fd44eeea513567d9822e39af164c67ced2158d1a
[kernel.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         union {
392                 struct logfs_segment_header sh;
393                 unsigned char c[ALIGN(sizeof(struct logfs_segment_header), 16)];
394         } u;
395         u64 ofs;
396         int err;
397
398         err = logfs_erase_segment(sb, area->a_segno, 1);
399         if (err)
400                 return err;
401
402         memset(&u, 0, sizeof(u));
403         u.sh.pad = 0;
404         u.sh.type = SEG_JOURNAL;
405         u.sh.level = 0;
406         u.sh.segno = cpu_to_be32(area->a_segno);
407         u.sh.ec = cpu_to_be32(area->a_erase_count);
408         u.sh.gec = cpu_to_be64(logfs_super(sb)->s_gec);
409         u.sh.crc = logfs_crc32(&u.sh, sizeof(u.sh), 4);
410
411         /* This causes a bug in segment.c.  Not yet. */
412         //logfs_set_segment_erased(sb, area->a_segno, area->a_erase_count, 0);
413
414         ofs = dev_ofs(sb, area->a_segno, 0);
415         area->a_used_bytes = sizeof(u);
416         logfs_buf_write(area, ofs, &u, sizeof(u));
417         return 0;
418 }
419
420 static size_t __logfs_write_header(struct logfs_super *super,
421                 struct logfs_journal_header *jh, size_t len, size_t datalen,
422                 u16 type, u8 compr)
423 {
424         jh->h_len       = cpu_to_be16(len);
425         jh->h_type      = cpu_to_be16(type);
426         jh->h_datalen   = cpu_to_be16(datalen);
427         jh->h_compr     = compr;
428         jh->h_pad[0]    = 'H';
429         jh->h_pad[1]    = 'E';
430         jh->h_pad[2]    = 'A';
431         jh->h_pad[3]    = 'D';
432         jh->h_pad[4]    = 'R';
433         jh->h_crc       = logfs_crc32(jh, len + sizeof(*jh), 4);
434         return ALIGN(len, 16) + sizeof(*jh);
435 }
436
437 static size_t logfs_write_header(struct logfs_super *super,
438                 struct logfs_journal_header *jh, size_t datalen, u16 type)
439 {
440         size_t len = datalen;
441
442         return __logfs_write_header(super, jh, len, datalen, type, COMPR_NONE);
443 }
444
445 static inline size_t logfs_journal_erasecount_size(struct logfs_super *super)
446 {
447         return LOGFS_JOURNAL_SEGS * sizeof(__be32);
448 }
449
450 static void *logfs_write_erasecount(struct super_block *sb, void *_ec,
451                 u16 *type, size_t *len)
452 {
453         struct logfs_super *super = logfs_super(sb);
454         struct logfs_je_journal_ec *ec = _ec;
455         int i;
456
457         journal_for_each(i)
458                 ec->ec[i] = cpu_to_be32(super->s_journal_ec[i]);
459         *type = JE_ERASECOUNT;
460         *len = logfs_journal_erasecount_size(super);
461         return ec;
462 }
463
464 static void account_shadow(void *_shadow, unsigned long _sb, u64 ignore,
465                 size_t ignore2)
466 {
467         struct logfs_shadow *shadow = _shadow;
468         struct super_block *sb = (void *)_sb;
469         struct logfs_super *super = logfs_super(sb);
470
471         /* consume new space */
472         super->s_free_bytes       -= shadow->new_len;
473         super->s_used_bytes       += shadow->new_len;
474         super->s_dirty_used_bytes -= shadow->new_len;
475
476         /* free up old space */
477         super->s_free_bytes       += shadow->old_len;
478         super->s_used_bytes       -= shadow->old_len;
479         super->s_dirty_free_bytes -= shadow->old_len;
480
481         logfs_set_segment_used(sb, shadow->old_ofs, -shadow->old_len);
482         logfs_set_segment_used(sb, shadow->new_ofs, shadow->new_len);
483
484         log_journal("account_shadow(%llx, %llx, %x) %llx->%llx %x->%x\n",
485                         shadow->ino, shadow->bix, shadow->gc_level,
486                         shadow->old_ofs, shadow->new_ofs,
487                         shadow->old_len, shadow->new_len);
488         mempool_free(shadow, super->s_shadow_pool);
489 }
490
491 static void account_shadows(struct super_block *sb)
492 {
493         struct logfs_super *super = logfs_super(sb);
494         struct inode *inode = super->s_master_inode;
495         struct logfs_inode *li = logfs_inode(inode);
496         struct shadow_tree *tree = &super->s_shadow_tree;
497
498         btree_grim_visitor64(&tree->new, (unsigned long)sb, account_shadow);
499         btree_grim_visitor64(&tree->old, (unsigned long)sb, account_shadow);
500         btree_grim_visitor32(&tree->segment_map, 0, NULL);
501         tree->no_shadowed_segments = 0;
502
503         if (li->li_block) {
504                 /*
505                  * We never actually use the structure, when attached to the
506                  * master inode.  But it is easier to always free it here than
507                  * to have checks in several places elsewhere when allocating
508                  * it.
509                  */
510                 li->li_block->ops->free_block(sb, li->li_block);
511         }
512         BUG_ON((s64)li->li_used_bytes < 0);
513 }
514
515 static void *__logfs_write_anchor(struct super_block *sb, void *_da,
516                 u16 *type, size_t *len)
517 {
518         struct logfs_super *super = logfs_super(sb);
519         struct logfs_je_anchor *da = _da;
520         struct inode *inode = super->s_master_inode;
521         struct logfs_inode *li = logfs_inode(inode);
522         int i;
523
524         da->da_height   = li->li_height;
525         da->da_last_ino = cpu_to_be64(super->s_last_ino);
526         da->da_size     = cpu_to_be64(i_size_read(inode));
527         da->da_used_bytes = cpu_to_be64(li->li_used_bytes);
528         for (i = 0; i < LOGFS_EMBEDDED_FIELDS; i++)
529                 da->da_data[i] = cpu_to_be64(li->li_data[i]);
530         *type = JE_ANCHOR;
531         *len = sizeof(*da);
532         return da;
533 }
534
535 static void *logfs_write_dynsb(struct super_block *sb, void *_dynsb,
536                 u16 *type, size_t *len)
537 {
538         struct logfs_super *super = logfs_super(sb);
539         struct logfs_je_dynsb *dynsb = _dynsb;
540
541         dynsb->ds_gec           = cpu_to_be64(super->s_gec);
542         dynsb->ds_sweeper       = cpu_to_be64(super->s_sweeper);
543         dynsb->ds_victim_ino    = cpu_to_be64(super->s_victim_ino);
544         dynsb->ds_rename_dir    = cpu_to_be64(super->s_rename_dir);
545         dynsb->ds_rename_pos    = cpu_to_be64(super->s_rename_pos);
546         dynsb->ds_used_bytes    = cpu_to_be64(super->s_used_bytes);
547         dynsb->ds_generation    = cpu_to_be32(super->s_generation);
548         *type = JE_DYNSB;
549         *len = sizeof(*dynsb);
550         return dynsb;
551 }
552
553 static void write_wbuf(struct super_block *sb, struct logfs_area *area,
554                 void *wbuf)
555 {
556         struct logfs_super *super = logfs_super(sb);
557         struct address_space *mapping = super->s_mapping_inode->i_mapping;
558         u64 ofs;
559         pgoff_t index;
560         int page_ofs;
561         struct page *page;
562
563         ofs = dev_ofs(sb, area->a_segno,
564                         area->a_used_bytes & ~(super->s_writesize - 1));
565         index = ofs >> PAGE_SHIFT;
566         page_ofs = ofs & (PAGE_SIZE - 1);
567
568         page = find_lock_page(mapping, index);
569         BUG_ON(!page);
570         memcpy(wbuf, page_address(page) + page_ofs, super->s_writesize);
571         unlock_page(page);
572 }
573
574 static void *logfs_write_area(struct super_block *sb, void *_a,
575                 u16 *type, size_t *len)
576 {
577         struct logfs_super *super = logfs_super(sb);
578         struct logfs_area *area = super->s_area[super->s_sum_index];
579         struct logfs_je_area *a = _a;
580
581         a->vim = VIM_DEFAULT;
582         a->gc_level = super->s_sum_index;
583         a->used_bytes = cpu_to_be32(area->a_used_bytes);
584         a->segno = cpu_to_be32(area->a_segno);
585         if (super->s_writesize > 1)
586                 write_wbuf(sb, area, a + 1);
587
588         *type = JE_AREA;
589         *len = sizeof(*a) + super->s_writesize;
590         return a;
591 }
592
593 static void *logfs_write_commit(struct super_block *sb, void *h,
594                 u16 *type, size_t *len)
595 {
596         struct logfs_super *super = logfs_super(sb);
597
598         *type = JE_COMMIT;
599         *len = super->s_no_je * sizeof(__be64);
600         return super->s_je_array;
601 }
602
603 static size_t __logfs_write_je(struct super_block *sb, void *buf, u16 type,
604                 size_t len)
605 {
606         struct logfs_super *super = logfs_super(sb);
607         void *header = super->s_compressed_je;
608         void *data = header + sizeof(struct logfs_journal_header);
609         ssize_t compr_len, pad_len;
610         u8 compr = COMPR_ZLIB;
611
612         if (len == 0)
613                 return logfs_write_header(super, header, 0, type);
614
615         BUG_ON(len > sb->s_blocksize);
616         compr_len = logfs_compress(buf, data, len, sb->s_blocksize);
617         if (compr_len < 0 || type == JE_ANCHOR) {
618                 memcpy(data, buf, len);
619                 compr_len = len;
620                 compr = COMPR_NONE;
621         }
622
623         pad_len = ALIGN(compr_len, 16);
624         memset(data + compr_len, 0, pad_len - compr_len);
625
626         return __logfs_write_header(super, header, compr_len, len, type, compr);
627 }
628
629 static s64 logfs_get_free_bytes(struct logfs_area *area, size_t *bytes,
630                 int must_pad)
631 {
632         u32 writesize = logfs_super(area->a_sb)->s_writesize;
633         s32 ofs;
634         int ret;
635
636         ret = logfs_open_area(area, *bytes);
637         if (ret)
638                 return -EAGAIN;
639
640         ofs = area->a_used_bytes;
641         area->a_used_bytes += *bytes;
642
643         if (must_pad) {
644                 area->a_used_bytes = ALIGN(area->a_used_bytes, writesize);
645                 *bytes = area->a_used_bytes - ofs;
646         }
647
648         return dev_ofs(area->a_sb, area->a_segno, ofs);
649 }
650
651 static int logfs_write_je_buf(struct super_block *sb, void *buf, u16 type,
652                 size_t buf_len)
653 {
654         struct logfs_super *super = logfs_super(sb);
655         struct logfs_area *area = super->s_journal_area;
656         struct logfs_journal_header *jh = super->s_compressed_je;
657         size_t len;
658         int must_pad = 0;
659         s64 ofs;
660
661         len = __logfs_write_je(sb, buf, type, buf_len);
662         if (jh->h_type == cpu_to_be16(JE_COMMIT))
663                 must_pad = 1;
664
665         ofs = logfs_get_free_bytes(area, &len, must_pad);
666         if (ofs < 0)
667                 return ofs;
668         logfs_buf_write(area, ofs, super->s_compressed_je, len);
669         BUG_ON(super->s_no_je >= MAX_JOURNAL_ENTRIES);
670         super->s_je_array[super->s_no_je++] = cpu_to_be64(ofs);
671         return 0;
672 }
673
674 static int logfs_write_je(struct super_block *sb,
675                 void* (*write)(struct super_block *sb, void *scratch,
676                         u16 *type, size_t *len))
677 {
678         void *buf;
679         size_t len;
680         u16 type;
681
682         buf = write(sb, logfs_super(sb)->s_je, &type, &len);
683         return logfs_write_je_buf(sb, buf, type, len);
684 }
685
686 int write_alias_journal(struct super_block *sb, u64 ino, u64 bix,
687                 level_t level, int child_no, __be64 val)
688 {
689         struct logfs_super *super = logfs_super(sb);
690         struct logfs_obj_alias *oa = super->s_je;
691         int err = 0, fill = super->s_je_fill;
692
693         log_aliases("logfs_write_obj_aliases #%x(%llx, %llx, %x, %x) %llx\n",
694                         fill, ino, bix, level, child_no, be64_to_cpu(val));
695         oa[fill].ino = cpu_to_be64(ino);
696         oa[fill].bix = cpu_to_be64(bix);
697         oa[fill].val = val;
698         oa[fill].level = (__force u8)level;
699         oa[fill].child_no = cpu_to_be16(child_no);
700         fill++;
701         if (fill >= sb->s_blocksize / sizeof(*oa)) {
702                 err = logfs_write_je_buf(sb, oa, JE_OBJ_ALIAS, sb->s_blocksize);
703                 fill = 0;
704         }
705
706         super->s_je_fill = fill;
707         return err;
708 }
709
710 static int logfs_write_obj_aliases(struct super_block *sb)
711 {
712         struct logfs_super *super = logfs_super(sb);
713         int err;
714
715         log_journal("logfs_write_obj_aliases: %d aliases to write\n",
716                         super->s_no_object_aliases);
717         super->s_je_fill = 0;
718         err = logfs_write_obj_aliases_pagecache(sb);
719         if (err)
720                 return err;
721
722         if (super->s_je_fill)
723                 err = logfs_write_je_buf(sb, super->s_je, JE_OBJ_ALIAS,
724                                 super->s_je_fill
725                                 * sizeof(struct logfs_obj_alias));
726         return err;
727 }
728
729 /*
730  * Write all journal entries.  The goto logic ensures that all journal entries
731  * are written whenever a new segment is used.  It is ugly and potentially a
732  * bit wasteful, but robustness is more important.  With this we can *always*
733  * erase all journal segments except the one containing the most recent commit.
734  */
735 void logfs_write_anchor(struct super_block *sb)
736 {
737         struct logfs_super *super = logfs_super(sb);
738         struct logfs_area *area = super->s_journal_area;
739         int i, err;
740
741         if (!(super->s_flags & LOGFS_SB_FLAG_DIRTY))
742                 return;
743         super->s_flags &= ~LOGFS_SB_FLAG_DIRTY;
744
745         BUG_ON(super->s_flags & LOGFS_SB_FLAG_SHUTDOWN);
746         mutex_lock(&super->s_journal_mutex);
747
748         /* Do this first or suffer corruption */
749         logfs_sync_segments(sb);
750         account_shadows(sb);
751
752 again:
753         super->s_no_je = 0;
754         for_each_area(i) {
755                 if (!super->s_area[i]->a_is_open)
756                         continue;
757                 super->s_sum_index = i;
758                 err = logfs_write_je(sb, logfs_write_area);
759                 if (err)
760                         goto again;
761         }
762         err = logfs_write_obj_aliases(sb);
763         if (err)
764                 goto again;
765         err = logfs_write_je(sb, logfs_write_erasecount);
766         if (err)
767                 goto again;
768         err = logfs_write_je(sb, __logfs_write_anchor);
769         if (err)
770                 goto again;
771         err = logfs_write_je(sb, logfs_write_dynsb);
772         if (err)
773                 goto again;
774         /*
775          * Order is imperative.  First we sync all writes, including the
776          * non-committed journal writes.  Then we write the final commit and
777          * sync the current journal segment.
778          * There is a theoretical bug here.  Syncing the journal segment will
779          * write a number of journal entries and the final commit.  All these
780          * are written in a single operation.  If the device layer writes the
781          * data back-to-front, the commit will precede the other journal
782          * entries, leaving a race window.
783          * Two fixes are possible.  Preferred is to fix the device layer to
784          * ensure writes happen front-to-back.  Alternatively we can insert
785          * another logfs_sync_area() super->s_devops->sync() combo before
786          * writing the commit.
787          */
788         /*
789          * On another subject, super->s_devops->sync is usually not necessary.
790          * Unless called from sys_sync or friends, a barrier would suffice.
791          */
792         super->s_devops->sync(sb);
793         err = logfs_write_je(sb, logfs_write_commit);
794         if (err)
795                 goto again;
796         log_journal("Write commit to %llx\n",
797                         be64_to_cpu(super->s_je_array[super->s_no_je - 1]));
798         logfs_sync_area(area);
799         BUG_ON(area->a_used_bytes != area->a_written_bytes);
800         super->s_devops->sync(sb);
801
802         mutex_unlock(&super->s_journal_mutex);
803         return;
804 }
805
806 void do_logfs_journal_wl_pass(struct super_block *sb)
807 {
808         struct logfs_super *super = logfs_super(sb);
809         struct logfs_area *area = super->s_journal_area;
810         struct btree_head32 *head = &super->s_reserved_segments;
811         u32 segno, ec;
812         int i, err;
813
814         log_journal("Journal requires wear-leveling.\n");
815         /* Drop old segments */
816         journal_for_each(i)
817                 if (super->s_journal_seg[i]) {
818                         btree_remove32(head, super->s_journal_seg[i]);
819                         logfs_set_segment_unreserved(sb,
820                                         super->s_journal_seg[i],
821                                         super->s_journal_ec[i]);
822                         super->s_journal_seg[i] = 0;
823                         super->s_journal_ec[i] = 0;
824                 }
825         /* Get new segments */
826         for (i = 0; i < super->s_no_journal_segs; i++) {
827                 segno = get_best_cand(sb, &super->s_reserve_list, &ec);
828                 super->s_journal_seg[i] = segno;
829                 super->s_journal_ec[i] = ec;
830                 logfs_set_segment_reserved(sb, segno);
831                 err = btree_insert32(head, segno, (void *)1, GFP_KERNEL);
832                 BUG_ON(err); /* mempool should prevent this */
833                 err = logfs_erase_segment(sb, segno, 1);
834                 BUG_ON(err); /* FIXME: remount-ro would be nicer */
835         }
836         /* Manually move journal_area */
837         freeseg(sb, area->a_segno);
838         area->a_segno = super->s_journal_seg[0];
839         area->a_is_open = 0;
840         area->a_used_bytes = 0;
841         /* Write journal */
842         logfs_write_anchor(sb);
843         /* Write superblocks */
844         err = logfs_write_sb(sb);
845         BUG_ON(err);
846 }
847
848 static const struct logfs_area_ops journal_area_ops = {
849         .get_free_segment       = journal_get_free_segment,
850         .get_erase_count        = journal_get_erase_count,
851         .erase_segment          = journal_erase_segment,
852 };
853
854 int logfs_init_journal(struct super_block *sb)
855 {
856         struct logfs_super *super = logfs_super(sb);
857         size_t bufsize = max_t(size_t, sb->s_blocksize, super->s_writesize)
858                 + MAX_JOURNAL_HEADER;
859         int ret = -ENOMEM;
860
861         mutex_init(&super->s_journal_mutex);
862         btree_init_mempool32(&super->s_reserved_segments, super->s_btree_pool);
863
864         super->s_je = kzalloc(bufsize, GFP_KERNEL);
865         if (!super->s_je)
866                 return ret;
867
868         super->s_compressed_je = kzalloc(bufsize, GFP_KERNEL);
869         if (!super->s_compressed_je)
870                 return ret;
871
872         super->s_master_inode = logfs_new_meta_inode(sb, LOGFS_INO_MASTER);
873         if (IS_ERR(super->s_master_inode))
874                 return PTR_ERR(super->s_master_inode);
875
876         ret = logfs_read_journal(sb);
877         if (ret)
878                 return -EIO;
879
880         reserve_sb_and_journal(sb);
881         logfs_calc_free(sb);
882
883         super->s_journal_area->a_ops = &journal_area_ops;
884         return 0;
885 }
886
887 void logfs_cleanup_journal(struct super_block *sb)
888 {
889         struct logfs_super *super = logfs_super(sb);
890
891         btree_grim_visitor32(&super->s_reserved_segments, 0, NULL);
892         destroy_meta_inode(super->s_master_inode);
893         super->s_master_inode = NULL;
894
895         kfree(super->s_compressed_je);
896         kfree(super->s_je);
897 }