super.c

   1
   2 /*
   3  * fs/lafs/super.c
   4  * Copyright (C) 2005-2009
   5  * Neil Brown <neilb@suse.de>
   6  * Released under the GPL, version 2
   7  */
   8
   9 #include        "lafs.h"
  10 #include        <linux/namei.h>
  11 #include        <linux/crc32.h>
  12 #include        <linux/statfs.h>
  13 #include        <linux/mount.h>
  14 #include        <linux/exportfs.h>
  15 #include        <linux/slab.h>
  16
  17 static struct super_operations lafs_sops;
  18 static const struct export_operations lafs_export_ops;
  19
  20 /*---------------------------------------------------------------------
  21  * Write out state and super blocks
  22  *  The super blocks only need to be written when the geometry of the
  23  *  array changes such as when a device is added, removed, or resized.
  24  *  So we don't bother with that just yet.
  25  *  The state block needs to be written - twice on each device - whenever
  26  *  a checkpoint is completed.  All copies are identical and the writes
  27  *  proceed in parallel.  There are 4 stateblock locations on each device.
  28  *  2 are typically less recent than the other two.  We over-write the
  29  *  less-recent copies.
  30  *  FIXME on a RAID4 we should pad the write to be a full stripe.
  31  *
  32  * Locking issues:  This is called from the checkpoint thread and so
  33  *  it does not race with anything else exclusive to that thread.
  34  *  The nonlog information needs to be reviewed once that functionality
  35  *  is implemented.
  36  */
  37
  38 int lafs_write_state(struct fs *fs)
  39 {
  40         struct lafs_state *st;
  41         int i, d;
  42
  43         fs->seq++;
  44         st = fs->state;
  45         st->seq = cpu_to_le32(fs->seq);
  46         st->nonlog_segment = cpu_to_le32(fs->nonlog_segment);
  47         st->nonlog_dev = cpu_to_le16(fs->nonlog_dev);
  48         st->nonlog_offset = cpu_to_le16(fs->nonlog_offset);
  49         st->nextyouth = cpu_to_le16(fs->youth_next);
  50         st->checkpointcluster = cpu_to_le64(fs->checkpointcluster);
  51         for (i = 0; i < fs->maxsnapshot; i++)
  52                 st->root_inodes[i] = cpu_to_le64(fs->ss[i].root_addr);
  53
  54         st->checksum = 0;
  55         st->checksum = crc32_le(0, (unsigned char *)st, fs->statesize);
  56
  57         for (d = 0; d < fs->devices ; d++)
  58                 for (i = (fs->seq & 1); i < 4 ; i += 2)
  59                         lafs_super_write(fs, d, fs->devs[d].stateaddr[i] >> 9,
  60                                          (char *)st, fs->statesize);
  61         lafs_super_wait(fs);
  62         /* FIXME what about a write error ??? */
  63         return 0;
  64 }
  65
  66 static int
  67 valid_devblock(struct lafs_dev *db, sector_t addr, sector_t size)
  68 {
  69         /* check that this devblock is valid, given that
  70          * it was found at sector 'addr'
  71          */
  72         u32 crc, crc2;
  73         u64 byteaddr;
  74         sector_t segsize;
  75         int i;
  76
  77         if (strncmp(db->idtag, "LaFS-DeviceBlock", 16) != 0)
  78                 return 0;
  79         if (strncmp(db->version, "AlphaDevel      ", 16) != 0)
  80                 return 0;
  81         /* uuid can be anything */
  82         crc = db->checksum;
  83         db->checksum = 0;
  84         crc2 = crc32_le(0, (unsigned char *)db, LAFS_DEVBLK_SIZE);
  85         db->checksum = crc;
  86         if (crc2 != crc) {
  87                 dprintk("%lx != %lx\n", (unsigned long)crc,
  88                         (unsigned long)crc2);
  89                 return 0;
  90         }
  91
  92         byteaddr = (u64)addr << 9; /* convert to byte */
  93         if (le64_to_cpu(db->devaddr[0]) != byteaddr &&
  94             le64_to_cpu(db->devaddr[1]) != byteaddr)
  95                 return 0;
  96
  97         if (db->statebits < 10 || db->statebits > 16)
  98                 return 0;
  99         if (db->blockbits < 9 || db->blockbits > 20)
 100                 return 0;
 101         if (le16_to_cpu(db->width) < 1 || le16_to_cpu(db->width) >= 512)
 102                 return 0;
 103         if (le32_to_cpu(db->stride) < 1)
 104                 return 0;
 105         /* devaddr[0] must be early, [1] must be late */
 106         if (le64_to_cpu(db->devaddr[0]) >=
 107             le64_to_cpu(db->segment_offset))
 108                 return 0;
 109
 110         if (le64_to_cpu(db->devaddr[1]) <
 111             le64_to_cpu(db->segment_offset) +
 112             ((((sector_t)le32_to_cpu(db->segment_count)
 113                * le32_to_cpu(db->segment_size)))
 114              << db->blockbits))
 115                 return 0;
 116
 117         /* 2 is an absolute minimum segment size, a few hundred is more
 118          * likely. We'll put a lower limit of 8, and an upper of 800000
 119          */
 120         if (le32_to_cpu(db->segment_size) < 8 ||
 121             le32_to_cpu(db->segment_size) > 800000)
 122                 return 0;
 123
 124         if (le32_to_cpu(db->segment_offset) >
 125             (le32_to_cpu(db->segment_size)<<db->blockbits) * 10)
 126                 return 0;
 127
 128         /* The 4 state blocks live before the first or after the last segment.
 129          * The distance from start of first to end of last is either:
 130          * - segment_count * segment_size  if width*stride <= segment_size
 131          * - (width-1) * stride + segment_size / width * segment_count
 132          *                if width * stride > segment_size
 133          */
 134         segsize = le32_to_cpu(db->segment_size);
 135         segsize *= le32_to_cpu(db->segment_count);
 136         if (le16_to_cpu(db->width) *  le32_to_cpu(db->stride)
 137             > le32_to_cpu(db->segment_size)) {
 138                 int stride = le32_to_cpu(db->stride);
 139                 int width = le16_to_cpu(db->width);
 140
 141                 sector_div(segsize, width);
 142                 segsize += (width - 1) * stride;
 143         }
 144         segsize <<= db->blockbits;
 145         for (i = 0; i < 4; i++) {
 146                 sector_t addr = le64_to_cpu(db->stateaddr[i]);
 147                 int offset = le32_to_cpu(db->segment_offset);
 148                 if (addr + (1<<db->statebits) > offset &&
 149                     addr < offset + segsize)
 150                         return 0;
 151                 if (addr + (1<<db->statebits) > (size << db->blockbits))
 152                         return 0;
 153         }
 154
 155         /* Check all segments fit within device */
 156         if (le32_to_cpu(db->segment_offset) + segsize > (size << db->blockbits))
 157                 return 0;
 158
 159         if (le32_to_cpu(db->level) > 10)
 160                 return 0;
 161
 162         /* I guess it look sane enough... */
 163         return 1;
 164 }
 165
 166 static int
 167 compare_dev(struct lafs_dev *orig, struct lafs_dev *new)
 168 {
 169         /* Both these are known to be valid.
 170          * Return:
 171          *   0 if they are for same filesystem, but 'new' is older
 172          *   1 if they are for same filesystem, and 'new' is newer
 173          *  -1 if they are for different filesystems
 174          */
 175         if (memcmp(orig->uuid, new->uuid, 16))
 176                 return -1;
 177         if (u32_after(le32_to_cpu(new->seq),
 178                       le32_to_cpu(orig->seq)))
 179                 return 1;
 180         return 0;
 181 }
 182
 183 static int
 184 valid_stateblock(struct lafs_state *st, struct lafs_dev *dv)
 185 {
 186         /* Given the 'dv' devblock, make sure 'st' is a valid
 187          * and consistent stateblock
 188          */
 189         u32 crc;
 190         if (strncmp(st->idtag, "LaFS-State-Block", 16) != 0)
 191                 return 0;
 192         if (strncmp(st->version, "AlphaDevel      ", 16) != 0)
 193                 return 0;
 194         crc = st->checksum;
 195         st->checksum = 0;
 196         if (crc32_le(0, (unsigned char *)st, 1<<dv->statebits) != crc)
 197                 return 0;
 198         st->checksum = crc;
 199
 200         if (memcmp(st->uuid, dv->uuid, 16))
 201                 return 0;
 202
 203         if (sizeof(*st) + le32_to_cpu(st->maxsnapshot) * 8
 204             > (1<<dv->statebits))
 205                 return 0;
 206
 207         return 1;
 208 }
 209
 210 static int
 211 compare_state(struct lafs_state *orig, struct lafs_state *new)
 212 {
 213         /* return 1 if 'new' is actually newer than 'orig'.
 214          * We already know they are both valid and have the same
 215          * uuid... I don't think there is anything else to be checked
 216          */
 217         return u32_after(le32_to_cpu(new->seq), le32_to_cpu(orig->seq));
 218 }
 219
 220 /*
 221  * Mount options.
 222  * As we can have multiple devices, things are slightly non-obvious.
 223  * The 'devname' can be either a device name, starting '/', or
 224  * a filesytem name (not starting '/').
 225  * The 'data' is a standard comma-separated list of options.
 226  * For 'mount' these are:
 227  *    dev=/dev/X
 228  *              - devices in addition to 'dev_name'
 229  *    new=/dev/X
 230  *              - A new device, with a superblock already present, to be added.
 231  *    incomplete
 232  *              - don't complain if not all devices are given
 233  *    ?? quota stuff, cleaning parameters,
 234  *
 235  * For 'remount', options are
 236  *    dev=  - add another device
 237  *    new=  - the device is being added.
 238  *
 239  */
 240
 241 struct options {
 242         int devcnt;
 243         int curr_dev;
 244         int statebits, blockbits;
 245         struct devent {
 246                 const char *dev;
 247                 int is_new;
 248                 int is_name;
 249                 struct block_device *bdev;
 250                 struct lafs_dev *devblock;
 251                 struct lafs_state *stateblock;
 252                 int devchoice, statechoice;
 253         } *devlist;
 254         const char *name;
 255 };
 256 static int
 257 count_devs(const char *name, char *data)
 258 {
 259         int cnt = 0;
 260         if (*name == '/')
 261                 cnt = 1;
 262         while (data && *data) {
 263                 if (strncmp(data, "dev=", 4) == 0)
 264                         cnt++;
 265                 if (strncmp(data, "new=", 4) == 0)
 266                         cnt++;
 267                 data = strchr(data, ',');
 268                 if (data)
 269                         data++;
 270         }
 271         return cnt;
 272 }
 273
 274 static int
 275 parse_opts(struct options *op, const char *name, char *data)
 276 {
 277         int dv = 0;
 278         char *p;
 279
 280         memset(op, 0, sizeof(*op));
 281         op->devcnt = count_devs(name, data);
 282         op->devlist = kzalloc(op->devcnt*sizeof(op->devlist[0]), GFP_KERNEL);
 283
 284         if (!op->devlist)
 285                 return -ENOMEM;
 286
 287         op->name = NULL;
 288         if (*name == '/') {
 289                 op->devlist[dv].is_name = 1;
 290                 op->devlist[dv++].dev = name;
 291         } else
 292                 op->name = name;
 293         while ((p = strsep(&data, ",")) != NULL) {
 294                 if (!*p)
 295                         continue;
 296                 if (strncmp(p, "dev=", 4) == 0)
 297                         op->devlist[dv++].dev = p+4;
 298                 else if (strncmp(p, "new=", 4) == 0) {
 299                         op->devlist[dv].is_new = 1;
 300                         op->devlist[dv++].dev = p+4;
 301                 } else {
 302                         printk(KERN_ERR
 303                                "LaFS: Unrecognised mount option \"%s\"\n", p);
 304                         return -EINVAL;
 305
 306                 }
 307         }
 308         op->devcnt = dv;
 309
 310         return 0;
 311 }
 312
 313 static int
 314 lafs_load_super(struct block_device *bdev, void *opv, int silent)
 315 {
 316         /* Find the devblock and the stateblock for this device
 317
 318          * Only do basic internal consistancy checks.  Inter-device
 319          * checks happen later
 320          */
 321         struct options *op = opv;
 322         struct devent *dv;
 323         struct page *pg;
 324         sector_t sect, dev_addr = 0, state_addr = 0;
 325         int err = 0;
 326         unsigned int n;
 327         int i;
 328         int have_dev = 0, have_state = 0;
 329         sector_t devsize;
 330
 331         dv = &op->devlist[op->curr_dev];
 332         BUG_ON(dv->devblock);
 333         BUG_ON(dv->stateblock);
 334
 335         n = queue_logical_block_size(bdev->bd_disk->queue);
 336         if (n < LAFS_DEVBLK_SIZE)
 337                 n = LAFS_DEVBLK_SIZE;
 338         BUG_ON(n > PAGE_SIZE);
 339         dv->devblock = kmalloc(n, GFP_KERNEL);
 340         if (!dv->devblock)
 341                 return -ENOMEM;
 342         pg = alloc_page(GFP_KERNEL);
 343         if (!pg)
 344                 return -ENOMEM;
 345
 346         devsize = i_size_read(bdev->bd_inode);
 347
 348         /* Now find a devblock, check the first two possible locations,
 349          * and the last two.  If two devblocks are found with different
 350          * uuids, we are confused!
 351          */
 352         sect = 0;
 353         for (i = 0; i < 4; i++) {
 354                 /* try to read block at 'sect' */
 355                 int ok = lafs_sync_page_io(bdev, sect, 0, n, pg, READ);
 356
 357                 if (ok && valid_devblock(page_address(pg), sect, devsize)) {
 358                         if (!have_dev) {
 359                                 have_dev = 1;
 360                                 memcpy(dv->devblock, page_address(pg), n);
 361                                 dev_addr = sect;
 362                         } else switch (compare_dev(dv->devblock,
 363                                                    page_address(pg))) {
 364                                 case 0: /* older, do nothing */
 365                                         break;
 366                                 case 1: /* newer, overwrite */
 367                                         memcpy(dv->devblock, page_address(pg), n);
 368                                         dev_addr = sect;
 369                                         break;
 370                                 default: /* inconsistent --- HELP */
 371                                         printk(KERN_ERR "LaFS: inconsistent device-blocks found.\n");
 372                                         err = -EINVAL;
 373                                         goto out;
 374                                 }
 375                 }
 376
 377                 if (i != 1)
 378                         sect += (n>>9);
 379                 else {
 380                         sect = devsize & ~(sector_t)(n-1);
 381                         sect >>= 9;
 382                         sect -= (n>>9)*2;
 383                 }
 384         }
 385         /* FIXME - we've lost the read error, if it was significant */
 386         err = -EINVAL;
 387         if (!have_dev) {
 388                 if (!silent)
 389                         printk(KERN_ERR "LaFS - no valid devblock found.\n");
 390                 goto out;
 391         }
 392
 393         /* OK, we have a valid devblock, that's nice.
 394          * Now we should be able to find some stateblocks.
 395          * The locations are in the devblock
 396          */
 397         n = le32_to_cpu(1<<dv->devblock->statebits);
 398         if ((n & (n-1)) ||
 399             n < queue_logical_block_size(bdev->bd_disk->queue) ||
 400             n > 128*1024) {
 401                 printk(KERN_ERR "LaFS: statesize of %u not acceptable.\n", n);
 402                 err = -EINVAL;
 403                 goto out;
 404         }
 405         dv->stateblock = kmalloc(n, GFP_KERNEL);
 406         err = -ENOMEM;
 407         if (!dv->stateblock)
 408                 goto out;
 409         for (i = 0; i < 4; i++) {
 410                 int ok;
 411                 sect = le64_to_cpu(dv->devblock->stateaddr[i])>>9;
 412                 ok = lafs_sync_page_io(bdev, sect, 0, n, pg, READ);
 413                 if (ok && valid_stateblock(page_address(pg), dv->devblock)) {
 414                         if (!have_state) {
 415                                 have_state = 1;
 416                                 memcpy(dv->stateblock, page_address(pg), n);
 417                                 state_addr = i;
 418                         } else if (compare_state(dv->stateblock,
 419                                                  page_address(pg))) {
 420                                 memcpy(dv->stateblock, page_address(pg), n);
 421                                 state_addr = i;
 422                         }
 423                 }
 424         }
 425
 426         if (have_state) {
 427                 err = 0;
 428                 dv->devchoice = dev_addr;
 429                 dv->statechoice = state_addr;
 430         } else {
 431                 err = -EINVAL;
 432                 if (!silent)
 433                         printk(KERN_ERR "LaFS: no valid stateblock found.\n");
 434         }
 435 out:
 436         page_cache_release(pg);
 437         return err;
 438 }
 439
 440 static int
 441 check_devs(struct options *op)
 442 {
 443         /* Check we have enough, that they are for the same
 444          * uuid, and they they don't overlap
 445          * Also check that 'seq' number of devblocks
 446          * are within '1'
 447          */
 448         int seqlo = le32_to_cpu(op->devlist[0].devblock->seq);
 449         int seqhi = le32_to_cpu(op->devlist[0].devblock->seq);
 450         int newdev = 0;
 451         int newstate = 0;
 452         int i, j;
 453
 454         for (i = 1; i < op->devcnt; i++) {
 455                 if (memcmp(op->devlist[0].stateblock->uuid,
 456                            op->devlist[i].stateblock->uuid,
 457                            16) != 0)
 458                         return -EINVAL;
 459
 460                 if (le32_to_cpu(op->devlist[i].devblock->seq) == seqlo)
 461                         ;
 462                 else if (le32_to_cpu(op->devlist[i].devblock->seq) == seqlo+1) {
 463                         newdev = i;
 464                         seqhi = seqlo+1;
 465                 } else if (le32_to_cpu(op->devlist[i].devblock->seq) == seqhi-1)
 466                         seqlo = seqhi-1;
 467                 else
 468                         return -EINVAL;
 469
 470                 if (u32_after(le32_to_cpu(op->devlist[i].stateblock->seq),
 471                               le32_to_cpu(op->devlist[newstate].
 472                                           stateblock->seq)))
 473                         newstate = i;
 474         }
 475         if (le32_to_cpu(op->devlist[newstate].stateblock->devices)
 476             != op->devcnt)
 477                 return -EINVAL;
 478
 479         op->statebits = op->devlist[0].devblock->statebits;
 480         op->blockbits = op->devlist[0].devblock->blockbits;
 481
 482         /* Now check devices don't overlap in start/size.
 483          * We do a simple quadratic search
 484          */
 485         for (i = 0; i < op->devcnt; i++)
 486                 for (j = 0; j < op->devcnt; j++)
 487                         if (i != j)
 488                                 if (le64_to_cpu(op->devlist[i].devblock->start) <
 489                                     le64_to_cpu(op->devlist[j].devblock->start) &&
 490
 491                                     le64_to_cpu(op->devlist[i].devblock->start)+
 492                                     le64_to_cpu(op->devlist[i].devblock->size) >
 493                                     le64_to_cpu(op->devlist[j].devblock->start))
 494                                         return -EINVAL;
 495         return newstate;
 496 }
 497
 498 /* we identify lafs superblocks by the filesystem uuid.  This means
 499  * that block-level snapshots cannot be mounted.  You should use
 500  * fs-level snapshots instead.
 501  */
 502 static int sb_test(struct super_block *sb, void *data)
 503 {
 504         struct sb_key *ptn = data;
 505         struct sb_key *sk = sb->s_fs_info;
 506         return memcmp(ptn->fs->state->uuid,
 507                       sk->fs->state->uuid, 16) == 0;
 508 }
 509
 510 static int sb_set(struct super_block *sb, void *data)
 511 {
 512         struct sb_key *ptn = data;
 513         sb->s_fs_info = ptn;
 514         return set_anon_super(sb, NULL);
 515 }
 516
 517
 518 static int
 519 lafs_load(struct fs *fs, struct options *op, int newest)
 520 {
 521         /* We seem to have a full set of devices for the filesystem.
 522          * Time to create our fs_info structure and fill it out.
 523          * This only includes information from the dev and state blocks.
 524          * Finding the root-inode comes a bit later.
 525          */
 526         struct lafs_state *st;
 527         int i;
 528         int err;
 529         struct sb_key *k;
 530
 531         st = fs->state = op->devlist[newest].stateblock;
 532         op->devlist[newest].stateblock = NULL;
 533 #ifdef DUMP
 534         dfs = fs;
 535 #endif
 536
 537         fs->seq = le32_to_cpu(st->seq);
 538         fs->levels = le32_to_cpu(st->levels);
 539         fs->devices = op->devcnt;
 540         fs->devs_loaded = fs->devices; /* FIXME use this or lose this */
 541         fs->statesize = 1 << op->statebits;
 542         fs->blocksize = 1 << op->blockbits;
 543         fs->blocksize_bits = op->blockbits;
 544
 545         fs->nonlog_segment = le32_to_cpu(st->nonlog_segment);
 546         fs->nonlog_dev = le16_to_cpu(st->nonlog_dev);
 547         fs->nonlog_offset = le16_to_cpu(st->nonlog_offset);
 548         fs->youth_next = le16_to_cpu(st->nextyouth);
 549         fs->checkpoint_youth = fs->youth_next;
 550         if (fs->youth_next < 8)
 551                 fs->youth_next = 8;
 552         fs->scan.first_free_pass = 1;
 553         fs->scan.free_dev = -1;
 554
 555         fs->maxsnapshot = le32_to_cpu(st->maxsnapshot);
 556
 557         fs->scan.free_usages = kmalloc(PAGE_SIZE, GFP_KERNEL);
 558         err = lafs_segtrack_init(fs->segtrack);
 559
 560         fs->ss = kzalloc(sizeof(struct snapshot)*fs->maxsnapshot, GFP_KERNEL);
 561         if (!fs->ss || !fs->scan.free_usages || err) {
 562                 if (!err)
 563                         err = -ENOMEM;
 564                 goto abort;
 565         }
 566
 567         fs->checkpointcluster = le64_to_cpu(st->checkpointcluster);
 568         for (i = 0; i < fs->maxsnapshot; i++) {
 569                 fs->ss[i].root_addr =
 570                         le64_to_cpu(st->root_inodes[i]);
 571                 dprintk("root inode %d are %llu\n",
 572                         i, fs->ss[i].root_addr);
 573         }
 574         INIT_LIST_HEAD(&fs->pending_orphans);
 575         INIT_LIST_HEAD(&fs->inode_index);
 576         INIT_LIST_HEAD(&fs->phase_leafs[0]);
 577         INIT_LIST_HEAD(&fs->phase_leafs[1]);
 578         INIT_LIST_HEAD(&fs->clean_leafs);
 579         INIT_LIST_HEAD(&fs->account_leafs);
 580         atomic_set(&fs->sb_writes_pending, 0);
 581         init_waitqueue_head(&fs->sb_writes_wait);
 582         init_waitqueue_head(&fs->async_complete);
 583         init_waitqueue_head(&fs->trunc_wait);
 584         mutex_init(&fs->cleaner.lock);
 585         spin_lock_init(&fs->stable_lock);
 586         spin_lock_init(&fs->alloc_lock);
 587         spin_lock_init(&fs->lock);
 588         init_waitqueue_head(&fs->phase_wait);
 589
 590         INIT_WORK(&fs->done_work, lafs_done_work);
 591
 592         /* FIXME add congention and unplug functions to this bdi */
 593         err = bdi_init(&fs->bdi);
 594         if (err)
 595                 goto abort;
 596
 597
 598         fs->phase_locked = 0;
 599         for (i = 0; i < WC_NUM; i++) {
 600                 int j;
 601                 mutex_init(&fs->wc[i].lock);
 602                 for (j = 0; j < 4 ; j++) {
 603                         atomic_set(&fs->wc[i].pending_cnt[j], 0);
 604                         INIT_LIST_HEAD(&fs->wc[i].pending_blocks[j]);
 605                 }
 606                 init_waitqueue_head(&fs->wc[i].pending_wait);
 607                 fs->wc[i].seg.dev = -1;
 608         }
 609
 610         fs->max_newsegs = 32; /* FIXME this should be configurable */
 611
 612         err = -ENOMEM;
 613         fs->devs = kzalloc(sizeof(struct fs_dev)*fs->devices, GFP_KERNEL);
 614         if (!fs->devs)
 615                 goto abort;
 616
 617         k = kzalloc(sizeof(*k), GFP_KERNEL);
 618         k->fs = fs;
 619         fs->prime_sb = sget(&lafs_fs_type, sb_test, sb_set, k);
 620         if (IS_ERR(fs->prime_sb)) {
 621                 kfree(k);
 622                 err = PTR_ERR(fs->prime_sb);
 623                 goto abort;
 624         }
 625         if (fs->prime_sb->s_root) {
 626                 /* filesystem with this uuid already exists */
 627                 deactivate_locked_super(fs->prime_sb);
 628                 kfree(k);
 629                 fs->prime_sb = NULL;
 630                 err = -EBUSY;
 631                 goto abort;
 632         }
 633         err = bdi_register_dev(&fs->bdi, fs->prime_sb->s_dev);
 634         if (err) {
 635                 deactivate_locked_super(fs->prime_sb);
 636                 kfree(k);
 637                 fs->prime_sb = NULL;
 638                 goto abort;
 639         }
 640         fs->prime_sb->s_bdi = &fs->bdi;
 641
 642         fs->prime_sb->s_blocksize = 1 << op->blockbits;
 643         fs->prime_sb->s_blocksize_bits = op->blockbits;
 644         fs->prime_sb->s_op = &lafs_sops;
 645         fs->prime_sb->s_export_op = &lafs_export_ops;
 646         fs->prime_sb->s_root = NULL;
 647
 648         /* We allow 29 bits for nanosecs, so they must be even. */
 649         fs->prime_sb->s_time_gran = 2;
 650
 651         for (i = 0; i < fs->devices; i++) {
 652                 struct fs_dev *dv = &fs->devs[i];
 653                 struct devent *de = &op->devlist[i];
 654                 int j;
 655                 dv->bdev = de->bdev;
 656                 de->bdev = NULL;
 657
 658                 dv->devblk = de->devblock;
 659                 de->devblock = NULL;
 660
 661                 dv->recent_dev = de->devchoice;
 662                 dv->recent_state = de->statechoice;
 663
 664                 dv->start = le64_to_cpu(dv->devblk->start);
 665                 dv->size = le64_to_cpu(dv->devblk->size);
 666                 dprintk("Dev %d seems to range %llu + %llu\n",
 667                         i, (unsigned long long)dv->start,
 668                         (unsigned long long)dv->size);
 669
 670                 dv->width = le16_to_cpu(dv->devblk->width);
 671                 dv->stride = le32_to_cpu(dv->devblk->stride);
 672                 dv->segment_size = le32_to_cpu(dv->devblk->segment_size);
 673                 dv->segment_offset = le32_to_cpu(dv->devblk->segment_offset);
 674                 dv->segment_count = le32_to_cpu(dv->devblk->segment_count);
 675                 dv->usage_inum = le32_to_cpu(dv->devblk->usage_inum);
 676                 dv->level = le16_to_cpu(dv->devblk->level);
 677
 678                 if (dv->segment_size > fs->max_segment)
 679                         fs->max_segment = dv->segment_size;
 680
 681                 if (dv->width * dv->stride <= dv->segment_size) {
 682                         dv->tables_per_seg = dv->segment_size /
 683                                 dv->width / dv->stride;
 684                         dv->rows_per_table = dv->stride;
 685                         dv->segment_stride = dv->segment_size;
 686                 } else {
 687                         dv->tables_per_seg = 1;
 688                         dv->rows_per_table = dv->segment_size / dv->width;
 689                         dv->segment_stride = dv->rows_per_table;
 690                 }
 691                 /* table size is the number of blocks in the segment usage
 692                  * file per snapshot
 693                  */
 694                 dv->tablesize = (dv->segment_count + (1<<(fs->blocksize_bits-1)) + 1)
 695                         >> (fs->blocksize_bits-1);
 696
 697                 for (j = 0; j < 2; j++)
 698                         dv->devaddr[j] = le64_to_cpu(dv->devblk->devaddr[j]);
 699                 for (j = 0; j < 4; j++)
 700                         dv->stateaddr[j] = le64_to_cpu(dv->devblk->stateaddr[j]);
 701         }
 702         return 0;
 703
 704 abort:
 705         bdi_destroy(&fs->bdi);
 706         kfree(fs->scan.free_usages);
 707         lafs_segtrack_free(fs->segtrack);
 708         kfree(fs->devs);
 709         kfree(fs->ss);
 710         kfree(fs);
 711         return -ENOMEM;
 712 }
 713
 714 static int show_orphans(struct fs *fs)
 715 {
 716         struct datablock *db;
 717         printk("Orphans:\n");
 718         list_for_each_entry(db, &fs->pending_orphans,
 719                             orphans) {
 720                 struct inode *ino = iget_my_inode(db);
 721                 printk("orphan=%s\n", strblk(&db->b));
 722                 if (ino)
 723                         lafs_print_tree(&LAFSI(ino)->iblock->b, 0);
 724                 iput(ino);
 725         }
 726         printk("cleaner active: %d %d\n", fs->cleaner.active,
 727                fs->scan.done);
 728         return 1; /* meaningless, but makes it easy to add to wait_event below */
 729 }
 730
 731 static void lafs_kill_sb(struct super_block *sb)
 732 {
 733         struct fs *fs = fs_from_sb(sb);
 734         /* Release the 'struct fs' */
 735         int i;
 736
 737         /* FIXME should I refcount this when there are multiple
 738          * filesets? How does that work?
 739          */
 740
 741         /* Delay final destruction of the root inode */
 742         /* FIXME all the sbs... */
 743         set_bit(I_Deleting, &LAFSI(fs->ss[0].root)->iflags);
 744
 745         /* FIXME I'm not sure we should be waiting for the
 746          * cleaner.  Maybe we should just release all tc->cleaning
 747          * blocks instead.
 748          */
 749         set_bit(CleanerDisabled, &fs->fsstate);
 750
 751         wait_event(fs->async_complete,
 752                    show_orphans(fs) &&
 753                    !test_bit(OrphansRunning, &fs->fsstate) &&
 754                    list_empty(&fs->pending_orphans) &&
 755                    fs->scan.done == 1 &&
 756                    fs->cleaner.active == 0);
 757
 758         if (LAFSI(fs->ss[0].root)->md.fs.accesstime) {
 759                 struct inode *i = LAFSI(fs->ss[0].root)->md.fs.accesstime;
 760                 LAFSI(fs->ss[0].root)->md.fs.accesstime = NULL;
 761                 iput(i);
 762         }
 763
 764         kill_anon_super(fs->prime_sb);
 765
 766         bdi_destroy(&fs->bdi);
 767
 768         for (i = 0; i < fs->devices; i++) {
 769                 struct fs_dev *dv = &fs->devs[i];
 770                 kfree(dv->devblk);
 771                 close_bdev_exclusive(dv->bdev, FMODE_READ|FMODE_WRITE);
 772         }
 773
 774         /* Final checkpoint will have cleared out the leafs lists,
 775          * so they should all be empty.
 776          */
 777         /* Lets see what is on the 'leaf' list? */
 778         for (i = 0; i < 2; i++) {
 779                 struct block *b;
 780                 dprintk("For phase %d\n", i);
 781         retry:
 782                 list_for_each_entry(b, &fs->phase_leafs[i], lru) {
 783                         /* FIXME this only OK for readonly mounts.
 784                          */
 785                         getref(b, MKREF(release));
 786                         lafs_refile(b, 0);
 787                         if (test_bit(B_Pinned, &b->flags)) {
 788                                 /* didn't fix the pincnt !! */
 789                                 printk("This was pinned: %s\n", strblk(b));
 790                                 lafs_print_tree(b, 1);
 791                                 BUG();
 792                         }
 793                         putref(b, MKREF(release));
 794                         goto retry;
 795                 }
 796         }
 797         BUG_ON(!list_empty(&fs->clean_leafs));
 798
 799         flush_scheduled_work();
 800         lafs_stop_thread(fs);
 801
 802         for (i = 0; i < 4; i++)
 803                 if (fs->cleaner.seg[i].chead)
 804                         put_page(fs->cleaner.seg[i].chead);
 805
 806         kfree(fs->state);
 807         kfree(fs->ss);
 808         kfree(fs->devs);
 809         lafs_segtrack_free(fs->segtrack);
 810         kfree(fs->scan.free_usages);
 811         kfree(fs->prime_sb->s_fs_info);
 812         kfree(fs);
 813 }
 814
 815 static void
 816 lafs_put_super(struct super_block *sb)
 817 {
 818         struct fs *fs = fs_from_sb(sb);
 819         int ss;
 820         struct lafs_inode *li;
 821
 822         lafs_checkpoint_lock(fs);
 823         lafs_checkpoint_start(fs);
 824         if (sb == fs->prime_sb)
 825                 /* Don't incorporate any more segusage/quota updates. */
 826                 set_bit(FinalCheckpoint, &fs->fsstate);
 827         lafs_checkpoint_unlock_wait(fs);
 828         lafs_cluster_wait_all(fs);
 829
 830         if (sb == fs->prime_sb) {
 831                 int d;
 832                 /* This is the main sb, not a snapshot or
 833                  * subordinate fs.
 834                  * Now that all inodes have been invalidated we can do
 835                  * the final checkpoint.
 836                  */
 837                 lafs_close_all_segments(fs);
 838                 lafs_empty_segment_table(fs);
 839                 lafs_seg_put_all(fs);
 840
 841                 iput(fs->orphans);
 842                 fs->orphans = NULL;
 843                 for (d=0; d < fs->devices; d++)
 844                         if (fs->devs[d].segsum) {
 845                                 iput(fs->devs[d].segsum);
 846                                 fs->devs[d].segsum = NULL;
 847                         }
 848         }
 849
 850         /* need to break a circular reference... */
 851         for (ss = 0; ss < fs->maxsnapshot; ss++)
 852                 if (fs->ss[ss].root &&
 853                     fs->ss[ss].root->i_sb == sb) {
 854                         dprintk("Putting ss %d\n", ss);
 855                         li = LAFSI(fs->ss[ss].root);
 856                         if (test_bit(B_Realloc, &li->dblock->b.flags))
 857                                 lafs_dump_tree();
 858                         iput(fs->ss[ss].root);
 859                         fs->ss[ss].root = NULL;
 860                         break;
 861                 }
 862 }
 863
 864 static int
 865 lafs_get_devs(struct fs *fs, struct options *op, int flags)
 866 {
 867         int err;
 868         int i;
 869
 870         for (i = 0; i < op->devcnt; i++) {
 871                 struct block_device *bdev;
 872                 op->curr_dev = i;
 873
 874                 bdev = open_bdev_exclusive(op->devlist[i].dev,
 875                                            FMODE_READ|FMODE_WRITE, fs);
 876                 err = PTR_ERR(bdev);
 877                 if (IS_ERR(bdev))
 878                         goto out;
 879                 err = lafs_load_super(bdev, op, flags & MS_SILENT ? 1 : 0);
 880                 if (err < 0)
 881                         goto out;
 882                 op->devlist[i].bdev = bdev;
 883         }
 884         return 0;
 885
 886 out:
 887         return err;
 888 }
 889
 890 static int
 891 lafs_get_sb(struct file_system_type *fs_type,
 892             int flags, const char *dev_name, void *data,
 893             struct vfsmount *mnt)
 894 {
 895         /* as we may have multiple devices, some in 'data', we cannot just
 896          * use get_sb_bdev, we need to roll-our-own.
 897          * We call get_sb_bdev on *each* bdev, and make sure the returned
 898          * superblocks are either all new, or all for the same filesystem.
 899          * If the later, we return the primary.
 900          * If the former, we init the filesystem copying static data
 901          * to all supers.
 902          * First we 'open_bdev_exclusive' each device, exclusive to lafs
 903          * Then we 'sget' a superblock that knows any/all the devices.
 904          * This may be pre-existing, or may be new
 905          * If new, it will be created knowing all devices.
 906          * If pre-existing, and don't have correct device list, error
 907          */
 908         struct options op;
 909         int err;
 910         int newest;
 911         struct fs *fs = kzalloc(sizeof(*fs), GFP_KERNEL);
 912         char *cdata = data;
 913         if (cdata == NULL)
 914                 cdata = "";
 915
 916         err = -ENOMEM;
 917         if (!fs)
 918                 goto out;
 919         err = parse_opts(&op, dev_name, cdata);
 920         if (err)
 921                 goto out;
 922
 923         /* We now have as list of device names.  We call open_bdev_exclusive
 924          * on each to collect some superblocks.
 925          */
 926         err = lafs_get_devs(fs, &op, flags);
 927         if (err)
 928                 goto out;
 929
 930         /* Each device has a valid dev and state block.  Hopefully they
 931          * are all for the same filesystem.  If they don't have the
 932          * same uuid, we will bale-out here.  We also check that we have
 933          * enough, and that they don't overlap.
 934          * While we are looking at state blocks, pick the newest.
 935          */
 936         newest = check_devs(&op);
 937         if (newest < 0) {
 938                 err = newest;
 939                 goto out;
 940         }
 941
 942         /* So they seem to be the same - better create our
 943          * 'fs' structure and fill it in
 944          */
 945         err = lafs_load(fs, &op, newest);
 946         if (err)
 947                 goto out;
 948
 949         /* Well, all the devices check out.  Now we need to find the
 950          * filesystem */
 951         err = lafs_mount(fs);
 952         if (err == 0)
 953                 err = lafs_start_thread(fs);
 954         if (err)
 955                 deactivate_locked_super(fs->prime_sb);
 956         else {
 957                 fs->prime_sb->s_flags |= MS_ACTIVE;
 958                 simple_set_mnt(mnt, fs->prime_sb);
 959         }
 960         /* And there you have it.  Filesystem all mounted, root dir found,
 961          * metadata files initialised, all pigs fed, and ready to fly!!!
 962          */
 963
 964 out:
 965         /* Now we clean up 'options'.  Anything that is wanted has
 966          * been moved into 'fs', so we just discard anything we find
 967          */
 968         if (op.devlist) {
 969                 int i;
 970                 for (i = 0; i < op.devcnt; i++) {
 971                         kfree(op.devlist[i].devblock);
 972                         kfree(op.devlist[i].stateblock);
 973                         if (op.devlist[i].bdev)
 974                                 close_bdev_exclusive(op.devlist[i].bdev,
 975                                                      FMODE_READ|FMODE_WRITE);
 976                 }
 977                 kfree(op.devlist);
 978         }
 979         return err;
 980 }
 981
 982 static int test_subset(struct super_block *sb, void *data)
 983 {
 984         struct sb_key *ptn = data;
 985         struct sb_key *k = sb->s_fs_info;
 986
 987         return ptn->fs == k->fs && ptn->root == k->root;
 988 }
 989
 990 static int set_subset(struct super_block *sb, void *data)
 991 {
 992         sb->s_fs_info = data;
 993         set_anon_super(sb, NULL);
 994         return 0;
 995 }
 996
 997 static struct file_system_type lafs_subset_fs_type;
 998 struct super_block *lafs_get_subset_sb(struct inode *ino)
 999 {
1000         /* ino must be a TypeInodeFile inode in the prime filesystem. */
1001         struct fs *fs = fs_from_inode(ino);
1002         struct super_block *sb;
1003         struct sb_key *k = kmalloc(sizeof(*k), GFP_KERNEL);
1004
1005         if (!k)
1006                 return ERR_PTR(-ENOMEM);
1007
1008         k->fs = fs;
1009         k->root = ino;
1010         sb = sget(&lafs_subset_fs_type, test_subset, set_subset, k);
1011         if (IS_ERR(sb)) {
1012                 kfree(k);
1013         } else if (sb->s_root) {
1014                 /* already allocated */
1015                 kfree(k);
1016         } else {
1017                 struct inode *rootdir, *imapfile;
1018                 int err = 0;
1019
1020                 igrab(ino);
1021                 sb->s_blocksize = fs->blocksize;
1022                 sb->s_blocksize_bits = fs->blocksize_bits;
1023                 sb->s_bdi = fs->prime_sb->s_bdi;
1024                 sb->s_op = &lafs_sops;
1025                 sb->s_export_op = &lafs_export_ops;
1026                 sb->s_time_gran = 2;
1027                 rootdir = lafs_iget(sb, 2, SYNC);
1028                 if (IS_ERR(rootdir) && PTR_ERR(rootdir) == -ENOENT) {
1029                         rootdir = lafs_new_inode(fs, sb, NULL,
1030                                                  TypeDir, 2, 0755, NULL);
1031                         /* FIXME could the inode get written before we set
1032                          * the link count ??*/
1033                         rootdir->i_nlink = 2;
1034                 }
1035                 if (IS_ERR(rootdir))
1036                         err = PTR_ERR(rootdir);
1037                 else {
1038                         sb->s_root = d_alloc_root(rootdir);
1039                         imapfile = lafs_iget(sb, 1, SYNC);
1040                         if (IS_ERR(imapfile) && PTR_ERR(imapfile) == -ENOENT)
1041                                 imapfile = lafs_new_inode(fs, sb, NULL,
1042                                                           TypeInodeMap, 1, 0, NULL);
1043
1044                         if (IS_ERR(imapfile))
1045                                 err = PTR_ERR(imapfile);
1046                         else
1047                                 iput(imapfile);
1048                 }
1049
1050                 if (!err) {
1051                         struct inode *atime = lafs_iget(sb, 3, SYNC);
1052                         if (!IS_ERR(atime)) {
1053                                 if (LAFSI(atime)->type != TypeAccessTime) {
1054                                         iput(atime);
1055                                         err = -EINVAL;
1056                                 } else
1057                                         LAFSI(ino)->md.fs.accesstime = atime;
1058                         } else if (PTR_ERR(atime) != -ENOENT)
1059                                 err = PTR_ERR(ino);
1060                 }
1061
1062                 if (!err) {
1063                         sb->s_op = fs->prime_sb->s_op;
1064                         sb->s_flags |= MS_ACTIVE;
1065                         atomic_inc(&fs->prime_sb->s_active);
1066                         igrab(ino);
1067                 } else {
1068                         deactivate_locked_super(sb);
1069                         sb = ERR_PTR(err);
1070                 }
1071         }
1072         return sb;
1073 }
1074
1075 static int
1076 lafs_get_subset(struct file_system_type *fs_type,
1077                 int flags, const char *dev_name, void *data,
1078                 struct vfsmount *mnt)
1079 {
1080         /* mount, possibly creating, a sub-fileset.
1081          * dev_name must be an absolute path that leads
1082          * to an object in a lafs file-system (or snapshot).
1083          * The object must be either an InodeFile or
1084          * an empty directory in the main file-system
1085          * with mode 0 (though that rule might change).
1086          * In the latter case we change the object to an
1087          * InodeFile
1088          * FIXME must require readonly for snapshots, and readwrite
1089          * to create.
1090          */
1091
1092         struct nameidata nd;
1093         int err;
1094         struct super_block *sb;
1095         struct inode *ino;
1096         struct fs *fs;
1097
1098         err = path_lookup(dev_name, LOOKUP_FOLLOW, &nd);
1099         if (err)
1100                 goto out_noput;
1101         sb = nd.path.dentry->d_sb;
1102         err = -EINVAL;
1103         if (sb->s_type != &lafs_fs_type &&
1104             sb->s_type != &lafs_snap_fs_type)
1105                 goto out;
1106         ino = nd.path.dentry->d_inode;
1107         if (LAFSI(ino)->type != TypeInodeFile &&
1108             LAFSI(ino)->type != TypeDir)
1109                 goto out;
1110         fs = fs_from_sb(sb);
1111         mutex_lock(&ino->i_mutex);
1112         if (LAFSI(ino)->type == TypeDir) {
1113                 struct datablock *inodb;
1114                 /* maybe convert this to TypeInodeFile */
1115                 if (sb->s_type != &lafs_fs_type)
1116                         goto out_unlock;
1117                 if (ino->i_size)
1118                         /* FIXME maybe I should run orphans */
1119                         goto out_unlock;
1120                 if ((ino->i_mode & 07777) != 0)
1121                         goto out_unlock;
1122                 inodb = lafs_inode_dblock(ino, SYNC, MKREF(make_subset));
1123                 err = PTR_ERR(inodb);
1124                 if (IS_ERR(inodb))
1125                         goto out_unlock;
1126                 lafs_iolock_block(&inodb->b);
1127                 set_bit(B_PinPending, &inodb->b.flags);
1128                 lafs_iounlock_block(&inodb->b);
1129                 lafs_checkpoint_lock(fs);
1130                 err = lafs_pin_dblock(inodb, ReleaseSpace);
1131                 if (!err) {
1132                         struct fs_md *md;
1133                         /* OK, we are good to go making this filesystem */
1134                         LAFSI(ino)->type = TypeInodeFile;
1135                         LAFSI(ino)->metadata_size = (sizeof(struct la_inode) +
1136                                                      sizeof(struct fs_metadata));
1137                         ino->i_op = &lafs_subset_ino_operations;
1138                         ino->i_fop = &lafs_subset_file_operations;
1139                         /* FIXME we lose md->parent here - what to do?? */
1140                         md = &LAFSI(ino)->md.fs;
1141                         md->usagetable = 0;
1142                         ino->i_mtime = current_fs_time(sb);
1143                         md->cblocks_used = 0;
1144                         md->pblocks_used = 0;
1145                         md->ablocks_used = 0;
1146                         md->blocks_allowed = 10000; /* FIXME */
1147                         md->blocks_unalloc = 0;
1148                         /* FIXME should I be using inode_init here */
1149                         md->creation_age = fs->wc[0].cluster_seq;
1150                         md->inodes_used = 0;
1151                         md->quota_inums[0] = 0;
1152                         md->quota_inums[1] = 0;
1153                         md->quota_inums[2] = 0;
1154                         md->quota_inodes[0] = NULL;
1155                         md->quota_inodes[1] = NULL;
1156                         md->quota_inodes[2] = NULL;
1157                         md->accesstime = NULL;
1158                         md->name = NULL;
1159                         lafs_dirty_dblock(inodb);
1160                         lafs_dirty_inode(ino);
1161                         /* We use a checkpoint to commit this change,
1162                          * it is too unusual to bother logging
1163                          */
1164                         lafs_checkpoint_start(fs);
1165                         lafs_checkpoint_unlock_wait(fs);
1166                 } else {
1167                         lafs_checkpoint_unlock(fs);
1168                 }
1169                 putdref(inodb, MKREF(make_subset));
1170                 if (err)
1171                         goto out_unlock;
1172         }
1173         err = 0;
1174         /* We have a TypeInodeFile so we can make a superblock */
1175         sb = lafs_get_subset_sb(ino);
1176         iput(ino);
1177
1178         if (IS_ERR(sb))
1179                 err = PTR_ERR(sb);
1180         else
1181                 simple_set_mnt(mnt, sb);
1182 out_unlock:
1183         mutex_unlock(&ino->i_mutex);
1184 out:
1185         path_put(&nd.path);
1186 out_noput:
1187         return err;
1188 }
1189
1190 static void lafs_kill_subset(struct super_block *sb)
1191 {
1192         struct sb_key *k = sb->s_fs_info;
1193         if (LAFSI(k->root)->md.fs.accesstime) {
1194                 iput(LAFSI(k->root)->md.fs.accesstime);
1195                 LAFSI(k->root)->md.fs.accesstime = NULL;
1196         }
1197         kill_anon_super(sb);
1198         iput(k->root);
1199         deactivate_super(k->fs->prime_sb);
1200         kfree(k);
1201 }
1202
1203 const struct file_operations lafs_subset_file_operations = {
1204 };
1205
1206 const struct inode_operations lafs_subset_ino_operations = {
1207 };
1208
1209
1210 struct file_system_type lafs_fs_type = {
1211         .owner          = THIS_MODULE,
1212         .name           = "lafs",
1213         .get_sb         = lafs_get_sb,
1214         .kill_sb        = lafs_kill_sb,
1215         .fs_flags       = FS_REQUIRES_DEV,
1216 };
1217
1218 static struct file_system_type lafs_subset_fs_type = {
1219         .owner          = THIS_MODULE,
1220         .name           = "lafs_subset",
1221         .get_sb         = lafs_get_subset,
1222         .kill_sb        = lafs_kill_subset,
1223 };
1224
1225 static int __init lafs_init(void)
1226 {
1227         int err;
1228
1229         BUILD_BUG_ON(B_NUM_FLAGS > 32);
1230
1231         err = lafs_ihash_init();
1232         err = err ?: register_filesystem(&lafs_fs_type);
1233         err = err ?: register_filesystem(&lafs_snap_fs_type);
1234         err = err ?: register_filesystem(&lafs_subset_fs_type);
1235         if (err)
1236                 goto out;
1237         return 0;
1238
1239 out:
1240         unregister_filesystem(&lafs_fs_type);
1241         unregister_filesystem(&lafs_snap_fs_type);
1242         unregister_filesystem(&lafs_subset_fs_type);
1243         lafs_ihash_free();
1244         return err;
1245 }
1246
1247 static void __exit lafs_exit(void)
1248 {
1249         unregister_filesystem(&lafs_fs_type);
1250         unregister_filesystem(&lafs_snap_fs_type);
1251         unregister_filesystem(&lafs_subset_fs_type);
1252         lafs_ihash_free();
1253 }
1254
1255 static struct inode *lafs_nfs_get_inode(struct super_block *sb,
1256                                         u64 ino, u32 generation)
1257 {
1258         struct inode *inode;
1259
1260         inode = lafs_iget(sb, ino, SYNC);
1261         if (IS_ERR(inode))
1262                 return ERR_CAST(inode);
1263         if (generation && inode->i_generation != generation) {
1264                 iput(inode);
1265                 return ERR_PTR(-ESTALE);
1266         }
1267
1268         return inode;
1269 }
1270
1271 static struct dentry *lafs_fh_to_dentry(struct super_block *sb, struct fid *fid,
1272                                         int fh_len, int fh_type)
1273 {
1274         return generic_fh_to_dentry(sb, fid, fh_len, fh_type,
1275                                     lafs_nfs_get_inode);
1276 }
1277
1278 static struct dentry *lafs_fh_to_parent(struct super_block *sb, struct fid *fid,
1279                                         int fh_len, int fh_type)
1280 {
1281         return generic_fh_to_parent(sb, fid, fh_len, fh_type,
1282                                     lafs_nfs_get_inode);
1283 }
1284
1285 static struct dentry *lafs_get_parent(struct dentry *child)
1286 {
1287         ino_t inum = LAFSI(child->d_inode)->md.file.parent;
1288         struct inode *inode = lafs_iget(child->d_inode->i_sb, inum, SYNC);
1289         if (IS_ERR(inode))
1290                 return ERR_CAST(inode);
1291         return d_obtain_alias(inode);
1292 }
1293
1294 static const struct export_operations lafs_export_ops = {
1295         .fh_to_dentry = lafs_fh_to_dentry,
1296         .fh_to_parent = lafs_fh_to_parent,
1297         .get_parent = lafs_get_parent,
1298 };
1299
1300 static struct inode *lafs_alloc_inode(struct super_block *sb)
1301 {
1302         struct lafs_inode *li;
1303         li = kmalloc(sizeof(*li), GFP_NOFS);
1304         if (!li)
1305                 return NULL;
1306         inode_init_once(&li->vfs_inode);
1307         li->vfs_inode.i_data.backing_dev_info = sb->s_bdi;
1308         li->iblock = NULL;
1309         li->dblock = NULL;
1310         li->update_cluster = 0;
1311         li->md.fs.name = NULL;
1312
1313         init_rwsem(&li->ind_sem);
1314         INIT_LIST_HEAD(&li->free_index);
1315
1316         return &li->vfs_inode;
1317 }
1318
1319 static void kfree_inode(struct rcu_head *head)
1320 {
1321         struct lafs_inode *lai = container_of(head, struct lafs_inode,
1322                                               md.rcu);
1323         if (lai->type == TypeInodeFile)
1324                 kfree(lai->md.fs.name);
1325         kfree(lai);
1326 }
1327
1328 void lafs_destroy_inode(struct inode *inode)
1329 {
1330         struct datablock *db;
1331
1332         BUG_ON(!list_empty(&inode->i_sb_list));
1333         // Cannot test i_list as dispose_list just does list_del
1334         db = lafs_inode_get_dblock(inode, MKREF(destroy));
1335
1336         if (db) {
1337                 set_bit(I_Destroyed, &LAFSI(inode)->iflags);
1338                 putdref(db, MKREF(destroy));
1339         } else {
1340                 spin_lock(&inode->i_data.private_lock);
1341                 if (LAFSI(inode)->iblock)
1342                         LAFS_BUG(atomic_read(&LAFSI(inode)->iblock->b.refcnt),
1343                                  &LAFSI(inode)->iblock->b);
1344                 /* FIXME could there be Async blocks keeps a refcount?
1345                  * we should free them
1346                  */
1347                 spin_unlock(&inode->i_data.private_lock);
1348                 lafs_release_index(&LAFSI(inode)->free_index);
1349                 call_rcu(&LAFSI(inode)->md.rcu,
1350                          kfree_inode);
1351         }
1352 }
1353
1354 static int lafs_sync_fs(struct super_block *sb, int wait)
1355 {
1356         if (!wait)
1357                 /* We only reach here if s_dirt was set, so it
1358                  * is reasonable to force a checkpoint.
1359                  */
1360                 lafs_checkpoint_start(fs_from_sb(sb));
1361         else
1362                 lafs_checkpoint_wait(fs_from_sb(sb));
1363         return 0;
1364 }
1365
1366 static int lafs_statfs(struct dentry *de, struct kstatfs *buf)
1367 {
1368         int i;
1369         u32 fsid;
1370         u32 *fsuuid;
1371         struct fs *fs = fs_from_inode(de->d_inode);
1372         struct lafs_inode *fsroot = LAFSI(ino_from_sb(de->d_inode->i_sb));
1373         struct lafs_inode *laroot = LAFSI(fs->ss[0].root);
1374
1375         fsid = 0;
1376         fsuuid = (u32 *)fs->state->uuid;
1377         for (i = 0; i < 16 / 4 ; i++) {
1378                 fsid ^= le32_to_cpu(fsuuid[i]);
1379                 buf->f_fsid.val[i/2] = fsid;
1380         }
1381         buf->f_fsid.val[1] ^= fsroot->vfs_inode.i_ino;
1382         buf->f_type = 0x4C614654; /* "LaFS" */
1383         buf->f_bsize = fs->blocksize;
1384         buf->f_blocks = fsroot->md.fs.blocks_allowed;
1385         if (buf->f_blocks == 0) {
1386                 /* should subtract usage of all other filesystems...*/
1387                 for (i = 0; i < fs->devs_loaded; i++)
1388                         buf->f_blocks += fs->devs[i].size;
1389         }
1390
1391         buf->f_files = 0;
1392         buf->f_ffree = 0;
1393         buf->f_namelen = 255;
1394         buf->f_frsize = 0;
1395
1396         spin_lock(&laroot->vfs_inode.i_lock);
1397         /* "bavail" is "blocks we could succeed in adding to the filesystem".
1398          * "bfree" is effectively total blocks - used blocks
1399          */
1400         buf->f_bavail = fs->free_blocks + fs->clean_reserved - fs->allocated_blocks;
1401         spin_unlock(&laroot->vfs_inode.i_lock);
1402         spin_lock(&fsroot->vfs_inode.i_lock);
1403         buf->f_bfree = buf->f_blocks - (fsroot->md.fs.cblocks_used +
1404                                         fsroot->md.fs.pblocks_used +
1405                                         fsroot->md.fs.ablocks_used);
1406         dprintk("df: tot=%ld free=%ld avail=%ld(%ld-%ld-%ld) cb=%ld pb=%ld ab=%ld\n",
1407                 (long)buf->f_blocks, (long)buf->f_bfree, (long)buf->f_bavail,
1408                 (long)fs->free_blocks, (long)fs->clean_reserved,
1409                 (long)fs->allocated_blocks,
1410                 (long)fsroot->md.fs.cblocks_used, (long)fsroot->md.fs.pblocks_used,
1411                 (long)fsroot->md.fs.ablocks_used);
1412         spin_unlock(&fsroot->vfs_inode.i_lock);
1413         return 0;
1414 }
1415
1416 /* FIXME we hold inode_lock while calling drop_inode, so
1417  * extra locking isn't really welcome....???
1418  */
1419 static void lafs_drop_inode(struct inode *inode)
1420 {
1421         struct fs *fs = fs_from_inode(inode);
1422         struct datablock *db;
1423
1424         /* This lock that we now hold on the inode could prevent
1425          * the cleaner from getting the inode.  So after
1426          * the complete the drop we might need to wake the cleaner.
1427          */
1428
1429         db = lafs_inode_get_dblock(inode, MKREF(drop));
1430
1431         generic_drop_inode(inode);
1432         if (db && test_bit(B_Async, &db->b.flags))
1433                 lafs_wake_thread(fs);
1434         if (db)
1435                 putdref(db, MKREF(drop));
1436 }
1437
1438 static struct super_operations lafs_sops = {
1439         .alloc_inode    = lafs_alloc_inode,
1440         .destroy_inode  = lafs_destroy_inode,  /* Inverse of 'alloc_inode' */
1441         /* Don't use read_inode */
1442         .dirty_inode    = lafs_dirty_inode,
1443         /* .write_inode not needed */
1444         /* put_inode ?? */
1445         .drop_inode     = lafs_drop_inode,
1446         /* drop_inode ?? */                     /* default will call delete or forget
1447                                                  * where 'forget' flushes and clears
1448                                                  */
1449
1450         .clear_inode    = lafs_clear_inode,    /* forget internal state of this inode */
1451         .delete_inode   = lafs_delete_inode,   /* remove this inode from filesystem */
1452         .put_super      = lafs_put_super,
1453         .sync_fs        = lafs_sync_fs,
1454         /* write_super_lockfs ?? */
1455         /* unlockfs ?? */
1456         .statfs         = lafs_statfs,
1457         /* remount_fs ?? */
1458 };
1459
1460 MODULE_AUTHOR("Neil Brown");
1461 MODULE_DESCRIPTION("LaFS - Log Structured File System");
1462 MODULE_LICENSE("GPL");
1463 module_init(lafs_init);
1464 module_exit(lafs_exit);
1465 int lafs_trace = 1;
1466 module_param(lafs_trace, int, 0644);
1467
1468 #ifdef DUMP
1469 struct fs *dfs;
1470 static int do_dump(const char *val, struct kernel_param *kp)
1471 {
1472         extern void lafs_dump_orphans(void);
1473         extern void lafs_dump_tree(void);
1474         extern void lafs_dump_cleanable(void);
1475         extern void lafs_dump_usage(void);
1476
1477         printk("Want dump of %s\n", val);
1478         if (strncmp(val, "orphan", 6) == 0)
1479                 lafs_dump_orphans();
1480         if (strncmp(val, "tree", 4) == 0)
1481                 lafs_dump_tree();
1482         if (strncmp(val, "cleanable", 9) == 0)
1483                 lafs_dump_cleanable();
1484         if (strncmp(val, "usage", 5) == 0)
1485                 lafs_dump_usage();
1486         return 0;
1487 }
1488
1489 static int get_dump(char *buffer, struct kernel_param *kp)
1490 {
1491         strcpy(buffer, "orphans,tree,cleanable,usage");
1492         return strlen(buffer);
1493 }
1494
1495 int arg;
1496 module_param_call(dump, do_dump, get_dump, &arg, 0775);
1497 #endif