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