fb0a613f885b768c12a49ef52f1ab0fac42e12be
[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 #include <linux/slab.h>
10
11 static void logfs_calc_free(struct super_block *sb)
12 {
13         struct logfs_super *super = logfs_super(sb);
14         u64 reserve, no_segs = super->s_no_segs;
15         s64 free;
16         int i;
17
18         /* superblock segments */
19         no_segs -= 2;
20         super->s_no_journal_segs = 0;
21         /* journal */
22         journal_for_each(i)
23                 if (super->s_journal_seg[i]) {
24                         no_segs--;
25                         super->s_no_journal_segs++;
26                 }
27
28         /* open segments plus one extra per level for GC */
29         no_segs -= 2 * super->s_total_levels;
30
31         free = no_segs * (super->s_segsize - LOGFS_SEGMENT_RESERVE);
32         free -= super->s_used_bytes;
33         /* just a bit extra */
34         free -= super->s_total_levels * 4096;
35
36         /* Bad blocks are 'paid' for with speed reserve - the filesystem
37          * simply gets slower as bad blocks accumulate.  Until the bad blocks
38          * exceed the speed reserve - then the filesystem gets smaller.
39          */
40         reserve = super->s_bad_segments + super->s_bad_seg_reserve;
41         reserve *= super->s_segsize - LOGFS_SEGMENT_RESERVE;
42         reserve = max(reserve, super->s_speed_reserve);
43         free -= reserve;
44         if (free < 0)
45                 free = 0;
46
47         super->s_free_bytes = free;
48 }
49
50 static void reserve_sb_and_journal(struct super_block *sb)
51 {
52         struct logfs_super *super = logfs_super(sb);
53         struct btree_head32 *head = &super->s_reserved_segments;
54         int i, err;
55
56         err = btree_insert32(head, seg_no(sb, super->s_sb_ofs[0]), (void *)1,
57                         GFP_KERNEL);
58         BUG_ON(err);
59
60         err = btree_insert32(head, seg_no(sb, super->s_sb_ofs[1]), (void *)1,
61                         GFP_KERNEL);
62         BUG_ON(err);
63
64         journal_for_each(i) {
65                 if (!super->s_journal_seg[i])
66                         continue;
67                 err = btree_insert32(head, super->s_journal_seg[i], (void *)1,
68                                 GFP_KERNEL);
69                 BUG_ON(err);
70         }
71 }
72
73 static void read_dynsb(struct super_block *sb,
74                 struct logfs_je_dynsb *dynsb)
75 {
76         struct logfs_super *super = logfs_super(sb);
77
78         super->s_gec            = be64_to_cpu(dynsb->ds_gec);
79         super->s_sweeper        = be64_to_cpu(dynsb->ds_sweeper);
80         super->s_victim_ino     = be64_to_cpu(dynsb->ds_victim_ino);
81         super->s_rename_dir     = be64_to_cpu(dynsb->ds_rename_dir);
82         super->s_rename_pos     = be64_to_cpu(dynsb->ds_rename_pos);
83         super->s_used_bytes     = be64_to_cpu(dynsb->ds_used_bytes);
84         super->s_generation     = be32_to_cpu(dynsb->ds_generation);
85 }
86
87 static void read_anchor(struct super_block *sb,
88                 struct logfs_je_anchor *da)
89 {
90         struct logfs_super *super = logfs_super(sb);
91         struct inode *inode = super->s_master_inode;
92         struct logfs_inode *li = logfs_inode(inode);
93         int i;
94
95         super->s_last_ino = be64_to_cpu(da->da_last_ino);
96         li->li_flags    = 0;
97         li->li_height   = da->da_height;
98         i_size_write(inode, be64_to_cpu(da->da_size));
99         li->li_used_bytes = be64_to_cpu(da->da_used_bytes);
100
101         for (i = 0; i < LOGFS_EMBEDDED_FIELDS; i++)
102                 li->li_data[i] = be64_to_cpu(da->da_data[i]);
103 }
104
105 static void read_erasecount(struct super_block *sb,
106                 struct logfs_je_journal_ec *ec)
107 {
108         struct logfs_super *super = logfs_super(sb);
109         int i;
110
111         journal_for_each(i)
112                 super->s_journal_ec[i] = be32_to_cpu(ec->ec[i]);
113 }
114
115 static int read_area(struct super_block *sb, struct logfs_je_area *a)
116 {
117         struct logfs_super *super = logfs_super(sb);
118         struct logfs_area *area = super->s_area[a->gc_level];
119         u64 ofs;
120         u32 writemask = ~(super->s_writesize - 1);
121
122         if (a->gc_level >= LOGFS_NO_AREAS)
123                 return -EIO;
124         if (a->vim != VIM_DEFAULT)
125                 return -EIO; /* TODO: close area and continue */
126
127         area->a_used_bytes = be32_to_cpu(a->used_bytes);
128         area->a_written_bytes = area->a_used_bytes & writemask;
129         area->a_segno = be32_to_cpu(a->segno);
130         if (area->a_segno)
131                 area->a_is_open = 1;
132
133         ofs = dev_ofs(sb, area->a_segno, area->a_written_bytes);
134         if (super->s_writesize > 1)
135                 logfs_buf_recover(area, ofs, a + 1, super->s_writesize);
136         else
137                 logfs_buf_recover(area, ofs, NULL, 0);
138         return 0;
139 }
140
141 static void *unpack(void *from, void *to)
142 {
143         struct logfs_journal_header *jh = from;
144         void *data = from + sizeof(struct logfs_journal_header);
145         int err;
146         size_t inlen, outlen;
147
148         inlen = be16_to_cpu(jh->h_len);
149         outlen = be16_to_cpu(jh->h_datalen);
150
151         if (jh->h_compr == COMPR_NONE)
152                 memcpy(to, data, inlen);
153         else {
154                 err = logfs_uncompress(data, to, inlen, outlen);
155                 BUG_ON(err);
156         }
157         return to;
158 }
159
160 static int __read_je_header(struct super_block *sb, u64 ofs,
161                 struct logfs_journal_header *jh)
162 {
163         struct logfs_super *super = logfs_super(sb);
164         size_t bufsize = max_t(size_t, sb->s_blocksize, super->s_writesize)
165                 + MAX_JOURNAL_HEADER;
166         u16 type, len, datalen;
167         int err;
168
169         /* read header only */
170         err = wbuf_read(sb, ofs, sizeof(*jh), jh);
171         if (err)
172                 return err;
173         type = be16_to_cpu(jh->h_type);
174         len = be16_to_cpu(jh->h_len);
175         datalen = be16_to_cpu(jh->h_datalen);
176         if (len > sb->s_blocksize)
177                 return -EIO;
178         if ((type < JE_FIRST) || (type > JE_LAST))
179                 return -EIO;
180         if (datalen > bufsize)
181                 return -EIO;
182         return 0;
183 }
184
185 static int __read_je_payload(struct super_block *sb, u64 ofs,
186                 struct logfs_journal_header *jh)
187 {
188         u16 len;
189         int err;
190
191         len = be16_to_cpu(jh->h_len);
192         err = wbuf_read(sb, ofs + sizeof(*jh), len, jh + 1);
193         if (err)
194                 return err;
195         if (jh->h_crc != logfs_crc32(jh, len + sizeof(*jh), 4)) {
196                 /* Old code was confused.  It forgot about the header length
197                  * and stopped calculating the crc 16 bytes before the end
198                  * of data - ick!
199                  * FIXME: Remove this hack once the old code is fixed.
200                  */
201                 if (jh->h_crc == logfs_crc32(jh, len, 4))
202                         WARN_ON_ONCE(1);
203                 else
204                         return -EIO;
205         }
206         return 0;
207 }
208
209 /*
210  * jh needs to be large enough to hold the complete entry, not just the header
211  */
212 static int __read_je(struct super_block *sb, u64 ofs,
213                 struct logfs_journal_header *jh)
214 {
215         int err;
216
217         err = __read_je_header(sb, ofs, jh);
218         if (err)
219                 return err;
220         return __read_je_payload(sb, ofs, jh);
221 }
222
223 static int read_je(struct super_block *sb, u64 ofs)
224 {
225         struct logfs_super *super = logfs_super(sb);
226         struct logfs_journal_header *jh = super->s_compressed_je;
227         void *scratch = super->s_je;
228         u16 type, datalen;
229         int err;
230
231         err = __read_je(sb, ofs, jh);
232         if (err)
233                 return err;
234         type = be16_to_cpu(jh->h_type);
235         datalen = be16_to_cpu(jh->h_datalen);
236
237         switch (type) {
238         case JE_DYNSB:
239                 read_dynsb(sb, unpack(jh, scratch));
240                 break;
241         case JE_ANCHOR:
242                 read_anchor(sb, unpack(jh, scratch));
243                 break;
244         case JE_ERASECOUNT:
245                 read_erasecount(sb, unpack(jh, scratch));
246                 break;
247         case JE_AREA:
248                 read_area(sb, unpack(jh, scratch));
249                 break;
250         case JE_OBJ_ALIAS:
251                 err = logfs_load_object_aliases(sb, unpack(jh, scratch),
252                                 datalen);
253                 break;
254         default:
255                 WARN_ON_ONCE(1);
256                 return -EIO;
257         }
258         return err;
259 }
260
261 static int logfs_read_segment(struct super_block *sb, u32 segno)
262 {
263         struct logfs_super *super = logfs_super(sb);
264         struct logfs_journal_header *jh = super->s_compressed_je;
265         u64 ofs, seg_ofs = dev_ofs(sb, segno, 0);
266         u32 h_ofs, last_ofs = 0;
267         u16 len, datalen, last_len = 0;
268         int i, err;
269
270         /* search for most recent commit */
271         for (h_ofs = 0; h_ofs < super->s_segsize; h_ofs += sizeof(*jh)) {
272                 ofs = seg_ofs + h_ofs;
273                 err = __read_je_header(sb, ofs, jh);
274                 if (err)
275                         continue;
276                 if (jh->h_type != cpu_to_be16(JE_COMMIT))
277                         continue;
278                 err = __read_je_payload(sb, ofs, jh);
279                 if (err)
280                         continue;
281                 len = be16_to_cpu(jh->h_len);
282                 datalen = be16_to_cpu(jh->h_datalen);
283                 if ((datalen > sizeof(super->s_je_array)) ||
284                                 (datalen % sizeof(__be64)))
285                         continue;
286                 last_ofs = h_ofs;
287                 last_len = datalen;
288                 h_ofs += ALIGN(len, sizeof(*jh)) - sizeof(*jh);
289         }
290         /* read commit */
291         if (last_ofs == 0)
292                 return -ENOENT;
293         ofs = seg_ofs + last_ofs;
294         log_journal("Read commit from %llx\n", ofs);
295         err = __read_je(sb, ofs, jh);
296         BUG_ON(err); /* We should have caught it in the scan loop already */
297         if (err)
298                 return err;
299         /* uncompress */
300         unpack(jh, super->s_je_array);
301         super->s_no_je = last_len / sizeof(__be64);
302         /* iterate over array */
303         for (i = 0; i < super->s_no_je; i++) {
304                 err = read_je(sb, be64_to_cpu(super->s_je_array[i]));
305                 if (err)
306                         return err;
307         }
308         super->s_journal_area->a_segno = segno;
309         return 0;
310 }
311
312 static u64 read_gec(struct super_block *sb, u32 segno)
313 {
314         struct logfs_segment_header sh;
315         __be32 crc;
316         int err;
317
318         if (!segno)
319                 return 0;
320         err = wbuf_read(sb, dev_ofs(sb, segno, 0), sizeof(sh), &sh);
321         if (err)
322                 return 0;
323         crc = logfs_crc32(&sh, sizeof(sh), 4);
324         if (crc != sh.crc) {
325                 WARN_ON(sh.gec != cpu_to_be64(0xffffffffffffffffull));
326                 /* Most likely it was just erased */
327                 return 0;
328         }
329         return be64_to_cpu(sh.gec);
330 }
331
332 static int logfs_read_journal(struct super_block *sb)
333 {
334         struct logfs_super *super = logfs_super(sb);
335         u64 gec[LOGFS_JOURNAL_SEGS], max;
336         u32 segno;
337         int i, max_i;
338
339         max = 0;
340         max_i = -1;
341         journal_for_each(i) {
342                 segno = super->s_journal_seg[i];
343                 gec[i] = read_gec(sb, super->s_journal_seg[i]);
344                 if (gec[i] > max) {
345                         max = gec[i];
346                         max_i = i;
347                 }
348         }
349         if (max_i == -1)
350                 return -EIO;
351         /* FIXME: Try older segments in case of error */
352         return logfs_read_segment(sb, super->s_journal_seg[max_i]);
353 }
354
355 /*
356  * First search the current segment (outer loop), then pick the next segment
357  * in the array, skipping any zero entries (inner loop).
358  */
359 static void journal_get_free_segment(struct logfs_area *area)
360 {
361         struct logfs_super *super = logfs_super(area->a_sb);
362         int i;
363
364         journal_for_each(i) {
365                 if (area->a_segno != super->s_journal_seg[i])
366                         continue;
367
368                 do {
369                         i++;
370                         if (i == LOGFS_JOURNAL_SEGS)
371                                 i = 0;
372                 } while (!super->s_journal_seg[i]);
373
374                 area->a_segno = super->s_journal_seg[i];
375                 area->a_erase_count = ++(super->s_journal_ec[i]);
376                 log_journal("Journal now at %x (ec %x)\n", area->a_segno,
377                                 area->a_erase_count);
378                 return;
379         }
380         BUG();
381 }
382
383 static void journal_get_erase_count(struct logfs_area *area)
384 {
385         /* erase count is stored globally and incremented in
386          * journal_get_free_segment() - nothing to do here */
387 }
388
389 static int journal_erase_segment(struct logfs_area *area)
390 {
391         struct super_block *sb = area->a_sb;
392         union {
393                 struct logfs_segment_header sh;
394                 unsigned char c[ALIGN(sizeof(struct logfs_segment_header), 16)];
395         } u;
396         u64 ofs;
397         int err;
398
399         err = logfs_erase_segment(sb, area->a_segno, 1);
400         if (err)
401                 return err;
402
403         memset(&u, 0, sizeof(u));
404         u.sh.pad = 0;
405         u.sh.type = SEG_JOURNAL;
406         u.sh.level = 0;
407         u.sh.segno = cpu_to_be32(area->a_segno);
408         u.sh.ec = cpu_to_be32(area->a_erase_count);
409         u.sh.gec = cpu_to_be64(logfs_super(sb)->s_gec);
410         u.sh.crc = logfs_crc32(&u.sh, sizeof(u.sh), 4);
411
412         /* This causes a bug in segment.c.  Not yet. */
413         //logfs_set_segment_erased(sb, area->a_segno, area->a_erase_count, 0);
414
415         ofs = dev_ofs(sb, area->a_segno, 0);
416         area->a_used_bytes = sizeof(u);
417         logfs_buf_write(area, ofs, &u, sizeof(u));
418         return 0;
419 }
420
421 static size_t __logfs_write_header(struct logfs_super *super,
422                 struct logfs_journal_header *jh, size_t len, size_t datalen,
423                 u16 type, u8 compr)
424 {
425         jh->h_len       = cpu_to_be16(len);
426         jh->h_type      = cpu_to_be16(type);
427         jh->h_datalen   = cpu_to_be16(datalen);
428         jh->h_compr     = compr;
429         jh->h_pad[0]    = 'H';
430         jh->h_pad[1]    = 'E';
431         jh->h_pad[2]    = 'A';
432         jh->h_pad[3]    = 'D';
433         jh->h_pad[4]    = 'R';
434         jh->h_crc       = logfs_crc32(jh, len + sizeof(*jh), 4);
435         return ALIGN(len, 16) + sizeof(*jh);
436 }
437
438 static size_t logfs_write_header(struct logfs_super *super,
439                 struct logfs_journal_header *jh, size_t datalen, u16 type)
440 {
441         size_t len = datalen;
442
443         return __logfs_write_header(super, jh, len, datalen, type, COMPR_NONE);
444 }
445
446 static inline size_t logfs_journal_erasecount_size(struct logfs_super *super)
447 {
448         return LOGFS_JOURNAL_SEGS * sizeof(__be32);
449 }
450
451 static void *logfs_write_erasecount(struct super_block *sb, void *_ec,
452                 u16 *type, size_t *len)
453 {
454         struct logfs_super *super = logfs_super(sb);
455         struct logfs_je_journal_ec *ec = _ec;
456         int i;
457
458         journal_for_each(i)
459                 ec->ec[i] = cpu_to_be32(super->s_journal_ec[i]);
460         *type = JE_ERASECOUNT;
461         *len = logfs_journal_erasecount_size(super);
462         return ec;
463 }
464
465 static void account_shadow(void *_shadow, unsigned long _sb, u64 ignore,
466                 size_t ignore2)
467 {
468         struct logfs_shadow *shadow = _shadow;
469         struct super_block *sb = (void *)_sb;
470         struct logfs_super *super = logfs_super(sb);
471
472         /* consume new space */
473         super->s_free_bytes       -= shadow->new_len;
474         super->s_used_bytes       += shadow->new_len;
475         super->s_dirty_used_bytes -= shadow->new_len;
476
477         /* free up old space */
478         super->s_free_bytes       += shadow->old_len;
479         super->s_used_bytes       -= shadow->old_len;
480         super->s_dirty_free_bytes -= shadow->old_len;
481
482         logfs_set_segment_used(sb, shadow->old_ofs, -shadow->old_len);
483         logfs_set_segment_used(sb, shadow->new_ofs, shadow->new_len);
484
485         log_journal("account_shadow(%llx, %llx, %x) %llx->%llx %x->%x\n",
486                         shadow->ino, shadow->bix, shadow->gc_level,
487                         shadow->old_ofs, shadow->new_ofs,
488                         shadow->old_len, shadow->new_len);
489         mempool_free(shadow, super->s_shadow_pool);
490 }
491
492 static void account_shadows(struct super_block *sb)
493 {
494         struct logfs_super *super = logfs_super(sb);
495         struct inode *inode = super->s_master_inode;
496         struct logfs_inode *li = logfs_inode(inode);
497         struct shadow_tree *tree = &super->s_shadow_tree;
498
499         btree_grim_visitor64(&tree->new, (unsigned long)sb, account_shadow);
500         btree_grim_visitor64(&tree->old, (unsigned long)sb, account_shadow);
501         btree_grim_visitor32(&tree->segment_map, 0, NULL);
502         tree->no_shadowed_segments = 0;
503
504         if (li->li_block) {
505                 /*
506                  * We never actually use the structure, when attached to the
507                  * master inode.  But it is easier to always free it here than
508                  * to have checks in several places elsewhere when allocating
509                  * it.
510                  */
511                 li->li_block->ops->free_block(sb, li->li_block);
512         }
513         BUG_ON((s64)li->li_used_bytes < 0);
514 }
515
516 static void *__logfs_write_anchor(struct super_block *sb, void *_da,
517                 u16 *type, size_t *len)
518 {
519         struct logfs_super *super = logfs_super(sb);
520         struct logfs_je_anchor *da = _da;
521         struct inode *inode = super->s_master_inode;
522         struct logfs_inode *li = logfs_inode(inode);
523         int i;
524
525         da->da_height   = li->li_height;
526         da->da_last_ino = cpu_to_be64(super->s_last_ino);
527         da->da_size     = cpu_to_be64(i_size_read(inode));
528         da->da_used_bytes = cpu_to_be64(li->li_used_bytes);
529         for (i = 0; i < LOGFS_EMBEDDED_FIELDS; i++)
530                 da->da_data[i] = cpu_to_be64(li->li_data[i]);
531         *type = JE_ANCHOR;
532         *len = sizeof(*da);
533         return da;
534 }
535
536 static void *logfs_write_dynsb(struct super_block *sb, void *_dynsb,
537                 u16 *type, size_t *len)
538 {
539         struct logfs_super *super = logfs_super(sb);
540         struct logfs_je_dynsb *dynsb = _dynsb;
541
542         dynsb->ds_gec           = cpu_to_be64(super->s_gec);
543         dynsb->ds_sweeper       = cpu_to_be64(super->s_sweeper);
544         dynsb->ds_victim_ino    = cpu_to_be64(super->s_victim_ino);
545         dynsb->ds_rename_dir    = cpu_to_be64(super->s_rename_dir);
546         dynsb->ds_rename_pos    = cpu_to_be64(super->s_rename_pos);
547         dynsb->ds_used_bytes    = cpu_to_be64(super->s_used_bytes);
548         dynsb->ds_generation    = cpu_to_be32(super->s_generation);
549         *type = JE_DYNSB;
550         *len = sizeof(*dynsb);
551         return dynsb;
552 }
553
554 static void write_wbuf(struct super_block *sb, struct logfs_area *area,
555                 void *wbuf)
556 {
557         struct logfs_super *super = logfs_super(sb);
558         struct address_space *mapping = super->s_mapping_inode->i_mapping;
559         u64 ofs;
560         pgoff_t index;
561         int page_ofs;
562         struct page *page;
563
564         ofs = dev_ofs(sb, area->a_segno,
565                         area->a_used_bytes & ~(super->s_writesize - 1));
566         index = ofs >> PAGE_SHIFT;
567         page_ofs = ofs & (PAGE_SIZE - 1);
568
569         page = find_lock_page(mapping, index);
570         BUG_ON(!page);
571         memcpy(wbuf, page_address(page) + page_ofs, super->s_writesize);
572         unlock_page(page);
573 }
574
575 static void *logfs_write_area(struct super_block *sb, void *_a,
576                 u16 *type, size_t *len)
577 {
578         struct logfs_super *super = logfs_super(sb);
579         struct logfs_area *area = super->s_area[super->s_sum_index];
580         struct logfs_je_area *a = _a;
581
582         a->vim = VIM_DEFAULT;
583         a->gc_level = super->s_sum_index;
584         a->used_bytes = cpu_to_be32(area->a_used_bytes);
585         a->segno = cpu_to_be32(area->a_segno);
586         if (super->s_writesize > 1)
587                 write_wbuf(sb, area, a + 1);
588
589         *type = JE_AREA;
590         *len = sizeof(*a) + super->s_writesize;
591         return a;
592 }
593
594 static void *logfs_write_commit(struct super_block *sb, void *h,
595                 u16 *type, size_t *len)
596 {
597         struct logfs_super *super = logfs_super(sb);
598
599         *type = JE_COMMIT;
600         *len = super->s_no_je * sizeof(__be64);
601         return super->s_je_array;
602 }
603
604 static size_t __logfs_write_je(struct super_block *sb, void *buf, u16 type,
605                 size_t len)
606 {
607         struct logfs_super *super = logfs_super(sb);
608         void *header = super->s_compressed_je;
609         void *data = header + sizeof(struct logfs_journal_header);
610         ssize_t compr_len, pad_len;
611         u8 compr = COMPR_ZLIB;
612
613         if (len == 0)
614                 return logfs_write_header(super, header, 0, type);
615
616         BUG_ON(len > sb->s_blocksize);
617         compr_len = logfs_compress(buf, data, len, sb->s_blocksize);
618         if (compr_len < 0 || type == JE_ANCHOR) {
619                 memcpy(data, buf, len);
620                 compr_len = len;
621                 compr = COMPR_NONE;
622         }
623
624         pad_len = ALIGN(compr_len, 16);
625         memset(data + compr_len, 0, pad_len - compr_len);
626
627         return __logfs_write_header(super, header, compr_len, len, type, compr);
628 }
629
630 static s64 logfs_get_free_bytes(struct logfs_area *area, size_t *bytes,
631                 int must_pad)
632 {
633         u32 writesize = logfs_super(area->a_sb)->s_writesize;
634         s32 ofs;
635         int ret;
636
637         ret = logfs_open_area(area, *bytes);
638         if (ret)
639                 return -EAGAIN;
640
641         ofs = area->a_used_bytes;
642         area->a_used_bytes += *bytes;
643
644         if (must_pad) {
645                 area->a_used_bytes = ALIGN(area->a_used_bytes, writesize);
646                 *bytes = area->a_used_bytes - ofs;
647         }
648
649         return dev_ofs(area->a_sb, area->a_segno, ofs);
650 }
651
652 static int logfs_write_je_buf(struct super_block *sb, void *buf, u16 type,
653                 size_t buf_len)
654 {
655         struct logfs_super *super = logfs_super(sb);
656         struct logfs_area *area = super->s_journal_area;
657         struct logfs_journal_header *jh = super->s_compressed_je;
658         size_t len;
659         int must_pad = 0;
660         s64 ofs;
661
662         len = __logfs_write_je(sb, buf, type, buf_len);
663         if (jh->h_type == cpu_to_be16(JE_COMMIT))
664                 must_pad = 1;
665
666         ofs = logfs_get_free_bytes(area, &len, must_pad);
667         if (ofs < 0)
668                 return ofs;
669         logfs_buf_write(area, ofs, super->s_compressed_je, len);
670         BUG_ON(super->s_no_je >= MAX_JOURNAL_ENTRIES);
671         super->s_je_array[super->s_no_je++] = cpu_to_be64(ofs);
672         return 0;
673 }
674
675 static int logfs_write_je(struct super_block *sb,
676                 void* (*write)(struct super_block *sb, void *scratch,
677                         u16 *type, size_t *len))
678 {
679         void *buf;
680         size_t len;
681         u16 type;
682
683         buf = write(sb, logfs_super(sb)->s_je, &type, &len);
684         return logfs_write_je_buf(sb, buf, type, len);
685 }
686
687 int write_alias_journal(struct super_block *sb, u64 ino, u64 bix,
688                 level_t level, int child_no, __be64 val)
689 {
690         struct logfs_super *super = logfs_super(sb);
691         struct logfs_obj_alias *oa = super->s_je;
692         int err = 0, fill = super->s_je_fill;
693
694         log_aliases("logfs_write_obj_aliases #%x(%llx, %llx, %x, %x) %llx\n",
695                         fill, ino, bix, level, child_no, be64_to_cpu(val));
696         oa[fill].ino = cpu_to_be64(ino);
697         oa[fill].bix = cpu_to_be64(bix);
698         oa[fill].val = val;
699         oa[fill].level = (__force u8)level;
700         oa[fill].child_no = cpu_to_be16(child_no);
701         fill++;
702         if (fill >= sb->s_blocksize / sizeof(*oa)) {
703                 err = logfs_write_je_buf(sb, oa, JE_OBJ_ALIAS, sb->s_blocksize);
704                 fill = 0;
705         }
706
707         super->s_je_fill = fill;
708         return err;
709 }
710
711 static int logfs_write_obj_aliases(struct super_block *sb)
712 {
713         struct logfs_super *super = logfs_super(sb);
714         int err;
715
716         log_journal("logfs_write_obj_aliases: %d aliases to write\n",
717                         super->s_no_object_aliases);
718         super->s_je_fill = 0;
719         err = logfs_write_obj_aliases_pagecache(sb);
720         if (err)
721                 return err;
722
723         if (super->s_je_fill)
724                 err = logfs_write_je_buf(sb, super->s_je, JE_OBJ_ALIAS,
725                                 super->s_je_fill
726                                 * sizeof(struct logfs_obj_alias));
727         return err;
728 }
729
730 /*
731  * Write all journal entries.  The goto logic ensures that all journal entries
732  * are written whenever a new segment is used.  It is ugly and potentially a
733  * bit wasteful, but robustness is more important.  With this we can *always*
734  * erase all journal segments except the one containing the most recent commit.
735  */
736 void logfs_write_anchor(struct super_block *sb)
737 {
738         struct logfs_super *super = logfs_super(sb);
739         struct logfs_area *area = super->s_journal_area;
740         int i, err;
741
742         if (!(super->s_flags & LOGFS_SB_FLAG_DIRTY))
743                 return;
744         super->s_flags &= ~LOGFS_SB_FLAG_DIRTY;
745
746         BUG_ON(super->s_flags & LOGFS_SB_FLAG_SHUTDOWN);
747         mutex_lock(&super->s_journal_mutex);
748
749         /* Do this first or suffer corruption */
750         logfs_sync_segments(sb);
751         account_shadows(sb);
752
753 again:
754         super->s_no_je = 0;
755         for_each_area(i) {
756                 if (!super->s_area[i]->a_is_open)
757                         continue;
758                 super->s_sum_index = i;
759                 err = logfs_write_je(sb, logfs_write_area);
760                 if (err)
761                         goto again;
762         }
763         err = logfs_write_obj_aliases(sb);
764         if (err)
765                 goto again;
766         err = logfs_write_je(sb, logfs_write_erasecount);
767         if (err)
768                 goto again;
769         err = logfs_write_je(sb, __logfs_write_anchor);
770         if (err)
771                 goto again;
772         err = logfs_write_je(sb, logfs_write_dynsb);
773         if (err)
774                 goto again;
775         /*
776          * Order is imperative.  First we sync all writes, including the
777          * non-committed journal writes.  Then we write the final commit and
778          * sync the current journal segment.
779          * There is a theoretical bug here.  Syncing the journal segment will
780          * write a number of journal entries and the final commit.  All these
781          * are written in a single operation.  If the device layer writes the
782          * data back-to-front, the commit will precede the other journal
783          * entries, leaving a race window.
784          * Two fixes are possible.  Preferred is to fix the device layer to
785          * ensure writes happen front-to-back.  Alternatively we can insert
786          * another logfs_sync_area() super->s_devops->sync() combo before
787          * writing the commit.
788          */
789         /*
790          * On another subject, super->s_devops->sync is usually not necessary.
791          * Unless called from sys_sync or friends, a barrier would suffice.
792          */
793         super->s_devops->sync(sb);
794         err = logfs_write_je(sb, logfs_write_commit);
795         if (err)
796                 goto again;
797         log_journal("Write commit to %llx\n",
798                         be64_to_cpu(super->s_je_array[super->s_no_je - 1]));
799         logfs_sync_area(area);
800         BUG_ON(area->a_used_bytes != area->a_written_bytes);
801         super->s_devops->sync(sb);
802
803         mutex_unlock(&super->s_journal_mutex);
804         return;
805 }
806
807 void do_logfs_journal_wl_pass(struct super_block *sb)
808 {
809         struct logfs_super *super = logfs_super(sb);
810         struct logfs_area *area = super->s_journal_area;
811         struct btree_head32 *head = &super->s_reserved_segments;
812         u32 segno, ec;
813         int i, err;
814
815         log_journal("Journal requires wear-leveling.\n");
816         /* Drop old segments */
817         journal_for_each(i)
818                 if (super->s_journal_seg[i]) {
819                         btree_remove32(head, super->s_journal_seg[i]);
820                         logfs_set_segment_unreserved(sb,
821                                         super->s_journal_seg[i],
822                                         super->s_journal_ec[i]);
823                         super->s_journal_seg[i] = 0;
824                         super->s_journal_ec[i] = 0;
825                 }
826         /* Get new segments */
827         for (i = 0; i < super->s_no_journal_segs; i++) {
828                 segno = get_best_cand(sb, &super->s_reserve_list, &ec);
829                 super->s_journal_seg[i] = segno;
830                 super->s_journal_ec[i] = ec;
831                 logfs_set_segment_reserved(sb, segno);
832                 err = btree_insert32(head, segno, (void *)1, GFP_KERNEL);
833                 BUG_ON(err); /* mempool should prevent this */
834                 err = logfs_erase_segment(sb, segno, 1);
835                 BUG_ON(err); /* FIXME: remount-ro would be nicer */
836         }
837         /* Manually move journal_area */
838         freeseg(sb, area->a_segno);
839         area->a_segno = super->s_journal_seg[0];
840         area->a_is_open = 0;
841         area->a_used_bytes = 0;
842         /* Write journal */
843         logfs_write_anchor(sb);
844         /* Write superblocks */
845         err = logfs_write_sb(sb);
846         BUG_ON(err);
847 }
848
849 static const struct logfs_area_ops journal_area_ops = {
850         .get_free_segment       = journal_get_free_segment,
851         .get_erase_count        = journal_get_erase_count,
852         .erase_segment          = journal_erase_segment,
853 };
854
855 int logfs_init_journal(struct super_block *sb)
856 {
857         struct logfs_super *super = logfs_super(sb);
858         size_t bufsize = max_t(size_t, sb->s_blocksize, super->s_writesize)
859                 + MAX_JOURNAL_HEADER;
860         int ret = -ENOMEM;
861
862         mutex_init(&super->s_journal_mutex);
863         btree_init_mempool32(&super->s_reserved_segments, super->s_btree_pool);
864
865         super->s_je = kzalloc(bufsize, GFP_KERNEL);
866         if (!super->s_je)
867                 return ret;
868
869         super->s_compressed_je = kzalloc(bufsize, GFP_KERNEL);
870         if (!super->s_compressed_je)
871                 return ret;
872
873         super->s_master_inode = logfs_new_meta_inode(sb, LOGFS_INO_MASTER);
874         if (IS_ERR(super->s_master_inode))
875                 return PTR_ERR(super->s_master_inode);
876
877         ret = logfs_read_journal(sb);
878         if (ret)
879                 return -EIO;
880
881         reserve_sb_and_journal(sb);
882         logfs_calc_free(sb);
883
884         super->s_journal_area->a_ops = &journal_area_ops;
885         return 0;
886 }
887
888 void logfs_cleanup_journal(struct super_block *sb)
889 {
890         struct logfs_super *super = logfs_super(sb);
891
892         btree_grim_visitor32(&super->s_reserved_segments, 0, NULL);
893         destroy_meta_inode(super->s_master_inode);
894         super->s_master_inode = NULL;
895
896         kfree(super->s_compressed_je);
897         kfree(super->s_je);
898 }