inode.c

   1
   2 /*
   3  * fs/lafs/inode.c
   4  * Copyright (C) 2005-2009
   5  * Neil Brown <neilb@suse.de>
   6  * Released under the GPL, version 2
   7  *
   8  * generic inode handling
   9  *
  10  */
  11
  12 #include        "lafs.h"
  13 #include <linux/random.h>
  14 #include <linux/delay.h>
  15 #include <linux/slab.h>
  16
  17 static void check_atime_ref(struct inode *ino, int async);
  18
  19 /* Supporting an async 'iget' - as required by the cleaner -
  20  * is slightly non-trivial.
  21  * iget*_locked will normally wait for any inode with one
  22  * of the flags I_FREEING I_CLEAR I_WILL_FREE I_NEW
  23  * to either be unhashed or have the flag cleared.
  24  * We cannot afford that wait in the cleaner as we could deadlock.
  25  * So we use iget5_locked and provide a test function that fails
  26  * if it finds the inode with any of those flags set.
  27  * If it does see the inode like that it sets a flag in the 'ikey'
  28  * that is passed in by reference so that it knows to continue
  29  * failing (for consistency) and so that the 'set' function
  30  * we provide can know to fail the 'set'.
  31  * The result of this is that if iget finds an inode it would
  32  * have to wait on, a flag is set and NULL is returned.
  33  * An unfortunate side effect is that an inode will be allocated
  34  * and then destroyed to no avail.
  35  * This is avoided by calling ilookup5 first.  This also allows
  36  * us to only allocate/load the data block if there really seems
  37  * to be a need.
  38  */
  39 struct ikey {
  40         ino_t inum;
  41         struct inode *fsys;
  42         bool was_busy;
  43 };
  44
  45 static int sync_itest(struct inode *inode, void *data)
  46 {
  47         struct ikey *ik = data;
  48
  49         if (inode->i_ino != ik->inum ||
  50             LAFSI(inode)->filesys != ik->fsys)
  51                 return 0;
  52         return 1;
  53 }
  54
  55 static int async_itest(struct inode *inode, void *data)
  56 {
  57         struct ikey *ik = data;
  58
  59         if (ik->was_busy)
  60                 /* found and is freeing */
  61                 return 0;
  62         if (!sync_itest(inode, data))
  63                 return 0;
  64         if (inode->i_state & (I_FREEING|I_CLEAR|I_WILL_FREE|I_NEW)) {
  65                 ik->was_busy = true;
  66                 return 0;
  67         }
  68         return 1;
  69 }
  70
  71 static int iset(struct inode *inode, void *data)
  72 {
  73         struct ikey *ik = data;
  74         if (ik->was_busy)
  75                 return -EBUSY;
  76         inode->i_ino = ik->inum;
  77         LAFSI(inode)->filesys = ik->fsys;
  78         return 0;
  79 }
  80
  81 struct inode *
  82 lafs_iget(struct inode *fsys, ino_t inum, int async)
  83 {
  84         /* find, and load if needed, this inum */
  85         struct inode *ino = NULL;
  86         struct inode *oldino;
  87         struct datablock *b = NULL;
  88         struct ikey ik = { .inum = inum, .fsys = fsys, };
  89         int err = 0;
  90         struct super_block *sb = fsys->i_sb;
  91
  92         if (async) {
  93                 /* We cannot afford to block on 'freeing_inode'
  94                  * So use iget5_locked and refuse to match such
  95                  * inodes.
  96                  * If the inode is 'freeing', inum gets set to NO_INO.
  97                  * ilookup5 is used first to avoid an unnecessary
  98                  * alloc/free if the inode is locked in some way.
  99                  */
 100                 while (!ino) {
 101
 102                         err = 0;
 103                         ino = ilookup5(sb, inum, async_itest, &ik);
 104                         if (ino)
 105                                 break;
 106
 107                         if (ik.was_busy)
 108                                 err = -EAGAIN;
 109
 110                         /* For async we will always want the dblock loaded,
 111                          * and we need to load it first as we cannot afford
 112                          * to fail -EAGAIN once we have an I_NEW inode.
 113                          */
 114                         if (!b)
 115                                 b = lafs_get_block(fsys, inum, NULL,
 116                                                    GFP_NOFS, MKREF(iget));
 117                         if (!b)
 118                                 return ERR_PTR(-ENOMEM);
 119
 120                         if (!err)
 121                                 err = lafs_read_block_async(b);
 122
 123                         if (!err) {
 124                                 /* Have the block, so safe to iget */
 125                                 ino = iget5_locked(sb, inum,
 126                                                    async_itest, iset,
 127                                                    &ik);
 128                                 if (!ino) {
 129                                         if (ik.was_busy)
 130                                                 err = -EAGAIN;
 131                                         else
 132                                                 err = -ENOMEM;
 133                                 }
 134                         }
 135                         if (err) {
 136                                 if (test_and_set_bit(B_Async, &b->b.flags)) {
 137                                         putdref(b, MKREF(iget));
 138                                         return ERR_PTR(err);
 139                                 }
 140                                 getdref(b, MKREF(async));
 141                         }
 142                 }
 143         } else
 144                 ino = iget5_locked(sb, inum, sync_itest, iset, &ik);
 145
 146         if (!ino) {
 147                 putdref(b, MKREF(iget));
 148                 return ERR_PTR(-ENOMEM);
 149         }
 150
 151         if (!(ino->i_state & I_NEW)) {
 152                 putdref(b, MKREF(iget));
 153                 if (ino->i_mode) {
 154                         check_atime_ref(ino, async);
 155                         return ino;
 156                 }
 157                 iput(ino);
 158                 return ERR_PTR(-ENOENT);
 159         }
 160
 161         igrab(LAFSI(ino)->filesys);
 162
 163         /* surprisingly the inode bdi does not default to the
 164          * super_blocks bdi...
 165          */
 166         ino->i_data.backing_dev_info = sb->s_bdi;
 167         /* Need to load block 'inum' from an inode file...
 168          */
 169         if (!b) {
 170                 b = lafs_get_block(fsys, inum, NULL, GFP_KERNEL, MKREF(iget));
 171                 if (!b)
 172                         err = -ENOMEM;
 173                 else
 174                         err = lafs_read_block(b);
 175         }
 176         if (err)
 177                 goto err;
 178
 179         oldino = rcu_my_inode(b);
 180         if (oldino) {
 181                 /* The inode is new, but the block thinks it has an
 182                  * old inode, so we must be in the process of destroying
 183                  * the old one.
 184                  * So fail the lookup without even looking at the content
 185                  * of the block (Which might not be clear yet).
 186                  */
 187                 spin_lock(&oldino->i_data.private_lock);
 188                 if (!test_bit(I_Deleting, &LAFSI(oldino)->iflags)) {
 189                         b->my_inode = NULL;
 190                         LAFSI(oldino)->dblock = NULL;
 191                         LAFSI(oldino)->iblock = NULL;
 192                 }
 193                 spin_unlock(&oldino->i_data.private_lock);
 194         }
 195         rcu_iput(oldino);
 196         if (b->my_inode) {
 197                 err = -ENOENT;
 198                 goto err;
 199         }
 200
 201         err = lafs_import_inode(ino, b);
 202         if (err) {
 203                 if (err != -ENOENT)
 204                         printk("lafs_import_inode failed %d\n", err);
 205                 goto err;
 206         }
 207         check_atime_ref(ino, async);
 208         unlock_new_inode(ino);
 209 out:
 210         if (b && test_and_clear_bit(B_Async, &b->b.flags)) {
 211                 putdref(b, MKREF(async));
 212                 lafs_wake_thread(fs_from_inode(fsys));
 213         }
 214         putdref(b, MKREF(iget));
 215         return ino;
 216 err:
 217         ino->i_nlink = 0;
 218         unlock_new_inode(ino);
 219         iput(ino);
 220         ino = ERR_PTR(err);
 221         goto out;
 222 }
 223
 224 struct inode *
 225 lafs_iget_fs(struct fs *fs, int fsnum, int inum, int async)
 226 {
 227         struct super_block *sb;
 228         struct inode *rv;
 229
 230         sb = fs->prime_sb;
 231
 232         if (fsnum) {
 233                 /* Need to locate or load the superblock for this
 234                  * subordinate filesystem
 235                  */
 236                 struct inode *filesys;
 237
 238                 filesys = lafs_iget(fs->ss[0].root, fsnum, async);
 239                 if (IS_ERR(filesys))
 240                         return filesys;
 241                 if (LAFSI(filesys)->type != TypeInodeFile) {
 242                         iput(filesys);
 243                         return ERR_PTR(-ENOENT);
 244                 }
 245                 rv = lafs_iget(filesys, inum, async);
 246         } else if (inum) {
 247                 rv = lafs_iget(fs->ss[0].root, inum, async);
 248                 if (!IS_ERR(rv))
 249                         atomic_inc(&sb->s_active);
 250         } else {
 251                 rv = igrab(fs->ss[0].root);
 252                 atomic_inc(&sb->s_active);
 253         }
 254         return rv;
 255 }
 256
 257 int __must_check
 258 lafs_import_inode(struct inode *ino, struct datablock *b)
 259 {
 260         struct la_inode *lai = map_dblock(b);
 261         struct lafs_inode *li = LAFSI(ino);
 262         int err = -ENOENT;
 263
 264         if (lai->filetype == 0) {
 265                 li->type = 0;
 266                 ino->i_mode = 0;
 267                 ino->i_nlink = 0;
 268                 goto out;
 269         }
 270
 271         ino->i_mode = S_IFREG;
 272         ino->i_nlink = 1; /* For special file, set nlink so they
 273                            * never appear unlinked */
 274
 275         err = -EINVAL;
 276
 277         LAFS_BUG(ino->i_ino != b->b.fileaddr, &b->b);
 278         li->cblocks = le32_to_cpu(lai->data_blocks);
 279         li->pblocks = li->ablocks = 0;
 280         li->vfs_inode.i_blocks = ((blkcnt_t)li->cblocks
 281                                   << (ino->i_sb->s_blocksize_bits - 9));
 282         li->ciblocks = le32_to_cpu(lai->index_blocks);
 283         li->piblocks = 0;
 284         li->iflags = 0;
 285
 286         ino->i_generation = le16_to_cpu(lai->generation);
 287         li->trunc_gen = lai->trunc_gen;
 288         li->flags = lai->flags;
 289         li->type = lai->filetype;
 290         li->metadata_size = le16_to_cpu(lai->metadata_size);
 291         li->depth = lai->depth;
 292
 293         dprintk("inode %lu type is %d\n", (unsigned long)ino->i_ino, li->type);
 294
 295         ino->i_data.a_ops = &lafs_file_aops;
 296         li->trunc_next = 0;
 297
 298         switch (li->type) {
 299         case TypeInodeFile:
 300         {
 301                 struct fs_md *i = &li->md.fs;
 302                 struct fs_metadata *l = &lai->metadata[0].fs;
 303                 int nlen;
 304
 305                 i->usagetable = le16_to_cpu(l->snapshot_usage_table);
 306                 decode_time(&ino->i_mtime, le64_to_cpu(l->update_time));
 307                 i->cblocks_used = le64_to_cpu(l->blocks_used);
 308                 i->pblocks_used = i->ablocks_used = 0;
 309                 i->blocks_allowed = le64_to_cpu(l->blocks_allowed);
 310                 i->blocks_unalloc = 0;
 311                 i->creation_age = le64_to_cpu(l->creation_age);
 312                 i->inodes_used = le32_to_cpu(l->inodes_used);
 313                 i->quota_inums[0] = le32_to_cpu(l->quota_inodes[0]);
 314                 i->quota_inums[1] = le32_to_cpu(l->quota_inodes[1]);
 315                 i->quota_inums[2] = le32_to_cpu(l->quota_inodes[2]);
 316                 i->quota_inodes[0] = i->quota_inodes[1]
 317                         = i->quota_inodes[2] = NULL;
 318                 nlen = li->metadata_size - offsetof(struct la_inode,
 319                                                     metadata[0].fs.name);
 320                 i->accesstime = NULL;
 321                 if (i->name)
 322                         kfree(i->name);
 323                 if (nlen == 0)
 324                         i->name = NULL;
 325                 else {
 326                         /* Need to unmap the dblock to kmalloc because
 327                          * the mapping makes us 'atomic'
 328                          */
 329                         unmap_dblock(b, lai);
 330                         i->name = kmalloc(nlen+1, GFP_KERNEL);
 331                         lai = map_dblock(b);
 332                         l = &lai->metadata[0].fs;
 333
 334                         err = -ENOMEM;
 335                         if (!i->name)
 336                                 goto out;
 337                         memcpy(i->name, l->name, nlen);
 338                         i->name[nlen] = 0;
 339                 }
 340                 /* Make this look like a directory */
 341                 ino->i_mode = S_IFDIR;
 342                 ino->i_uid = 0;
 343                 ino->i_gid = 0;
 344                 ino->i_size = 0;
 345                 ino->i_op = &lafs_subset_ino_operations;
 346                 ino->i_fop = &lafs_subset_file_operations;
 347                 break;
 348         }
 349
 350         case TypeInodeMap:
 351         {
 352                 struct inodemap_md *m = &li->md.inodemap;
 353                 struct inodemap_metadata *s = &lai->metadata[0].inodemap;
 354                 m->size = le32_to_cpu(s->size);
 355                 m->thisblock = NoBlock;
 356                 m->nextbit = 0;
 357                 break;
 358         }
 359
 360         case TypeSegmentMap:
 361         {
 362                 struct su_md *m = &li->md.segmentusage;
 363                 struct su_metadata *s = &lai->metadata[0].segmentusage;
 364                 m->table_size = le32_to_cpu(s->table_size);
 365                 break;
 366         }
 367
 368         case TypeQuota:
 369         {
 370                 struct quota_md *m = &li->md.quota;
 371                 struct quota_metadata *s = &lai->metadata[0].quota;
 372                 m->gracetime = le32_to_cpu(s->gracetime);
 373                 m->graceunits = le32_to_cpu(s->graceunits);
 374                 break;
 375         }
 376         case TypeOrphanList:
 377         {
 378                 struct orphan_md *m = &li->md.orphan;
 379                 /* This will be set via lafs_count_orphans */
 380                 m->nextfree = 0;
 381                 m->reserved = 0;
 382                 break;
 383         }
 384         case TypeAccessTime:
 385                 break;
 386
 387         default: /* TypeBase or larger */
 388         {
 389                 struct file_md *i = &li->md.file;
 390                 struct file_metadata *l = &lai->metadata[0].file;
 391                 struct dir_metadata *d = &lai->metadata[0].dir;
 392                 struct special_metadata *s = &lai->metadata[0].special;
 393
 394                 if (li->type < TypeBase)
 395                         goto out;
 396                 i->flags = le16_to_cpu(l->flags);
 397                 ino->i_mode = le16_to_cpu(l->mode);
 398                 ino->i_uid = le32_to_cpu(l->userid);
 399                 ino->i_gid = le32_to_cpu(l->groupid);
 400                 i->treeid = le32_to_cpu(l->treeid);
 401                 i->creationtime = le64_to_cpu(l->creationtime);
 402                 decode_time(&ino->i_mtime, le64_to_cpu(l->modifytime));
 403                 decode_time(&ino->i_ctime, le64_to_cpu(l->ctime));
 404                 decode_time(&i->i_accesstime, le64_to_cpu(l->accesstime));
 405                 ino->i_atime = i->i_accesstime;
 406                 i->atime_offset = 0; /* Will be filled-in later probably */
 407                 lafs_add_atime_offset(&ino->i_atime, i->atime_offset);
 408                 ino->i_size = le64_to_cpu(l->size);
 409                 i->parent = le32_to_cpu(l->parent);
 410                 ino->i_nlink = le32_to_cpu(l->linkcount);
 411                 if (ino->i_nlink == 0 && list_empty(&b->orphans) &&
 412                     fs_from_inode(ino)->rolled) {
 413                         /* This block should already be on the orphan
 414                          * list, otherwise there is a filesystem
 415                          * inconsistency.
 416                          * Either the orphan file is wrong, or the
 417                          * linkcount is wrong.
 418                          * It is safest to assume the later - either
 419                          * way an FS check would be needed to fix it.
 420                          * Note: while roll-forward is happening, this
 421                          * situation is perfectly possible and is handled
 422                          * correctly.
 423                          */
 424                         /* FIXME set a superblock flag requesting
 425                          * directory linkage checking
 426                          */
 427                         ino->i_nlink = 1;
 428                 }
 429
 430                 dprintk("  mode = 0%o uid %d size %lld\n",
 431                         ino->i_mode, ino->i_uid, ino->i_size);
 432                 switch (li->type) {
 433                 case TypeFile:
 434                         ino->i_op = &lafs_file_ino_operations;
 435                         ino->i_fop = &lafs_file_file_operations;
 436                         ino->i_mode = (ino->i_mode & 07777)  | S_IFREG;
 437                         break;
 438                 case TypeDir:
 439                         i->seed = le32_to_cpu(d->hash_seed);
 440                         ino->i_op = &lafs_dir_ino_operations;
 441                         ino->i_fop = &lafs_dir_file_operations;
 442                         ino->i_mode = (ino->i_mode & 07777)  | S_IFDIR;
 443                         {
 444                                 u32 *b = (u32 *)lai;
 445                                 dprintk("Hmm. %d %d %d\n",
 446                                         (int)b[24],
 447                                         (int)b[25],
 448                                         (int)b[26]);
 449                         }
 450                         break;
 451                 case TypeSymlink:
 452                         ino->i_op = &lafs_link_ino_operations;
 453                         ino->i_mode = (ino->i_mode & 07777)  | S_IFLNK;
 454                         break;
 455                 case TypeSpecial:
 456                         /* the data had better be in the inode ... */
 457                         ino->i_rdev = MKDEV(le32_to_cpu(s->major),
 458                                             le32_to_cpu(s->minor));
 459                         ino->i_op = &lafs_special_ino_operations;
 460                         init_special_inode(ino, ino->i_mode, ino->i_rdev);
 461                         break;
 462                 }
 463                 break;
 464         }
 465         }
 466
 467         ino->i_blkbits = ino->i_sb->s_blocksize_bits;
 468         /* FIXME i_blocks and i_byte - used for quota?? */
 469         err = 0;
 470
 471         /* Note: no refcount yet.  Either will remove the reference to the
 472          * other when freed
 473          */
 474         li->dblock = b;
 475         rcu_assign_pointer(b->my_inode, ino);
 476
 477 out:
 478         if (err && li->type)
 479                 printk("inode %lu type is %d\n",
 480                        (unsigned long)ino->i_ino, li->type);
 481         unmap_dblock(b, lai);
 482         return err;
 483 }
 484
 485 static void check_atime_ref(struct inode *ino, int async)
 486 {
 487         /* If there is an time file in this filesystem the inode
 488          * should hold a reference to the relevant block in
 489          * that file.
 490          */
 491         struct inode *root, *at;
 492         u32 bnum;
 493         struct datablock *b;
 494         if (async)
 495                 /* Never bother for async lookups */
 496                 return;
 497         if (LAFSI(ino)->type < TypeBase)
 498                 return;
 499         if (test_bit(I_AccessTime, &LAFSI(ino)->iflags))
 500                 return;
 501         root = LAFSI(ino)->filesys;
 502         at = LAFSI(root)->md.fs.accesstime;
 503         if (at == NULL)
 504                 return;
 505
 506         if (LAFSI(ino)->md.file.atime_offset)
 507                 LAFSI(ino)->md.file.atime_offset = 0;
 508
 509         /* "* 2" to get byte number, then shift to get block
 510          * number
 511          */
 512         bnum = ino->i_ino >> (at->i_blkbits-1);
 513         b = lafs_get_block(at, bnum, NULL, GFP_NOFS, MKREF(atime));
 514         if (b) {
 515                 if (lafs_read_block(b) == 0) {
 516                         u16 *atp;
 517                         int i;
 518                         atp = map_dblock(b);
 519                         i = (ino->i_ino * 2) & ((1<<at->i_blkbits)-1);
 520                         LAFSI(ino)->md.file.atime_offset = le16_to_cpu(atp[i]);
 521                         set_bit(I_AccessTime, &LAFSI(ino)->iflags);
 522                         unmap_dblock(b, atp);
 523                         lafs_add_atime_offset(&ino->i_atime,
 524                                               LAFSI(ino)->md.file.atime_offset);
 525                 } else
 526                         putdref(b, MKREF(atime));
 527         }
 528 }
 529
 530 void lafs_add_atime_offset(struct timespec *atime, int offset)
 531 {
 532         int expon;
 533         time_t mantissa;
 534         if (offset == 0)
 535                 return;
 536
 537         expon = offset & 0x1f;
 538         if (expon)
 539                 mantissa = (offset >> 5) | 0x800;
 540         else
 541                 mantissa = (offset >> 5);
 542         if (expon >= 11) {
 543                 /* seconds */
 544                 mantissa <<= expon-11;
 545                 atime->tv_sec += mantissa;
 546         } else {
 547                 /* milliseconds */
 548                 if (expon)
 549                         mantissa <<= expon-1;
 550                 timespec_add_ns(atime, (s64)mantissa * 1000000);
 551         }
 552 }
 553
 554 static int normalise(int *mantissa)
 555 {
 556         /* Shift down until value can be stored in 12 bits:
 557          * Top bit will be '1', so only 11 bits needed.
 558          * Not used on values below 2048.
 559          */
 560         int shift = 0;
 561         while (*mantissa >= 4096) {
 562                 *mantissa >>= 1;
 563                 shift ++;
 564         }
 565         return shift;
 566 }
 567
 568 static int update_atime_delta(struct inode *ino)
 569 {
 570         /* calculate new delta to show the difference between
 571          * i_atime and i_accesstime
 572          */
 573         int rv;
 574         if (LAFSI(ino)->type < TypeBase)
 575                 return 0;
 576         if (timespec_compare(&ino->i_atime,
 577                              &LAFSI(ino)->md.file.i_accesstime) <= 0) {
 578                 /* We cannot store negative delta so if i_atime is in the
 579                  * past, just store zero
 580                  */
 581                 rv = 0;
 582         } else {
 583                 struct timespec diff;
 584                 int shift;
 585
 586                 diff = timespec_sub(ino->i_atime,
 587                                     LAFSI(ino)->md.file.i_accesstime);
 588                 if (diff.tv_sec >= 2048) {
 589                         /* Just store the seconds */
 590                         rv = diff.tv_sec;
 591                         shift = normalise(&rv) + 11;
 592                 } else {
 593                         /* Store the milliseconds */
 594                         int rv = diff.tv_nsec / 1000000;
 595                         rv += diff.tv_sec * 1000;
 596                         if (rv >= 2048)
 597                                 shift = normalise(&rv) + 1;
 598                         else
 599                                 shift = 0;
 600                 }
 601                 if (shift > 31)
 602                         rv = 0xFFFF;
 603                 else {
 604                         rv &= 0x7ff;
 605                         rv <<= 5;
 606                         rv |= shift;
 607                 }
 608         }
 609         if (LAFSI(ino)->md.file.atime_offset == rv)
 610                 return 0;
 611
 612         LAFSI(ino)->md.file.atime_offset = rv;
 613         return 1;
 614 }
 615
 616 static void store_atime_delta(struct inode *ino)
 617 {
 618         struct inode *at;
 619         u32 bnum;
 620         struct datablock *b;
 621         u16 *atp;
 622         int i;
 623
 624         if (!test_bit(I_AccessTime, &LAFSI(ino)->iflags))
 625                 /* sorry, nothing we can do here */
 626                 return;
 627
 628         /* We own a reference, so this lookup must succeed */
 629         at = LAFSI(LAFSI(ino)->filesys)->md.fs.accesstime;
 630         bnum = ino->i_ino >> (at->i_blkbits-1);
 631         b = lafs_get_block(at, bnum, NULL, GFP_NOFS, MKREF(store_atime));
 632         BUG_ON(!b);
 633         atp = map_dblock(b);
 634         i = (ino->i_ino * 2) & ((1<<at->i_blkbits)-1);
 635         if (le16_to_cpu(atp[i]) != LAFSI(ino)->md.file.atime_offset) {
 636                 atp[i] = cpu_to_le16(LAFSI(ino)->md.file.atime_offset);
 637                 lafs_dirty_dblock(b);
 638         }
 639         unmap_dblock(b, atp);
 640         putdref(b, MKREF(store_atime));
 641 }
 642
 643 void lafs_inode_checkpin(struct inode *ino)
 644 {
 645         /* Make sure I_Pinned is set correctly.
 646          * It should be set precisely if i_nlink is non-zero,
 647          * and ->iblock is B_Pinned.
 648          * When it is set, we own a reference to the inode.
 649          *
 650          * This needs to be called whenever we change
 651          * i_nlink, and whenever we pin or unpin an InoIdx
 652          * block.
 653          */
 654         if (ino->i_nlink == 0) {
 655                 /* I_Pinned should not be set */
 656                 if (test_and_clear_bit(I_Pinned, &LAFSI(ino)->iflags)) {
 657                         if (ino->i_sb->s_type == &lafs_fs_type)
 658                                 iput(ino);
 659                         else
 660                                 lafs_iput_fs(ino);
 661                 }
 662         } else {
 663                 /* Need to check if iblock is Pinned. */
 664                 struct indexblock *ib = NULL;
 665                 if (LAFSI(ino)->iblock) {
 666                         spin_lock(&ino->i_data.private_lock);
 667                         ib = LAFSI(ino)->iblock;
 668                         if (ib && !test_bit(B_Pinned, &ib->b.flags))
 669                                 ib = NULL;
 670                         spin_unlock(&ino->i_data.private_lock);
 671                 }
 672                 if (ib) {
 673                         if (!test_and_set_bit(I_Pinned, &LAFSI(ino)->iflags)) {
 674                                 if (ino->i_sb->s_type == &lafs_fs_type)
 675                                         igrab(ino);
 676                                 else
 677                                         lafs_igrab_fs(ino);
 678                         }
 679                 } else {
 680                         if (test_and_clear_bit(I_Pinned, &LAFSI(ino)->iflags)) {
 681                         if (ino->i_sb->s_type == &lafs_fs_type)
 682                                 iput(ino);
 683                         else
 684                                 lafs_iput_fs(ino);
 685                         }
 686                 }
 687         }
 688 }
 689
 690 struct datablock *lafs_inode_get_dblock(struct inode *ino, REFARG)
 691 {
 692         struct datablock *db;
 693
 694         spin_lock(&ino->i_data.private_lock);
 695         db = LAFSI(ino)->dblock;
 696         if (db) {
 697                 if (db->b.inode == ino)
 698                         getdref_locked(db, REF);
 699                 else {
 700                         spin_lock_nested(&db->b.inode->i_data.private_lock, 1);
 701                         getdref_locked(db, REF);
 702                         spin_unlock(&db->b.inode->i_data.private_lock);
 703                 }
 704         }
 705         spin_unlock(&ino->i_data.private_lock);
 706         return db;
 707 }
 708
 709 struct datablock *lafs_inode_dblock(struct inode *ino, int async, REFARG)
 710 {
 711         struct datablock *db;
 712         int err;
 713
 714         db = lafs_inode_get_dblock(ino, REF);
 715
 716         if (!db)
 717                 db = lafs_get_block(LAFSI(ino)->filesys, ino->i_ino, NULL,
 718                                     GFP_KERNEL, REF);
 719         if (!db)
 720                 return ERR_PTR(-ENOMEM);
 721
 722         LAFSI(ino)->dblock = db;
 723         rcu_assign_pointer(db->my_inode, ino);
 724         if (async)
 725                 err = lafs_read_block_async(db);
 726         else
 727                 err = lafs_read_block(db);
 728         if (err == 0)
 729                 return db;
 730
 731         putdref(db, REF);
 732         return ERR_PTR(err);
 733 }
 734
 735 void lafs_inode_init(struct datablock *b, int type, int mode, struct inode *dir)
 736 {
 737         /* A new block has been allocated in an inode file to hold an
 738          * inode.  We get to fill in initial values so that when
 739          * 'iget' calls lafs_import_inode, the correct inode is
 740          * loaded.
 741          */
 742         struct fs *fs = fs_from_inode(b->b.inode);
 743         struct la_inode *lai = map_dblock(b);
 744         int size;
 745
 746         lai->data_blocks = cpu_to_le32(0);
 747         lai->index_blocks = cpu_to_le32(0);
 748         get_random_bytes(&lai->generation, sizeof(lai->generation));
 749         lai->depth = 1;
 750         lai->trunc_gen = 0;
 751         lai->filetype = type;
 752         lai->flags = 0;
 753
 754         switch(type) {
 755         case TypeInodeFile:
 756         {
 757                 struct fs_metadata *l = &lai->metadata[0].fs;
 758                 size = sizeof(struct fs_metadata);
 759                 l->update_time = 0;
 760                 l->blocks_used = 0;
 761                 l->blocks_allowed = 0;
 762                 l->creation_age = fs->wc[0].cluster_seq;
 763                 l->inodes_used = 0;
 764                 l->quota_inodes[0] = 0;
 765                 l->quota_inodes[1] = 0;
 766                 l->quota_inodes[2] = 0;
 767                 l->snapshot_usage_table = 0;
 768                 l->pad = 0;
 769                 /* name will be zero length and not used */
 770                 break;
 771         }
 772         case TypeInodeMap:
 773         {
 774                 struct inodemap_metadata *l = &lai->metadata[0].inodemap;
 775                 l->size = 0;
 776                 size = sizeof(struct inodemap_metadata);
 777                 break;
 778         }
 779         case TypeSegmentMap:
 780                 size = sizeof(struct su_metadata);
 781                 break;
 782         case TypeQuota:
 783                 size = sizeof(struct quota_metadata);
 784                 break;
 785         case TypeOrphanList:
 786                 size = 0;
 787                 break;
 788         case TypeAccessTime:
 789                 size = 0;
 790                 break;
 791         default:
 792         {
 793                 struct file_metadata *l = &lai->metadata[0].file;
 794                 struct timespec now = CURRENT_TIME;
 795
 796                 l->flags = cpu_to_le16(0);
 797                 l->userid = cpu_to_le32(current->cred->fsuid);
 798                 if (dir && (dir->i_mode & S_ISGID)) {
 799                         l->groupid = cpu_to_le32(dir->i_gid);
 800                         if (type == TypeDir)
 801                                 mode |= S_ISGID;
 802                 } else
 803                         l->groupid = cpu_to_le32(current->cred->fsgid);
 804                 if (dir && LAFSI(dir)->md.file.treeid)
 805                         l->treeid = cpu_to_le32(LAFSI(dir)->md.file.treeid);
 806                 else
 807                         l->treeid = l->userid;
 808
 809                 l->mode = cpu_to_le16(mode);
 810                 l->creationtime = encode_time(&now);
 811                 l->modifytime = l->creationtime;
 812                 l->ctime = l->creationtime;
 813                 l->accesstime = l->creationtime;
 814                 l->size = 0;
 815                 l->parent = dir ? cpu_to_le32(dir->i_ino) : 0;
 816                 l->linkcount = 0;
 817                 l->attrinode = 0;
 818                 if (type == TypeDir) {
 819                         struct dir_metadata *l = &lai->metadata[0].dir;
 820                         u32 seed;
 821                         get_random_bytes(&seed,
 822                                          sizeof(seed));
 823                         seed = (seed & ~7) | 1;
 824                         l->hash_seed = cpu_to_le32(seed);
 825                         size = sizeof(struct dir_metadata);
 826                 } else if (type == TypeSpecial) {
 827                         struct special_metadata *s = &lai->metadata[0].special;
 828                         s->major = s->minor = 0;
 829                         size = sizeof(struct special_metadata);
 830                 } else
 831                         size = sizeof(struct file_metadata);
 832         }
 833         }
 834         size += sizeof(struct la_inode);
 835         lai->metadata_size = cpu_to_le32(size);
 836         memset(((char *)lai)+size, 0, fs->blocksize-size);
 837         *(u16 *)(((char *)lai)+size) = cpu_to_le16(IBLK_EXTENT);
 838
 839         unmap_dblock(b, lai);
 840         set_bit(B_Valid, &b->b.flags);
 841         LAFS_BUG(!test_bit(B_Pinned, &b->b.flags), &b->b);
 842         lafs_dirty_dblock(b);
 843 }
 844
 845 static int inode_map_free(struct fs *fs, struct inode *fsys, u32 inum);
 846
 847 void lafs_evict_inode(struct inode *ino)
 848 {
 849         struct fs *fs = fs_from_inode(ino);
 850         struct lafs_inode *li = LAFSI(ino);
 851
 852         if (ino->i_mode == 0) {
 853                 /* There never was an inode here,
 854                  * so nothing to do.
 855                  * We just call end_writeback to get the
 856                  * flags set properly.
 857                  */
 858                 end_writeback(ino);
 859                 return;
 860         }
 861
 862         dprintk("EVICT INODE %d\n", (int)ino->i_ino);
 863
 864
 865         /* Normal truncation holds an igrab, so we cannot be
 866          * deleted until any truncation finishes
 867          */
 868         BUG_ON(test_bit(I_Trunc, &LAFSI(ino)->iflags));
 869
 870         if (ino->i_nlink == 0) {
 871                 struct datablock *b =
 872                         lafs_inode_dblock(ino, SYNC, MKREF(delete_inode));
 873                 i_size_write(ino, 0);
 874                 truncate_inode_pages(&ino->i_data, 0);
 875                 LAFSI(ino)->trunc_next = 0;
 876                 set_bit(I_Deleting, &LAFSI(ino)->iflags);
 877                 set_bit(I_Trunc, &LAFSI(ino)->iflags);
 878                 lafs_igrab_fs(ino);
 879                 if (!IS_ERR(b)) {
 880                         set_bit(B_Claimed, &b->b.flags);
 881                         lafs_add_orphan(fs, b);
 882                         dprintk("PUNCH hole for %d\n", (int)b->b.fileaddr);
 883                         putdref(b, MKREF(delete_inode));
 884                 }
 885                 inode_map_free(fs, LAFSI(ino)->filesys,  ino->i_ino);
 886         } else
 887                 truncate_inode_pages(&ino->i_data, 0);
 888         end_writeback(ino);
 889
 890         dprintk("CLEAR INODE %d\n", (int)ino->i_ino);
 891
 892         li->type = 0;
 893
 894         /* Now is a good time to break the linkage between
 895          * inode and dblock - but not if the file is
 896          * being deleted
 897          */
 898         if (!test_bit(I_Deleting, &li->iflags)) {
 899                 struct datablock *db;
 900                 spin_lock(&ino->i_data.private_lock);
 901                 db = li->dblock;
 902                 if (db) {
 903                         struct indexblock *ib = li->iblock;
 904                         LAFS_BUG(ib && atomic_read(&ib->b.refcnt), &db->b);
 905                         db->my_inode = NULL;
 906                         li->dblock = NULL;
 907                         li->iblock = NULL;
 908                 }
 909                 spin_unlock(&ino->i_data.private_lock);
 910         }
 911
 912         /* FIXME release quota inodes if filesystem */
 913 }
 914
 915 static int prune(void *data, u32 addr, u64 paddr, int len)
 916 {
 917         /* This whole index block is being pruned, just account
 918          * for everything and it will be cleared afterwards
 919          */
 920         struct indexblock *ib = data;
 921         struct inode *ino = ib->b.inode;
 922         struct fs *fs = fs_from_inode(ino);
 923         int ph = !!test_bit(B_Phase1, &ib->b.flags);
 924         int i;
 925         dprintk("PRUNE %d for %d at %lld\n", addr, len, (long long)paddr);
 926         if (paddr == 0 || len == 0)
 927                 return 0;
 928         for (i = 0 ; i < len ; i++)
 929                 lafs_summary_update(fs, ino, paddr+i, 0, 0, ph, 0);
 930         return len;
 931 }
 932
 933 static int prune_some(void *data, u32 addr, u64 paddr, int len)
 934 {
 935         /* Part of this index block is being pruned.  Copy
 936          * what addresses we can into uninc_table so that
 937          * it can be 'incorporated'
 938          * We should probably share some code with
 939          * lafs_allocated_block??
 940          */
 941         struct indexblock *ib = data;
 942         struct inode *ino = ib->b.inode;
 943         struct fs *fs = fs_from_inode(ino);
 944         int ph = !!test_bit(B_Phase1, &ib->b.flags);
 945         int i;
 946
 947         if (paddr == 0 || len == 0)
 948                 return 0;
 949         dprintk("PRUNE2 %d for %d at %lld\n", addr, len, (long long)paddr);
 950         for (i = 0 ; i < len ; i++) {
 951                 /* FIXME should allow longer truncation ranges in uninc_table
 952                  * as they are easy to handle.
 953                  */
 954                 struct addr *a;
 955                 if (addr + i < LAFSI(ino)->trunc_next)
 956                         continue;
 957                 spin_lock(&ino->i_data.private_lock);
 958                 a = &ib->uninc_table.pending_addr
 959                         [ib->uninc_table.pending_cnt - 1];
 960                 if (ib->uninc_table.pending_cnt <
 961                     ARRAY_SIZE(ib->uninc_table.pending_addr)) {
 962                         a++;
 963                         a->fileaddr = addr + i;
 964                         a->physaddr = 0;
 965                         a->cnt = 1;
 966                         LAFS_BUG(!test_bit(B_Pinned, &ib->b.flags), &ib->b);
 967                         ib->uninc_table.pending_cnt++;
 968                 } else {
 969                         spin_unlock(&ino->i_data.private_lock);
 970                         break;
 971                 }
 972                 spin_unlock(&ino->i_data.private_lock);
 973                 lafs_summary_update(fs, ino, paddr+i, 0, 0, ph, 0);
 974         }
 975         return i;
 976 }
 977
 978 int lafs_inode_handle_orphan(struct datablock *b)
 979 {
 980         /* Don't need rcu protection for my_inode run_orphan
 981          * holds a reference
 982          */
 983         struct indexblock *ib, *ib2;
 984         struct inode *ino = b->my_inode;
 985         struct fs *fs = fs_from_inode(ino);
 986         u32 trunc_next, next_trunc;
 987         int loop_cnt = 20;
 988         int err = -ENOMEM;
 989
 990         if (!test_bit(I_Trunc, &LAFSI(ino)->iflags)) {
 991                 if (test_bit(I_Deleting, &LAFSI(ino)->iflags)) {
 992                         LAFS_BUG(ino->i_nlink, &b->b);
 993                         if (LAFSI(ino)->cblocks +
 994                             LAFSI(ino)->pblocks +
 995                             LAFSI(ino)->ablocks +
 996                             LAFSI(ino)->ciblocks +
 997                             LAFSI(ino)->piblocks)
 998                         printk("Deleting inode %lu: %ld+%ld+%ld %ld+%ld\n",
 999                                ino->i_ino,
1000                                LAFSI(ino)->cblocks,
1001                                LAFSI(ino)->pblocks,
1002                                LAFSI(ino)->ablocks,
1003                                LAFSI(ino)->ciblocks,
1004                                LAFSI(ino)->piblocks);
1005                         BUG_ON(LAFSI(ino)->cblocks +
1006                                LAFSI(ino)->pblocks +
1007                                LAFSI(ino)->ablocks +
1008                                LAFSI(ino)->ciblocks +
1009                                LAFSI(ino)->piblocks);
1010                         if (lafs_erase_dblock_async(b))
1011                                 lafs_orphan_release(fs, b);
1012                 } else if (ino->i_nlink || LAFSI(ino)->type == 0)
1013                         lafs_orphan_release(fs, b);
1014                 else
1015                         lafs_orphan_forget(fs, b);
1016                 return 0;
1017         }
1018
1019         ib = lafs_make_iblock(ino, ADOPT, SYNC, MKREF(inode_handle_orphan));
1020         if (IS_ERR(ib))
1021                 return PTR_ERR(ib);
1022
1023         /* Here is the guts of 'truncate'.  We find the next leaf index
1024          * block and discard all the addresses there-in.
1025          */
1026         trunc_next = LAFSI(ino)->trunc_next;
1027
1028         if (trunc_next == 0xFFFFFFFF) {
1029                 /* truncate has finished in that all data blocks
1030                  * have been removed and all index block are either
1031                  * gone or pending incorporation at which point they will
1032                  * go.
1033                  * If we hit a phase change, we will need to postpone
1034                  * the rest of the cleaning until it completes.
1035                  * If there is a checkpoint happening, then all the work
1036                  * that we can do now, it will do for us.  So just
1037                  * let it.
1038                  */
1039                 struct indexblock *tmp;
1040                 struct indexblock *next;
1041                 u32 lastaddr;
1042
1043                 if (!test_bit(B_Pinned, &ib->b.flags)) {
1044                         /* must be finished */
1045                         LAFS_BUG(test_bit(B_Dirty, &ib->b.flags), &ib->b);
1046                         clear_bit(I_Trunc, &LAFSI(ino)->iflags);
1047                         lafs_iput_fs(ino);
1048                         wake_up(&fs->trunc_wait);
1049                         err = -ERESTARTSYS;
1050                         goto out2;
1051                 }
1052                 if (fs->checkpointing) {
1053                         /* This cannot happen with current code,
1054                          * but leave it in case we ever have
1055                          * orphan handling parallel with checkpoints
1056                          */
1057                         err = -EBUSY; /* Try again after the checkpoint */
1058                         goto out2;
1059                 }
1060
1061                 lastaddr = (i_size_read(ino) +
1062                             fs->blocksize - 1)
1063                         >> fs->blocksize_bits;
1064                 /* Find a Pinned descendent of ib which has no
1065                  * Pinned descendents and no PrimaryRef dependent
1066                  * (so take the last).
1067                  * Prefer blocks that are beyond EOF (again, take the last).
1068                  * If there are none, descend the last block that
1069                  * is not after EOF and look at its children.
1070                  */
1071                 ib2 = next = ib;
1072                 spin_lock(&ib->b.inode->i_data.private_lock);
1073                 while (next) {
1074                         ib2 = next;
1075                         next = NULL;
1076                         list_for_each_entry(tmp, &ib2->children, b.siblings) {
1077                                 if (!test_bit(B_Index, &tmp->b.flags) ||
1078                                     !test_bit(B_Pinned, &tmp->b.flags))
1079                                         continue;
1080                                 if (next == NULL ||
1081                                     tmp->b.fileaddr > next->b.fileaddr)
1082                                         next = tmp;
1083                         }
1084                 }
1085                 if (ib2->b.fileaddr < lastaddr) {
1086                         /* Must be all done */
1087                         spin_unlock(&ib->b.inode->i_data.private_lock);
1088                         clear_bit(I_Trunc, &LAFSI(ino)->iflags);
1089                         lafs_iput_fs(ino);
1090                         wake_up(&fs->trunc_wait);
1091                         err = -ERESTARTSYS;
1092                         goto out2;
1093                 }
1094                 getiref(ib2, MKREF(inode_handle_orphan2));
1095                 spin_unlock(&ib->b.inode->i_data.private_lock);
1096
1097                 /* ib2 is an index block beyond EOF with no
1098                  * Pinned children.
1099                  * Incorporating it should unpin it.
1100                  */
1101                 if (!list_empty(&ib2->children)) {
1102                         lafs_print_tree(&ib2->b, 3);
1103                         LAFS_BUG(1, &ib2->b);
1104                 }
1105
1106                 if (!lafs_iolock_written_async(&ib2->b)) {
1107                         putiref(ib2, MKREF(inode_handle_orphan2));
1108                         err = -EAGAIN;
1109                         goto out2;
1110                 }
1111                 while (ib2->uninc_table.pending_cnt || ib2->uninc)
1112                         lafs_incorporate(fs, ib2);
1113
1114                 if (test_bit(B_Dirty, &ib2->b.flags) ||
1115                     test_bit(B_Realloc, &ib2->b.flags))
1116                         lafs_cluster_allocate(&ib2->b, 0);
1117                 else
1118                         lafs_iounlock_block(&ib2->b);
1119
1120                 if (!list_empty(&ib2->b.siblings)) {
1121                         printk("looping on %s\n", strblk(&ib2->b));
1122                         loop_cnt--;
1123                         if (loop_cnt < 0)
1124                                 BUG();
1125                 }
1126                 putiref(ib2, MKREF(inode_handle_orphan2));
1127                 err = -ERESTARTSYS;
1128                 if (ib->uninc) {
1129                         if (lafs_iolock_written_async(&ib->b)) {
1130                                 while (ib->uninc)
1131                                         lafs_incorporate(fs, ib);
1132                                 lafs_iounlock_block(&ib->b);
1133                         } else
1134                                 err = -EAGAIN;
1135                 }
1136         out2:
1137                 putiref(ib, MKREF(inode_handle_orphan));
1138                 return err;
1139         }
1140
1141         putiref(ib, MKREF(inode_handle_orphan));
1142
1143         ib = lafs_leaf_find(ino, trunc_next, ADOPT, &next_trunc,
1144                             ASYNC, MKREF(inode_handle_orphan3));
1145         if (IS_ERR(ib))
1146                 return PTR_ERR(ib);
1147         /* now hold an iolock on ib */
1148
1149         /* Ok, trunc_next seems to refer to a block that exists.
1150          * We need to erase it..
1151          *
1152          * So we open up the index block ourselves, call
1153          * lafs_summary_update with each block address, and then
1154          * erase the block.
1155          */
1156
1157         if (LAFSI(ino)->depth == 0) {
1158                 /* Nothing to truncate */
1159                 clear_bit(I_Trunc, &LAFSI(ino)->iflags);
1160                 lafs_iput_fs(ino);
1161                 if (test_bit(B_Pinned, &ib->b.flags))
1162                         /* Need to move the dirtiness which keeps this
1163                          * pinned to the data block.
1164                          */
1165                         lafs_cluster_allocate(&ib->b, 0);
1166                 else
1167                         lafs_iounlock_block(&ib->b);
1168                 err = -ERESTARTSYS;
1169                 goto out_put;
1170         }
1171
1172         lafs_checkpoint_lock(fs);
1173         err = lafs_reserve_block(&ib->b, ReleaseSpace);
1174         if (err < 0)
1175                 goto out;
1176
1177         if (!test_bit(B_Valid, &ib->b.flags) &&
1178             test_bit(B_InoIdx, &ib->b.flags)) {
1179                 /* still invalid, just re-erase to remove
1180                  * pinning */
1181                 LAFSI(ino)->trunc_next = next_trunc;
1182                 lafs_cluster_allocate(&ib->b, 0);
1183                 err = -ERESTARTSYS;
1184                 goto out_unlocked;
1185         }
1186
1187         lafs_pin_block(&ib->b);
1188
1189         /* It might be that this can happen, in which case
1190          * we simply update trunc_next and loop.  But I'd like
1191          * to be sure before I implement that
1192          */
1193         if (!test_bit(B_Valid, &ib->b.flags)) {
1194                 printk("Not Valid: %s\n", strblk(&ib->b));
1195                 printk("depth = %d\n", LAFSI(ino)->depth);
1196                 if (test_bit(B_InoIdx, &ib->b.flags))
1197                         printk("DB: %s\n", strblk(&LAFSI(ib->b.inode)->dblock->b));
1198                 LAFSI(ino)->trunc_next = next_trunc;
1199                 //BUG_ON(!test_bit(B_Valid, &ib->b.flags));
1200                 err = -ERESTARTSYS;
1201                 goto out;
1202         }
1203
1204         if (ib->b.fileaddr < trunc_next &&
1205             lafs_leaf_next(ib, 0) < trunc_next) {
1206                 /* We only want to truncate part of this index block.
1207                  * So we copy addresses into uninc_table and then
1208                  * call lafs_incorporate.
1209                  * This might cause the index tree to grow, so we
1210                  * cannot trust next_trunc
1211                  */
1212                 if (ib->uninc_table.pending_cnt == 0 &&
1213                     ib->uninc == NULL) {
1214                         lafs_dirty_iblock(ib, 0);
1215                         /* FIXME this just removes 8 blocks at a time,
1216                          * which is not enough
1217                          */
1218                         lafs_walk_leaf_index(ib, prune_some, ib);
1219                 }
1220                 if (test_bit(B_Dirty, &ib->b.flags))
1221                         lafs_incorporate(fs, ib);
1222                 err = -ERESTARTSYS;
1223                 goto out;
1224         }
1225         LAFSI(ino)->trunc_next = next_trunc;
1226
1227         while (ib->uninc_table.pending_cnt || ib->uninc) {
1228                 /* There should be no Realloc data blocks here
1229                  * but index blocks might be realloc still.
1230                  */
1231                 LAFS_BUG(!test_bit(B_Dirty, &ib->b.flags) &&
1232                          !test_bit(B_Realloc, &ib->b.flags), &ib->b);
1233                 lafs_incorporate(fs, ib);
1234         }
1235         if (test_bit(B_InoIdx, &ib->b.flags) ||
1236             !test_bit(B_PhysValid, &ib->b.flags) ||
1237             ib->b.physaddr != 0) {
1238                 lafs_walk_leaf_index(ib, prune, ib);
1239                 lafs_clear_index(ib);
1240                 lafs_dirty_iblock(ib, 0);
1241         }
1242         if (test_bit(B_Dirty, &ib->b.flags))
1243                 lafs_incorporate(fs, ib);
1244         if (!list_empty(&ib->children))
1245                 lafs_print_tree(&ib->b, 2);
1246         LAFS_BUG(!list_empty(&ib->children), &ib->b);
1247         err = -ERESTARTSYS;
1248 out:
1249         lafs_iounlock_block(&ib->b);
1250 out_unlocked:
1251         lafs_checkpoint_unlock(fs);
1252 out_put:
1253         putiref(ib, MKREF(inode_handle_orphan3));
1254         return err;
1255 }
1256
1257 void lafs_dirty_inode(struct inode *ino)
1258 {
1259         /* this is called in one of three cases:
1260          * 1/ by lafs internally when dblock or iblock is pinned and
1261          *    ready to be dirtied
1262          * 2/ by writeout before requesting a write - to update mtime
1263          * 3/ by read to update atime
1264          *
1265          * We want to handle atime updates carefully as they may not change
1266          * the stored inode itself.
1267          * For all other updates, the inode dblock exists and is pinned.
1268          * In those cases we will be updating the inode and so can store
1269          * the atime exactly.
1270          * For an atime update, the dblock may not exists, or may not be
1271          * Pinned.  If it isn't then we don't want to make the inode dirty
1272          * but only want to update the delta stored in the atime file.
1273          * The block for that should already be pinned.
1274          *
1275          *
1276          * We mustn't update the data block as it could be in
1277          * writeout and we cannot always wait safely.
1278          * So require that anyone who really cares, dirties the datablock
1279          * or a child themselves.
1280          * When cluster_allocate eventually gets called, it will update
1281          * the datablock from the inode.
1282          * If an update has to wait for the next phase, lock_dblock
1283          * (e.g. in setattr) will do that.
1284          *
1285          * We also use this opportunity to update the filesystem modify time.
1286          */
1287         struct timespec now;
1288         struct inode *filesys;
1289         int atime_only = 1;
1290
1291         if (LAFSI(ino)->dblock) {
1292                 struct datablock *db;
1293                 spin_lock(&ino->i_data.private_lock);
1294                 db = LAFSI(ino)->dblock;
1295                 if (db && test_bit(B_Pinned, &db->b.flags))
1296                         atime_only = 0;
1297                 spin_unlock(&ino->i_data.private_lock);
1298         }
1299
1300         if (atime_only) {
1301                 if (update_atime_delta(ino))
1302                         store_atime_delta(ino);
1303                 return;
1304         }
1305
1306         set_bit(I_Dirty, &LAFSI(ino)->iflags);
1307         ino->i_sb->s_dirt = 1;
1308
1309         if (LAFSI(ino)->type < TypeBase)
1310                 return;
1311         LAFSI(ino)->md.file.i_accesstime = ino->i_atime;
1312         if (LAFSI(ino)->md.file.atime_offset) {
1313                 LAFSI(ino)->md.file.atime_offset = 0;
1314                 store_atime_delta(ino);
1315         }
1316
1317         now = current_fs_time(ino->i_sb);
1318         filesys = LAFSI(ino)->filesys;
1319         if (!timespec_equal(&filesys->i_mtime, &now)) {
1320                 filesys->i_mtime = now;
1321                 set_bit(I_Dirty, &LAFSI(filesys)->iflags);
1322         }
1323 }
1324
1325 int lafs_sync_inode(struct inode *ino, int wait)
1326 {
1327         /* fsync has been called on this file so we need
1328          * to sync any inode updates to the next cluster.
1329          *
1330          * If we cannot create an update record,
1331          * we wait for a phase change, which writes everything
1332          * out.
1333          */
1334         struct datablock *b;
1335         struct fs *fs = fs_from_inode(ino);
1336         struct update_handle uh;
1337         int err;
1338
1339         if (wait) {
1340                 if (LAFSI(ino)->update_cluster > 1)
1341                         lafs_cluster_wait(fs, LAFSI(ino)->update_cluster);
1342                 if (LAFSI(ino)->update_cluster == 1) {
1343                         lafs_checkpoint_lock(fs);
1344                         lafs_checkpoint_unlock_wait(fs);
1345                 }
1346                 return 0;
1347         }
1348
1349         LAFSI(ino)->update_cluster = 0;
1350         if (!test_bit(I_Dirty, &LAFSI(ino)->iflags))
1351                 return 0;
1352         b = lafs_inode_dblock(ino, SYNC, MKREF(write_inode));
1353         if (IS_ERR(b))
1354                 return PTR_ERR(b);
1355
1356         lafs_iolock_written(&b->b);
1357         lafs_inode_fillblock(ino);
1358         lafs_iounlock_block(&b->b);
1359
1360         err = lafs_cluster_update_prepare(&uh, fs, LAFS_INODE_LOG_SIZE);
1361         if (err)
1362                 lafs_cluster_update_abort(&uh);
1363         else {
1364                 lafs_checkpoint_lock(fs);
1365                 if (lafs_cluster_update_pin(&uh) == 0) {
1366                         if (test_and_clear_bit(B_Dirty, &b->b.flags))
1367                                 lafs_space_return(fs, 1);
1368                         LAFSI(ino)->update_cluster =
1369                                 lafs_cluster_update_commit
1370                                 (&uh, b, LAFS_INODE_LOG_START,
1371                                  LAFS_INODE_LOG_SIZE);
1372                 } else
1373                         lafs_cluster_update_abort(&uh);
1374                 lafs_checkpoint_unlock(fs);
1375         }
1376         if (test_bit(B_Dirty, &b->b.flags)) {
1377                 /* FIXME need to write out the data block...
1378                  * Is that just lafs_cluster_allocate ?
1379                  */
1380         }
1381
1382         if (LAFSI(ino)->update_cluster == 0) {
1383                 lafs_checkpoint_lock(fs);
1384                 if (test_bit(B_Dirty, &b->b.flags))
1385                         LAFSI(ino)->update_cluster = 1;
1386                 lafs_checkpoint_start(fs);
1387                 lafs_checkpoint_unlock(fs);
1388         }
1389         putdref(b, MKREF(write_inode));
1390         return 0; /* FIXME should I return some error message??? */
1391 }
1392
1393 void lafs_inode_fillblock(struct inode *ino)
1394 {
1395         /* copy data from ino into the related data block */
1396
1397         struct lafs_inode *li = LAFSI(ino);
1398         struct datablock *db = li->dblock;
1399         struct la_inode *lai;
1400
1401         clear_bit(I_Dirty, &LAFSI(ino)->iflags);
1402
1403         lai = map_dblock(db);
1404         lai->data_blocks = cpu_to_le32(li->cblocks);
1405         lai->index_blocks = cpu_to_le32(li->ciblocks);
1406         lai->generation = cpu_to_le16(ino->i_generation);
1407         lai->trunc_gen = li->trunc_gen;
1408         lai->flags = li->flags;
1409         lai->filetype = li->type;
1410         if (lai->metadata_size != cpu_to_le16(li->metadata_size)) {
1411                 /* Changing metadata size is wierd.
1412                  * We will need to handle this somehow for xattrs
1413                  * For now we just want to cope with
1414                  * Dir -> InodeFile changes, and that guarantees us
1415                  * there is no index info - so just clear the index
1416                  * area.
1417                  */
1418                 u16 *s = (u16*)(((char*)lai) + li->metadata_size);
1419                 BUG_ON(li->type != TypeInodeFile);
1420                 lai->metadata_size = cpu_to_le16(li->metadata_size);
1421                 memset(s, 0, ino->i_sb->s_blocksize - li->metadata_size);
1422                 *s = cpu_to_le16(IBLK_INDIRECT);
1423         }
1424         lai->depth = li->depth;
1425
1426         switch (li->type) {
1427         case TypeInodeFile:
1428         {
1429                 struct fs_md *i = &li->md.fs;
1430                 struct fs_metadata *l = &lai->metadata[0].fs;
1431                 int nlen;
1432
1433                 l->snapshot_usage_table = cpu_to_le16(i->usagetable);
1434                 l->update_time = cpu_to_le64(encode_time(&ino->i_mtime));
1435                 l->blocks_used = cpu_to_le64(i->cblocks_used);
1436                 l->blocks_allowed = cpu_to_le64(i->blocks_allowed);
1437                 l->creation_age = cpu_to_le64(i->creation_age);
1438                 l->inodes_used = cpu_to_le32(i->inodes_used);
1439                 l->quota_inodes[0] = cpu_to_le32(i->quota_inums[0]);
1440                 l->quota_inodes[1] = cpu_to_le32(i->quota_inums[1]);
1441                 l->quota_inodes[2] = cpu_to_le32(i->quota_inums[2]);
1442                 nlen = lai->metadata_size - offsetof(struct la_inode,
1443                                                      metadata[0].fs.name);
1444                 memset(l->name, 0, nlen);
1445                 if (i->name == NULL)
1446                         nlen = 0;
1447                 else if (strlen(i->name) < nlen)
1448                         nlen = strlen(i->name);
1449                 memcpy(l->name, i->name, nlen);
1450                 break;
1451         }
1452
1453         case TypeInodeMap:
1454         {
1455                 struct inodemap_md *m = &li->md.inodemap;
1456                 struct inodemap_metadata *s = &lai->metadata[0].inodemap;
1457                 s->size = cpu_to_le32(m->size);
1458                 break;
1459         }
1460
1461         case TypeSegmentMap:
1462         {
1463                 struct su_md *m = &li->md.segmentusage;
1464                 struct su_metadata *s = &lai->metadata[0].segmentusage;
1465                 s->table_size = cpu_to_le32(m->table_size);
1466                 break;
1467         }
1468
1469         case TypeQuota:
1470         {
1471                 struct quota_md *m = &li->md.quota;
1472                 struct quota_metadata *s = &lai->metadata[0].quota;
1473                 s->gracetime = cpu_to_le32(m->gracetime);
1474                 s->graceunits = cpu_to_le32(m->graceunits);
1475                 break;
1476         }
1477         case TypeOrphanList:
1478         case TypeAccessTime:
1479                 break;
1480
1481         default: /* TypeBase or larger */
1482         {
1483                 struct file_md *i = &li->md.file;
1484                 struct file_metadata *l = &lai->metadata[0].file;
1485                 struct dir_metadata *d = &lai->metadata[0].dir;
1486                 struct special_metadata *s = &lai->metadata[0].special;
1487
1488                 if (li->type < TypeBase)
1489                         break;
1490                 l->flags = cpu_to_le16(i->flags);
1491                 l->mode = cpu_to_le16(ino->i_mode);
1492                 l->userid = cpu_to_le32(ino->i_uid);
1493                 l->groupid = cpu_to_le32(ino->i_gid);
1494                 l->treeid = cpu_to_le32(i->treeid);
1495                 l->creationtime = cpu_to_le64(i->creationtime);
1496                 l->modifytime = cpu_to_le64(encode_time(&ino->i_mtime));
1497                 l->ctime = cpu_to_le64(encode_time(&ino->i_ctime));
1498                 l->accesstime = cpu_to_le64(encode_time(&i->i_accesstime));
1499                 l->size = cpu_to_le64(ino->i_size);
1500                 l->parent = cpu_to_le32(i->parent);
1501                 l->linkcount = cpu_to_le32(ino->i_nlink);
1502
1503                 switch (li->type) {
1504                 case TypeFile:
1505                         break;
1506                 case TypeDir:
1507                         d->hash_seed = cpu_to_le32(i->seed);
1508                         break;
1509                 case TypeSymlink:
1510                         break;
1511                 case TypeSpecial:
1512                         s->major = cpu_to_le32(MAJOR(ino->i_rdev));
1513                         s->minor = cpu_to_le32(MINOR(ino->i_rdev));
1514                         break;
1515                 }
1516         }
1517         }
1518         unmap_dblock(db, lai);
1519 }
1520
1521 /*-----------------------------------------------------------------------
1522  * Inode allocate map handling.
1523  * Inode 1 of each fileset is a bitmap of free inode numbers.
1524  * Whenever the file is extended in size, new bits are set to one.  They
1525  * are then cleared when the inode is allocated.  When a block becomes
1526  * full of zeros, we don't need to store it any more.
1527  *
1528  * We don't clear the bit until we are committed to creating an inode
1529  * This means we cannot clear it straight away, so two different threads
1530  * might see the same inode number as being available.  We have two
1531  * approaches to guard against this.
1532  * Firstly we have a 'current' pointer into the inodemap file and
1533  * increase that past the inode we return.  This discourages multiple
1534  * hits but as the pointer would need to be rewound occasionally it
1535  * isn't a guarantee.  The guarantee against multiple allocations is done
1536  * via a flag in the block representing an inode.  This is set
1537  * while an inode is being allocated.
1538  */
1539
1540 /* inode number allocation has the prealloc/pin/commit/abort structure
1541  * so it can be committed effectively
1542  */
1543
1544 static int
1545 choose_free_inum(struct fs *fs, struct inode *fsys, u32 *inump,
1546                  struct datablock **bp, int *restarted)
1547 {
1548         struct inode *im = lafs_iget(fsys, 1, SYNC);
1549         loff_t bnum;
1550         struct datablock *b;
1551         char *buf;
1552         int err;
1553         int bit;
1554
1555         if (*bp) {
1556                 struct inode *i = (*bp)->b.inode;
1557                 putdref(*bp, MKREF(cfi_map));
1558                 iput(i);
1559                 *bp = NULL;
1560         }
1561
1562         mutex_lock_nested(&im->i_mutex, I_MUTEX_QUOTA);
1563 retry:
1564         bnum = LAFSI(im)->md.inodemap.thisblock;
1565
1566         if (bnum == NoBlock ||
1567             LAFSI(im)->md.inodemap.nextbit >= (fs->blocksize<<3)) {
1568                 if (bnum == NoBlock)
1569                         bnum = LAFSI(im)->md.inodemap.size;
1570
1571                 if (bnum+1 < LAFSI(im)->md.inodemap.size)
1572                         bnum++;
1573                 else if (!*restarted) {
1574                         bnum = 0;
1575                         *restarted = 1;
1576                 } else {
1577                         /* Need to add a new block to the file */
1578                         bnum = LAFSI(im)->md.inodemap.size;
1579                         b = lafs_get_block(im, bnum, NULL, GFP_KERNEL,
1580                                            MKREF(cfi_map));
1581                         err = -ENOMEM;
1582                         if (!b)
1583                                 goto abort;
1584                         lafs_iolock_written(&b->b);
1585                         set_bit(B_PinPending, &b->b.flags);
1586                         lafs_iounlock_block(&b->b);
1587                 retry2:
1588                         lafs_checkpoint_lock(fs);
1589                         err = lafs_pin_dblock(b, NewSpace);
1590                         if (err == -EAGAIN) {
1591                                 lafs_checkpoint_unlock_wait(fs);
1592                                 goto retry2;
1593                         }
1594                         if (err < 0)
1595                                 goto abort_unlock;
1596
1597                         buf = map_dblock(b);
1598                         /* Set block to "all are free" */
1599                         memset(buf, 0xff, fs->blocksize);
1600                         unmap_dblock(b, buf);
1601                         set_bit(B_Valid, &b->b.flags);
1602                         LAFSI(im)->md.inodemap.size = bnum+1;
1603                         lafs_dirty_inode(im);
1604                         lafs_dirty_dblock(b);
1605                         lafs_checkpoint_unlock(fs);
1606                         putdref(b, MKREF(cfi_map));
1607                 }
1608                 b = NULL;
1609                 err = lafs_find_next(im, &bnum);
1610                 if (err < 0)
1611                         goto abort;
1612                 if (err == 0)
1613                         bnum = 0;
1614
1615                 LAFSI(im)->md.inodemap.nextbit = 0;
1616                 LAFSI(im)->md.inodemap.thisblock = bnum;
1617                 goto retry;
1618         }
1619         b = lafs_get_block(im, bnum, NULL, GFP_KERNEL, MKREF(cfi_map));
1620         err = -ENOSPC;
1621         if (!b)
1622                 goto abort;
1623         err = lafs_find_block(b, NOADOPT);
1624         if (err)
1625                 goto abort;
1626         if (b->b.physaddr == 0 && !test_bit(B_Valid, &b->b.flags)) {
1627                 LAFSI(im)->md.inodemap.nextbit =
1628                         (fs->blocksize<<3) + 1;
1629                 putdref(b,MKREF(cfi_map));
1630                 goto retry;
1631         }
1632         err = lafs_read_block(b);
1633         if (err)
1634                 goto abort;
1635
1636         bit = LAFSI(im)->md.inodemap.nextbit;
1637         LAFSI(im)->md.inodemap.thisblock = bnum;
1638         buf = map_dblock(b);
1639         while (bnum == 0 && bit < 16) {
1640                 /* Never return an inum below 16 - they are special */
1641                 if (!generic_test_le_bit(bit, (unsigned long *)buf))
1642                         generic___clear_le_bit(bit, (unsigned long *)buf);
1643                 bit++;
1644         }
1645
1646         bit = generic_find_next_le_bit((unsigned long *)buf,
1647                                        fs->blocksize<<3, bit);
1648         unmap_dblock(b, buf);
1649         LAFSI(im)->md.inodemap.nextbit = bit+1;
1650         if (bit >= fs->blocksize<<3) {
1651                 putdref(b,MKREF(cfi_map));
1652                 goto retry;
1653         }
1654         mutex_unlock(&im->i_mutex);
1655         *bp = b;
1656         *inump = bit + (bnum << (im->i_blkbits + 3));
1657         return 0;
1658
1659 abort_unlock:
1660         lafs_checkpoint_unlock(fs);
1661 abort:
1662         putdref(b, MKREF(cfi_map));
1663         *bp = NULL;
1664         mutex_unlock(&im->i_mutex);
1665         iput(im);
1666         return err;
1667 }
1668
1669 struct inode_map_new_info {
1670         struct datablock *ib, *mb;
1671 };
1672
1673 static int
1674 inode_map_new_prepare(struct fs *fs, int inum, struct inode *fsys,
1675                       struct inode_map_new_info *imni)
1676 {
1677         int choice = inum;
1678         int restarted = 0;
1679         int err = 0;
1680         struct datablock *b;
1681
1682         imni->ib = imni->mb = NULL;
1683 retry:
1684         if (inum == 0)
1685                 /* choose a possibly-free inode number */
1686                 err = choose_free_inum(fs, fsys, &choice,
1687                                        &imni->mb, &restarted);
1688         if (err)
1689                 return err;
1690
1691         b = lafs_get_block(fsys, choice, NULL, GFP_KERNEL,
1692                            MKREF(cfi_ino));
1693         if (!b)
1694                 return -ENOMEM;
1695
1696         if (test_and_set_bit(B_Claimed, &b->b.flags)) {
1697                 putdref(b, MKREF(cfi_ino));
1698                 if (inum)
1699                         return -EEXIST;
1700                 goto retry;
1701         }
1702         if (imni->mb) {
1703                 lafs_iolock_written(&imni->mb->b);
1704                 set_bit(B_PinPending, &imni->mb->b.flags);
1705                 lafs_iounlock_block(&imni->mb->b);
1706         }
1707         set_bit(B_PinPending, &b->b.flags);
1708         b->my_inode = NULL;
1709         imni->ib = b;
1710         return 0;
1711 }
1712
1713 static int
1714 inode_map_new_pin(struct inode_map_new_info *imni)
1715 {
1716         int err = 0;
1717         if (imni->mb)
1718                 err = lafs_pin_dblock(imni->mb, NewSpace);
1719         err = err ?: lafs_pin_dblock(imni->ib, NewSpace);
1720         return err;
1721 }
1722
1723 static void
1724 inode_map_new_commit(struct inode_map_new_info *imni)
1725 {
1726         unsigned long *buf;
1727
1728         if (imni->mb) {
1729                 int blksize = imni->ib->b.inode->i_sb->s_blocksize;
1730                 int bit = imni->ib->b.fileaddr & (blksize*8 - 1);
1731                 int hole = 0;
1732                 struct inode *ino = imni->mb->b.inode;
1733
1734                 mutex_lock_nested(&ino->i_mutex, I_MUTEX_QUOTA);
1735                 buf = map_dblock(imni->mb);
1736                 generic___clear_le_bit(bit, buf);
1737                 if (buf[blksize/sizeof(*buf)-1] == 0 &&
1738                     generic_find_next_le_bit(buf, blksize*8, 0) == blksize*8)
1739                         /* block is empty, punch a hole */
1740                         hole = 1;
1741
1742                 unmap_dblock(imni->mb, buf);
1743                 if (hole)
1744                         lafs_erase_dblock(imni->mb);
1745                 else
1746                         lafs_dirty_dblock(imni->mb);
1747
1748                 putdref(imni->mb, MKREF(cfi_map));
1749                 mutex_unlock(&ino->i_mutex);
1750                 iput(ino);
1751         }
1752         putdref(imni->ib, MKREF(cfi_ino));
1753 }
1754
1755 static void
1756 inode_map_new_abort(struct inode_map_new_info *imni)
1757 {
1758         if (imni->ib) {
1759                 clear_bit(B_Claimed, &imni->ib->b.flags);
1760                 clear_bit(B_PinPending, &imni->ib->b.flags);
1761                 lafs_orphan_release(fs_from_inode(imni->ib->b.inode),
1762                                     imni->ib);
1763         }
1764         putdref(imni->ib, MKREF(cfi_ino));
1765         if (imni->mb) {
1766                 struct inode *ino = imni->mb->b.inode;
1767                 putdref(imni->mb, MKREF(cfi_map));
1768                 iput(ino);
1769         }
1770 }
1771
1772 struct inode *
1773 lafs_new_inode(struct fs *fs, struct inode *fsys, struct inode *dir,
1774                int type, int inum, int mode, struct datablock **inodbp)
1775 {
1776         /* allocate and instantiate a new inode.  If inum is non-zero,
1777          * choose any number, otherwise we are creating a special inode
1778          * and have to use the given number.
1779          * This creation is committed independently of any name that might
1780          * subsequently be given to the inode.  So we register it as an
1781          * orphan so that it will be cleaned up if the name isn't
1782          * successfully created
1783          *
1784          */
1785         struct inode *ino;
1786         struct datablock *b;
1787         struct inode_map_new_info imni;
1788         struct update_handle ui;
1789         int err;
1790
1791         err = inode_map_new_prepare(fs, inum, fsys, &imni);
1792         err = lafs_cluster_update_prepare(&ui, fs, sizeof(struct la_inode))
1793                 ?: err;
1794         if (err == 0)
1795                 err = lafs_make_orphan(fs, imni.ib);
1796         if (err)
1797                 goto abort;
1798 retry:
1799         lafs_checkpoint_lock(fs);
1800
1801         err = inode_map_new_pin(&imni);
1802
1803         if (err == -EAGAIN) {
1804                 lafs_checkpoint_unlock_wait(fs);
1805                 goto retry;
1806         }
1807         if (err < 0)
1808                 goto abort_unlock;
1809
1810         b = getdref(imni.ib, MKREF(inode_new));
1811
1812         lafs_iolock_block(&b->b); /* make sure we don't race with the cleaner
1813                                    * and zero this inode while trying to load it
1814                                    */
1815         lafs_inode_init(b, type, mode, dir);
1816         lafs_iounlock_block(&b->b);
1817
1818         inode_map_new_commit(&imni);
1819         ino = lafs_iget(fsys, b->b.fileaddr, SYNC);
1820         if (IS_ERR(ino)) {
1821                 lafs_cluster_update_abort(&ui);
1822                 LAFS_BUG(1, &b->b);
1823         } else
1824                 lafs_cluster_update_commit(&ui, b, 0,
1825                                            LAFSI(ino)->metadata_size);
1826         LAFS_BUG(LAFSI(ino)->dblock != b, &b->b);
1827         LAFS_BUG(b->my_inode != ino, &b->b);
1828         lafs_checkpoint_unlock(fs);
1829
1830         if (inodbp)
1831                 *inodbp = b;
1832         else
1833                 putdref(b, MKREF(inode_new));
1834         return ino;
1835
1836 abort_unlock:
1837         lafs_checkpoint_unlock(fs);
1838         err = -ENOSPC;
1839 abort:
1840         inode_map_new_abort(&imni);
1841         lafs_cluster_update_abort(&ui);
1842         dprintk("After abort %d: %s\n", err, strblk(&imni.ib->b));
1843         return ERR_PTR(err);
1844 }
1845
1846 static int inode_map_free(struct fs *fs, struct inode *fsys, u32 inum)
1847 {
1848         struct inode *im = lafs_iget(fsys, 1, SYNC);
1849         int bit;
1850         unsigned long *buf;
1851         struct datablock *b;
1852         u32 bnum;
1853         int err;
1854
1855         mutex_lock_nested(&im->i_mutex, I_MUTEX_QUOTA);
1856
1857         bnum = inum >> (3 + fs->blocksize_bits);
1858         bit = inum - (bnum << (3 + fs->blocksize_bits));
1859         b = lafs_get_block(im, bnum, NULL, GFP_KERNEL, MKREF(inode_map_free));
1860         if (!b) {
1861                 mutex_unlock(&im->i_mutex);
1862                 iput(im);
1863                 return -ENOMEM;
1864         }
1865         err = lafs_read_block(b);
1866         if (err) {
1867                 putdref(b, MKREF(inode_map_free));
1868                 mutex_unlock(&im->i_mutex);
1869                 iput(im);
1870                 return err;
1871         }
1872         lafs_iolock_written(&b->b);
1873         set_bit(B_PinPending, &b->b.flags);
1874         lafs_iounlock_block(&b->b);
1875 retry:
1876         lafs_checkpoint_lock(fs);
1877         err = lafs_pin_dblock(b, ReleaseSpace);
1878         if (err == -EAGAIN) {
1879                 lafs_checkpoint_unlock_wait(fs);
1880                 goto retry;
1881         }
1882         BUG_ON(err < 0);
1883         buf = map_dblock(b);
1884         generic___set_le_bit(bit, buf);
1885         unmap_dblock(b, buf);
1886         lafs_dirty_dblock(b);
1887         putdref(b, MKREF(inode_map_free));
1888         lafs_checkpoint_unlock(fs);
1889         mutex_unlock(&im->i_mutex);
1890         iput(im);
1891         return 0;
1892 }
1893
1894 int lafs_inode_inuse(struct fs *fs, struct inode *fsys, u32 inum)
1895 {
1896         /* This is used during roll-forward to register a newly created
1897          * inode in the inode map
1898          */
1899         struct inode *im = lafs_iget(fsys, 1, SYNC);
1900         int bit;
1901         unsigned long *buf;
1902         struct datablock *b;
1903         u32 bnum;
1904         int err;
1905
1906         mutex_lock_nested(&im->i_mutex, I_MUTEX_QUOTA);
1907
1908         bnum = inum >> (3 + fs->blocksize_bits);
1909         bit = inum - (bnum << (3 + fs->blocksize_bits));
1910         if (bnum > LAFSI(im)->md.inodemap.size) {
1911                 /* inum to unbelievably big */
1912                 mutex_unlock(&im->i_mutex);
1913                 iput(im);
1914                 return -EINVAL;
1915         }
1916         b = lafs_get_block(im, bnum, NULL, GFP_KERNEL, MKREF(inode_map_free));
1917         if (!b) {
1918                 mutex_unlock(&im->i_mutex);
1919                 iput(im);
1920                 return -ENOMEM;
1921         }
1922
1923         err = lafs_read_block(b);
1924         if (err) {
1925                 putdref(b, MKREF(inode_map_free));
1926                 mutex_unlock(&im->i_mutex);
1927                 iput(im);
1928                 return err;
1929         }
1930
1931         lafs_iolock_written(&b->b);
1932         set_bit(B_PinPending, &b->b.flags);
1933         lafs_iounlock_block(&b->b);
1934 retry:
1935         lafs_checkpoint_lock(fs);
1936         err = lafs_pin_dblock(b, CleanSpace);
1937         if (err == -EAGAIN) {
1938                 lafs_checkpoint_unlock_wait(fs);
1939                 goto retry;
1940         }
1941         BUG_ON(err < 0);
1942         buf = map_dblock(b);
1943         if (bnum == LAFSI(im)->md.inodemap.size) {
1944                 /* need to add a new block to the file */
1945                 memset(buf, 0xff, fs->blocksize);
1946                 LAFSI(im)->md.inodemap.size = bnum + 1;
1947                 lafs_dirty_inode(im);
1948         }
1949         generic___clear_le_bit(bit, buf);
1950         unmap_dblock(b, buf);
1951         lafs_dirty_dblock(b);
1952         putdref(b, MKREF(inode_map_free));
1953         lafs_checkpoint_unlock(fs);
1954         mutex_unlock(&im->i_mutex);
1955         iput(im);
1956         return 0;
1957 }
1958
1959
1960
1961 int lafs_setattr(struct dentry *dentry, struct iattr *attr)
1962 {
1963         int err;
1964         struct inode *ino = dentry->d_inode;
1965         struct fs *fs = fs_from_inode(ino);
1966         struct datablock *db;
1967
1968         err = inode_change_ok(ino, attr);
1969         db = lafs_inode_dblock(ino, SYNC, MKREF(setattr));
1970         if (IS_ERR(db))
1971                 err = PTR_ERR(db);
1972         if (err)
1973                 return err;
1974
1975         /* We don't need iolock_written here as we don't
1976          * actually change the inode block yet
1977          */
1978         lafs_iolock_block(&db->b);
1979         set_bit(B_PinPending, &db->b.flags);
1980         lafs_iounlock_block(&db->b);
1981
1982         /* FIXME quota stuff */
1983
1984 again:
1985         lafs_checkpoint_lock(fs);
1986         err = lafs_pin_dblock(db, ReleaseSpace);
1987         if (err == -EAGAIN) {
1988                 lafs_checkpoint_unlock_wait(fs);
1989                 goto again;
1990         }
1991
1992         if (!err) {
1993                 if ((attr->ia_valid & ATTR_SIZE) &&
1994                     attr->ia_size != i_size_read(ino))
1995                         truncate_setsize(ino, attr->ia_size);
1996                 setattr_copy(ino, attr);
1997                 mark_inode_dirty(ino);
1998
1999                 lafs_dirty_dblock(db);
2000         }
2001         clear_bit(B_PinPending, &db->b.flags);
2002         putdref(db, MKREF(setattr));
2003         lafs_checkpoint_unlock(fs);
2004
2005         return err;
2006 }
2007
2008 void lafs_truncate(struct inode *ino)
2009 {
2010         /* Want to truncate this file.
2011          * i_size has already been changed, and the address space
2012          * has been cleaned up.
2013          * So just start the background truncate
2014          */
2015         struct fs *fs = fs_from_inode(ino);
2016         struct datablock *db = lafs_inode_dblock(ino, SYNC, MKREF(trunc));
2017         loff_t trunc_block;
2018         DEFINE_WAIT(wq);
2019
2020         if (IS_ERR(db))
2021                 return;
2022
2023         trunc_block = ((i_size_read(ino) + fs->blocksize - 1)
2024                        >> fs->blocksize_bits);
2025         /* We hold i_mutex, so regular orphan processing cannot
2026          * contine - we have to push it forward ourselves.
2027          */
2028         while (test_bit(I_Trunc, &LAFSI(ino)->iflags) &&
2029                LAFSI(ino)->trunc_next < trunc_block) {
2030                 prepare_to_wait(&fs->async_complete, &wq,
2031                                 TASK_UNINTERRUPTIBLE);
2032                 lafs_inode_handle_orphan(db);
2033                 if (test_bit(B_Orphan, &db->b.flags))
2034                         schedule();
2035         }
2036         finish_wait(&fs->async_complete, &wq);
2037
2038         /* There is nothing we can do about errors here.  The
2039          * most likely are ENOMEM which itself is very unlikely.
2040          * If this doesn't get registered as an orphan .... maybe
2041          * it will have to wait until something else truncates it.
2042          */
2043         lafs_make_orphan(fs, db);
2044
2045         if (!test_and_set_bit(I_Trunc, &LAFSI(ino)->iflags))
2046                 lafs_igrab_fs(ino);
2047         if (trunc_block == 0)
2048                 LAFSI(ino)->trunc_gen++;
2049         LAFSI(ino)->trunc_next = trunc_block;
2050         putdref(db, MKREF(trunc));
2051 }
2052
2053 const struct inode_operations lafs_special_ino_operations = {
2054         .setattr        = lafs_setattr,
2055         .getattr        = lafs_getattr,
2056         .truncate       = lafs_truncate,
2057 };