Merge branch 'master' into export-slabh
[kernel.git] / fs / logfs / readwrite.c
1 /*
2  * fs/logfs/readwrite.c
3  *
4  * As should be obvious for Linux kernel code, license is GPLv2
5  *
6  * Copyright (c) 2005-2008 Joern Engel <joern@logfs.org>
7  *
8  *
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
18  */
19 #include "logfs.h"
20 #include <linux/sched.h>
21 #include <linux/slab.h>
22
23 static u64 adjust_bix(u64 bix, level_t level)
24 {
25         switch (level) {
26         case 0:
27                 return bix;
28         case LEVEL(1):
29                 return max_t(u64, bix, I0_BLOCKS);
30         case LEVEL(2):
31                 return max_t(u64, bix, I1_BLOCKS);
32         case LEVEL(3):
33                 return max_t(u64, bix, I2_BLOCKS);
34         case LEVEL(4):
35                 return max_t(u64, bix, I3_BLOCKS);
36         case LEVEL(5):
37                 return max_t(u64, bix, I4_BLOCKS);
38         default:
39                 WARN_ON(1);
40                 return bix;
41         }
42 }
43
44 static inline u64 maxbix(u8 height)
45 {
46         return 1ULL << (LOGFS_BLOCK_BITS * height);
47 }
48
49 /**
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
53  * index.
54  *
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.
58  */
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)
63 {
64         return INDIRECT_BIT | (1ULL << LEVEL_SHIFT);
65 }
66
67 pgoff_t logfs_pack_index(u64 bix, level_t level)
68 {
69         pgoff_t index;
70
71         BUG_ON(bix >= INDIRECT_BIT);
72         if (level == 0)
73                 return bix;
74
75         index  = INDIRECT_BIT;
76         index |= (__force long)level << LEVEL_SHIFT;
77         index |= bix >> ((__force u8)level * LOGFS_BLOCK_BITS);
78         return index;
79 }
80
81 void logfs_unpack_index(pgoff_t index, u64 *bix, level_t *level)
82 {
83         u8 __level;
84
85         if (!(index & INDIRECT_BIT)) {
86                 *bix = index;
87                 *level = 0;
88                 return;
89         }
90
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);
95         return;
96 }
97 #undef ARCH_SHIFT
98 #undef INDIRECT_BIT
99 #undef LEVEL_SHIFT
100
101 /*
102  * Time is stored as nanoseconds since the epoch.
103  */
104 static struct timespec be64_to_timespec(__be64 betime)
105 {
106         return ns_to_timespec(be64_to_cpu(betime));
107 }
108
109 static __be64 timespec_to_be64(struct timespec tsp)
110 {
111         return cpu_to_be64((u64)tsp.tv_sec * NSEC_PER_SEC + tsp.tv_nsec);
112 }
113
114 static void logfs_disk_to_inode(struct logfs_disk_inode *di, struct inode*inode)
115 {
116         struct logfs_inode *li = logfs_inode(inode);
117         int i;
118
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);
131
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 */
136         case S_IFIFO:
137                 inode->i_rdev = be64_to_cpu(di->di_data[0]);
138                 break;
139         case S_IFDIR:   /* fall through */
140         case S_IFREG:   /* fall through */
141         case S_IFLNK:
142                 for (i = 0; i < LOGFS_EMBEDDED_FIELDS; i++)
143                         li->li_data[i] = be64_to_cpu(di->di_data[i]);
144                 break;
145         default:
146                 BUG();
147         }
148 }
149
150 static void logfs_inode_to_disk(struct inode *inode, struct logfs_disk_inode*di)
151 {
152         struct logfs_inode *li = logfs_inode(inode);
153         int i;
154
155         di->di_mode     = cpu_to_be16(inode->i_mode);
156         di->di_height   = li->li_height;
157         di->di_pad      = 0;
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);
168
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 */
173         case S_IFIFO:
174                 di->di_data[0] = cpu_to_be64(inode->i_rdev);
175                 break;
176         case S_IFDIR:   /* fall through */
177         case S_IFREG:   /* fall through */
178         case S_IFLNK:
179                 for (i = 0; i < LOGFS_EMBEDDED_FIELDS; i++)
180                         di->di_data[i] = cpu_to_be64(li->li_data[i]);
181                 break;
182         default:
183                 BUG();
184         }
185 }
186
187 static void __logfs_set_blocks(struct inode *inode)
188 {
189         struct super_block *sb = inode->i_sb;
190         struct logfs_inode *li = logfs_inode(inode);
191
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;
195 }
196
197 void logfs_set_blocks(struct inode *inode, u64 bytes)
198 {
199         struct logfs_inode *li = logfs_inode(inode);
200
201         li->li_used_bytes = bytes;
202         __logfs_set_blocks(inode);
203 }
204
205 static void prelock_page(struct super_block *sb, struct page *page, int lock)
206 {
207         struct logfs_super *super = logfs_super(sb);
208
209         BUG_ON(!PageLocked(page));
210         if (lock) {
211                 BUG_ON(PagePreLocked(page));
212                 SetPagePreLocked(page);
213         } else {
214                 /* We are in GC path. */
215                 if (PagePreLocked(page))
216                         super->s_lock_count++;
217                 else
218                         SetPagePreLocked(page);
219         }
220 }
221
222 static void preunlock_page(struct super_block *sb, struct page *page, int lock)
223 {
224         struct logfs_super *super = logfs_super(sb);
225
226         BUG_ON(!PageLocked(page));
227         if (lock)
228                 ClearPagePreLocked(page);
229         else {
230                 /* We are in GC path. */
231                 BUG_ON(!PagePreLocked(page));
232                 if (super->s_lock_count)
233                         super->s_lock_count--;
234                 else
235                         ClearPagePreLocked(page);
236         }
237 }
238
239 /*
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
242  * s_write_mutex.
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.
246  */
247 static void logfs_get_wblocks(struct super_block *sb, struct page *page,
248                 int lock)
249 {
250         struct logfs_super *super = logfs_super(sb);
251
252         if (page)
253                 prelock_page(sb, page, lock);
254
255         if (lock) {
256                 mutex_lock(&super->s_write_mutex);
257                 logfs_gc_pass(sb);
258                 /* FIXME: We also have to check for shadowed space
259                  * and mempool fill grade */
260         }
261 }
262
263 static void logfs_put_wblocks(struct super_block *sb, struct page *page,
264                 int lock)
265 {
266         struct logfs_super *super = logfs_super(sb);
267
268         if (page)
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. */
272         if (lock)
273                 mutex_unlock(&super->s_write_mutex);
274 }
275
276 static struct page *logfs_get_read_page(struct inode *inode, u64 bix,
277                 level_t level)
278 {
279         return find_or_create_page(inode->i_mapping,
280                         logfs_pack_index(bix, level), GFP_NOFS);
281 }
282
283 static void logfs_put_read_page(struct page *page)
284 {
285         unlock_page(page);
286         page_cache_release(page);
287 }
288
289 static void logfs_lock_write_page(struct page *page)
290 {
291         int loop = 0;
292
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);
297                         BUG();
298                 }
299                 if (PagePreLocked(page)) {
300                         /* Holder of page lock is waiting for us, it
301                          * is safe to use this page. */
302                         break;
303                 }
304                 /* Some other process has this page locked and has
305                  * nothing to do with us.  Wait for it to finish.
306                  */
307                 schedule();
308         }
309         BUG_ON(!PageLocked(page));
310 }
311
312 static struct page *logfs_get_write_page(struct inode *inode, u64 bix,
313                 level_t level)
314 {
315         struct address_space *mapping = inode->i_mapping;
316         pgoff_t index = logfs_pack_index(bix, level);
317         struct page *page;
318         int err;
319
320 repeat:
321         page = find_get_page(mapping, index);
322         if (!page) {
323                 page = __page_cache_alloc(GFP_NOFS);
324                 if (!page)
325                         return NULL;
326                 err = add_to_page_cache_lru(page, mapping, index, GFP_NOFS);
327                 if (unlikely(err)) {
328                         page_cache_release(page);
329                         if (err == -EEXIST)
330                                 goto repeat;
331                         return NULL;
332                 }
333         } else logfs_lock_write_page(page);
334         BUG_ON(!PageLocked(page));
335         return page;
336 }
337
338 static void logfs_unlock_write_page(struct page *page)
339 {
340         if (!PagePreLocked(page))
341                 unlock_page(page);
342 }
343
344 static void logfs_put_write_page(struct page *page)
345 {
346         logfs_unlock_write_page(page);
347         page_cache_release(page);
348 }
349
350 static struct page *logfs_get_page(struct inode *inode, u64 bix, level_t level,
351                 int rw)
352 {
353         if (rw == READ)
354                 return logfs_get_read_page(inode, bix, level);
355         else
356                 return logfs_get_write_page(inode, bix, level);
357 }
358
359 static void logfs_put_page(struct page *page, int rw)
360 {
361         if (rw == READ)
362                 logfs_put_read_page(page);
363         else
364                 logfs_put_write_page(page);
365 }
366
367 static unsigned long __get_bits(u64 val, int skip, int no)
368 {
369         u64 ret = val;
370
371         ret >>= skip * no;
372         ret <<= 64 - no;
373         ret >>= 64 - no;
374         return ret;
375 }
376
377 static unsigned long get_bits(u64 val, level_t skip)
378 {
379         return __get_bits(val, (__force int)skip, LOGFS_BLOCK_BITS);
380 }
381
382 static inline void init_shadow_tree(struct super_block *sb,
383                 struct shadow_tree *tree)
384 {
385         struct logfs_super *super = logfs_super(sb);
386
387         btree_init_mempool64(&tree->new, super->s_btree_pool);
388         btree_init_mempool64(&tree->old, super->s_btree_pool);
389 }
390
391 static void indirect_write_block(struct logfs_block *block)
392 {
393         struct page *page;
394         struct inode *inode;
395         int ret;
396
397         page = block->page;
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);
402         /*
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.
408          *
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
413          * entirely tasteles.
414          */
415         BUG_ON(ret);
416 }
417
418 static void inode_write_block(struct logfs_block *block)
419 {
420         struct inode *inode;
421         int ret;
422
423         inode = block->inode;
424         if (inode->i_ino == LOGFS_INO_MASTER)
425                 logfs_write_anchor(inode->i_sb);
426         else {
427                 ret = __logfs_write_inode(inode, 0);
428                 /* see indirect_write_block comment */
429                 BUG_ON(ret);
430         }
431 }
432
433 static gc_level_t inode_block_level(struct logfs_block *block)
434 {
435         BUG_ON(block->inode->i_ino == LOGFS_INO_MASTER);
436         return GC_LEVEL(LOGFS_MAX_LEVELS);
437 }
438
439 static gc_level_t indirect_block_level(struct logfs_block *block)
440 {
441         struct page *page;
442         struct inode *inode;
443         u64 bix;
444         level_t level;
445
446         page = block->page;
447         inode = page->mapping->host;
448         logfs_unpack_index(page->index, &bix, &level);
449         return expand_level(inode->i_ino, level);
450 }
451
452 /*
453  * This silences a false, yet annoying gcc warning.  I hate it when my editor
454  * jumps into bitops.h each time I recompile this file.
455  * TODO: Complain to gcc folks about this and upgrade compiler.
456  */
457 static unsigned long fnb(const unsigned long *addr,
458                 unsigned long size, unsigned long offset)
459 {
460         return find_next_bit(addr, size, offset);
461 }
462
463 static __be64 inode_val0(struct inode *inode)
464 {
465         struct logfs_inode *li = logfs_inode(inode);
466         u64 val;
467
468         /*
469          * Explicit shifting generates good code, but must match the format
470          * of the structure.  Add some paranoia just in case.
471          */
472         BUILD_BUG_ON(offsetof(struct logfs_disk_inode, di_mode) != 0);
473         BUILD_BUG_ON(offsetof(struct logfs_disk_inode, di_height) != 2);
474         BUILD_BUG_ON(offsetof(struct logfs_disk_inode, di_flags) != 4);
475
476         val =   (u64)inode->i_mode << 48 |
477                 (u64)li->li_height << 40 |
478                 (u64)li->li_flags;
479         return cpu_to_be64(val);
480 }
481
482 static int inode_write_alias(struct super_block *sb,
483                 struct logfs_block *block, write_alias_t *write_one_alias)
484 {
485         struct inode *inode = block->inode;
486         struct logfs_inode *li = logfs_inode(inode);
487         unsigned long pos;
488         u64 ino , bix;
489         __be64 val;
490         level_t level;
491         int err;
492
493         for (pos = 0; ; pos++) {
494                 pos = fnb(block->alias_map, LOGFS_BLOCK_FACTOR, pos);
495                 if (pos >= LOGFS_EMBEDDED_FIELDS + INODE_POINTER_OFS)
496                         return 0;
497
498                 switch (pos) {
499                 case INODE_HEIGHT_OFS:
500                         val = inode_val0(inode);
501                         break;
502                 case INODE_USED_OFS:
503                         val = cpu_to_be64(li->li_used_bytes);;
504                         break;
505                 case INODE_SIZE_OFS:
506                         val = cpu_to_be64(i_size_read(inode));
507                         break;
508                 case INODE_POINTER_OFS ... INODE_POINTER_OFS + LOGFS_EMBEDDED_FIELDS - 1:
509                         val = cpu_to_be64(li->li_data[pos - INODE_POINTER_OFS]);
510                         break;
511                 default:
512                         BUG();
513                 }
514
515                 ino = LOGFS_INO_MASTER;
516                 bix = inode->i_ino;
517                 level = LEVEL(0);
518                 err = write_one_alias(sb, ino, bix, level, pos, val);
519                 if (err)
520                         return err;
521         }
522 }
523
524 static int indirect_write_alias(struct super_block *sb,
525                 struct logfs_block *block, write_alias_t *write_one_alias)
526 {
527         unsigned long pos;
528         struct page *page = block->page;
529         u64 ino , bix;
530         __be64 *child, val;
531         level_t level;
532         int err;
533
534         for (pos = 0; ; pos++) {
535                 pos = fnb(block->alias_map, LOGFS_BLOCK_FACTOR, pos);
536                 if (pos >= LOGFS_BLOCK_FACTOR)
537                         return 0;
538
539                 ino = page->mapping->host->i_ino;
540                 logfs_unpack_index(page->index, &bix, &level);
541                 child = kmap_atomic(page, KM_USER0);
542                 val = child[pos];
543                 kunmap_atomic(child, KM_USER0);
544                 err = write_one_alias(sb, ino, bix, level, pos, val);
545                 if (err)
546                         return err;
547         }
548 }
549
550 int logfs_write_obj_aliases_pagecache(struct super_block *sb)
551 {
552         struct logfs_super *super = logfs_super(sb);
553         struct logfs_block *block;
554         int err;
555
556         list_for_each_entry(block, &super->s_object_alias, alias_list) {
557                 err = block->ops->write_alias(sb, block, write_alias_journal);
558                 if (err)
559                         return err;
560         }
561         return 0;
562 }
563
564 void __free_block(struct super_block *sb, struct logfs_block *block)
565 {
566         BUG_ON(!list_empty(&block->item_list));
567         list_del(&block->alias_list);
568         mempool_free(block, logfs_super(sb)->s_block_pool);
569 }
570
571 static void inode_free_block(struct super_block *sb, struct logfs_block *block)
572 {
573         struct inode *inode = block->inode;
574
575         logfs_inode(inode)->li_block = NULL;
576         __free_block(sb, block);
577 }
578
579 static void indirect_free_block(struct super_block *sb,
580                 struct logfs_block *block)
581 {
582         ClearPagePrivate(block->page);
583         block->page->private = 0;
584         __free_block(sb, block);
585 }
586
587
588 static struct logfs_block_ops inode_block_ops = {
589         .write_block = inode_write_block,
590         .block_level = inode_block_level,
591         .free_block = inode_free_block,
592         .write_alias = inode_write_alias,
593 };
594
595 struct logfs_block_ops indirect_block_ops = {
596         .write_block = indirect_write_block,
597         .block_level = indirect_block_level,
598         .free_block = indirect_free_block,
599         .write_alias = indirect_write_alias,
600 };
601
602 struct logfs_block *__alloc_block(struct super_block *sb,
603                 u64 ino, u64 bix, level_t level)
604 {
605         struct logfs_super *super = logfs_super(sb);
606         struct logfs_block *block;
607
608         block = mempool_alloc(super->s_block_pool, GFP_NOFS);
609         memset(block, 0, sizeof(*block));
610         INIT_LIST_HEAD(&block->alias_list);
611         INIT_LIST_HEAD(&block->item_list);
612         block->sb = sb;
613         block->ino = ino;
614         block->bix = bix;
615         block->level = level;
616         return block;
617 }
618
619 static void alloc_inode_block(struct inode *inode)
620 {
621         struct logfs_inode *li = logfs_inode(inode);
622         struct logfs_block *block;
623
624         if (li->li_block)
625                 return;
626
627         block = __alloc_block(inode->i_sb, LOGFS_INO_MASTER, inode->i_ino, 0);
628         block->inode = inode;
629         li->li_block = block;
630         block->ops = &inode_block_ops;
631 }
632
633 void initialize_block_counters(struct page *page, struct logfs_block *block,
634                 __be64 *array, int page_is_empty)
635 {
636         u64 ptr;
637         int i, start;
638
639         block->partial = 0;
640         block->full = 0;
641         start = 0;
642         if (page->index < first_indirect_block()) {
643                 /* Counters are pointless on level 0 */
644                 return;
645         }
646         if (page->index == first_indirect_block()) {
647                 /* Skip unused pointers */
648                 start = I0_BLOCKS;
649                 block->full = I0_BLOCKS;
650         }
651         if (!page_is_empty) {
652                 for (i = start; i < LOGFS_BLOCK_FACTOR; i++) {
653                         ptr = be64_to_cpu(array[i]);
654                         if (ptr)
655                                 block->partial++;
656                         if (ptr & LOGFS_FULLY_POPULATED)
657                                 block->full++;
658                 }
659         }
660 }
661
662 static void alloc_data_block(struct inode *inode, struct page *page)
663 {
664         struct logfs_block *block;
665         u64 bix;
666         level_t level;
667
668         if (PagePrivate(page))
669                 return;
670
671         logfs_unpack_index(page->index, &bix, &level);
672         block = __alloc_block(inode->i_sb, inode->i_ino, bix, level);
673         block->page = page;
674         SetPagePrivate(page);
675         page->private = (unsigned long)block;
676         block->ops = &indirect_block_ops;
677 }
678
679 static void alloc_indirect_block(struct inode *inode, struct page *page,
680                 int page_is_empty)
681 {
682         struct logfs_block *block;
683         __be64 *array;
684
685         if (PagePrivate(page))
686                 return;
687
688         alloc_data_block(inode, page);
689
690         block = logfs_block(page);
691         array = kmap_atomic(page, KM_USER0);
692         initialize_block_counters(page, block, array, page_is_empty);
693         kunmap_atomic(array, KM_USER0);
694 }
695
696 static void block_set_pointer(struct page *page, int index, u64 ptr)
697 {
698         struct logfs_block *block = logfs_block(page);
699         __be64 *array;
700         u64 oldptr;
701
702         BUG_ON(!block);
703         array = kmap_atomic(page, KM_USER0);
704         oldptr = be64_to_cpu(array[index]);
705         array[index] = cpu_to_be64(ptr);
706         kunmap_atomic(array, KM_USER0);
707         SetPageUptodate(page);
708
709         block->full += !!(ptr & LOGFS_FULLY_POPULATED)
710                 - !!(oldptr & LOGFS_FULLY_POPULATED);
711         block->partial += !!ptr - !!oldptr;
712 }
713
714 static u64 block_get_pointer(struct page *page, int index)
715 {
716         __be64 *block;
717         u64 ptr;
718
719         block = kmap_atomic(page, KM_USER0);
720         ptr = be64_to_cpu(block[index]);
721         kunmap_atomic(block, KM_USER0);
722         return ptr;
723 }
724
725 static int logfs_read_empty(struct page *page)
726 {
727         zero_user_segment(page, 0, PAGE_CACHE_SIZE);
728         return 0;
729 }
730
731 static int logfs_read_direct(struct inode *inode, struct page *page)
732 {
733         struct logfs_inode *li = logfs_inode(inode);
734         pgoff_t index = page->index;
735         u64 block;
736
737         block = li->li_data[index];
738         if (!block)
739                 return logfs_read_empty(page);
740
741         return logfs_segment_read(inode, page, block, index, 0);
742 }
743
744 static int logfs_read_loop(struct inode *inode, struct page *page,
745                 int rw_context)
746 {
747         struct logfs_inode *li = logfs_inode(inode);
748         u64 bix, bofs = li->li_data[INDIRECT_INDEX];
749         level_t level, target_level;
750         int ret;
751         struct page *ipage;
752
753         logfs_unpack_index(page->index, &bix, &target_level);
754         if (!bofs)
755                 return logfs_read_empty(page);
756
757         if (bix >= maxbix(li->li_height))
758                 return logfs_read_empty(page);
759
760         for (level = LEVEL(li->li_height);
761                         (__force u8)level > (__force u8)target_level;
762                         level = SUBLEVEL(level)){
763                 ipage = logfs_get_page(inode, bix, level, rw_context);
764                 if (!ipage)
765                         return -ENOMEM;
766
767                 ret = logfs_segment_read(inode, ipage, bofs, bix, level);
768                 if (ret) {
769                         logfs_put_read_page(ipage);
770                         return ret;
771                 }
772
773                 bofs = block_get_pointer(ipage, get_bits(bix, SUBLEVEL(level)));
774                 logfs_put_page(ipage, rw_context);
775                 if (!bofs)
776                         return logfs_read_empty(page);
777         }
778
779         return logfs_segment_read(inode, page, bofs, bix, 0);
780 }
781
782 static int logfs_read_block(struct inode *inode, struct page *page,
783                 int rw_context)
784 {
785         pgoff_t index = page->index;
786
787         if (index < I0_BLOCKS)
788                 return logfs_read_direct(inode, page);
789         return logfs_read_loop(inode, page, rw_context);
790 }
791
792 static int logfs_exist_loop(struct inode *inode, u64 bix)
793 {
794         struct logfs_inode *li = logfs_inode(inode);
795         u64 bofs = li->li_data[INDIRECT_INDEX];
796         level_t level;
797         int ret;
798         struct page *ipage;
799
800         if (!bofs)
801                 return 0;
802         if (bix >= maxbix(li->li_height))
803                 return 0;
804
805         for (level = LEVEL(li->li_height); level != 0; level = SUBLEVEL(level)) {
806                 ipage = logfs_get_read_page(inode, bix, level);
807                 if (!ipage)
808                         return -ENOMEM;
809
810                 ret = logfs_segment_read(inode, ipage, bofs, bix, level);
811                 if (ret) {
812                         logfs_put_read_page(ipage);
813                         return ret;
814                 }
815
816                 bofs = block_get_pointer(ipage, get_bits(bix, SUBLEVEL(level)));
817                 logfs_put_read_page(ipage);
818                 if (!bofs)
819                         return 0;
820         }
821
822         return 1;
823 }
824
825 int logfs_exist_block(struct inode *inode, u64 bix)
826 {
827         struct logfs_inode *li = logfs_inode(inode);
828
829         if (bix < I0_BLOCKS)
830                 return !!li->li_data[bix];
831         return logfs_exist_loop(inode, bix);
832 }
833
834 static u64 seek_holedata_direct(struct inode *inode, u64 bix, int data)
835 {
836         struct logfs_inode *li = logfs_inode(inode);
837
838         for (; bix < I0_BLOCKS; bix++)
839                 if (data ^ (li->li_data[bix] == 0))
840                         return bix;
841         return I0_BLOCKS;
842 }
843
844 static u64 seek_holedata_loop(struct inode *inode, u64 bix, int data)
845 {
846         struct logfs_inode *li = logfs_inode(inode);
847         __be64 *rblock;
848         u64 increment, bofs = li->li_data[INDIRECT_INDEX];
849         level_t level;
850         int ret, slot;
851         struct page *page;
852
853         BUG_ON(!bofs);
854
855         for (level = LEVEL(li->li_height); level != 0; level = SUBLEVEL(level)) {
856                 increment = 1 << (LOGFS_BLOCK_BITS * ((__force u8)level-1));
857                 page = logfs_get_read_page(inode, bix, level);
858                 if (!page)
859                         return bix;
860
861                 ret = logfs_segment_read(inode, page, bofs, bix, level);
862                 if (ret) {
863                         logfs_put_read_page(page);
864                         return bix;
865                 }
866
867                 slot = get_bits(bix, SUBLEVEL(level));
868                 rblock = kmap_atomic(page, KM_USER0);
869                 while (slot < LOGFS_BLOCK_FACTOR) {
870                         if (data && (rblock[slot] != 0))
871                                 break;
872                         if (!data && !(be64_to_cpu(rblock[slot]) & LOGFS_FULLY_POPULATED))
873                                 break;
874                         slot++;
875                         bix += increment;
876                         bix &= ~(increment - 1);
877                 }
878                 if (slot >= LOGFS_BLOCK_FACTOR) {
879                         kunmap_atomic(rblock, KM_USER0);
880                         logfs_put_read_page(page);
881                         return bix;
882                 }
883                 bofs = be64_to_cpu(rblock[slot]);
884                 kunmap_atomic(rblock, KM_USER0);
885                 logfs_put_read_page(page);
886                 if (!bofs) {
887                         BUG_ON(data);
888                         return bix;
889                 }
890         }
891         return bix;
892 }
893
894 /**
895  * logfs_seek_hole - find next hole starting at a given block index
896  * @inode:              inode to search in
897  * @bix:                block index to start searching
898  *
899  * Returns next hole.  If the file doesn't contain any further holes, the
900  * block address next to eof is returned instead.
901  */
902 u64 logfs_seek_hole(struct inode *inode, u64 bix)
903 {
904         struct logfs_inode *li = logfs_inode(inode);
905
906         if (bix < I0_BLOCKS) {
907                 bix = seek_holedata_direct(inode, bix, 0);
908                 if (bix < I0_BLOCKS)
909                         return bix;
910         }
911
912         if (!li->li_data[INDIRECT_INDEX])
913                 return bix;
914         else if (li->li_data[INDIRECT_INDEX] & LOGFS_FULLY_POPULATED)
915                 bix = maxbix(li->li_height);
916         else {
917                 bix = seek_holedata_loop(inode, bix, 0);
918                 if (bix < maxbix(li->li_height))
919                         return bix;
920                 /* Should not happen anymore.  But if some port writes semi-
921                  * corrupt images (as this one used to) we might run into it.
922                  */
923                 WARN_ON_ONCE(bix == maxbix(li->li_height));
924         }
925
926         return bix;
927 }
928
929 static u64 __logfs_seek_data(struct inode *inode, u64 bix)
930 {
931         struct logfs_inode *li = logfs_inode(inode);
932
933         if (bix < I0_BLOCKS) {
934                 bix = seek_holedata_direct(inode, bix, 1);
935                 if (bix < I0_BLOCKS)
936                         return bix;
937         }
938
939         if (bix < maxbix(li->li_height)) {
940                 if (!li->li_data[INDIRECT_INDEX])
941                         bix = maxbix(li->li_height);
942                 else
943                         return seek_holedata_loop(inode, bix, 1);
944         }
945
946         return bix;
947 }
948
949 /**
950  * logfs_seek_data - find next data block after a given block index
951  * @inode:              inode to search in
952  * @bix:                block index to start searching
953  *
954  * Returns next data block.  If the file doesn't contain any further data
955  * blocks, the last block in the file is returned instead.
956  */
957 u64 logfs_seek_data(struct inode *inode, u64 bix)
958 {
959         struct super_block *sb = inode->i_sb;
960         u64 ret, end;
961
962         ret = __logfs_seek_data(inode, bix);
963         end = i_size_read(inode) >> sb->s_blocksize_bits;
964         if (ret >= end)
965                 ret = max(bix, end);
966         return ret;
967 }
968
969 static int logfs_is_valid_direct(struct logfs_inode *li, u64 bix, u64 ofs)
970 {
971         return pure_ofs(li->li_data[bix]) == ofs;
972 }
973
974 static int __logfs_is_valid_loop(struct inode *inode, u64 bix,
975                 u64 ofs, u64 bofs)
976 {
977         struct logfs_inode *li = logfs_inode(inode);
978         level_t level;
979         int ret;
980         struct page *page;
981
982         for (level = LEVEL(li->li_height); level != 0; level = SUBLEVEL(level)){
983                 page = logfs_get_write_page(inode, bix, level);
984                 BUG_ON(!page);
985
986                 ret = logfs_segment_read(inode, page, bofs, bix, level);
987                 if (ret) {
988                         logfs_put_write_page(page);
989                         return 0;
990                 }
991
992                 bofs = block_get_pointer(page, get_bits(bix, SUBLEVEL(level)));
993                 logfs_put_write_page(page);
994                 if (!bofs)
995                         return 0;
996
997                 if (pure_ofs(bofs) == ofs)
998                         return 1;
999         }
1000         return 0;
1001 }
1002
1003 static int logfs_is_valid_loop(struct inode *inode, u64 bix, u64 ofs)
1004 {
1005         struct logfs_inode *li = logfs_inode(inode);
1006         u64 bofs = li->li_data[INDIRECT_INDEX];
1007
1008         if (!bofs)
1009                 return 0;
1010
1011         if (bix >= maxbix(li->li_height))
1012                 return 0;
1013
1014         if (pure_ofs(bofs) == ofs)
1015                 return 1;
1016
1017         return __logfs_is_valid_loop(inode, bix, ofs, bofs);
1018 }
1019
1020 static int __logfs_is_valid_block(struct inode *inode, u64 bix, u64 ofs)
1021 {
1022         struct logfs_inode *li = logfs_inode(inode);
1023
1024         if ((inode->i_nlink == 0) && atomic_read(&inode->i_count) == 1)
1025                 return 0;
1026
1027         if (bix < I0_BLOCKS)
1028                 return logfs_is_valid_direct(li, bix, ofs);
1029         return logfs_is_valid_loop(inode, bix, ofs);
1030 }
1031
1032 /**
1033  * logfs_is_valid_block - check whether this block is still valid
1034  *
1035  * @sb  - superblock
1036  * @ofs - block physical offset
1037  * @ino - block inode number
1038  * @bix - block index
1039  * @level - block level
1040  *
1041  * Returns 0 if the block is invalid, 1 if it is valid and 2 if it will
1042  * become invalid once the journal is written.
1043  */
1044 int logfs_is_valid_block(struct super_block *sb, u64 ofs, u64 ino, u64 bix,
1045                 gc_level_t gc_level)
1046 {
1047         struct logfs_super *super = logfs_super(sb);
1048         struct inode *inode;
1049         int ret, cookie;
1050
1051         /* Umount closes a segment with free blocks remaining.  Those
1052          * blocks are by definition invalid. */
1053         if (ino == -1)
1054                 return 0;
1055
1056         LOGFS_BUG_ON((u64)(u_long)ino != ino, sb);
1057
1058         inode = logfs_safe_iget(sb, ino, &cookie);
1059         if (IS_ERR(inode))
1060                 goto invalid;
1061
1062         ret = __logfs_is_valid_block(inode, bix, ofs);
1063         logfs_safe_iput(inode, cookie);
1064         if (ret)
1065                 return ret;
1066
1067 invalid:
1068         /* Block is nominally invalid, but may still sit in the shadow tree,
1069          * waiting for a journal commit.
1070          */
1071         if (btree_lookup64(&super->s_shadow_tree.old, ofs))
1072                 return 2;
1073         return 0;
1074 }
1075
1076 int logfs_readpage_nolock(struct page *page)
1077 {
1078         struct inode *inode = page->mapping->host;
1079         int ret = -EIO;
1080
1081         ret = logfs_read_block(inode, page, READ);
1082
1083         if (ret) {
1084                 ClearPageUptodate(page);
1085                 SetPageError(page);
1086         } else {
1087                 SetPageUptodate(page);
1088                 ClearPageError(page);
1089         }
1090         flush_dcache_page(page);
1091
1092         return ret;
1093 }
1094
1095 static int logfs_reserve_bytes(struct inode *inode, int bytes)
1096 {
1097         struct logfs_super *super = logfs_super(inode->i_sb);
1098         u64 available = super->s_free_bytes + super->s_dirty_free_bytes
1099                         - super->s_dirty_used_bytes - super->s_dirty_pages;
1100
1101         if (!bytes)
1102                 return 0;
1103
1104         if (available < bytes)
1105                 return -ENOSPC;
1106
1107         if (available < bytes + super->s_root_reserve &&
1108                         !capable(CAP_SYS_RESOURCE))
1109                 return -ENOSPC;
1110
1111         return 0;
1112 }
1113
1114 int get_page_reserve(struct inode *inode, struct page *page)
1115 {
1116         struct logfs_super *super = logfs_super(inode->i_sb);
1117         int ret;
1118
1119         if (logfs_block(page) && logfs_block(page)->reserved_bytes)
1120                 return 0;
1121
1122         logfs_get_wblocks(inode->i_sb, page, WF_LOCK);
1123         ret = logfs_reserve_bytes(inode, 6 * LOGFS_MAX_OBJECTSIZE);
1124         if (!ret) {
1125                 alloc_data_block(inode, page);
1126                 logfs_block(page)->reserved_bytes += 6 * LOGFS_MAX_OBJECTSIZE;
1127                 super->s_dirty_pages += 6 * LOGFS_MAX_OBJECTSIZE;
1128         }
1129         logfs_put_wblocks(inode->i_sb, page, WF_LOCK);
1130         return ret;
1131 }
1132
1133 /*
1134  * We are protected by write lock.  Push victims up to superblock level
1135  * and release transaction when appropriate.
1136  */
1137 /* FIXME: This is currently called from the wrong spots. */
1138 static void logfs_handle_transaction(struct inode *inode,
1139                 struct logfs_transaction *ta)
1140 {
1141         struct logfs_super *super = logfs_super(inode->i_sb);
1142
1143         if (!ta)
1144                 return;
1145         logfs_inode(inode)->li_block->ta = NULL;
1146
1147         if (inode->i_ino != LOGFS_INO_MASTER) {
1148                 BUG(); /* FIXME: Yes, this needs more thought */
1149                 /* just remember the transaction until inode is written */
1150                 //BUG_ON(logfs_inode(inode)->li_transaction);
1151                 //logfs_inode(inode)->li_transaction = ta;
1152                 return;
1153         }
1154
1155         switch (ta->state) {
1156         case CREATE_1: /* fall through */
1157         case UNLINK_1:
1158                 BUG_ON(super->s_victim_ino);
1159                 super->s_victim_ino = ta->ino;
1160                 break;
1161         case CREATE_2: /* fall through */
1162         case UNLINK_2:
1163                 BUG_ON(super->s_victim_ino != ta->ino);
1164                 super->s_victim_ino = 0;
1165                 /* transaction ends here - free it */
1166                 kfree(ta);
1167                 break;
1168         case CROSS_RENAME_1:
1169                 BUG_ON(super->s_rename_dir);
1170                 BUG_ON(super->s_rename_pos);
1171                 super->s_rename_dir = ta->dir;
1172                 super->s_rename_pos = ta->pos;
1173                 break;
1174         case CROSS_RENAME_2:
1175                 BUG_ON(super->s_rename_dir != ta->dir);
1176                 BUG_ON(super->s_rename_pos != ta->pos);
1177                 super->s_rename_dir = 0;
1178                 super->s_rename_pos = 0;
1179                 kfree(ta);
1180                 break;
1181         case TARGET_RENAME_1:
1182                 BUG_ON(super->s_rename_dir);
1183                 BUG_ON(super->s_rename_pos);
1184                 BUG_ON(super->s_victim_ino);
1185                 super->s_rename_dir = ta->dir;
1186                 super->s_rename_pos = ta->pos;
1187                 super->s_victim_ino = ta->ino;
1188                 break;
1189         case TARGET_RENAME_2:
1190                 BUG_ON(super->s_rename_dir != ta->dir);
1191                 BUG_ON(super->s_rename_pos != ta->pos);
1192                 BUG_ON(super->s_victim_ino != ta->ino);
1193                 super->s_rename_dir = 0;
1194                 super->s_rename_pos = 0;
1195                 break;
1196         case TARGET_RENAME_3:
1197                 BUG_ON(super->s_rename_dir);
1198                 BUG_ON(super->s_rename_pos);
1199                 BUG_ON(super->s_victim_ino != ta->ino);
1200                 super->s_victim_ino = 0;
1201                 kfree(ta);
1202                 break;
1203         default:
1204                 BUG();
1205         }
1206 }
1207
1208 /*
1209  * Not strictly a reservation, but rather a check that we still have enough
1210  * space to satisfy the write.
1211  */
1212 static int logfs_reserve_blocks(struct inode *inode, int blocks)
1213 {
1214         return logfs_reserve_bytes(inode, blocks * LOGFS_MAX_OBJECTSIZE);
1215 }
1216
1217 struct write_control {
1218         u64 ofs;
1219         long flags;
1220 };
1221
1222 static struct logfs_shadow *alloc_shadow(struct inode *inode, u64 bix,
1223                 level_t level, u64 old_ofs)
1224 {
1225         struct logfs_super *super = logfs_super(inode->i_sb);
1226         struct logfs_shadow *shadow;
1227
1228         shadow = mempool_alloc(super->s_shadow_pool, GFP_NOFS);
1229         memset(shadow, 0, sizeof(*shadow));
1230         shadow->ino = inode->i_ino;
1231         shadow->bix = bix;
1232         shadow->gc_level = expand_level(inode->i_ino, level);
1233         shadow->old_ofs = old_ofs & ~LOGFS_FULLY_POPULATED;
1234         return shadow;
1235 }
1236
1237 static void free_shadow(struct inode *inode, struct logfs_shadow *shadow)
1238 {
1239         struct logfs_super *super = logfs_super(inode->i_sb);
1240
1241         mempool_free(shadow, super->s_shadow_pool);
1242 }
1243
1244 /**
1245  * fill_shadow_tree - Propagate shadow tree changes due to a write
1246  * @inode:      Inode owning the page
1247  * @page:       Struct page that was written
1248  * @shadow:     Shadow for the current write
1249  *
1250  * Writes in logfs can result in two semi-valid objects.  The old object
1251  * is still valid as long as it can be reached by following pointers on
1252  * the medium.  Only when writes propagate all the way up to the journal
1253  * has the new object safely replaced the old one.
1254  *
1255  * To handle this problem, a struct logfs_shadow is used to represent
1256  * every single write.  It is attached to the indirect block, which is
1257  * marked dirty.  When the indirect block is written, its shadows are
1258  * handed up to the next indirect block (or inode).  Untimately they
1259  * will reach the master inode and be freed upon journal commit.
1260  *
1261  * This function handles a single step in the propagation.  It adds the
1262  * shadow for the current write to the tree, along with any shadows in
1263  * the page's tree, in case it was an indirect block.  If a page is
1264  * written, the inode parameter is left NULL, if an inode is written,
1265  * the page parameter is left NULL.
1266  */
1267 static void fill_shadow_tree(struct inode *inode, struct page *page,
1268                 struct logfs_shadow *shadow)
1269 {
1270         struct logfs_super *super = logfs_super(inode->i_sb);
1271         struct logfs_block *block = logfs_block(page);
1272         struct shadow_tree *tree = &super->s_shadow_tree;
1273
1274         if (PagePrivate(page)) {
1275                 if (block->alias_map)
1276                         super->s_no_object_aliases -= bitmap_weight(
1277                                         block->alias_map, LOGFS_BLOCK_FACTOR);
1278                 logfs_handle_transaction(inode, block->ta);
1279                 block->ops->free_block(inode->i_sb, block);
1280         }
1281         if (shadow) {
1282                 if (shadow->old_ofs)
1283                         btree_insert64(&tree->old, shadow->old_ofs, shadow,
1284                                         GFP_NOFS);
1285                 else
1286                         btree_insert64(&tree->new, shadow->new_ofs, shadow,
1287                                         GFP_NOFS);
1288
1289                 super->s_dirty_used_bytes += shadow->new_len;
1290                 super->s_dirty_free_bytes += shadow->old_len;
1291         }
1292 }
1293
1294 static void logfs_set_alias(struct super_block *sb, struct logfs_block *block,
1295                 long child_no)
1296 {
1297         struct logfs_super *super = logfs_super(sb);
1298
1299         if (block->inode && block->inode->i_ino == LOGFS_INO_MASTER) {
1300                 /* Aliases in the master inode are pointless. */
1301                 return;
1302         }
1303
1304         if (!test_bit(child_no, block->alias_map)) {
1305                 set_bit(child_no, block->alias_map);
1306                 super->s_no_object_aliases++;
1307         }
1308         list_move_tail(&block->alias_list, &super->s_object_alias);
1309 }
1310
1311 /*
1312  * Object aliases can and often do change the size and occupied space of a
1313  * file.  So not only do we have to change the pointers, we also have to
1314  * change inode->i_size and li->li_used_bytes.  Which is done by setting
1315  * another two object aliases for the inode itself.
1316  */
1317 static void set_iused(struct inode *inode, struct logfs_shadow *shadow)
1318 {
1319         struct logfs_inode *li = logfs_inode(inode);
1320
1321         if (shadow->new_len == shadow->old_len)
1322                 return;
1323
1324         alloc_inode_block(inode);
1325         li->li_used_bytes += shadow->new_len - shadow->old_len;
1326         __logfs_set_blocks(inode);
1327         logfs_set_alias(inode->i_sb, li->li_block, INODE_USED_OFS);
1328         logfs_set_alias(inode->i_sb, li->li_block, INODE_SIZE_OFS);
1329 }
1330
1331 static int logfs_write_i0(struct inode *inode, struct page *page,
1332                 struct write_control *wc)
1333 {
1334         struct logfs_shadow *shadow;
1335         u64 bix;
1336         level_t level;
1337         int full, err = 0;
1338
1339         logfs_unpack_index(page->index, &bix, &level);
1340         if (wc->ofs == 0)
1341                 if (logfs_reserve_blocks(inode, 1))
1342                         return -ENOSPC;
1343
1344         shadow = alloc_shadow(inode, bix, level, wc->ofs);
1345         if (wc->flags & WF_WRITE)
1346                 err = logfs_segment_write(inode, page, shadow);
1347         if (wc->flags & WF_DELETE)
1348                 logfs_segment_delete(inode, shadow);
1349         if (err) {
1350                 free_shadow(inode, shadow);
1351                 return err;
1352         }
1353
1354         set_iused(inode, shadow);
1355         full = 1;
1356         if (level != 0) {
1357                 alloc_indirect_block(inode, page, 0);
1358                 full = logfs_block(page)->full == LOGFS_BLOCK_FACTOR;
1359         }
1360         fill_shadow_tree(inode, page, shadow);
1361         wc->ofs = shadow->new_ofs;
1362         if (wc->ofs && full)
1363                 wc->ofs |= LOGFS_FULLY_POPULATED;
1364         return 0;
1365 }
1366
1367 static int logfs_write_direct(struct inode *inode, struct page *page,
1368                 long flags)
1369 {
1370         struct logfs_inode *li = logfs_inode(inode);
1371         struct write_control wc = {
1372                 .ofs = li->li_data[page->index],
1373                 .flags = flags,
1374         };
1375         int err;
1376
1377         alloc_inode_block(inode);
1378
1379         err = logfs_write_i0(inode, page, &wc);
1380         if (err)
1381                 return err;
1382
1383         li->li_data[page->index] = wc.ofs;
1384         logfs_set_alias(inode->i_sb, li->li_block,
1385                         page->index + INODE_POINTER_OFS);
1386         return 0;
1387 }
1388
1389 static int ptr_change(u64 ofs, struct page *page)
1390 {
1391         struct logfs_block *block = logfs_block(page);
1392         int empty0, empty1, full0, full1;
1393
1394         empty0 = ofs == 0;
1395         empty1 = block->partial == 0;
1396         if (empty0 != empty1)
1397                 return 1;
1398
1399         /* The !! is necessary to shrink result to int */
1400         full0 = !!(ofs & LOGFS_FULLY_POPULATED);
1401         full1 = block->full == LOGFS_BLOCK_FACTOR;
1402         if (full0 != full1)
1403                 return 1;
1404         return 0;
1405 }
1406
1407 static int __logfs_write_rec(struct inode *inode, struct page *page,
1408                 struct write_control *this_wc,
1409                 pgoff_t bix, level_t target_level, level_t level)
1410 {
1411         int ret, page_empty = 0;
1412         int child_no = get_bits(bix, SUBLEVEL(level));
1413         struct page *ipage;
1414         struct write_control child_wc = {
1415                 .flags = this_wc->flags,
1416         };
1417
1418         ipage = logfs_get_write_page(inode, bix, level);
1419         if (!ipage)
1420                 return -ENOMEM;
1421
1422         if (this_wc->ofs) {
1423                 ret = logfs_segment_read(inode, ipage, this_wc->ofs, bix, level);
1424                 if (ret)
1425                         goto out;
1426         } else if (!PageUptodate(ipage)) {
1427                 page_empty = 1;
1428                 logfs_read_empty(ipage);
1429         }
1430
1431         child_wc.ofs = block_get_pointer(ipage, child_no);
1432
1433         if ((__force u8)level-1 > (__force u8)target_level)
1434                 ret = __logfs_write_rec(inode, page, &child_wc, bix,
1435                                 target_level, SUBLEVEL(level));
1436         else
1437                 ret = logfs_write_i0(inode, page, &child_wc);
1438
1439         if (ret)
1440                 goto out;
1441
1442         alloc_indirect_block(inode, ipage, page_empty);
1443         block_set_pointer(ipage, child_no, child_wc.ofs);
1444         /* FIXME: first condition seems superfluous */
1445         if (child_wc.ofs || logfs_block(ipage)->partial)
1446                 this_wc->flags |= WF_WRITE;
1447         /* the condition on this_wc->ofs ensures that we won't consume extra
1448          * space for indirect blocks in the future, which we cannot reserve */
1449         if (!this_wc->ofs || ptr_change(this_wc->ofs, ipage))
1450                 ret = logfs_write_i0(inode, ipage, this_wc);
1451         else
1452                 logfs_set_alias(inode->i_sb, logfs_block(ipage), child_no);
1453 out:
1454         logfs_put_write_page(ipage);
1455         return ret;
1456 }
1457
1458 static int logfs_write_rec(struct inode *inode, struct page *page,
1459                 pgoff_t bix, level_t target_level, long flags)
1460 {
1461         struct logfs_inode *li = logfs_inode(inode);
1462         struct write_control wc = {
1463                 .ofs = li->li_data[INDIRECT_INDEX],
1464                 .flags = flags,
1465         };
1466         int ret;
1467
1468         alloc_inode_block(inode);
1469
1470         if (li->li_height > (__force u8)target_level)
1471                 ret = __logfs_write_rec(inode, page, &wc, bix, target_level,
1472                                 LEVEL(li->li_height));
1473         else
1474                 ret = logfs_write_i0(inode, page, &wc);
1475         if (ret)
1476                 return ret;
1477
1478         if (li->li_data[INDIRECT_INDEX] != wc.ofs) {
1479                 li->li_data[INDIRECT_INDEX] = wc.ofs;
1480                 logfs_set_alias(inode->i_sb, li->li_block,
1481                                 INDIRECT_INDEX + INODE_POINTER_OFS);
1482         }
1483         return ret;
1484 }
1485
1486 void logfs_add_transaction(struct inode *inode, struct logfs_transaction *ta)
1487 {
1488         alloc_inode_block(inode);
1489         logfs_inode(inode)->li_block->ta = ta;
1490 }
1491
1492 void logfs_del_transaction(struct inode *inode, struct logfs_transaction *ta)
1493 {
1494         struct logfs_block *block = logfs_inode(inode)->li_block;
1495
1496         if (block && block->ta)
1497                 block->ta = NULL;
1498 }
1499
1500 static int grow_inode(struct inode *inode, u64 bix, level_t level)
1501 {
1502         struct logfs_inode *li = logfs_inode(inode);
1503         u8 height = (__force u8)level;
1504         struct page *page;
1505         struct write_control wc = {
1506                 .flags = WF_WRITE,
1507         };
1508         int err;
1509
1510         BUG_ON(height > 5 || li->li_height > 5);
1511         while (height > li->li_height || bix >= maxbix(li->li_height)) {
1512                 page = logfs_get_write_page(inode, I0_BLOCKS + 1,
1513                                 LEVEL(li->li_height + 1));
1514                 if (!page)
1515                         return -ENOMEM;
1516                 logfs_read_empty(page);
1517                 alloc_indirect_block(inode, page, 1);
1518                 block_set_pointer(page, 0, li->li_data[INDIRECT_INDEX]);
1519                 err = logfs_write_i0(inode, page, &wc);
1520                 logfs_put_write_page(page);
1521                 if (err)
1522                         return err;
1523                 li->li_data[INDIRECT_INDEX] = wc.ofs;
1524                 wc.ofs = 0;
1525                 li->li_height++;
1526                 logfs_set_alias(inode->i_sb, li->li_block, INODE_HEIGHT_OFS);
1527         }
1528         return 0;
1529 }
1530
1531 static int __logfs_write_buf(struct inode *inode, struct page *page, long flags)
1532 {
1533         struct logfs_super *super = logfs_super(inode->i_sb);
1534         pgoff_t index = page->index;
1535         u64 bix;
1536         level_t level;
1537         int err;
1538
1539         flags |= WF_WRITE | WF_DELETE;
1540         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1541
1542         logfs_unpack_index(index, &bix, &level);
1543         if (logfs_block(page) && logfs_block(page)->reserved_bytes)
1544                 super->s_dirty_pages -= logfs_block(page)->reserved_bytes;
1545
1546         if (index < I0_BLOCKS)
1547                 return logfs_write_direct(inode, page, flags);
1548
1549         bix = adjust_bix(bix, level);
1550         err = grow_inode(inode, bix, level);
1551         if (err)
1552                 return err;
1553         return logfs_write_rec(inode, page, bix, level, flags);
1554 }
1555
1556 int logfs_write_buf(struct inode *inode, struct page *page, long flags)
1557 {
1558         struct super_block *sb = inode->i_sb;
1559         int ret;
1560
1561         logfs_get_wblocks(sb, page, flags & WF_LOCK);
1562         ret = __logfs_write_buf(inode, page, flags);
1563         logfs_put_wblocks(sb, page, flags & WF_LOCK);
1564         return ret;
1565 }
1566
1567 static int __logfs_delete(struct inode *inode, struct page *page)
1568 {
1569         long flags = WF_DELETE;
1570
1571         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1572
1573         if (page->index < I0_BLOCKS)
1574                 return logfs_write_direct(inode, page, flags);
1575         return logfs_write_rec(inode, page, page->index, 0, flags);
1576 }
1577
1578 int logfs_delete(struct inode *inode, pgoff_t index,
1579                 struct shadow_tree *shadow_tree)
1580 {
1581         struct super_block *sb = inode->i_sb;
1582         struct page *page;
1583         int ret;
1584
1585         page = logfs_get_read_page(inode, index, 0);
1586         if (!page)
1587                 return -ENOMEM;
1588
1589         logfs_get_wblocks(sb, page, 1);
1590         ret = __logfs_delete(inode, page);
1591         logfs_put_wblocks(sb, page, 1);
1592
1593         logfs_put_read_page(page);
1594
1595         return ret;
1596 }
1597
1598 int logfs_rewrite_block(struct inode *inode, u64 bix, u64 ofs,
1599                 gc_level_t gc_level, long flags)
1600 {
1601         level_t level = shrink_level(gc_level);
1602         struct page *page;
1603         int err;
1604
1605         page = logfs_get_write_page(inode, bix, level);
1606         if (!page)
1607                 return -ENOMEM;
1608
1609         err = logfs_segment_read(inode, page, ofs, bix, level);
1610         if (!err) {
1611                 if (level != 0)
1612                         alloc_indirect_block(inode, page, 0);
1613                 err = logfs_write_buf(inode, page, flags);
1614                 if (!err && shrink_level(gc_level) == 0) {
1615                         /* Rewrite cannot mark the inode dirty but has to
1616                          * write it immediatly.
1617                          * Q: Can't we just create an alias for the inode
1618                          * instead?  And if not, why not?
1619                          */
1620                         if (inode->i_ino == LOGFS_INO_MASTER)
1621                                 logfs_write_anchor(inode->i_sb);
1622                         else {
1623                                 err = __logfs_write_inode(inode, flags);
1624                         }
1625                 }
1626         }
1627         logfs_put_write_page(page);
1628         return err;
1629 }
1630
1631 static int truncate_data_block(struct inode *inode, struct page *page,
1632                 u64 ofs, struct logfs_shadow *shadow, u64 size)
1633 {
1634         loff_t pageofs = page->index << inode->i_sb->s_blocksize_bits;
1635         u64 bix;
1636         level_t level;
1637         int err;
1638
1639         /* Does truncation happen within this page? */
1640         if (size <= pageofs || size - pageofs >= PAGE_SIZE)
1641                 return 0;
1642
1643         logfs_unpack_index(page->index, &bix, &level);
1644         BUG_ON(level != 0);
1645
1646         err = logfs_segment_read(inode, page, ofs, bix, level);
1647         if (err)
1648                 return err;
1649
1650         zero_user_segment(page, size - pageofs, PAGE_CACHE_SIZE);
1651         return logfs_segment_write(inode, page, shadow);
1652 }
1653
1654 static int logfs_truncate_i0(struct inode *inode, struct page *page,
1655                 struct write_control *wc, u64 size)
1656 {
1657         struct logfs_shadow *shadow;
1658         u64 bix;
1659         level_t level;
1660         int err = 0;
1661
1662         logfs_unpack_index(page->index, &bix, &level);
1663         BUG_ON(level != 0);
1664         shadow = alloc_shadow(inode, bix, level, wc->ofs);
1665
1666         err = truncate_data_block(inode, page, wc->ofs, shadow, size);
1667         if (err) {
1668                 free_shadow(inode, shadow);
1669                 return err;
1670         }
1671
1672         logfs_segment_delete(inode, shadow);
1673         set_iused(inode, shadow);
1674         fill_shadow_tree(inode, page, shadow);
1675         wc->ofs = shadow->new_ofs;
1676         return 0;
1677 }
1678
1679 static int logfs_truncate_direct(struct inode *inode, u64 size)
1680 {
1681         struct logfs_inode *li = logfs_inode(inode);
1682         struct write_control wc;
1683         struct page *page;
1684         int e;
1685         int err;
1686
1687         alloc_inode_block(inode);
1688
1689         for (e = I0_BLOCKS - 1; e >= 0; e--) {
1690                 if (size > (e+1) * LOGFS_BLOCKSIZE)
1691                         break;
1692
1693                 wc.ofs = li->li_data[e];
1694                 if (!wc.ofs)
1695                         continue;
1696
1697                 page = logfs_get_write_page(inode, e, 0);
1698                 if (!page)
1699                         return -ENOMEM;
1700                 err = logfs_segment_read(inode, page, wc.ofs, e, 0);
1701                 if (err) {
1702                         logfs_put_write_page(page);
1703                         return err;
1704                 }
1705                 err = logfs_truncate_i0(inode, page, &wc, size);
1706                 logfs_put_write_page(page);
1707                 if (err)
1708                         return err;
1709
1710                 li->li_data[e] = wc.ofs;
1711         }
1712         return 0;
1713 }
1714
1715 /* FIXME: these need to become per-sb once we support different blocksizes */
1716 static u64 __logfs_step[] = {
1717         1,
1718         I1_BLOCKS,
1719         I2_BLOCKS,
1720         I3_BLOCKS,
1721 };
1722
1723 static u64 __logfs_start_index[] = {
1724         I0_BLOCKS,
1725         I1_BLOCKS,
1726         I2_BLOCKS,
1727         I3_BLOCKS
1728 };
1729
1730 static inline u64 logfs_step(level_t level)
1731 {
1732         return __logfs_step[(__force u8)level];
1733 }
1734
1735 static inline u64 logfs_factor(u8 level)
1736 {
1737         return __logfs_step[level] * LOGFS_BLOCKSIZE;
1738 }
1739
1740 static inline u64 logfs_start_index(level_t level)
1741 {
1742         return __logfs_start_index[(__force u8)level];
1743 }
1744
1745 static void logfs_unpack_raw_index(pgoff_t index, u64 *bix, level_t *level)
1746 {
1747         logfs_unpack_index(index, bix, level);
1748         if (*bix <= logfs_start_index(SUBLEVEL(*level)))
1749                 *bix = 0;
1750 }
1751
1752 static int __logfs_truncate_rec(struct inode *inode, struct page *ipage,
1753                 struct write_control *this_wc, u64 size)
1754 {
1755         int truncate_happened = 0;
1756         int e, err = 0;
1757         u64 bix, child_bix, next_bix;
1758         level_t level;
1759         struct page *page;
1760         struct write_control child_wc = { /* FIXME: flags */ };
1761
1762         logfs_unpack_raw_index(ipage->index, &bix, &level);
1763         err = logfs_segment_read(inode, ipage, this_wc->ofs, bix, level);
1764         if (err)
1765                 return err;
1766
1767         for (e = LOGFS_BLOCK_FACTOR - 1; e >= 0; e--) {
1768                 child_bix = bix + e * logfs_step(SUBLEVEL(level));
1769                 next_bix = child_bix + logfs_step(SUBLEVEL(level));
1770                 if (size > next_bix * LOGFS_BLOCKSIZE)
1771                         break;
1772
1773                 child_wc.ofs = pure_ofs(block_get_pointer(ipage, e));
1774                 if (!child_wc.ofs)
1775                         continue;
1776
1777                 page = logfs_get_write_page(inode, child_bix, SUBLEVEL(level));
1778                 if (!page)
1779                         return -ENOMEM;
1780
1781                 if ((__force u8)level > 1)
1782                         err = __logfs_truncate_rec(inode, page, &child_wc, size);
1783                 else
1784                         err = logfs_truncate_i0(inode, page, &child_wc, size);
1785                 logfs_put_write_page(page);
1786                 if (err)
1787                         return err;
1788
1789                 truncate_happened = 1;
1790                 alloc_indirect_block(inode, ipage, 0);
1791                 block_set_pointer(ipage, e, child_wc.ofs);
1792         }
1793
1794         if (!truncate_happened) {
1795                 printk("ineffectual truncate (%lx, %lx, %llx)\n", inode->i_ino, ipage->index, size);
1796                 return 0;
1797         }
1798
1799         this_wc->flags = WF_DELETE;
1800         if (logfs_block(ipage)->partial)
1801                 this_wc->flags |= WF_WRITE;
1802
1803         return logfs_write_i0(inode, ipage, this_wc);
1804 }
1805
1806 static int logfs_truncate_rec(struct inode *inode, u64 size)
1807 {
1808         struct logfs_inode *li = logfs_inode(inode);
1809         struct write_control wc = {
1810                 .ofs = li->li_data[INDIRECT_INDEX],
1811         };
1812         struct page *page;
1813         int err;
1814
1815         alloc_inode_block(inode);
1816
1817         if (!wc.ofs)
1818                 return 0;
1819
1820         page = logfs_get_write_page(inode, 0, LEVEL(li->li_height));
1821         if (!page)
1822                 return -ENOMEM;
1823
1824         err = __logfs_truncate_rec(inode, page, &wc, size);
1825         logfs_put_write_page(page);
1826         if (err)
1827                 return err;
1828
1829         if (li->li_data[INDIRECT_INDEX] != wc.ofs)
1830                 li->li_data[INDIRECT_INDEX] = wc.ofs;
1831         return 0;
1832 }
1833
1834 static int __logfs_truncate(struct inode *inode, u64 size)
1835 {
1836         int ret;
1837
1838         if (size >= logfs_factor(logfs_inode(inode)->li_height))
1839                 return 0;
1840
1841         ret = logfs_truncate_rec(inode, size);
1842         if (ret)
1843                 return ret;
1844
1845         return logfs_truncate_direct(inode, size);
1846 }
1847
1848 int logfs_truncate(struct inode *inode, u64 size)
1849 {
1850         struct super_block *sb = inode->i_sb;
1851         int err;
1852
1853         logfs_get_wblocks(sb, NULL, 1);
1854         err = __logfs_truncate(inode, size);
1855         if (!err)
1856                 err = __logfs_write_inode(inode, 0);
1857         logfs_put_wblocks(sb, NULL, 1);
1858
1859         if (!err)
1860                 err = vmtruncate(inode, size);
1861
1862         /* I don't trust error recovery yet. */
1863         WARN_ON(err);
1864         return err;
1865 }
1866
1867 static void move_page_to_inode(struct inode *inode, struct page *page)
1868 {
1869         struct logfs_inode *li = logfs_inode(inode);
1870         struct logfs_block *block = logfs_block(page);
1871
1872         if (!block)
1873                 return;
1874
1875         log_blockmove("move_page_to_inode(%llx, %llx, %x)\n",
1876                         block->ino, block->bix, block->level);
1877         BUG_ON(li->li_block);
1878         block->ops = &inode_block_ops;
1879         block->inode = inode;
1880         li->li_block = block;
1881
1882         block->page = NULL;
1883         page->private = 0;
1884         ClearPagePrivate(page);
1885 }
1886
1887 static void move_inode_to_page(struct page *page, struct inode *inode)
1888 {
1889         struct logfs_inode *li = logfs_inode(inode);
1890         struct logfs_block *block = li->li_block;
1891
1892         if (!block)
1893                 return;
1894
1895         log_blockmove("move_inode_to_page(%llx, %llx, %x)\n",
1896                         block->ino, block->bix, block->level);
1897         BUG_ON(PagePrivate(page));
1898         block->ops = &indirect_block_ops;
1899         block->page = page;
1900         page->private = (unsigned long)block;
1901         SetPagePrivate(page);
1902
1903         block->inode = NULL;
1904         li->li_block = NULL;
1905 }
1906
1907 int logfs_read_inode(struct inode *inode)
1908 {
1909         struct super_block *sb = inode->i_sb;
1910         struct logfs_super *super = logfs_super(sb);
1911         struct inode *master_inode = super->s_master_inode;
1912         struct page *page;
1913         struct logfs_disk_inode *di;
1914         u64 ino = inode->i_ino;
1915
1916         if (ino << sb->s_blocksize_bits > i_size_read(master_inode))
1917                 return -ENODATA;
1918         if (!logfs_exist_block(master_inode, ino))
1919                 return -ENODATA;
1920
1921         page = read_cache_page(master_inode->i_mapping, ino,
1922                         (filler_t *)logfs_readpage, NULL);
1923         if (IS_ERR(page))
1924                 return PTR_ERR(page);
1925
1926         di = kmap_atomic(page, KM_USER0);
1927         logfs_disk_to_inode(di, inode);
1928         kunmap_atomic(di, KM_USER0);
1929         move_page_to_inode(inode, page);
1930         page_cache_release(page);
1931         return 0;
1932 }
1933
1934 /* Caller must logfs_put_write_page(page); */
1935 static struct page *inode_to_page(struct inode *inode)
1936 {
1937         struct inode *master_inode = logfs_super(inode->i_sb)->s_master_inode;
1938         struct logfs_disk_inode *di;
1939         struct page *page;
1940
1941         BUG_ON(inode->i_ino == LOGFS_INO_MASTER);
1942
1943         page = logfs_get_write_page(master_inode, inode->i_ino, 0);
1944         if (!page)
1945                 return NULL;
1946
1947         di = kmap_atomic(page, KM_USER0);
1948         logfs_inode_to_disk(inode, di);
1949         kunmap_atomic(di, KM_USER0);
1950         move_inode_to_page(page, inode);
1951         return page;
1952 }
1953
1954 /* Cheaper version of write_inode.  All changes are concealed in
1955  * aliases, which are moved back.  No write to the medium happens.
1956  */
1957 void logfs_clear_inode(struct inode *inode)
1958 {
1959         struct super_block *sb = inode->i_sb;
1960         struct logfs_inode *li = logfs_inode(inode);
1961         struct logfs_block *block = li->li_block;
1962         struct page *page;
1963
1964         /* Only deleted files may be dirty at this point */
1965         BUG_ON(inode->i_state & I_DIRTY && inode->i_nlink);
1966         if (!block)
1967                 return;
1968         if ((logfs_super(sb)->s_flags & LOGFS_SB_FLAG_SHUTDOWN)) {
1969                 block->ops->free_block(inode->i_sb, block);
1970                 return;
1971         }
1972
1973         BUG_ON(inode->i_ino < LOGFS_RESERVED_INOS);
1974         page = inode_to_page(inode);
1975         BUG_ON(!page); /* FIXME: Use emergency page */
1976         logfs_put_write_page(page);
1977 }
1978
1979 static int do_write_inode(struct inode *inode)
1980 {
1981         struct super_block *sb = inode->i_sb;
1982         struct inode *master_inode = logfs_super(sb)->s_master_inode;
1983         loff_t size = (inode->i_ino + 1) << inode->i_sb->s_blocksize_bits;
1984         struct page *page;
1985         int err;
1986
1987         BUG_ON(inode->i_ino == LOGFS_INO_MASTER);
1988         /* FIXME: lock inode */
1989
1990         if (i_size_read(master_inode) < size)
1991                 i_size_write(master_inode, size);
1992
1993         /* TODO: Tell vfs this inode is clean now */
1994
1995         page = inode_to_page(inode);
1996         if (!page)
1997                 return -ENOMEM;
1998
1999         /* FIXME: transaction is part of logfs_block now.  Is that enough? */
2000         err = logfs_write_buf(master_inode, page, 0);
2001         logfs_put_write_page(page);
2002         return err;
2003 }
2004
2005 static void logfs_mod_segment_entry(struct super_block *sb, u32 segno,
2006                 int write,
2007                 void (*change_se)(struct logfs_segment_entry *, long),
2008                 long arg)
2009 {
2010         struct logfs_super *super = logfs_super(sb);
2011         struct inode *inode;
2012         struct page *page;
2013         struct logfs_segment_entry *se;
2014         pgoff_t page_no;
2015         int child_no;
2016
2017         page_no = segno >> (sb->s_blocksize_bits - 3);
2018         child_no = segno & ((sb->s_blocksize >> 3) - 1);
2019
2020         inode = super->s_segfile_inode;
2021         page = logfs_get_write_page(inode, page_no, 0);
2022         BUG_ON(!page); /* FIXME: We need some reserve page for this case */
2023         if (!PageUptodate(page))
2024                 logfs_read_block(inode, page, WRITE);
2025
2026         if (write)
2027                 alloc_indirect_block(inode, page, 0);
2028         se = kmap_atomic(page, KM_USER0);
2029         change_se(se + child_no, arg);
2030         if (write) {
2031                 logfs_set_alias(sb, logfs_block(page), child_no);
2032                 BUG_ON((int)be32_to_cpu(se[child_no].valid) > super->s_segsize);
2033         }
2034         kunmap_atomic(se, KM_USER0);
2035
2036         logfs_put_write_page(page);
2037 }
2038
2039 static void __get_segment_entry(struct logfs_segment_entry *se, long _target)
2040 {
2041         struct logfs_segment_entry *target = (void *)_target;
2042
2043         *target = *se;
2044 }
2045
2046 void logfs_get_segment_entry(struct super_block *sb, u32 segno,
2047                 struct logfs_segment_entry *se)
2048 {
2049         logfs_mod_segment_entry(sb, segno, 0, __get_segment_entry, (long)se);
2050 }
2051
2052 static void __set_segment_used(struct logfs_segment_entry *se, long increment)
2053 {
2054         u32 valid;
2055
2056         valid = be32_to_cpu(se->valid);
2057         valid += increment;
2058         se->valid = cpu_to_be32(valid);
2059 }
2060
2061 void logfs_set_segment_used(struct super_block *sb, u64 ofs, int increment)
2062 {
2063         struct logfs_super *super = logfs_super(sb);
2064         u32 segno = ofs >> super->s_segshift;
2065
2066         if (!increment)
2067                 return;
2068
2069         logfs_mod_segment_entry(sb, segno, 1, __set_segment_used, increment);
2070 }
2071
2072 static void __set_segment_erased(struct logfs_segment_entry *se, long ec_level)
2073 {
2074         se->ec_level = cpu_to_be32(ec_level);
2075 }
2076
2077 void logfs_set_segment_erased(struct super_block *sb, u32 segno, u32 ec,
2078                 gc_level_t gc_level)
2079 {
2080         u32 ec_level = ec << 4 | (__force u8)gc_level;
2081
2082         logfs_mod_segment_entry(sb, segno, 1, __set_segment_erased, ec_level);
2083 }
2084
2085 static void __set_segment_reserved(struct logfs_segment_entry *se, long ignore)
2086 {
2087         se->valid = cpu_to_be32(RESERVED);
2088 }
2089
2090 void logfs_set_segment_reserved(struct super_block *sb, u32 segno)
2091 {
2092         logfs_mod_segment_entry(sb, segno, 1, __set_segment_reserved, 0);
2093 }
2094
2095 static void __set_segment_unreserved(struct logfs_segment_entry *se,
2096                 long ec_level)
2097 {
2098         se->valid = 0;
2099         se->ec_level = cpu_to_be32(ec_level);
2100 }
2101
2102 void logfs_set_segment_unreserved(struct super_block *sb, u32 segno, u32 ec)
2103 {
2104         u32 ec_level = ec << 4;
2105
2106         logfs_mod_segment_entry(sb, segno, 1, __set_segment_unreserved,
2107                         ec_level);
2108 }
2109
2110 int __logfs_write_inode(struct inode *inode, long flags)
2111 {
2112         struct super_block *sb = inode->i_sb;
2113         int ret;
2114
2115         logfs_get_wblocks(sb, NULL, flags & WF_LOCK);
2116         ret = do_write_inode(inode);
2117         logfs_put_wblocks(sb, NULL, flags & WF_LOCK);
2118         return ret;
2119 }
2120
2121 static int do_delete_inode(struct inode *inode)
2122 {
2123         struct super_block *sb = inode->i_sb;
2124         struct inode *master_inode = logfs_super(sb)->s_master_inode;
2125         struct page *page;
2126         int ret;
2127
2128         page = logfs_get_write_page(master_inode, inode->i_ino, 0);
2129         if (!page)
2130                 return -ENOMEM;
2131
2132         move_inode_to_page(page, inode);
2133
2134         logfs_get_wblocks(sb, page, 1);
2135         ret = __logfs_delete(master_inode, page);
2136         logfs_put_wblocks(sb, page, 1);
2137
2138         logfs_put_write_page(page);
2139         return ret;
2140 }
2141
2142 /*
2143  * ZOMBIE inodes have already been deleted before and should remain dead,
2144  * if it weren't for valid checking.  No need to kill them again here.
2145  */
2146 void logfs_delete_inode(struct inode *inode)
2147 {
2148         struct logfs_inode *li = logfs_inode(inode);
2149
2150         if (!(li->li_flags & LOGFS_IF_ZOMBIE)) {
2151                 li->li_flags |= LOGFS_IF_ZOMBIE;
2152                 if (i_size_read(inode) > 0)
2153                         logfs_truncate(inode, 0);
2154                 do_delete_inode(inode);
2155         }
2156         truncate_inode_pages(&inode->i_data, 0);
2157         clear_inode(inode);
2158 }
2159
2160 void btree_write_block(struct logfs_block *block)
2161 {
2162         struct inode *inode;
2163         struct page *page;
2164         int err, cookie;
2165
2166         inode = logfs_safe_iget(block->sb, block->ino, &cookie);
2167         page = logfs_get_write_page(inode, block->bix, block->level);
2168
2169         err = logfs_readpage_nolock(page);
2170         BUG_ON(err);
2171         BUG_ON(!PagePrivate(page));
2172         BUG_ON(logfs_block(page) != block);
2173         err = __logfs_write_buf(inode, page, 0);
2174         BUG_ON(err);
2175         BUG_ON(PagePrivate(page) || page->private);
2176
2177         logfs_put_write_page(page);
2178         logfs_safe_iput(inode, cookie);
2179 }
2180
2181 /**
2182  * logfs_inode_write - write inode or dentry objects
2183  *
2184  * @inode:              parent inode (ifile or directory)
2185  * @buf:                object to write (inode or dentry)
2186  * @n:                  object size
2187  * @_pos:               object number (file position in blocks/objects)
2188  * @flags:              write flags
2189  * @lock:               0 if write lock is already taken, 1 otherwise
2190  * @shadow_tree:        shadow below this inode
2191  *
2192  * FIXME: All caller of this put a 200-300 byte variable on the stack,
2193  * only to call here and do a memcpy from that stack variable.  A good
2194  * example of wasted performance and stack space.
2195  */
2196 int logfs_inode_write(struct inode *inode, const void *buf, size_t count,
2197                 loff_t bix, long flags, struct shadow_tree *shadow_tree)
2198 {
2199         loff_t pos = bix << inode->i_sb->s_blocksize_bits;
2200         int err;
2201         struct page *page;
2202         void *pagebuf;
2203
2204         BUG_ON(pos & (LOGFS_BLOCKSIZE-1));
2205         BUG_ON(count > LOGFS_BLOCKSIZE);
2206         page = logfs_get_write_page(inode, bix, 0);
2207         if (!page)
2208                 return -ENOMEM;
2209
2210         pagebuf = kmap_atomic(page, KM_USER0);
2211         memcpy(pagebuf, buf, count);
2212         flush_dcache_page(page);
2213         kunmap_atomic(pagebuf, KM_USER0);
2214
2215         if (i_size_read(inode) < pos + LOGFS_BLOCKSIZE)
2216                 i_size_write(inode, pos + LOGFS_BLOCKSIZE);
2217
2218         err = logfs_write_buf(inode, page, flags);
2219         logfs_put_write_page(page);
2220         return err;
2221 }
2222
2223 int logfs_open_segfile(struct super_block *sb)
2224 {
2225         struct logfs_super *super = logfs_super(sb);
2226         struct inode *inode;
2227
2228         inode = logfs_read_meta_inode(sb, LOGFS_INO_SEGFILE);
2229         if (IS_ERR(inode))
2230                 return PTR_ERR(inode);
2231         super->s_segfile_inode = inode;
2232         return 0;
2233 }
2234
2235 int logfs_init_rw(struct super_block *sb)
2236 {
2237         struct logfs_super *super = logfs_super(sb);
2238         int min_fill = 3 * super->s_no_blocks;
2239
2240         INIT_LIST_HEAD(&super->s_object_alias);
2241         mutex_init(&super->s_write_mutex);
2242         super->s_block_pool = mempool_create_kmalloc_pool(min_fill,
2243                         sizeof(struct logfs_block));
2244         super->s_shadow_pool = mempool_create_kmalloc_pool(min_fill,
2245                         sizeof(struct logfs_shadow));
2246         return 0;
2247 }
2248
2249 void logfs_cleanup_rw(struct super_block *sb)
2250 {
2251         struct logfs_super *super = logfs_super(sb);
2252
2253         destroy_meta_inode(super->s_segfile_inode);
2254         if (super->s_block_pool)
2255                 mempool_destroy(super->s_block_pool);
2256         if (super->s_shadow_pool)
2257                 mempool_destroy(super->s_shadow_pool);
2258 }