = logfs_inode(inode);
920 if (bix < I0_BLOCKS) {
921 bix = seek_holedata_direct(inode, bix, 1);
926 if (bix < maxbix(li->li_height)) {
927 if (!li->li_data[INDIRECT_INDEX])
928 bix = maxbix(li->li_height);
930 return seek_holedata_loop(inode, bix, 1);
937 * logfs_seek_data - find next data block after a given block index
938 * @inode: inode to search in
939 * @bix: block index to start searching
941 * Returns next data block. If the file doesn't contain any further data
942 * blocks, the last block in the file is returned instead.
944 u64 logfs_seek_data(struct inode *inode, u64 bix)
946 struct super_block *sb = inode->i_sb;
949 ret = __logfs_seek_data(inode, bix);
950 end = i_size_read(inode) >> sb->s_blocksize_bits;
956 static int logfs_is_valid_direct(struct logfs_inode *li, u64 bix, u64 ofs)
958 return pure_ofs(li->li_data[bix]) == ofs;
961 static int __logfs_is_valid_loop(struct inode *inode, u64 bix,
964 struct logfs_inode *li = logfs_inode(inode);
969 for (level = LEVEL(li->li_height); level != 0; level = SUBLEVEL(level)){
970 page = logfs_get_write_page(inode, bix, level);
973 ret = logfs_segment_read(inode, page, bofs, bix, level);
975 logfs_put_write_page(page);
979 bofs = block_get_pointer(page, get_bits(bix, SUBLEVEL(level)));
980 logfs_put_write_page(page);
984 if (pure_ofs(bofs) == ofs)
990 static int logfs_is_valid_loop(struct inode *inode, u64 bix, u64 ofs)
992 struct logfs_inode *li = logfs_inode(inode);
993 u64 bofs = li->li_data[INDIRECT_INDEX];
998 if (bix >= maxbix(li->li_height))
1001 if (pure_ofs(bofs) == ofs)
1004 return __logfs_is_valid_loop(inode, bix, ofs, bofs);
1007 static int __logfs_is_valid_block(struct inode *inode, u64 bix, u64 ofs)
1009 struct logfs_inode *li = logfs_inode(inode);
1011 if ((inode->i_nlink == 0) && atomic_read(&inode->i_count) == 1)
1014 if (bix < I0_BLOCKS)
1015 return logfs_is_valid_direct(li, bix, ofs);
1016 return logfs_is_valid_loop(inode, bix, ofs);
1020 * logfs_is_valid_block - check whether this block is still valid
1023 * @ofs - block physical offset
1024 * @ino - block inode number
1025 * @bix - block index
1026 * @level - block level
1028 * Returns 0 if the block is invalid, 1 if it is valid and 2 if it will
1029 * become invalid once the journal is written.
1031 int logfs_is_valid_block(struct super_block *sb, u64 ofs, u64 ino, u64 bix,
1032 gc_level_t gc_level)
1034 struct logfs_super *super = logfs_super(sb);
1035 struct inode *inode;
1038 /* Umount closes a segment with free blocks remaining. Those
1039 * blocks are by definition invalid. */
1043 LOGFS_BUG_ON((u64)(u_long)ino != ino, sb);
1045 inode = logfs_safe_iget(sb, ino, &cookie);
1049 ret = __logfs_is_valid_block(inode, bix, ofs);
1050 logfs_safe_iput(inode, cookie);
1055 /* Block is nominally invalid, but may still sit in the shadow tree,
1056 * waiting for a journal commit.
1058 if (btree_lookup64(&super->s_shadow_tree.old, ofs))
1063 int logfs_readpage_nolock(struct page *page)
1065 struct inode *inode = page->mapping->host;
1068 ret = logfs_read_block(inode, page, READ);
1071 ClearPageUptodate(page);
1074 SetPageUptodate(page);
1075 ClearPageError(page);
1077 flush_dcache_page(page);
1082 static int logfs_reserve_bytes(struct inode *inode, int bytes)
1084 struct logfs_super *super = logfs_super(inode->i_sb);
1085 u64 available = super->s_free_bytes + super->s_dirty_free_bytes
1086 - super->s_dirty_used_bytes - super->s_dirty_pages;
1091 if (available < bytes)
1094 if (available < bytes + super->s_root_reserve &&
1095 !capable(CAP_SYS_RESOURCE))
1101 int get_page_reserve(struct inode *inode, struct page *page)
1103 struct logfs_super *super = logfs_super(inode->i_sb);
1104 struct logfs_block *block = logfs_block(page);
1107 if (block && block->reserved_bytes)
1110 logfs_get_wblocks(inode->i_sb, page, WF_LOCK);
1111 while ((ret = logfs_reserve_bytes(inode, 6 * LOGFS_MAX_OBJECTSIZE)) &&
1112 !list_empty(&super->s_writeback_list)) {
1113 block = list_entry(super->s_writeback_list.next,
1114 struct logfs_block, alias_list);
1115 block->ops->write_block(block);
1118 alloc_data_block(inode, page);
1119 block = logfs_block(page);
1120 block->reserved_bytes += 6 * LOGFS_MAX_OBJECTSIZE;
1121 super->s_dirty_pages += 6 * LOGFS_MAX_OBJECTSIZE;
1122 list_move_tail(&block->alias_list, &super->s_writeback_list);
1124 logfs_put_wblocks(inode->i_sb, page, WF_LOCK);
1129 * We are protected by write lock. Push victims up to superblock level
1130 * and release transaction when appropriate.
1132 /* FIXME: This is currently called from the wrong spots. */
1133 static void logfs_handle_transaction(struct inode *inode,
1134 struct logfs_transaction *ta)
1136 struct logfs_super *super = logfs_super(inode->i_sb);
1140 logfs_inode(inode)->li_block->ta = NULL;
1142 if (inode->i_ino != LOGFS_INO_MASTER) {
1143 BUG(); /* FIXME: Yes, this needs more thought */
1144 /* just remember the transaction until inode is written */
1145 //BUG_ON(logfs_inode(inode)->li_transaction);
1146 //logfs_inode(inode)->li_transaction = ta;
1150 switch (ta->state) {
1151 case CREATE_1: /* fall through */
1153 BUG_ON(super->s_victim_ino);
1154 super->s_victim_ino = ta->ino;
1156 case CREATE_2: /* fall through */
1158 BUG_ON(super->s_victim_ino != ta->ino);
1159 super->s_victim_ino = 0;
1160 /* transaction ends here - free it */
1163 case CROSS_RENAME_1:
1164 BUG_ON(super->s_rename_dir);
1165 BUG_ON(super->s_rename_pos);
1166 super->s_rename_dir = ta->dir;
1167 super->s_rename_pos = ta->pos;
1169 case CROSS_RENAME_2:
1170 BUG_ON(super->s_rename_dir != ta->dir);
1171 BUG_ON(super->s_rename_pos != ta->pos);
1172 super->s_rename_dir = 0;
1173 super->s_rename_pos = 0;
1176 case TARGET_RENAME_1:
1177 BUG_ON(super->s_rename_dir);
1178 BUG_ON(super->s_rename_pos);
1179 BUG_ON(super->s_victim_ino);
1180 super->s_rename_dir = ta->dir;
1181 super->s_rename_pos = ta->pos;
1182 super->s_victim_ino = ta->ino;
1184 case TARGET_RENAME_2:
1185 BUG_ON(super->s_rename_dir != ta->dir);
1186 BUG_ON(super->s_rename_pos != ta->pos);
1187 BUG_ON(super->s_victim_ino != ta->ino);
1188 super->s_rename_dir = 0;
1189 super->s_rename_pos = 0;
1191 case TARGET_RENAME_3:
1192 BUG_ON(super->s_rename_dir);
1193 BUG_ON(super->s_rename_pos);
1194 BUG_ON(super->s_victim_ino != ta->ino);
1195 super->s_victim_ino = 0;
1204 * Not strictly a reservation, but rather a check that we still have enough
1205 * space to satisfy the write.
1207 static int logfs_reserve_blocks(struct inode *inode, int blocks)
1209 return logfs_reserve_bytes(inode, blocks * LOGFS_MAX_OBJECTSIZE);
1212 struct write_control {
1217 static struct logfs_shadow *alloc_shadow(struct inode *inode, u64 bix,
1218 level_t level, u64 old_ofs)
1220 struct logfs_super *super = logfs_super(inode->i_sb);
1221 struct logfs_shadow *shadow;
1223 shadow = mempool_alloc(super->s_shadow_pool, GFP_NOFS);
1224 memset(shadow, 0, sizeof(*shadow));
1225 shadow->ino = inode->i_ino;
1227 shadow->gc_level = expand_level(inode->i_ino, level);
1228 shadow->old_ofs = old_ofs & ~LOGFS_FULLY_POPULATED;
1232 static void free_shadow(struct inode *inode, struct logfs_shadow *shadow)
1234 struct logfs_super *super = logfs_super(inode->i_sb);
1236 mempool_free(shadow, super->s_shadow_pool);
1239 static void mark_segment(struct shadow_tree *tree, u32 segno)
1243 if (!btree_lookup32(&tree->segment_map, segno)) {
1244 err = btree_insert32(&tree->segment_map, segno, (void *)1,
1247 tree->no_shadowed_segments++;
1252 * fill_shadow_tree - Propagate shadow tree changes due to a write
1253 * @inode: Inode owning the page
1254 * @page: Struct page that was written
1255 * @shadow: Shadow for the current write
1257 * Writes in logfs can result in two semi-valid objects. The old object
1258 * is still valid as long as it can be reached by following pointers on
1259 * the medium. Only when writes propagate all the way up to the journal
1260 * has the new object safely replaced the old one.
1262 * To handle this problem, a struct logfs_shadow is used to represent
1263 * every single write. It is attached to the indirect block, which is
1264 * marked dirty. When the indirect block is written, its shadows are
1265 * handed up to the next indirect block (or inode). Untimately they
1266 * will reach the master inode and be freed upon journal commit.
1268 * This function handles a single step in the propagation. It adds the
1269 * shadow for the current write to the tree, along with any shadows in
1270 * the page's tree, in case it was an indirect block. If a page is
1271 * written, the inode parameter is left NULL, if an inode is written,
1272 * the page parameter is left NULL.
1274 static void fill_shadow_tree(struct inode *inode, struct page *page,
1275 struct logfs_shadow *shadow)
1277 struct logfs_super *super = logfs_super(inode->i_sb);
1278 struct logfs_block *block = logfs_block(page);
1279 struct shadow_tree *tree = &super->s_shadow_tree;
1281 if (PagePrivate(page)) {
1282 if (block->alias_map)
1283 super->s_no_object_aliases -= bitmap_weight(
1284 block->alias_map, LOGFS_BLOCK_FACTOR);
1285 logfs_handle_transaction(inode, block->ta);
1286 block->ops->free_block(inode->i_sb, block);
1289 if (shadow->old_ofs)
1290 btree_insert64(&tree->old, shadow->old_ofs, shadow,
1293 btree_insert64(&tree->new, shadow->new_ofs, shadow,
1296 super->s_dirty_used_bytes += shadow->new_len;
1297 super->s_dirty_free_bytes += shadow->old_len;
1298 mark_segment(tree, shadow->old_ofs >> super->s_segshift);
1299 mark_segment(tree, shadow->new_ofs >> super->s_segshift);
1303 static void logfs_set_alias(struct super_block *sb, struct logfs_block *block,
1306 struct logfs_super *super = logfs_super(sb);
1308 if (block->inode && block->inode->i_ino == LOGFS_INO_MASTER) {
1309 /* Aliases in the master inode are pointless. */
1313 if (!test_bit(child_no, block->alias_map)) {
1314 set_bit(child_no, block->alias_map);
1315 super->s_no_object_aliases++;
1317 list_move_tail(&block->alias_list, &super->s_object_alias);
1321 * Object aliases can and often do change the size and occupied space of a
1322 * file. So not only do we have to change the pointers, we also have to
1323 * change inode->i_size and li->li_used_bytes. Which is done by setting
1324 * another two object aliases for the inode itself.
1326 static void set_iused(struct inode *inode, struct logfs_shadow *shadow)
1328 struct logfs_inode *li = logfs_inode(inode);
1330 if (shadow->new_len == shadow->old_len)
1333 alloc_inode_block(inode);
1334 li->li_used_bytes += shadow->new_len - shadow->old_len;
1335 __logfs_set_blocks(inode);
1336 logfs_set_alias(inode->i_sb, li->li_block, INODE_USED_OFS);
1337 logfs_set_alias(inode->i_sb, li->li_block, INODE_SIZE_OFS);
1340 static int logfs_write_i0(struct inode *inode, struct page *page,
1341 struct write_control *wc)
1343 struct logfs_shadow *shadow;
1348 logfs_unpack_index(page->index, &bix, &level);
1350 if (logfs_reserve_blocks(inode, 1))
1353 shadow = alloc_shadow(inode, bix, level, wc->ofs);
1354 if (wc->flags & WF_WRITE)
1355 err = logfs_segment_write(inode, page, shadow);
1356 if (wc->flags & WF_DELETE)
1357 logfs_segment_delete(inode, shadow);
1359 free_shadow(inode, shadow);
1363 set_iused(inode, shadow);
1366 alloc_indirect_block(inode, page, 0);
1367 full = logfs_block(page)->full == LOGFS_BLOCK_FACTOR;
1369 fill_shadow_tree(inode, page, shadow);
1370 wc->ofs = shadow->new_ofs;
1371 if (wc->ofs && full)
1372 wc->ofs |= LOGFS_FULLY_POPULATED;
1376 static int logfs_write_direct(struct inode *inode, struct page *page,
1379 struct logfs_inode *li = logfs_inode(inode);
1380 struct write_control wc = {
1381 .ofs = li->li_data[page->index],
1386 alloc_inode_block(inode);
1388 err = logfs_write_i0(inode, page, &wc);
1392 li->li_data[page->index] = wc.ofs;
1393 logfs_set_alias(inode->i_sb, li->li_block,
1394 page->index + INODE_POINTER_OFS);
1398 static int ptr_change(u64 ofs, struct page *page)
1400 struct logfs_block *block = logfs_block(page);
1401 int empty0, empty1, full0, full1;
1404 empty1 = block->partial == 0;
1405 if (empty0 != empty1)
1408 /* The !! is necessary to shrink result to int */
1409 full0 = !!(ofs & LOGFS_FULLY_POPULATED);
1410 full1 = block->full == LOGFS_BLOCK_FACTOR;
1416 static int __logfs_write_rec(struct inode *inode, struct page *page,
1417 struct write_control *this_wc,
1418 pgoff_t bix, level_t target_level, level_t level)
1420 int ret, page_empty = 0;
1421 int child_no = get_bits(bix, SUBLEVEL(level));
1423 struct write_control child_wc = {
1424 .flags = this_wc->flags,
1427 ipage = logfs_get_write_page(inode, bix, level);
1432 ret = logfs_segment_read(inode, ipage, this_wc->ofs, bix, level);
1435 } else if (!PageUptodate(ipage)) {
1437 logfs_read_empty(ipage);
1440 child_wc.ofs = block_get_pointer(ipage, child_no);
1442 if ((__force u8)level-1 > (__force u8)target_level)
1443 ret = __logfs_write_rec(inode, page, &child_wc, bix,
1444 target_level, SUBLEVEL(level));
1446 ret = logfs_write_i0(inode, page, &child_wc);
1451 alloc_indirect_block(inode, ipage, page_empty);
1452 block_set_pointer(ipage, child_no, child_wc.ofs);
1453 /* FIXME: first condition seems superfluous */
1454 if (child_wc.ofs || logfs_block(ipage)->partial)
1455 this_wc->flags |= WF_WRITE;
1456 /* the condition on this_wc->ofs ensures that we won't consume extra
1457 * space for indirect blocks in the future, which we cannot reserve */
1458 if (!this_wc->ofs || ptr_change(this_wc->ofs, ipage))
1459 ret = logfs_write_i0(inode, ipage, this_wc);
1461 logfs_set_alias(inode->i_sb, logfs_block(ipage), child_no);
1463 logfs_put_write_page(ipage);
1467 static int logfs_write_rec(struct inode *inode, struct page *page,
1468 pgoff_t bix, level_t target_level, long flags)
1470 struct logfs_inode *li = logfs_inode(inode);
1471 struct write_control wc = {
1472 .ofs = li->li_data[INDIRECT_INDEX],
1477 alloc_inode_block(inode);
1479 if (li->li_height > (__force u8)target_level)
1480 ret = __logfs_write_rec(inode, page, &wc, bix, target_level,
1481 LEVEL(li->li_height));
1483 ret = logfs_write_i0(inode, page, &wc);
1487 if (li->li_data[INDIRECT_INDEX] != wc.ofs) {
1488 li->li_data[INDIRECT_INDEX] = wc.ofs;
1489 logfs_set_alias(inode->i_sb, li->li_block,
1490 INDIRECT_INDEX + INODE_POINTER_OFS);
1495 void logfs_add_transaction(struct inode *inode, struct logfs_transaction *ta)
1497 alloc_inode_block(inode);
1498 logfs_inode(inode)->li_block->ta = ta;
1501 void logfs_del_transaction(struct inode *inode, struct logfs_transaction *ta)
1503 struct logfs_block *block = logfs_inode(inode)->li_block;
1505 if (block && block->ta)
1509 static int grow_inode(struct inode *inode, u64 bix, level_t level)
1511 struct logfs_inode *li = logfs_inode(inode);
1512 u8 height = (__force u8)level;
1514 struct write_control wc = {
1519 BUG_ON(height > 5 || li->li_height > 5);
1520 while (height > li->li_height || bix >= maxbix(li->li_height)) {
1521 page = logfs_get_write_page(inode, I0_BLOCKS + 1,
1522 LEVEL(li->li_height + 1));
1525 logfs_read_empty(page);
1526 alloc_indirect_block(inode, page, 1);
1527 block_set_pointer(page, 0, li->li_data[INDIRECT_INDEX]);
1528 err = logfs_write_i0(inode, page, &wc);
1529 logfs_put_write_page(page);
1532 li->li_data[INDIRECT_INDEX] = wc.ofs;
1535 logfs_set_alias(inode->i_sb, li->li_block, INODE_HEIGHT_OFS);
1540 static int __logfs_write_buf(struct inode *inode, struct page *page, long flags)
1542 struct logfs_super *super = logfs_super(inode->i_sb);
1543 pgoff_t index = page->index;
1548 flags |= WF_WRITE | WF_DELETE;
1549 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1551 logfs_unpack_index(index, &bix, &level);
1552 if (logfs_block(page) && logfs_block(page)->reserved_bytes)
1553 super->s_dirty_pages -= logfs_block(page)->reserved_bytes;
1555 if (index < I0_BLOCKS)
1556 return logfs_write_direct(inode, page, flags);
1558 bix = adjust_bix(bix, level);
1559 err = grow_inode(inode, bix, level);
1562 return logfs_write_rec(inode, page, bix, level, flags);
1565 int logfs_write_buf(struct inode *inode, struct page *page, long flags)
1567 struct super_block *sb = inode->i_sb;
1570 logfs_get_wblocks(sb, page, flags & WF_LOCK);
1571 ret = __logfs_write_buf(inode, page, flags);
1572 logfs_put_wblocks(sb, page, flags & WF_LOCK);
1576 static int __logfs_delete(struct inode *inode, struct page *page)
1578 long flags = WF_DELETE;
1580 inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1582 if (page->index < I0_BLOCKS)
1583 return logfs_write_direct(inode, page, flags);
1584 return logfs_write_rec(inode, page, page->index, 0, flags);
1587 int logfs_delete(struct inode *inode, pgoff_t index,
1588 struct shadow_tree *shadow_tree)
1590 struct super_block *sb = inode->i_sb;
1594 page = logfs_get_read_page(inode, index, 0);
1598 logfs_get_wblocks(sb, page, 1);
1599 ret = __logfs_delete(inode, page);
1600 logfs_put_wblocks(sb, page, 1);
1602 logfs_put_read_page(page);
1607 int logfs_rewrite_block(struct inode *inode, u64 bix, u64 ofs,
1608 gc_level_t gc_level, long flags)
1610 level_t level = shrink_level(gc_level);
1614 page = logfs_get_write_page(inode, bix, level);
1618 err = logfs_segment_read(inode, page, ofs, bix, level);
1621 alloc_indirect_block(inode, page, 0);
1622 err = logfs_write_buf(inode, page, flags);
1623 if (!err && shrink_level(gc_level) == 0) {
1624 /* Rewrite cannot mark the inode dirty but has to
1625 * write it immediately.
1626 * Q: Can't we just create an alias for the inode
1627 * instead? And if not, why not?
1629 if (inode->i_ino == LOGFS_INO_MASTER)
1630 logfs_write_anchor(inode->i_sb);
1632 err = __logfs_write_inode(inode, flags);
1636 logfs_put_write_page(page);
1640 static int truncate_data_block(struct inode *inode, struct page *page,
1641 u64 ofs, struct logfs_shadow *shadow, u64 size)
1643 loff_t pageofs = page->index << inode->i_sb->s_blocksize_bits;
1648 /* Does truncation happen within this page? */
1649 if (size <= pageofs || size - pageofs >= PAGE_SIZE)
1652 logfs_unpack_index(page->index, &bix, &level);
1655 err = logfs_segment_read(inode, page, ofs, bix, level);
1659 zero_user_segment(page, size - pageofs, PAGE_CACHE_SIZE);
1660 return logfs_segment_write(inode, page, shadow);
1663 static int logfs_truncate_i0(struct inode *inode, struct page *page,
1664 struct write_control *wc, u64 size)
1666 struct logfs_shadow *shadow;
1671 logfs_unpack_index(page->index, &bix, &level);
1673 shadow = alloc_shadow(inode, bix, level, wc->ofs);
1675 err = truncate_data_block(inode, page, wc->ofs, shadow, size);
1677 free_shadow(inode, shadow);
1681 logfs_segment_delete(inode, shadow);
1682 set_iused(inode, shadow);
1683 fill_shadow_tree(inode, page, shadow);
1684 wc->ofs = shadow->new_ofs;
1688 static int logfs_truncate_direct(struct inode *inode, u64 size)
1690 struct logfs_inode *li = logfs_inode(inode);
1691 struct write_control wc;
1696 alloc_inode_block(inode);
1698 for (e = I0_BLOCKS - 1; e >= 0; e--) {
1699 if (size > (e+1) * LOGFS_BLOCKSIZE)
1702 wc.ofs = li->li_data[e];
1706 page = logfs_get_write_page(inode, e, 0);
1709 err = logfs_segment_read(inode, page, wc.ofs, e, 0);
1711 logfs_put_write_page(page);
1714 err = logfs_truncate_i0(inode, page, &wc, size);
1715 logfs_put_write_page(page);
1719 li->li_data[e] = wc.ofs;
1724 /* FIXME: these need to become per-sb once we support different blocksizes */
1725 static u64 __logfs_step[] = {
1732 static u64 __logfs_start_index[] = {
1739 static inline u64 logfs_step(level_t level)
1741 return __logfs_step[(__force u8)level];
1744 static inline u64 logfs_factor(u8 level)
1746 return __logfs_step[level] * LOGFS_BLOCKSIZE;
1749 static inline u64 logfs_start_index(level_t level)
1751 return __logfs_start_index[(__force u8)level];
1754 static void logfs_unpack_raw_index(pgoff_t index, u64 *bix, level_t *level)
1756 logfs_unpack_index(index, bix, level);
1757 if (*bix <= logfs_start_index(SUBLEVEL(*level)))
1761 static int __logfs_truncate_rec(struct inode *inode, struct page *ipage,
1762 struct write_control *this_wc, u64 size)
1764 int truncate_happened = 0;
1766 u64 bix, child_bix, next_bix;
1769 struct write_control child_wc = { /* FIXME: flags */ };
1771 logfs_unpack_raw_index(ipage->index, &bix, &level);
1772 err = logfs_segment_read(inode, ipage, this_wc->ofs, bix, level);
1776 for (e = LOGFS_BLOCK_FACTOR - 1; e >= 0; e--) {
1777 child_bix = bix + e * logfs_step(SUBLEVEL(level));
1778 next_bix = child_bix + logfs_step(SUBLEVEL(level));
1779 if (size > next_bix * LOGFS_BLOCKSIZE)
1782 child_wc.ofs = pure_ofs(block_get_pointer(ipage, e));
1786 page = logfs_get_write_page(inode, child_bix, SUBLEVEL(level));
1790 if ((__force u8)level > 1)
1791 err = __logfs_truncate_rec(inode, page, &child_wc, size);
1793 err = logfs_truncate_i0(inode, page, &child_wc, size);
1794 logfs_put_write_page(page);
1798 truncate_happened = 1;
1799 alloc_indirect_block(inode, ipage, 0);
1800 block_set_pointer(ipage, e, child_wc.ofs);
1803 if (!truncate_happened) {
1804 printk("ineffectual truncate (%lx, %lx, %llx)\n", inode->i_ino, ipage->index, size);
1808 this_wc->flags = WF_DELETE;
1809 if (logfs_block(ipage)->partial)
1810 this_wc->flags |= WF_WRITE;
1812 return logfs_write_i0(inode, ipage, this_wc);
1815 static int logfs_truncate_rec(struct inode *inode, u64 size)
1817 struct logfs_inode *li = logfs_inode(inode);
1818 struct write_control wc = {
1819 .ofs = li->li_data[INDIRECT_INDEX],
1824 alloc_inode_block(inode);
1829 page = logfs_get_write_page(inode, 0, LEVEL(li->li_height));
1833 err = __logfs_truncate_rec(inode, page, &wc, size);
1834 logfs_put_write_page(page);
1838 if (li->li_data[INDIRECT_INDEX] != wc.ofs)
1839 li->li_data[INDIRECT_INDEX] = wc.ofs;
1843 static int __logfs_truncate(struct inode *inode, u64 size)
1847 if (size >= logfs_factor(logfs_inode(inode)->li_height))
1850 ret = logfs_truncate_rec(inode, size);
1854 return logfs_truncate_direct(inode, size);
1858 * Truncate, by changing the segment file, can consume a fair amount
1859 * of resources. So back off from time to time and do some GC.
1860 * 8 or 2048 blocks should be well within safety limits even if
1861 * every single block resided in a different segment.
1863 #define TRUNCATE_STEP (8 * 1024 * 1024)
1864 int logfs_truncate(struct inode *inode, u64 target)
1866 struct super_block *sb = inode->i_sb;
1867 u64 size = i_size_read(inode);
1870 size = ALIGN(size, TRUNCATE_STEP);
1871 while (size > target) {
1872 if (size > TRUNCATE_STEP)
1873 size -= TRUNCATE_STEP;
1879 logfs_get_wblocks(sb, NULL, 1);
1880 err = __logfs_truncate(inode, size);
1882 err = __logfs_write_inode(inode, 0);
1883 logfs_put_wblocks(sb, NULL, 1);
1887 err = vmtruncate(inode, target);
1889 /* I don't trust error recovery yet. */
1894 static void move_page_to_inode(struct inode *inode, struct page *page)
1896 struct logfs_inode *li = logfs_inode(inode);
1897 struct logfs_block *block = logfs_block(page);
1902 log_blockmove("move_page_to_inode(%llx, %llx, %x)\n",
1903 block->ino, block->bix, block->level);
1904 BUG_ON(li->li_block);
1905 block->ops = &inode_block_ops;
1906 block->inode = inode;
1907 li->li_block = block;
1910 if (PagePrivate(page)) {
1911 ClearPagePrivate(page);
1912 page_cache_release(page);
1913 set_page_private(page, 0);
1917 static void move_inode_to_page(struct page *page, struct inode *inode)
1919 struct logfs_inode *li = logfs_inode(inode);
1920 struct logfs_block *block = li->li_block;
1925 log_blockmove("move_inode_to_page(%llx, %llx, %x)\n",
1926 block->ino, block->bix, block->level);
1927 BUG_ON(PagePrivate(page));
1928 block->ops = &indirect_block_ops;
1931 if (!PagePrivate(page)) {
1932 SetPagePrivate(page);
1933 page_cache_get(page);
1934 set_page_private(page, (unsigned long) block);
1937 block->inode = NULL;
1938 li->li_block = NULL;
1941 int logfs_read_inode(struct inode *inode)
1943 struct super_block *sb = inode->i_sb;
1944 struct logfs_super *super = logfs_super(sb);
1945 struct inode *master_inode = super->s_master_inode;
1947 struct logfs_disk_inode *di;
1948 u64 ino = inode->i_ino;
1950 if (ino << sb->s_blocksize_bits > i_size_read(master_inode))
1952 if (!logfs_exist_block(master_inode, ino))
1955 page = read_cache_page(master_inode->i_mapping, ino,
1956 (filler_t *)logfs_readpage, NULL);
1958 return PTR_ERR(page);
1960 di = kmap_atomic(page, KM_USER0);
1961 logfs_disk_to_inode(di, inode);
1962 kunmap_atomic(di, KM_USER0);
1963 move_page_to_inode(inode, page);
1964 page_cache_release(page);
1968 /* Caller must logfs_put_write_page(page); */
1969 static struct page *inode_to_page(struct inode *inode)
1971 struct inode *master_inode = logfs_super(inode->i_sb)->s_master_inode;
1972 struct logfs_disk_inode *di;
1975 BUG_ON(inode->i_ino == LOGFS_INO_MASTER);
1977 page = logfs_get_write_page(master_inode, inode->i_ino, 0);
1981 di = kmap_atomic(page, KM_USER0);
1982 logfs_inode_to_disk(inode, di);
1983 kunmap_atomic(di, KM_USER0);
1984 move_inode_to_page(page, inode);
1988 static int do_write_inode(struct inode *inode)
1990 struct super_block *sb = inode->i_sb;
1991 struct inode *master_inode = logfs_super(sb)->s_master_inode;
1992 loff_t size = (inode->i_ino + 1) << inode->i_sb->s_blocksize_bits;
1996 BUG_ON(inode->i_ino == LOGFS_INO_MASTER);
1997 /* FIXME: lock inode */
1999 if (i_size_read(master_inode) < size)
2000 i_size_write(master_inode, size);
2002 /* TODO: Tell vfs this inode is clean now */
2004 page = inode_to_page(inode);
2008 /* FIXME: transaction is part of logfs_block now. Is that enough? */
2009 err = logfs_write_buf(master_inode, page, 0);
2011 move_page_to_inode(inode, page);
2013 logfs_put_write_page(page);
2017 static void logfs_mod_segment_entry(struct super_block *sb, u32 segno,
2019 void (*change_se)(struct logfs_segment_entry *, long),
2022 struct logfs_super *super = logfs_super(sb);
2023 struct inode *inode;
2025 struct logfs_segment_entry *se;
2029 page_no = segno >> (sb->s_blocksize_bits - 3);
2030 child_no = segno & ((sb->s_blocksize >> 3) - 1);
2032 inode = super->s_segfile_inode;
2033 page = logfs_get_write_page(inode, page_no, 0);
2034 BUG_ON(!page); /* FIXME: We need some reserve page for this case */
2035 if (!PageUptodate(page))
2036 logfs_read_block(inode, page, WRITE);
2039 alloc_indirect_block(inode, page, 0);
2040 se = kmap_atomic(page, KM_USER0);
2041 change_se(se + child_no, arg);
2043 logfs_set_alias(sb, logfs_block(page), child_no);
2044 BUG_ON((int)be32_to_cpu(se[child_no].valid) > super->s_segsize);
2046 kunmap_atomic(se, KM_USER0);
2048 logfs_put_write_page(page);
2051 static void __get_segment_entry(struct logfs_segment_entry *se, long _target)
2053 struct logfs_segment_entry *target = (void *)_target;
2058 void logfs_get_segment_entry(struct super_block *sb, u32 segno,
2059 struct logfs_segment_entry *se)
2061 logfs_mod_segment_entry(sb, segno, 0, __get_segment_entry, (long)se);
2064 static void __set_segment_used(struct logfs_segment_entry *se, long increment)
2068 valid = be32_to_cpu(se->valid);
2070 se->valid = cpu_to_be32(valid);
2073 void logfs_set_segment_used(struct super_block *sb, u64 ofs, int increment)
2075 struct logfs_super *super = logfs_super(sb);
2076 u32 segno = ofs >> super->s_segshift;
2081 logfs_mod_segment_entry(sb, segno, 1, __set_segment_used, increment);
2084 static void __set_segment_erased(struct logfs_segment_entry *se, long ec_level)
2086 se->ec_level = cpu_to_be32(ec_level);
2089 void logfs_set_segment_erased(struct super_block *sb, u32 segno, u32 ec,
2090 gc_level_t gc_level)
2092 u32 ec_level = ec << 4 | (__force u8)gc_level;
2094 logfs_mod_segment_entry(sb, segno, 1, __set_segment_erased, ec_level);
2097 static void __set_segment_reserved(struct logfs_segment_entry *se, long ignore)
2099 se->valid = cpu_to_be32(RESERVED);
2102 void logfs_set_segment_reserved(struct super_block *sb, u32 segno)
2104 logfs_mod_segment_entry(sb, segno, 1, __set_segment_reserved, 0);
2107 static void __set_segment_unreserved(struct logfs_segment_entry *se,
2111 se->ec_level = cpu_to_be32(ec_level);
2114 void logfs_set_segment_unreserved(struct super_block *sb, u32 segno, u32 ec)
2116 u32 ec_level = ec << 4;
2118 logfs_mod_segment_entry(sb, segno, 1, __set_segment_unreserved,
2122 int __logfs_write_inode(struct inode *inode, long flags)
2124 struct super_block *sb = inode->i_sb;
2127 logfs_get_wblocks(sb, NULL, flags & WF_LOCK);
2128 ret = do_write_inode(inode);
2129 logfs_put_wblocks(sb, NULL, flags & WF_LOCK);
2133 static int do_delete_inode(struct inode *inode)
2135 struct super_block *sb = inode->i_sb;
2136 struct inode *master_inode = logfs_super(sb)->s_master_inode;
2140 page = logfs_get_write_page(master_inode, inode->i_ino, 0);
2144 move_inode_to_page(page, inode);
2146 logfs_get_wblocks(sb, page, 1);
2147 ret = __logfs_delete(master_inode, page);
2148 logfs_put_wblocks(sb, page, 1);
2150 logfs_put_write_page(page);
2155 * ZOMBIE inodes have already been deleted before and should remain dead,
2156 * if it weren't for valid checking. No need to kill them again here.
2158 void logfs_evict_inode(struct inode *inode)
2160 struct super_block *sb = inode->i_sb;
2161 struct logfs_inode *li = logfs_inode(inode);
2162 struct logfs_block *block = li->li_block;
2165 if (!inode->i_nlink) {
2166 if (!(li->li_flags & LOGFS_IF_ZOMBIE)) {
2167 li->li_flags |= LOGFS_IF_ZOMBIE;
2168 if (i_size_read(inode) > 0)
2169 logfs_truncate(inode, 0);
2170 do_delete_inode(inode);
2173 truncate_inode_pages(&inode->i_data, 0);
2174 end_writeback(inode);
2176 /* Cheaper version of write_inode. All changes are concealed in
2177 * aliases, which are moved back. No write to the medium happens.
2179 /* Only deleted files may be dirty at this point */
2180 BUG_ON(inode->i_state & I_DIRTY && inode->i_nlink);
2183 if ((logfs_super(sb)->s_flags & LOGFS_SB_FLAG_SHUTDOWN)) {
2184 block->ops->free_block(inode->i_sb, block);
2188 BUG_ON(inode->i_ino < LOGFS_RESERVED_INOS);
2189 page = inode_to_page(inode);
2190 BUG_ON(!page); /* FIXME: Use emergency page */
2191 logfs_put_write_page(page);
2194 void btree_write_block(struct logfs_block *block)
2196 struct inode *inode;
2200 inode = logfs_safe_iget(block->sb, block->ino, &cookie);
2201 page = logfs_get_write_page(inode, block->bix, block->level);
2203 err = logfs_readpage_nolock(page);
2205 BUG_ON(!PagePrivate(page));
2206 BUG_ON(logfs_block(page) != block);
2207 err = __logfs_write_buf(inode, page, 0);
2209 BUG_ON(PagePrivate(page) || page->private);
2211 logfs_put_write_page(page);
2212 logfs_safe_iput(inode, cookie);
2216 * logfs_inode_write - write inode or dentry objects
2218 * @inode: parent inode (ifile or directory)
2219 * @buf: object to write (inode or dentry)
2221 * @_pos: object number (file position in blocks/objects)
2222 * @flags: write flags
2223 * @lock: 0 if write lock is already taken, 1 otherwise
2224 * @shadow_tree: shadow below this inode
2226 * FIXME: All caller of this put a 200-300 byte variable on the stack,
2227 * only to call here and do a memcpy from that stack variable. A good
2228 * example of wasted performance and stack space.
2230 int logfs_inode_write(struct inode *inode, const void *buf, size_t count,
2231 loff_t bix, long flags, struct shadow_tree *shadow_tree)
2233 loff_t pos = bix << inode->i_sb->s_blocksize_bits;
2238 BUG_ON(pos & (LOGFS_BLOCKSIZE-1));
2239 BUG_ON(count > LOGFS_BLOCKSIZE);
2240 page = logfs_get_write_page(inode, bix, 0);
2244 pagebuf = kmap_atomic(page, KM_USER0);
2245 memcpy(pagebuf, buf, count);
2246 flush_dcache_page(page);
2247 kunmap_atomic(pagebuf, KM_USER0);
2249 if (i_size_read(inode) < pos + LOGFS_BLOCKSIZE)
2250 i_size_write(inode, pos + LOGFS_BLOCKSIZE);
2252 err = logfs_write_buf(inode, page, flags);
2253 logfs_put_write_page(page);
2257 int logfs_open_segfile(struct super_block *sb)
2259 struct logfs_super *super = logfs_super(sb);
2260 struct inode *inode;
2262 inode = logfs_read_meta_inode(sb, LOGFS_INO_SEGFILE);
2264 return PTR_ERR(inode);
2265 super->s_segfile_inode = inode;
2269 int logfs_init_rw(struct super_block *sb)
2271 struct logfs_super *super = logfs_super(sb);
2272 int min_fill = 3 * super->s_no_blocks;
2274 INIT_LIST_HEAD(&super->s_object_alias);
2275 INIT_LIST_HEAD(&super->s_writeback_list);
2276 mutex_init(&super->s_write_mutex);
2277 super->s_block_pool = mempool_create_kmalloc_pool(min_fill,
2278 sizeof(struct logfs_block));
2279 super->s_shadow_pool = mempool_create_kmalloc_pool(min_fill,
2280 sizeof(struct logfs_shadow));
2284 void logfs_cleanup_rw(struct super_block *sb)
2286 struct logfs_super *super = logfs_super(sb);
2288 logfs_mempool_destroy(super->s_block_pool);
2289 logfs_mempool_destroy(super->s_shadow_pool);