[LogFS] Prevent memory corruption on large deletes
[linux-2.6.git] / fs / logfs / journal.c
1 /*
2  * fs/logfs/journal.c   - journal handling code
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 #include "logfs.h"
9
10 static void logfs_calc_free(struct super_block *sb)
11 {
12         struct logfs_super *super = logfs_super(sb);
13         u64 reserve, no_segs = super->s_no_segs;
14         s64 free;
15         int i;
16
17         /* superblock segments */
18         no_segs -= 2;
19         super->s_no_journal_segs = 0;
20         /* journal */
21         journal_for_each(i)
22                 if (super->s_journal_seg[i]) {
23                         no_segs--;
24                         super->s_no_journal_segs++;
25                 }
26
27         /* open segments plus one extra per level for GC */
28         no_segs -= 2 * super->s_total_levels;
29
30         free = no_segs * (super->s_segsize - LOGFS_SEGMENT_RESERVE);
31         free -= super->s_used_bytes;
32         /* just a bit extra */
33         free -= super->s_total_levels * 4096;
34
35         /* Bad blocks are 'paid' for with speed reserve - the filesystem
36          * simply gets slower as bad blocks accumulate.  Until the bad blocks
37          * exceed the speed reserve - then the filesystem gets smaller.
38          */
39         reserve = super->s_bad_segments + super->s_bad_seg_reserve;
40         reserve *= super->s_segsize - LOGFS_SEGMENT_RESERVE;
41         reserve = max(reserve, super->s_speed_reserve);
42         free -= reserve;
43         if (free < 0)
44                 free = 0;
45
46         super->s_free_bytes = free;
47 }
48
49 static void reserve_sb_and_journal(struct super_block *sb)
50 {
51         struct logfs_super *super = logfs_super(sb);
52         struct btree_head32 *head = &super->s_reserved_segments;
53         int i, err;
54
55         err = btree_insert32(head, seg_no(sb, super->s_sb_ofs[0]), (void *)1,
56                         GFP_KERNEL);
57         BUG_ON(err);
58
59         err = btree_insert32(head, seg_no(sb, super->s_sb_ofs[1]), (void *)1,
60                         GFP_KERNEL);
61         BUG_ON(err);
62
63         journal_for_each(i) {
64                 if (!super->s_journal_seg[i])
65                         continue;
66                 err = btree_insert32(head, super->s_journal_seg[i], (void *)1,
67                                 GFP_KERNEL);
68                 BUG_ON(err);
69         }
70 }
71
72 static void read_dynsb(struct super_block *sb,
73                 struct logfs_je_dynsb *dynsb)
74 {
75         struct logfs_super *super = logfs_super(sb);
76
77         super->s_gec            = be64_to_cpu(dynsb->ds_gec);
78         super->s_sweeper        = be64_to_cpu(dynsb->ds_sweeper);
79         super->s_victim_ino     = be64_to_cpu(dynsb->ds_victim_ino);
80         super->s_rename_dir     = be64_to_cpu(dynsb->ds_rename_dir);
81         super->s_rename_pos     = be64_to_cpu(dynsb->ds_rename_pos);
82         super->s_used_bytes     = be64_to_cpu(dynsb->ds_used_bytes);
83         super->s_generation     = be32_to_cpu(dynsb->ds_generation);
84 }
85
86 static void read_anchor(struct super_block *sb,
87                 struct logfs_je_anchor *da)
88 {
89         struct logfs_super *super = logfs_super(sb);
90         struct inode *inode = super->s_master_inode;
91         struct logfs_inode *li = logfs_inode(inode);
92         int i;
93
94         super->s_last_ino = be64_to_cpu(da->da_last_ino);
95         li->li_flags    = 0;
96         li->li_height   = da->da_height;
97         i_size_write(inode, be64_to_cpu(da->da_size));
98         li->li_used_bytes = be64_to_cpu(da->da_used_bytes);
99
100         for (i = 0; i < LOGFS_EMBEDDED_FIELDS; i++)
101                 li->li_data[i] = be64_to_cpu(da->da_data[i]);
102 }
103
104 static void read_erasecount(struct super_block *sb,
105                 struct logfs_je_journal_ec *ec)
106 {
107         struct logfs_super *super = logfs_super(sb);
108         int i;
109
110         journal_for_each(i)
111                 super->s_journal_ec[i] = be32_to_cpu(ec->ec[i]);
112 }
113
114 static int read_area(struct super_block *sb, struct logfs_je_area *a)
115 {
116         struct logfs_super *super = logfs_super(sb);
117         struct logfs_area *area = super->s_area[a->gc_level];
118         u64 ofs;
119         u32 writemask = ~(super->s_writesize - 1);
120
121         if (a->gc_level >= LOGFS_NO_AREAS)
122                 return -EIO;
123         if (a->vim != VIM_DEFAULT)
124                 return -EIO; /* TODO: close area and continue */
125
126         area->a_used_bytes = be32_to_cpu(a->used_bytes);
127         area->a_written_bytes = area->a_used_bytes & writemask;
128         area->a_segno = be32_to_cpu(a->segno);
129         if (area->a_segno)
130                 area->a_is_open = 1;
131
132         ofs = dev_ofs(sb, area->a_segno, area->a_written_bytes);
133         if (super->s_writesize > 1)
134                 logfs_buf_recover(area, ofs, a + 1, super->s_writesize);
135         else
136                 logfs_buf_recover(area, ofs, NULL, 0);
137         return 0;
138 }
139
140 static void *unpack(void *from, void *to)
141 {
142         struct logfs_journal_header *jh = from;
143         void *data = from + sizeof(struct logfs_journal_header);
144         int err;
145         size_t inlen, outlen;
146
147         inlen = be16_to_cpu(jh->h_len);
148         outlen = be16_to_cpu(jh->h_datalen);
149
150         if (jh->h_compr == COMPR_NONE)
151                 memcpy(to, data, inlen);
152         else {
153                 err = logfs_uncompress(data, to, inlen, outlen);
154                 BUG_ON(err);
155         }
156         return to;
157 }
158
159 static int __read_je_header(struct super_block *sb, u64 ofs,
160                 struct logfs_journal_header *jh)
161 {
162         struct logfs_super *super = logfs_super(sb);
163         size_t bufsize = max_t(size_t, sb->s_blocksize, super->s_writesize)
164                 + MAX_JOURNAL_HEADER;
165         u16 type, len, datalen;
166         int err;
167
168         /* read header only */
169         err = wbuf_read(sb, ofs, sizeof(*jh), jh);
170         if (err)
171                 return err;
172         type = be16_to_cpu(jh->h_type);
173         len = be16_to_cpu(jh->h_len);
174         datalen = be16_to_cpu(jh->h_datalen);
175         if (len > sb->s_blocksize)
176                 return -EIO;
177         if ((type < JE_FIRST) || (type > JE_LAST))
178                 return -EIO;
179         if (datalen > bufsize)
180                 return -EIO;
181         return 0;
182 }
183
184 static int __read_je_payload(struct super_block *sb, u64 ofs,
185                 struct logfs_journal_header *jh)
186 {
187         u16 len;
188         int err;
189
190         len = be16_to_cpu(jh->h_len);
191         err = wbuf_read(sb, ofs + sizeof(*jh), len, jh + 1);
192         if (err)
193                 return err;
194         if (jh->h_crc != logfs_crc32(jh, len + sizeof(*jh), 4)) {
195                 /* Old code was confused.  It forgot about the header length
196                  * and stopped calculating the crc 16 bytes before the end
197                  * of data - ick!
198                  * FIXME: Remove this hack once the old code is fixed.
199                  */
200                 if (jh->h_crc == logfs_crc32(jh, len, 4))
201                         WARN_ON_ONCE(1);
202                 else
203                         return -EIO;
204         }
205         return 0;
206 }
207
208 /*
209  * jh needs to be large enough to hold the complete entry, not just the header
210  */
211 static int __read_je(struct super_block *sb, u64 ofs,
212                 struct logfs_journal_header *jh)
213 {
214         int err;
215
216         err = __read_je_header(sb, ofs, jh);
217         if (err)
218                 return err;
219         return __read_je_payload(sb, ofs, jh);
220 }
221
222 static int read_je(struct super_block *sb, u64 ofs)
223 {
224         struct logfs_super *super = logfs_super(sb);
225         struct logfs_journal_header *jh = super->s_compressed_je;
226         void *scratch = super->s_je;
227         u16 type, datalen;
228         int err;
229
230         err = __read_je(sb, ofs, jh);
231         if (err)
232                 return err;
233         type = be16_to_cpu(jh->h_type);
234         datalen = be16_to_cpu(jh->h_datalen);
235
236         switch (type) {
237         case JE_DYNSB:
238                 read_dynsb(sb, unpack(jh, scratch));
239                 break;
240         case JE_ANCHOR:
241                 read_anchor(sb, unpack(jh, scratch));
242                 break;
243         case JE_ERASECOUNT:
244                 read_erasecount(sb, unpack(jh, scratch));
245                 break;
246         case JE_AREA:
247                 read_area(sb, unpack(jh, scratch));
248                 break;
249         case JE_OBJ_ALIAS:
250                 err = logfs_load_object_aliases(sb, unpack(jh, scratch),
251                                 datalen);
252                 break;
253         default:
254                 WARN_ON_ONCE(1);
255                 return -EIO;
256         }
257         return err;
258 }
259
260 static int logfs_read_segment(struct super_block *sb, u32 segno)
261 {
262         struct logfs_super *super = logfs_super(sb);
263         struct logfs_journal_header *jh = super->s_compressed_je;
264         u64 ofs, seg_ofs = dev_ofs(sb, segno, 0);
265         u32 h_ofs, last_ofs = 0;
266         u16 len, datalen, last_len = 0;
267         int i, err;
268
269         /* search for most recent commit */
270         for (h_ofs = 0; h_ofs < super->s_segsize; h_ofs += sizeof(*jh)) {
271                 ofs = seg_ofs + h_ofs;
272                 err = __read_je_header(sb, ofs, jh);
273                 if (err)
274                         continue;
275                 if (jh->h_type != cpu_to_be16(JE_COMMIT))
276                         continue;
277                 err = __read_je_payload(sb, ofs, jh);
278                 if (err)
279                         continue;
280                 len = be16_to_cpu(jh->h_len);
281                 datalen = be16_to_cpu(jh->h_datalen);
282                 if ((datalen > sizeof(super->s_je_array)) ||
283                                 (datalen % sizeof(__be64)))
284                         continue;
285                 last_ofs = h_ofs;
286                 last_len = datalen;
287                 h_ofs += ALIGN(len, sizeof(*jh)) - sizeof(*jh);
288         }
289         /* read commit */
290         if (last_ofs == 0)
291                 return -ENOENT;
292         ofs = seg_ofs + last_ofs;
293         log_journal("Read commit from %llx\n", ofs);
294         err = __read_je(sb, ofs, jh);
295         BUG_ON(err); /* We should have caught it in the scan loop already */
296         if (err)
297                 return err;
298         /* uncompress */
299         unpack(jh, super->s_je_array);
300         super->s_no_je = last_len / sizeof(__be64);
301         /* iterate over array */
302         for (i = 0; i < super->s_no_je; i++) {
303                 err = read_je(sb, be64_to_cpu(super->s_je_array[i]));
304                 if (err)
305                         return err;
306         }
307         super->s_journal_area->a_segno = segno;
308         return 0;
309 }
310
311 static u64 read_gec(struct super_block *sb, u32 segno)
312 {
313         struct logfs_segment_header sh;
314         __be32 crc;
315         int err;
316
317         if (!segno)
318                 return 0;
319         err = wbuf_read(sb, dev_ofs(sb, segno, 0), sizeof(sh), &sh);
320         if (err)
321                 return 0;
322         crc = logfs_crc32(&sh, sizeof(sh), 4);
323         if (crc != sh.crc) {
324                 WARN_ON(sh.gec != cpu_to_be64(0xffffffffffffffffull));
325                 /* Most likely it was just erased */
326                 return 0;
327         }
328         return be64_to_cpu(sh.gec);
329 }
330
331 static int logfs_read_journal(struct super_block *sb)
332 {
333         struct logfs_super *super = logfs_super(sb);
334         u64 gec[LOGFS_JOURNAL_SEGS], max;
335         u32 segno;
336         int i, max_i;
337
338         max = 0;
339         max_i = -1;
340         journal_for_each(i) {
341                 segno = super->s_journal_seg[i];
342                 gec[i] = read_gec(sb, super->s_journal_seg[i]);
343                 if (gec[i] > max) {
344                         max = gec[i];
345                         max_i = i;
346                 }
347         }
348         if (max_i == -1)
349                 return -EIO;
350         /* FIXME: Try older segments in case of error */
351         return logfs_read_segment(sb, super->s_journal_seg[max_i]);
352 }
353
354 /*
355  * First search the current segment (outer loop), then pick the next segment
356  * in the array, skipping any zero entries (inner loop).
357  */
358 static void journal_get_free_segment(struct logfs_area *area)
359 {
360         struct logfs_super *super = logfs_super(area->a_sb);
361         int i;
362
363         journal_for_each(i) {
364                 if (area->a_segno != super->s_journal_seg[i])
365                         continue;
366
367                 do {
368                         i++;
369                         if (i == LOGFS_JOURNAL_SEGS)
370                                 i = 0;
371                 } while (!super->s_journal_seg[i]);
372
373                 area->a_segno = super->s_journal_seg[i];
374                 area->a_erase_count = ++(super->s_journal_ec[i]);
375                 log_journal("Journal now at %x (ec %x)\n", area->a_segno,
376                                 area->a_erase_count);
377                 return;
378         }
379         BUG();
380 }
381
382 static void journal_get_erase_count(struct logfs_area *area)
383 {
384         /* erase count is stored globally and incremented in
385          * journal_get_free_segment() - nothing to do here */
386 }
387
388 static int journal_erase_segment(struct logfs_area *area)
389 {
390         struct super_block *sb = area->a_sb;
391         struct logfs_segment_header sh;
392         u64 ofs;
393         int err;
394
395         err = logfs_erase_segment(sb, area->a_segno, 1);
396         if (err)
397                 return err;
398
399         sh.pad = 0;
400         sh.type = SEG_JOURNAL;
401         sh.level = 0;
402         sh.segno = cpu_to_be32(area->a_segno);
403         sh.ec = cpu_to_be32(area->a_erase_count);
404         sh.gec = cpu_to_be64(logfs_super(sb)->s_gec);
405         sh.crc = logfs_crc32(&sh, sizeof(sh), 4);
406
407         /* This causes a bug in segment.c.  Not yet. */
408         //logfs_set_segment_erased(sb, area->a_segno, area->a_erase_count, 0);
409
410         ofs = dev_ofs(sb, area->a_segno, 0);
411         area->a_used_bytes = ALIGN(sizeof(sh), 16);
412         logfs_buf_write(area, ofs, &sh, sizeof(sh));
413         return 0;
414 }
415
416 static size_t __logfs_write_header(struct logfs_super *super,
417                 struct logfs_journal_header *jh, size_t len, size_t datalen,
418                 u16 type, u8 compr)
419 {
420         jh->h_len       = cpu_to_be16(len);
421         jh->h_type      = cpu_to_be16(type);
422         jh->h_datalen   = cpu_to_be16(datalen);
423         jh->h_compr     = compr;
424         jh->h_pad[0]    = 'H';
425         jh->h_pad[1]    = 'E';
426         jh->h_pad[2]    = 'A';
427         jh->h_pad[3]    = 'D';
428         jh->h_pad[4]    = 'R';
429         jh->h_crc       = logfs_crc32(jh, len + sizeof(*jh), 4);
430         return ALIGN(len, 16) + sizeof(*jh);
431 }
432
433 static size_t logfs_write_header(struct logfs_super *super,
434                 struct logfs_journal_header *jh, size_t datalen, u16 type)
435 {
436         size_t len = datalen;
437
438         return __logfs_write_header(super, jh, len, datalen, type, COMPR_NONE);
439 }
440
441 static inline size_t logfs_journal_erasecount_size(struct logfs_super *super)
442 {
443         return LOGFS_JOURNAL_SEGS * sizeof(__be32);
444 }
445
446 static void *logfs_write_erasecount(struct super_block *sb, void *_ec,
447                 u16 *type, size_t *len)
448 {
449         struct logfs_super *super = logfs_super(sb);
450         struct logfs_je_journal_ec *ec = _ec;
451         int i;
452
453         journal_for_each(i)
454                 ec->ec[i] = cpu_to_be32(super->s_journal_ec[i]);
455         *type = JE_ERASECOUNT;
456         *len = logfs_journal_erasecount_size(super);
457         return ec;
458 }
459
460 static void account_shadow(void *_shadow, unsigned long _sb, u64 ignore,
461                 size_t ignore2)
462 {
463         struct logfs_shadow *shadow = _shadow;
464         struct super_block *sb = (void *)_sb;
465         struct logfs_super *super = logfs_super(sb);
466
467         /* consume new space */
468         super->s_free_bytes       -= shadow->new_len;
469         super->s_used_bytes       += shadow->new_len;
470         super->s_dirty_used_bytes -= shadow->new_len;
471
472         /* free up old space */
473         super->s_free_bytes       += shadow->old_len;
474         super->s_used_bytes       -= shadow->old_len;
475         super->s_dirty_free_bytes -= shadow->old_len;
476
477         logfs_set_segment_used(sb, shadow->old_ofs, -shadow->old_len);
478         logfs_set_segment_used(sb, shadow->new_ofs, shadow->new_len);
479
480         log_journal("account_shadow(%llx, %llx, %x) %llx->%llx %x->%x\n",
481                         shadow->ino, shadow->bix, shadow->gc_level,
482                         shadow->old_ofs, shadow->new_ofs,
483                         shadow->old_len, shadow->new_len);
484         mempool_free(shadow, super->s_shadow_pool);
485 }
486
487 static void account_shadows(struct super_block *sb)
488 {
489         struct logfs_super *super = logfs_super(sb);
490         struct inode *inode = super->s_master_inode;
491         struct logfs_inode *li = logfs_inode(inode);
492         struct shadow_tree *tree = &super->s_shadow_tree;
493
494         btree_grim_visitor64(&tree->new, (unsigned long)sb, account_shadow);
495         btree_grim_visitor64(&tree->old, (unsigned long)sb, account_shadow);
496         btree_grim_visitor32(&tree->segment_map, 0, NULL);
497         tree->no_shadowed_segments = 0;
498
499         if (li->li_block) {
500                 /*
501                  * We never actually use the structure, when attached to the
502                  * master inode.  But it is easier to always free it here than
503                  * to have checks in several places elsewhere when allocating
504                  * it.
505                  */
506                 li->li_block->ops->free_block(sb, li->li_block);
507         }
508         BUG_ON((s64)li->li_used_bytes < 0);
509 }
510
511 static void *__logfs_write_anchor(struct super_block *sb, void *_da,
512                 u16 *type, size_t *len)
513 {
514         struct logfs_super *super = logfs_super(sb);
515         struct logfs_je_anchor *da = _da;
516         struct inode *inode = super->s_master_inode;
517         struct logfs_inode *li = logfs_inode(inode);
518         int i;
519
520         da->da_height   = li->li_height;
521         da->da_last_ino = cpu_to_be64(super->s_last_ino);
522         da->da_size     = cpu_to_be64(i_size_read(inode));
523         da->da_used_bytes = cpu_to_be64(li->li_used_bytes);
524         for (i = 0; i < LOGFS_EMBEDDED_FIELDS; i++)
525                 da->da_data[i] = cpu_to_be64(li->li_data[i]);
526         *type = JE_ANCHOR;
527         *len = sizeof(*da);
528         return da;
529 }
530
531 static void *logfs_write_dynsb(struct super_block *sb, void *_dynsb,
532                 u16 *type, size_t *len)
533 {
534         struct logfs_super *super = logfs_super(sb);
535         struct logfs_je_dynsb *dynsb = _dynsb;
536
537         dynsb->ds_gec           = cpu_to_be64(super->s_gec);
538         dynsb->ds_sweeper       = cpu_to_be64(super->s_sweeper);
539         dynsb->ds_victim_ino    = cpu_to_be64(super->s_victim_ino);
540         dynsb->ds_rename_dir    = cpu_to_be64(super->s_rename_dir);
541         dynsb->ds_rename_pos    = cpu_to_be64(super->s_rename_pos);
542         dynsb->ds_used_bytes    = cpu_to_be64(super->s_used_bytes);
543         dynsb->ds_generation    = cpu_to_be32(super->s_generation);
544         *type = JE_DYNSB;
545         *len = sizeof(*dynsb);
546         return dynsb;
547 }
548
549 static void write_wbuf(struct super_block *sb, struct logfs_area *area,
550                 void *wbuf)
551 {
552         struct logfs_super *super = logfs_super(sb);
553         struct address_space *mapping = super->s_mapping_inode->i_mapping;
554         u64 ofs;
555         pgoff_t index;
556         int page_ofs;
557         struct page *page;
558
559         ofs = dev_ofs(sb, area->a_segno,
560                         area->a_used_bytes & ~(super->s_writesize - 1));
561         index = ofs >> PAGE_SHIFT;
562         page_ofs = ofs & (PAGE_SIZE - 1);
563
564         page = find_lock_page(mapping, index);
565         BUG_ON(!page);
566         memcpy(wbuf, page_address(page) + page_ofs, super->s_writesize);
567         unlock_page(page);
568 }
569
570 static void *logfs_write_area(struct super_block *sb, void *_a,
571                 u16 *type, size_t *len)
572 {
573         struct logfs_super *super = logfs_super(sb);
574         struct logfs_area *area = super->s_area[super->s_sum_index];
575         struct logfs_je_area *a = _a;
576
577         a->vim = VIM_DEFAULT;
578         a->gc_level = super->s_sum_index;
579         a->used_bytes = cpu_to_be32(area->a_used_bytes);
580         a->segno = cpu_to_be32(area->a_segno);
581         if (super->s_writesize > 1)
582                 write_wbuf(sb, area, a + 1);
583
584         *type = JE_AREA;
585         *len = sizeof(*a) + super->s_writesize;
586         return a;
587 }
588
589 static void *logfs_write_commit(struct super_block *sb, void *h,
590                 u16 *type, size_t *len)
591 {
592         struct logfs_super *super = logfs_super(sb);
593
594         *type = JE_COMMIT;
595         *len = super->s_no_je * sizeof(__be64);
596         return super->s_je_array;
597 }
598
599 static size_t __logfs_write_je(struct super_block *sb, void *buf, u16 type,
600                 size_t len)
601 {
602         struct logfs_super *super = logfs_super(sb);
603         void *header = super->s_compressed_je;
604         void *data = header + sizeof(struct logfs_journal_header);
605         ssize_t compr_len, pad_len;
606         u8 compr = COMPR_ZLIB;
607
608         if (len == 0)
609                 return logfs_write_header(super, header, 0, type);
610
611         compr_len = logfs_compress(buf, data, len, sb->s_blocksize);
612         if (compr_len < 0 || type == JE_ANCHOR) {
613                 BUG_ON(len > sb->s_blocksize);
614                 memcpy(data, buf, len);
615                 compr_len = len;
616                 compr = COMPR_NONE;
617         }
618
619         pad_len = ALIGN(compr_len, 16);
620         memset(data + compr_len, 0, pad_len - compr_len);
621
622         return __logfs_write_header(super, header, compr_len, len, type, compr);
623 }
624
625 static s64 logfs_get_free_bytes(struct logfs_area *area, size_t *bytes,
626                 int must_pad)
627 {
628         u32 writesize = logfs_super(area->a_sb)->s_writesize;
629         s32 ofs;
630         int ret;
631
632         ret = logfs_open_area(area, *bytes);
633         if (ret)
634                 return -EAGAIN;
635
636         ofs = area->a_used_bytes;
637         area->a_used_bytes += *bytes;
638
639         if (must_pad) {
640                 area->a_used_bytes = ALIGN(area->a_used_bytes, writesize);
641                 *bytes = area->a_used_bytes - ofs;
642         }
643
644         return dev_ofs(area->a_sb, area->a_segno, ofs);
645 }
646
647 static int logfs_write_je_buf(struct super_block *sb, void *buf, u16 type,
648                 size_t buf_len)
649 {
650         struct logfs_super *super = logfs_super(sb);
651         struct logfs_area *area = super->s_journal_area;
652         struct logfs_journal_header *jh = super->s_compressed_je;
653         size_t len;
654         int must_pad = 0;
655         s64 ofs;
656
657         len = __logfs_write_je(sb, buf, type, buf_len);
658         if (jh->h_type == cpu_to_be16(JE_COMMIT))
659                 must_pad = 1;
660
661         ofs = logfs_get_free_bytes(area, &len, must_pad);
662         if (ofs < 0)
663                 return ofs;
664         logfs_buf_write(area, ofs, super->s_compressed_je, len);
665         BUG_ON(super->s_no_je >= MAX_JOURNAL_ENTRIES);
666         super->s_je_array[super->s_no_je++] = cpu_to_be64(ofs);
667         return 0;
668 }
669
670 static int logfs_write_je(struct super_block *sb,
671                 void* (*write)(struct super_block *sb, void *scratch,
672                         u16 *type, size_t *len))
673 {
674         void *buf;
675         size_t len;
676         u16 type;
677
678         buf = write(sb, logfs_super(sb)->s_je, &type, &len);
679         return logfs_write_je_buf(sb, buf, type, len);
680 }
681
682 int write_alias_journal(struct super_block *sb, u64 ino, u64 bix,
683                 level_t level, int child_no, __be64 val)
684 {
685         struct logfs_super *super = logfs_super(sb);
686         struct logfs_obj_alias *oa = super->s_je;
687         int err = 0, fill = super->s_je_fill;
688
689         log_aliases("logfs_write_obj_aliases #%x(%llx, %llx, %x, %x) %llx\n",
690                         fill, ino, bix, level, child_no, be64_to_cpu(val));
691         oa[fill].ino = cpu_to_be64(ino);
692         oa[fill].bix = cpu_to_be64(bix);
693         oa[fill].val = val;
694         oa[fill].level = (__force u8)level;
695         oa[fill].child_no = cpu_to_be16(child_no);
696         fill++;
697         if (fill >= sb->s_blocksize / sizeof(*oa)) {
698                 err = logfs_write_je_buf(sb, oa, JE_OBJ_ALIAS, sb->s_blocksize);
699                 fill = 0;
700         }
701
702         super->s_je_fill = fill;
703         return err;
704 }
705
706 static int logfs_write_obj_aliases(struct super_block *sb)
707 {
708         struct logfs_super *super = logfs_super(sb);
709         int err;
710
711         log_journal("logfs_write_obj_aliases: %d aliases to write\n",
712                         super->s_no_object_aliases);
713         super->s_je_fill = 0;
714         err = logfs_write_obj_aliases_pagecache(sb);
715         if (err)
716                 return err;
717
718         if (super->s_je_fill)
719                 err = logfs_write_je_buf(sb, super->s_je, JE_OBJ_ALIAS,
720                                 super->s_je_fill
721                                 * sizeof(struct logfs_obj_alias));
722         return err;
723 }
724
725 /*
726  * Write all journal entries.  The goto logic ensures that all journal entries
727  * are written whenever a new segment is used.  It is ugly and potentially a
728  * bit wasteful, but robustness is more important.  With this we can *always*
729  * erase all journal segments except the one containing the most recent commit.
730  */
731 void logfs_write_anchor(struct super_block *sb)
732 {
733         struct logfs_super *super = logfs_super(sb);
734         struct logfs_area *area = super->s_journal_area;
735         int i, err;
736
737         if (!(super->s_flags & LOGFS_SB_FLAG_DIRTY))
738                 return;
739         super->s_flags &= ~LOGFS_SB_FLAG_DIRTY;
740
741         BUG_ON(super->s_flags & LOGFS_SB_FLAG_SHUTDOWN);
742         mutex_lock(&super->s_journal_mutex);
743
744         /* Do this first or suffer corruption */
745         logfs_sync_segments(sb);
746         account_shadows(sb);
747
748 again:
749         super->s_no_je = 0;
750         for_each_area(i) {
751                 if (!super->s_area[i]->a_is_open)
752                         continue;
753                 super->s_sum_index = i;
754                 err = logfs_write_je(sb, logfs_write_area);
755                 if (err)
756                         goto again;
757         }
758         err = logfs_write_obj_aliases(sb);
759         if (err)
760                 goto again;
761         err = logfs_write_je(sb, logfs_write_erasecount);
762         if (err)
763                 goto again;
764         err = logfs_write_je(sb, __logfs_write_anchor);
765         if (err)
766                 goto again;
767         err = logfs_write_je(sb, logfs_write_dynsb);
768         if (err)
769                 goto again;
770         /*
771          * Order is imperative.  First we sync all writes, including the
772          * non-committed journal writes.  Then we write the final commit and
773          * sync the current journal segment.
774          * There is a theoretical bug here.  Syncing the journal segment will
775          * write a number of journal entries and the final commit.  All these
776          * are written in a single operation.  If the device layer writes the
777          * data back-to-front, the commit will precede the other journal
778          * entries, leaving a race window.
779          * Two fixes are possible.  Preferred is to fix the device layer to
780          * ensure writes happen front-to-back.  Alternatively we can insert
781          * another logfs_sync_area() super->s_devops->sync() combo before
782          * writing the commit.
783          */
784         /*
785          * On another subject, super->s_devops->sync is usually not necessary.
786          * Unless called from sys_sync or friends, a barrier would suffice.
787          */
788         super->s_devops->sync(sb);
789         err = logfs_write_je(sb, logfs_write_commit);
790         if (err)
791                 goto again;
792         log_journal("Write commit to %llx\n",
793                         be64_to_cpu(super->s_je_array[super->s_no_je - 1]));
794         logfs_sync_area(area);
795         BUG_ON(area->a_used_bytes != area->a_written_bytes);
796         super->s_devops->sync(sb);
797
798         mutex_unlock(&super->s_journal_mutex);
799         return;
800 }
801
802 void do_logfs_journal_wl_pass(struct super_block *sb)
803 {
804         struct logfs_super *super = logfs_super(sb);
805         struct logfs_area *area = super->s_journal_area;
806         struct btree_head32 *head = &super->s_reserved_segments;
807         u32 segno, ec;
808         int i, err;
809
810         log_journal("Journal requires wear-leveling.\n");
811         /* Drop old segments */
812         journal_for_each(i)
813                 if (super->s_journal_seg[i]) {
814                         btree_remove32(head, super->s_journal_seg[i]);
815                         logfs_set_segment_unreserved(sb,
816                                         super->s_journal_seg[i],
817                                         super->s_journal_ec[i]);
818                         super->s_journal_seg[i] = 0;
819                         super->s_journal_ec[i] = 0;
820                 }
821         /* Get new segments */
822         for (i = 0; i < super->s_no_journal_segs; i++) {
823                 segno = get_best_cand(sb, &super->s_reserve_list, &ec);
824                 super->s_journal_seg[i] = segno;
825                 super->s_journal_ec[i] = ec;
826                 logfs_set_segment_reserved(sb, segno);
827                 err = btree_insert32(head, segno, (void *)1, GFP_KERNEL);
828                 BUG_ON(err); /* mempool should prevent this */
829                 err = logfs_erase_segment(sb, segno, 1);
830                 BUG_ON(err); /* FIXME: remount-ro would be nicer */
831         }
832         /* Manually move journal_area */
833         freeseg(sb, area->a_segno);
834         area->a_segno = super->s_journal_seg[0];
835         area->a_is_open = 0;
836         area->a_used_bytes = 0;
837         /* Write journal */
838         logfs_write_anchor(sb);
839         /* Write superblocks */
840         err = logfs_write_sb(sb);
841         BUG_ON(err);
842 }
843
844 static const struct logfs_area_ops journal_area_ops = {
845         .get_free_segment       = journal_get_free_segment,
846         .get_erase_count        = journal_get_erase_count,
847         .erase_segment          = journal_erase_segment,
848 };
849
850 int logfs_init_journal(struct super_block *sb)
851 {
852         struct logfs_super *super = logfs_super(sb);
853         size_t bufsize = max_t(size_t, sb->s_blocksize, super->s_writesize)
854                 + MAX_JOURNAL_HEADER;
855         int ret = -ENOMEM;
856
857         mutex_init(&super->s_journal_mutex);
858         btree_init_mempool32(&super->s_reserved_segments, super->s_btree_pool);
859
860         super->s_je = kzalloc(bufsize, GFP_KERNEL);
861         if (!super->s_je)
862                 return ret;
863
864         super->s_compressed_je = kzalloc(bufsize, GFP_KERNEL);
865         if (!super->s_compressed_je)
866                 return ret;
867
868         super->s_master_inode = logfs_new_meta_inode(sb, LOGFS_INO_MASTER);
869         if (IS_ERR(super->s_master_inode))
870                 return PTR_ERR(super->s_master_inode);
871
872         ret = logfs_read_journal(sb);
873         if (ret)
874                 return -EIO;
875
876         reserve_sb_and_journal(sb);
877         logfs_calc_free(sb);
878
879         super->s_journal_area->a_ops = &journal_area_ops;
880         return 0;
881 }
882
883 void logfs_cleanup_journal(struct super_block *sb)
884 {
885         struct logfs_super *super = logfs_super(sb);
886
887         btree_grim_visitor32(&super->s_reserved_segments, 0, NULL);
888         destroy_meta_inode(super->s_master_inode);
889         super->s_master_inode = NULL;
890
891         kfree(super->s_compressed_je);
892         kfree(super->s_je);
893 }