SAFE public projects git trees. - safe/jmp/linux-2.6/blob - fs/logfs/readwrite.c

   1 /*
   2  * fs/logfs/readwrite.c
   3  *
   4  * As should be obvious for Linux kernel code, license is GPLv2
   5  *
   6  * Copyright (c) 2005-2008 Joern Engel <joern@logfs.org>
   7  *
   8  *
   9  * Actually contains five sets of very similar functions:
  10  * read         read blocks from a file
  11  * seek_hole    find next hole
  12  * seek_data    find next data block
  13  * valid        check whether a block still belongs to a file
  14  * write        write blocks to a file
  15  * delete       delete a block (for directories and ifile)
  16  * rewrite      move existing blocks of a file to a new location (gc helper)
  17  * truncate     truncate a file
  18  */
  19 #include "logfs.h"
  20 #include <linux/sched.h>
  21
  22 static u64 adjust_bix(u64 bix, level_t level)
  23 {
  24         switch (level) {
  25         case 0:
  26                 return bix;
  27         case LEVEL(1):
  28                 return max_t(u64, bix, I0_BLOCKS);
  29         case LEVEL(2):
  30                 return max_t(u64, bix, I1_BLOCKS);
  31         case LEVEL(3):
  32                 return max_t(u64, bix, I2_BLOCKS);
  33         case LEVEL(4):
  34                 return max_t(u64, bix, I3_BLOCKS);
  35         case LEVEL(5):
  36                 return max_t(u64, bix, I4_BLOCKS);
  37         default:
  38                 WARN_ON(1);
  39                 return bix;
  40         }
  41 }
  42
  43 static inline u64 maxbix(u8 height)
  44 {
  45         return 1ULL << (LOGFS_BLOCK_BITS * height);
  46 }
  47
  48 /**
  49  * The inode address space is cut in two halves.  Lower half belongs to data
  50  * pages, upper half to indirect blocks.  If the high bit (INDIRECT_BIT) is
  51  * set, the actual block index (bix) and level can be derived from the page
  52  * index.
  53  *
  54  * The lowest three bits of the block index are set to 0 after packing and
  55  * unpacking.  Since the lowest n bits (9 for 4KiB blocksize) are ignored
  56  * anyway this is harmless.
  57  */
  58 #define ARCH_SHIFT      (BITS_PER_LONG - 32)
  59 #define INDIRECT_BIT    (0x80000000UL << ARCH_SHIFT)
  60 #define LEVEL_SHIFT     (28 + ARCH_SHIFT)
  61 static inline pgoff_t first_indirect_block(void)
  62 {
  63         return INDIRECT_BIT | (1ULL << LEVEL_SHIFT);
  64 }
  65
  66 pgoff_t logfs_pack_index(u64 bix, level_t level)
  67 {
  68         pgoff_t index;
  69
  70         BUG_ON(bix >= INDIRECT_BIT);
  71         if (level == 0)
  72                 return bix;
  73
  74         index  = INDIRECT_BIT;
  75         index |= (__force long)level << LEVEL_SHIFT;
  76         index |= bix >> ((__force u8)level * LOGFS_BLOCK_BITS);
  77         return index;
  78 }
  79
  80 void logfs_unpack_index(pgoff_t index, u64 *bix, level_t *level)
  81 {
  82         u8 __level;
  83
  84         if (!(index & INDIRECT_BIT)) {
  85                 *bix = index;
  86                 *level = 0;
  87                 return;
  88         }
  89
  90         __level = (index & ~INDIRECT_BIT) >> LEVEL_SHIFT;
  91         *level = LEVEL(__level);
  92         *bix = (index << (__level * LOGFS_BLOCK_BITS)) & ~INDIRECT_BIT;
  93         *bix = adjust_bix(*bix, *level);
  94         return;
  95 }
  96 #undef ARCH_SHIFT
  97 #undef INDIRECT_BIT
  98 #undef LEVEL_SHIFT
  99
 100 /*
 101  * Time is stored as nanoseconds since the epoch.
 102  */
 103 static struct timespec be64_to_timespec(__be64 betime)
 104 {
 105         return ns_to_timespec(be64_to_cpu(betime));
 106 }
 107
 108 static __be64 timespec_to_be64(struct timespec tsp)
 109 {
 110         return cpu_to_be64((u64)tsp.tv_sec * NSEC_PER_SEC + tsp.tv_nsec);
 111 }
 112
 113 static void logfs_disk_to_inode(struct logfs_disk_inode *di, struct inode*inode)
 114 {
 115         struct logfs_inode *li = logfs_inode(inode);
 116         int i;
 117
 118         inode->i_mode   = be16_to_cpu(di->di_mode);
 119         li->li_height   = di->di_height;
 120         li->li_flags    = be32_to_cpu(di->di_flags);
 121         inode->i_uid    = be32_to_cpu(di->di_uid);
 122         inode->i_gid    = be32_to_cpu(di->di_gid);
 123         inode->i_size   = be64_to_cpu(di->di_size);
 124         logfs_set_blocks(inode, be64_to_cpu(di->di_used_bytes));
 125         inode->i_atime  = be64_to_timespec(di->di_atime);
 126         inode->i_ctime  = be64_to_timespec(di->di_ctime);
 127         inode->i_mtime  = be64_to_timespec(di->di_mtime);
 128         inode->i_nlink  = be32_to_cpu(di->di_refcount);
 129         inode->i_generation = be32_to_cpu(di->di_generation);
 130
 131         switch (inode->i_mode & S_IFMT) {
 132         case S_IFSOCK:  /* fall through */
 133         case S_IFBLK:   /* fall through */
 134         case S_IFCHR:   /* fall through */
 135         case S_IFIFO:
 136                 inode->i_rdev = be64_to_cpu(di->di_data[0]);
 137                 break;
 138         case S_IFDIR:   /* fall through */
 139         case S_IFREG:   /* fall through */
 140         case S_IFLNK:
 141                 for (i = 0; i < LOGFS_EMBEDDED_FIELDS; i++)
 142                         li->li_data[i] = be64_to_cpu(di->di_data[i]);
 143                 break;
 144         default:
 145                 BUG();
 146         }
 147 }
 148
 149 static void logfs_inode_to_disk(struct inode *inode, struct logfs_disk_inode*di)
 150 {
 151         struct logfs_inode *li = logfs_inode(inode);
 152         int i;
 153
 154         di->di_mode     = cpu_to_be16(inode->i_mode);
 155         di->di_height   = li->li_height;
 156         di->di_pad      = 0;
 157         di->di_flags    = cpu_to_be32(li->li_flags);
 158         di->di_uid      = cpu_to_be32(inode->i_uid);
 159         di->di_gid      = cpu_to_be32(inode->i_gid);
 160         di->di_size     = cpu_to_be64(i_size_read(inode));
 161         di->di_used_bytes = cpu_to_be64(li->li_used_bytes);
 162         di->di_atime    = timespec_to_be64(inode->i_atime);
 163         di->di_ctime    = timespec_to_be64(inode->i_ctime);
 164         di->di_mtime    = timespec_to_be64(inode->i_mtime);
 165         di->di_refcount = cpu_to_be32(inode->i_nlink);
 166         di->di_generation = cpu_to_be32(inode->i_generation);
 167
 168         switch (inode->i_mode & S_IFMT) {
 169         case S_IFSOCK:  /* fall through */
 170         case S_IFBLK:   /* fall through */
 171         case S_IFCHR:   /* fall through */
 172         case S_IFIFO:
 173                 di->di_data[0] = cpu_to_be64(inode->i_rdev);
 174                 break;
 175         case S_IFDIR:   /* fall through */
 176         case S_IFREG:   /* fall through */
 177         case S_IFLNK:
 178                 for (i = 0; i < LOGFS_EMBEDDED_FIELDS; i++)
 179                         di->di_data[i] = cpu_to_be64(li->li_data[i]);
 180                 break;
 181         default:
 182                 BUG();
 183         }
 184 }
 185
 186 static void __logfs_set_blocks(struct inode *inode)
 187 {
 188         struct super_block *sb = inode->i_sb;
 189         struct logfs_inode *li = logfs_inode(inode);
 190
 191         inode->i_blocks = ULONG_MAX;
 192         if (li->li_used_bytes >> sb->s_blocksize_bits < ULONG_MAX)
 193                 inode->i_blocks = ALIGN(li->li_used_bytes, 512) >> 9;
 194 }
 195
 196 void logfs_set_blocks(struct inode *inode, u64 bytes)
 197 {
 198         struct logfs_inode *li = logfs_inode(inode);
 199
 200         li->li_used_bytes = bytes;
 201         __logfs_set_blocks(inode);
 202 }
 203
 204 static void prelock_page(struct super_block *sb, struct page *page, int lock)
 205 {
 206         struct logfs_super *super = logfs_super(sb);
 207
 208         BUG_ON(!PageLocked(page));
 209         if (lock) {
 210                 BUG_ON(PagePreLocked(page));
 211                 SetPagePreLocked(page);
 212         } else {
 213                 /* We are in GC path. */
 214                 if (PagePreLocked(page))
 215                         super->s_lock_count++;
 216                 else
 217                         SetPagePreLocked(page);
 218         }
 219 }
 220
 221 static void preunlock_page(struct super_block *sb, struct page *page, int lock)
 222 {
 223         struct logfs_super *super = logfs_super(sb);
 224
 225         BUG_ON(!PageLocked(page));
 226         if (lock)
 227                 ClearPagePreLocked(page);
 228         else {
 229                 /* We are in GC path. */
 230                 BUG_ON(!PagePreLocked(page));
 231                 if (super->s_lock_count)
 232                         super->s_lock_count--;
 233                 else
 234                         ClearPagePreLocked(page);
 235         }
 236 }
 237
 238 /*
 239  * Logfs is prone to an AB-BA deadlock where one task tries to acquire
 240  * s_write_mutex with a locked page and GC tries to get that page while holding
 241  * s_write_mutex.
 242  * To solve this issue logfs will ignore the page lock iff the page in question
 243  * is waiting for s_write_mutex.  We annotate this fact by setting PG_pre_locked
 244  * in addition to PG_locked.
 245  */
 246 static void logfs_get_wblocks(struct super_block *sb, struct page *page,
 247                 int lock)
 248 {
 249         struct logfs_super *super = logfs_super(sb);
 250
 251         if (page)
 252                 prelock_page(sb, page, lock);
 253
 254         if (lock) {
 255                 mutex_lock(&super->s_write_mutex);
 256                 logfs_gc_pass(sb);
 257                 /* FIXME: We also have to check for shadowed space
 258                  * and mempool fill grade */
 259         }
 260 }
 261
 262 static void logfs_put_wblocks(struct super_block *sb, struct page *page,
 263                 int lock)
 264 {
 265         struct logfs_super *super = logfs_super(sb);
 266
 267         if (page)
 268                 preunlock_page(sb, page, lock);
 269         /* Order matters - we must clear PG_pre_locked before releasing
 270          * s_write_mutex or we could race against another task. */
 271         if (lock)
 272                 mutex_unlock(&super->s_write_mutex);
 273 }
 274
 275 static struct page *logfs_get_read_page(struct inode *inode, u64 bix,
 276                 level_t level)
 277 {
 278         return find_or_create_page(inode->i_mapping,
 279                         logfs_pack_index(bix, level), GFP_NOFS);
 280 }
 281
 282 static void logfs_put_read_page(struct page *page)
 283 {
 284         unlock_page(page);
 285         page_cache_release(page);
 286 }
 287
 288 static void logfs_lock_write_page(struct page *page)
 289 {
 290         int loop = 0;
 291
 292         while (unlikely(!trylock_page(page))) {
 293                 if (loop++ > 0x1000) {
 294                         /* Has been observed once so far... */
 295                         printk(KERN_ERR "stack at %p\n", &loop);
 296                         BUG();
 297                 }
 298                 if (PagePreLocked(page)) {
 299                         /* Holder of page lock is waiting for us, it
 300                          * is safe to use this page. */
 301                         break;
 302                 }
 303                 /* Some other process has this page locked and has
 304                  * nothing to do with us.  Wait for it to finish.
 305                  */
 306                 schedule();
 307         }
 308         BUG_ON(!PageLocked(page));
 309 }
 310
 311 static struct page *logfs_get_write_page(struct inode *inode, u64 bix,
 312                 level_t level)
 313 {
 314         struct address_space *mapping = inode->i_mapping;
 315         pgoff_t index = logfs_pack_index(bix, level);
 316         struct page *page;
 317         int err;
 318
 319 repeat:
 320         page = find_get_page(mapping, index);
 321         if (!page) {
 322                 page = __page_cache_alloc(GFP_NOFS);
 323                 if (!page)
 324                         return NULL;
 325                 err = add_to_page_cache_lru(page, mapping, index, GFP_NOFS);
 326                 if (unlikely(err)) {
 327                         page_cache_release(page);
 328                         if (err == -EEXIST)
 329                                 goto repeat;
 330                         return NULL;
 331                 }
 332         } else logfs_lock_write_page(page);
 333         BUG_ON(!PageLocked(page));
 334         return page;
 335 }
 336
 337 static void logfs_unlock_write_page(struct page *page)
 338 {
 339         if (!PagePreLocked(page))
 340                 unlock_page(page);
 341 }
 342
 343 static void logfs_put_write_page(struct page *page)
 344 {
 345         logfs_unlock_write_page(page);
 346         page_cache_release(page);
 347 }
 348
 349 static struct page *logfs_get_page(struct inode *inode, u64 bix, level_t level,
 350                 int rw)
 351 {
 352         if (rw == READ)
 353                 return logfs_get_read_page(inode, bix, level);
 354         else
 355                 return logfs_get_write_page(inode, bix, level);
 356 }
 357
 358 static void logfs_put_page(struct page *page, int rw)
 359 {
 360         if (rw == READ)
 361                 logfs_put_read_page(page);
 362         else
 363                 logfs_put_write_page(page);
 364 }
 365
 366 static unsigned long __get_bits(u64 val, int skip, int no)
 367 {
 368         u64 ret = val;
 369
 370         ret >>= skip * no;
 371         ret <<= 64 - no;
 372         ret >>= 64 - no;
 373         return ret;
 374 }
 375
 376 static unsigned long get_bits(u64 val, level_t skip)
 377 {
 378         return __get_bits(val, (__force int)skip, LOGFS_BLOCK_BITS);
 379 }
 380
 381 static inline void init_shadow_tree(struct super_block *sb,
 382                 struct shadow_tree *tree)
 383 {
 384         struct logfs_super *super = logfs_super(sb);
 385
 386         btree_init_mempool64(&tree->new, super->s_btree_pool);
 387         btree_init_mempool64(&tree->old, super->s_btree_pool);
 388 }
 389
 390 static void indirect_write_block(struct logfs_block *block)
 391 {
 392         struct page *page;
 393         struct inode *inode;
 394         int ret;
 395
 396         page = block->page;
 397         inode = page->mapping->host;
 398         logfs_lock_write_page(page);
 399         ret = logfs_write_buf(inode, page, 0);
 400         logfs_unlock_write_page(page);
 401         /*
 402          * This needs some rework.  Unless you want your filesystem to run
 403          * completely synchronously (you don't), the filesystem will always
 404          * report writes as 'successful' before the actual work has been
 405          * done.  The actual work gets done here and this is where any errors
 406          * will show up.  And there isn't much we can do about it, really.
 407          *
 408          * Some attempts to fix the errors (move from bad blocks, retry io,...)
 409          * have already been done, so anything left should be either a broken
 410          * device or a bug somewhere in logfs itself.  Being relatively new,
 411          * the odds currently favor a bug, so for now the line below isn't
 412          * entirely tasteles.
 413          */
 414         BUG_ON(ret);
 415 }
 416
 417 static void inode_write_block(struct logfs_block *block)
 418 {
 419         struct inode *inode;
 420         int ret;
 421
 422         inode = block->inode;
 423         if (inode->i_ino == LOGFS_INO_MASTER)
 424                 logfs_write_anchor(inode->i_sb);
 425         else {
 426                 ret = __logfs_write_inode(inode, 0);
 427                 /* see indirect_write_block comment */
 428                 BUG_ON(ret);
 429         }
 430 }
 431
 432 /*
 433  * This silences a false, yet annoying gcc warning.  I hate it when my editor
 434  * jumps into bitops.h each time I recompile this file.
 435  * TODO: Complain to gcc folks about this and upgrade compiler.
 436  */
 437 static unsigned long fnb(const unsigned long *addr,
 438                 unsigned long size, unsigned long offset)
 439 {
 440         return find_next_bit(addr, size, offset);
 441 }
 442
 443 static __be64 inode_val0(struct inode *inode)
 444 {
 445         struct logfs_inode *li = logfs_inode(inode);
 446         u64 val;
 447
 448         /*
 449          * Explicit shifting generates good code, but must match the format
 450          * of the structure.  Add some paranoia just in case.
 451          */
 452         BUILD_BUG_ON(offsetof(struct logfs_disk_inode, di_mode) != 0);
 453         BUILD_BUG_ON(offsetof(struct logfs_disk_inode, di_height) != 2);
 454         BUILD_BUG_ON(offsetof(struct logfs_disk_inode, di_flags) != 4);
 455
 456         val =   (u64)inode->i_mode << 48 |
 457                 (u64)li->li_height << 40 |
 458                 (u64)li->li_flags;
 459         return cpu_to_be64(val);
 460 }
 461
 462 static int inode_write_alias(struct super_block *sb,
 463                 struct logfs_block *block, write_alias_t *write_one_alias)
 464 {
 465         struct inode *inode = block->inode;
 466         struct logfs_inode *li = logfs_inode(inode);
 467         unsigned long pos;
 468         u64 ino , bix;
 469         __be64 val;
 470         level_t level;
 471         int err;
 472
 473         for (pos = 0; ; pos++) {
 474                 pos = fnb(block->alias_map, LOGFS_BLOCK_FACTOR, pos);
 475                 if (pos >= LOGFS_EMBEDDED_FIELDS + INODE_POINTER_OFS)
 476                         return 0;
 477
 478                 switch (pos) {
 479                 case INODE_HEIGHT_OFS:
 480                         val = inode_val0(inode);
 481                         break;
 482                 case INODE_USED_OFS:
 483                         val = cpu_to_be64(li->li_used_bytes);;
 484                         break;
 485                 case INODE_SIZE_OFS:
 486                         val = cpu_to_be64(i_size_read(inode));
 487                         break;
 488                 case INODE_POINTER_OFS ... INODE_POINTER_OFS + LOGFS_EMBEDDED_FIELDS - 1:
 489                         val = cpu_to_be64(li->li_data[pos - INODE_POINTER_OFS]);
 490                         break;
 491                 default:
 492                         BUG();
 493                 }
 494
 495                 ino = LOGFS_INO_MASTER;
 496                 bix = inode->i_ino;
 497                 level = LEVEL(0);
 498                 err = write_one_alias(sb, ino, bix, level, pos, val);
 499                 if (err)
 500                         return err;
 501         }
 502 }
 503
 504 static int indirect_write_alias(struct super_block *sb,
 505                 struct logfs_block *block, write_alias_t *write_one_alias)
 506 {
 507         unsigned long pos;
 508         struct page *page = block->page;
 509         u64 ino , bix;
 510         __be64 *child, val;
 511         level_t level;
 512         int err;
 513
 514         for (pos = 0; ; pos++) {
 515                 pos = fnb(block->alias_map, LOGFS_BLOCK_FACTOR, pos);
 516                 if (pos >= LOGFS_BLOCK_FACTOR)
 517                         return 0;
 518
 519                 ino = page->mapping->host->i_ino;
 520                 logfs_unpack_index(page->index, &bix, &level);
 521                 child = kmap_atomic(page, KM_USER0);
 522                 val = child[pos];
 523                 kunmap_atomic(child, KM_USER0);
 524                 err = write_one_alias(sb, ino, bix, level, pos, val);
 525                 if (err)
 526                         return err;
 527         }
 528 }
 529
 530 int logfs_write_obj_aliases_pagecache(struct super_block *sb)
 531 {
 532         struct logfs_super *super = logfs_super(sb);
 533         struct logfs_block *block;
 534         int err;
 535
 536         list_for_each_entry(block, &super->s_object_alias, alias_list) {
 537                 err = block->ops->write_alias(sb, block, write_alias_journal);
 538                 if (err)
 539                         return err;
 540         }
 541         return 0;
 542 }
 543
 544 void __free_block(struct super_block *sb, struct logfs_block *block)
 545 {
 546         BUG_ON(!list_empty(&block->item_list));
 547         list_del(&block->alias_list);
 548         mempool_free(block, logfs_super(sb)->s_block_pool);
 549 }
 550
 551 static void inode_free_block(struct super_block *sb, struct logfs_block *block)
 552 {
 553         struct inode *inode = block->inode;
 554
 555         logfs_inode(inode)->li_block = NULL;
 556         __free_block(sb, block);
 557 }
 558
 559 static void indirect_free_block(struct super_block *sb,
 560                 struct logfs_block *block)
 561 {
 562         ClearPagePrivate(block->page);
 563         block->page->private = 0;
 564         __free_block(sb, block);
 565 }
 566
 567
 568 static struct logfs_block_ops inode_block_ops = {
 569         .write_block = inode_write_block,
 570         .free_block = inode_free_block,
 571         .write_alias = inode_write_alias,
 572 };
 573
 574 struct logfs_block_ops indirect_block_ops = {
 575         .write_block = indirect_write_block,
 576         .free_block = indirect_free_block,
 577         .write_alias = indirect_write_alias,
 578 };
 579
 580 struct logfs_block *__alloc_block(struct super_block *sb,
 581                 u64 ino, u64 bix, level_t level)
 582 {
 583         struct logfs_super *super = logfs_super(sb);
 584         struct logfs_block *block;
 585
 586         block = mempool_alloc(super->s_block_pool, GFP_NOFS);
 587         memset(block, 0, sizeof(*block));
 588         INIT_LIST_HEAD(&block->alias_list);
 589         INIT_LIST_HEAD(&block->item_list);
 590         block->sb = sb;
 591         block->ino = ino;
 592         block->bix = bix;
 593         block->level = level;
 594         return block;
 595 }
 596
 597 static void alloc_inode_block(struct inode *inode)
 598 {
 599         struct logfs_inode *li = logfs_inode(inode);
 600         struct logfs_block *block;
 601
 602         if (li->li_block)
 603                 return;
 604
 605         block = __alloc_block(inode->i_sb, LOGFS_INO_MASTER, inode->i_ino, 0);
 606         block->inode = inode;
 607         li->li_block = block;
 608         block->ops = &inode_block_ops;
 609 }
 610
 611 void initialize_block_counters(struct page *page, struct logfs_block *block,
 612                 __be64 *array, int page_is_empty)
 613 {
 614         u64 ptr;
 615         int i, start;
 616
 617         block->partial = 0;
 618         block->full = 0;
 619         start = 0;
 620         if (page->index < first_indirect_block()) {
 621                 /* Counters are pointless on level 0 */
 622                 return;
 623         }
 624         if (page->index == first_indirect_block()) {
 625                 /* Skip unused pointers */
 626                 start = I0_BLOCKS;
 627                 block->full = I0_BLOCKS;
 628         }
 629         if (!page_is_empty) {
 630                 for (i = start; i < LOGFS_BLOCK_FACTOR; i++) {
 631                         ptr = be64_to_cpu(array[i]);
 632                         if (ptr)
 633                                 block->partial++;
 634                         if (ptr & LOGFS_FULLY_POPULATED)
 635                                 block->full++;
 636                 }
 637         }
 638 }
 639
 640 static void alloc_data_block(struct inode *inode, struct page *page)
 641 {
 642         struct logfs_block *block;
 643         u64 bix;
 644         level_t level;
 645
 646         if (PagePrivate(page))
 647                 return;
 648
 649         logfs_unpack_index(page->index, &bix, &level);
 650         block = __alloc_block(inode->i_sb, inode->i_ino, bix, level);
 651         block->page = page;
 652         SetPagePrivate(page);
 653         page->private = (unsigned long)block;
 654         block->ops = &indirect_block_ops;
 655 }
 656
 657 static void alloc_indirect_block(struct inode *inode, struct page *page,
 658                 int page_is_empty)
 659 {
 660         struct logfs_block *block;
 661         __be64 *array;
 662
 663         if (PagePrivate(page))
 664                 return;
 665
 666         alloc_data_block(inode, page);
 667
 668         block = logfs_block(page);
 669         array = kmap_atomic(page, KM_USER0);
 670         initialize_block_counters(page, block, array, page_is_empty);
 671         kunmap_atomic(array, KM_USER0);
 672 }
 673
 674 static void block_set_pointer(struct page *page, int index, u64 ptr)
 675 {
 676         struct logfs_block *block = logfs_block(page);
 677         __be64 *array;
 678         u64 oldptr;
 679
 680         BUG_ON(!block);
 681         array = kmap_atomic(page, KM_USER0);
 682         oldptr = be64_to_cpu(array[index]);
 683         array[index] = cpu_to_be64(ptr);
 684         kunmap_atomic(array, KM_USER0);
 685         SetPageUptodate(page);
 686
 687         block->full += !!(ptr & LOGFS_FULLY_POPULATED)
 688                 - !!(oldptr & LOGFS_FULLY_POPULATED);
 689         block->partial += !!ptr - !!oldptr;
 690 }
 691
 692 static u64 block_get_pointer(struct page *page, int index)
 693 {
 694         __be64 *block;
 695         u64 ptr;
 696
 697         block = kmap_atomic(page, KM_USER0);
 698         ptr = be64_to_cpu(block[index]);
 699         kunmap_atomic(block, KM_USER0);
 700         return ptr;
 701 }
 702
 703 static int logfs_read_empty(struct page *page)
 704 {
 705         zero_user_segment(page, 0, PAGE_CACHE_SIZE);
 706         return 0;
 707 }
 708
 709 static int logfs_read_direct(struct inode *inode, struct page *page)
 710 {
 711         struct logfs_inode *li = logfs_inode(inode);
 712         pgoff_t index = page->index;
 713         u64 block;
 714
 715         block = li->li_data[index];
 716         if (!block)
 717                 return logfs_read_empty(page);
 718
 719         return logfs_segment_read(inode, page, block, index, 0);
 720 }
 721
 722 static int logfs_read_loop(struct inode *inode, struct page *page,
 723                 int rw_context)
 724 {
 725         struct logfs_inode *li = logfs_inode(inode);
 726         u64 bix, bofs = li->li_data[INDIRECT_INDEX];
 727         level_t level, target_level;
 728         int ret;
 729         struct page *ipage;
 730
 731         logfs_unpack_index(page->index, &bix, &target_level);
 732         if (!bofs)
 733                 return logfs_read_empty(page);
 734
 735         if (bix >= maxbix(li->li_height))
 736                 return logfs_read_empty(page);
 737
 738         for (level = LEVEL(li->li_height);
 739                         (__force u8)level > (__force u8)target_level;
 740                         level = SUBLEVEL(level)){
 741                 ipage = logfs_get_page(inode, bix, level, rw_context);
 742                 if (!ipage)
 743                         return -ENOMEM;
 744
 745                 ret = logfs_segment_read(inode, ipage, bofs, bix, level);
 746                 if (ret) {
 747                         logfs_put_read_page(ipage);
 748                         return ret;
 749                 }
 750
 751                 bofs = block_get_pointer(ipage, get_bits(bix, SUBLEVEL(level)));
 752                 logfs_put_page(ipage, rw_context);
 753                 if (!bofs)
 754                         return logfs_read_empty(page);
 755         }
 756
 757         return logfs_segment_read(inode, page, bofs, bix, 0);
 758 }
 759
 760 static int logfs_read_block(struct inode *inode, struct page *page,
 761                 int rw_context)
 762 {
 763         pgoff_t index = page->index;
 764
 765         if (index < I0_BLOCKS)
 766                 return logfs_read_direct(inode, page);
 767         return logfs_read_loop(inode, page, rw_context);
 768 }
 769
 770 static int logfs_exist_loop(struct inode *inode, u64 bix)
 771 {
 772         struct logfs_inode *li = logfs_inode(inode);
 773         u64 bofs = li->li_data[INDIRECT_INDEX];
 774         level_t level;
 775         int ret;
 776         struct page *ipage;
 777
 778         if (!bofs)
 779                 return 0;
 780         if (bix >= maxbix(li->li_height))
 781                 return 0;
 782
 783         for (level = LEVEL(li->li_height); level != 0; level = SUBLEVEL(level)) {
 784                 ipage = logfs_get_read_page(inode, bix, level);
 785                 if (!ipage)
 786                         return -ENOMEM;
 787
 788                 ret = logfs_segment_read(inode, ipage, bofs, bix, level);
 789                 if (ret) {
 790                         logfs_put_read_page(ipage);
 791                         return ret;
 792                 }
 793
 794                 bofs = block_get_pointer(ipage, get_bits(bix, SUBLEVEL(level)));
 795                 logfs_put_read_page(ipage);
 796                 if (!bofs)
 797                         return 0;
 798         }
 799
 800         return 1;
 801 }
 802
 803 int logfs_exist_block(struct inode *inode, u64 bix)
 804 {
 805         struct logfs_inode *li = logfs_inode(inode);
 806
 807         if (bix < I0_BLOCKS)
 808                 return !!li->li_data[bix];
 809         return logfs_exist_loop(inode, bix);
 810 }
 811
 812 static u64 seek_holedata_direct(struct inode *inode, u64 bix, int data)
 813 {
 814         struct logfs_inode *li = logfs_inode(inode);
 815
 816         for (; bix < I0_BLOCKS; bix++)
 817                 if (data ^ (li->li_data[bix] == 0))
 818                         return bix;
 819         return I0_BLOCKS;
 820 }
 821
 822 static u64 seek_holedata_loop(struct inode *inode, u64 bix, int data)
 823 {
 824         struct logfs_inode *li = logfs_inode(inode);
 825         __be64 *rblock;
 826         u64 increment, bofs = li->li_data[INDIRECT_INDEX];
 827         level_t level;
 828         int ret, slot;
 829         struct page *page;
 830
 831         BUG_ON(!bofs);
 832
 833         for (level = LEVEL(li->li_height); level != 0; level = SUBLEVEL(level)) {
 834                 increment = 1 << (LOGFS_BLOCK_BITS * ((__force u8)level-1));
 835                 page = logfs_get_read_page(inode, bix, level);
 836                 if (!page)
 837                         return bix;
 838
 839                 ret = logfs_segment_read(inode, page, bofs, bix, level);
 840                 if (ret) {
 841                         logfs_put_read_page(page);
 842                         return bix;
 843                 }
 844
 845                 slot = get_bits(bix, SUBLEVEL(level));
 846                 rblock = kmap_atomic(page, KM_USER0);
 847                 while (slot < LOGFS_BLOCK_FACTOR) {
 848                         if (data && (rblock[slot] != 0))
 849                                 break;
 850                         if (!data && !(be64_to_cpu(rblock[slot]) & LOGFS_FULLY_POPULATED))
 851                                 break;
 852                         slot++;
 853                         bix += increment;
 854                         bix &= ~(increment - 1);
 855                 }
 856                 if (slot >= LOGFS_BLOCK_FACTOR) {
 857                         kunmap_atomic(rblock, KM_USER0);
 858                         logfs_put_read_page(page);
 859                         return bix;
 860                 }
 861                 bofs = be64_to_cpu(rblock[slot]);
 862                 kunmap_atomic(rblock, KM_USER0);
 863                 logfs_put_read_page(page);
 864                 if (!bofs) {
 865                         BUG_ON(data);
 866                         return bix;
 867                 }
 868         }
 869         return bix;
 870 }
 871
 872 /**
 873  * logfs_seek_hole - find next hole starting at a given block index
 874  * @inode:              inode to search in
 875  * @bix:                block index to start searching
 876  *
 877  * Returns next hole.  If the file doesn't contain any further holes, the
 878  * block address next to eof is returned instead.
 879  */
 880 u64 logfs_seek_hole(struct inode *inode, u64 bix)
 881 {
 882         struct logfs_inode *li = logfs_inode(inode);
 883
 884         if (bix < I0_BLOCKS) {
 885                 bix = seek_holedata_direct(inode, bix, 0);
 886                 if (bix < I0_BLOCKS)
 887                         return bix;
 888         }
 889
 890         if (!li->li_data[INDIRECT_INDEX])
 891                 return bix;
 892         else if (li->li_data[INDIRECT_INDEX] & LOGFS_FULLY_POPULATED)
 893                 bix = maxbix(li->li_height);
 894         else {
 895                 bix = seek_holedata_loop(inode, bix, 0);
 896                 if (bix < maxbix(li->li_height))
 897                         return bix;
 898                 /* Should not happen anymore.  But if some port writes semi-
 899                  * corrupt images (as this one used to) we might run into it.
 900                  */
 901                 WARN_ON_ONCE(bix == maxbix(li->li_height));
 902         }
 903
 904         return bix;
 905 }
 906
 907 static u64 __logfs_seek_data(struct inode *inode, u64 bix)
 908 {
 909         struct logfs_inode *li = logfs_inode(inode);
 910
 911         if (bix < I0_BLOCKS) {
 912                 bix = seek_holedata_direct(inode, bix, 1);
 913                 if (bix < I0_BLOCKS)
 914                         return bix;
 915         }
 916
 917         if (bix < maxbix(li->li_height)) {
 918                 if (!li->li_data[INDIRECT_INDEX])
 919                         bix = maxbix(li->li_height);
 920                 else
 921                         return seek_holedata_loop(inode, bix, 1);
 922         }
 923
 924         return bix;
 925 }
 926
 927 /**
 928  * logfs_seek_data - find next data block after a given block index
 929  * @inode:              inode to search in
 930  * @bix:                block index to start searching
 931  *
 932  * Returns next data block.  If the file doesn't contain any further data
 933  * blocks, the last block in the file is returned instead.
 934  */
 935 u64 logfs_seek_data(struct inode *inode, u64 bix)
 936 {
 937         struct super_block *sb = inode->i_sb;
 938         u64 ret, end;
 939
 940         ret = __logfs_seek_data(inode, bix);
 941         end = i_size_read(inode) >> sb->s_blocksize_bits;
 942         if (ret >= end)
 943                 ret = max(bix, end);
 944         return ret;
 945 }
 946
 947 static int logfs_is_valid_direct(struct logfs_inode *li, u64 bix, u64 ofs)
 948 {
 949         return pure_ofs(li->li_data[bix]) == ofs;
 950 }
 951
 952 static int __logfs_is_valid_loop(struct inode *inode, u64 bix,
 953                 u64 ofs, u64 bofs)
 954 {
 955         struct logfs_inode *li = logfs_inode(inode);
 956         level_t level;
 957         int ret;
 958         struct page *page;
 959
 960         for (level = LEVEL(li->li_height); level != 0; level = SUBLEVEL(level)){
 961                 page = logfs_get_write_page(inode, bix, level);
 962                 BUG_ON(!page);
 963
 964                 ret = logfs_segment_read(inode, page, bofs, bix, level);
 965                 if (ret) {
 966                         logfs_put_write_page(page);
 967                         return 0;
 968                 }
 969
 970                 bofs = block_get_pointer(page, get_bits(bix, SUBLEVEL(level)));
 971                 logfs_put_write_page(page);
 972                 if (!bofs)
 973                         return 0;
 974
 975                 if (pure_ofs(bofs) == ofs)
 976                         return 1;
 977         }
 978         return 0;
 979 }
 980
 981 static int logfs_is_valid_loop(struct inode *inode, u64 bix, u64 ofs)
 982 {
 983         struct logfs_inode *li = logfs_inode(inode);
 984         u64 bofs = li->li_data[INDIRECT_INDEX];
 985
 986         if (!bofs)
 987                 return 0;
 988
 989         if (bix >= maxbix(li->li_height))
 990                 return 0;
 991
 992         if (pure_ofs(bofs) == ofs)
 993                 return 1;
 994
 995         return __logfs_is_valid_loop(inode, bix, ofs, bofs);
 996 }
 997
 998 static int __logfs_is_valid_block(struct inode *inode, u64 bix, u64 ofs)
 999 {
1000         struct logfs_inode *li = logfs_inode(inode);
1001
1002         if ((inode->i_nlink == 0) && atomic_read(&inode->i_count) == 1)
1003                 return 0;
1004
1005         if (bix < I0_BLOCKS)
1006                 return logfs_is_valid_direct(li, bix, ofs);
1007         return logfs_is_valid_loop(inode, bix, ofs);
1008 }
1009
1010 /**
1011  * logfs_is_valid_block - check whether this block is still valid
1012  *
1013  * @sb  - superblock
1014  * @ofs - block physical offset
1015  * @ino - block inode number
1016  * @bix - block index
1017  * @level - block level
1018  *
1019  * Returns 0 if the block is invalid, 1 if it is valid and 2 if it will
1020  * become invalid once the journal is written.
1021  */
1022 int logfs_is_valid_block(struct super_block *sb, u64 ofs, u64 ino, u64 bix,
1023                 gc_level_t gc_level)
1024 {
1025         struct logfs_super *super = logfs_super(sb);
1026         struct inode *inode;
1027         int ret, cookie;
1028
1029         /* Umount closes a segment with free blocks remaining.  Those
1030          * blocks are by definition invalid. */
1031         if (ino == -1)
1032                 return 0;
1033
1034         LOGFS_BUG_ON((u64)(u_long)ino != ino, sb);
1035
1036         inode = logfs_safe_iget(sb, ino, &cookie);
1037         if (IS_ERR(inode))
1038                 goto invalid;
1039
1040         ret = __logfs_is_valid_block(inode, bix, ofs);
1041         logfs_safe_iput(inode, cookie);
1042         if (ret)
1043                 return ret;
1044
1045 invalid:
1046         /* Block is nominally invalid, but may still sit in the shadow tree,
1047          * waiting for a journal commit.
1048          */
1049         if (btree_lookup64(&super->s_shadow_tree.old, ofs))
1050                 return 2;
1051         return 0;
1052 }
1053
1054 int logfs_readpage_nolock(struct page *page)
1055 {
1056         struct inode *inode = page->mapping->host;
1057         int ret = -EIO;
1058
1059         ret = logfs_read_block(inode, page, READ);
1060
1061         if (ret) {
1062                 ClearPageUptodate(page);
1063                 SetPageError(page);
1064         } else {
1065                 SetPageUptodate(page);
1066                 ClearPageError(page);
1067         }
1068         flush_dcache_page(page);
1069
1070         return ret;
1071 }
1072
1073 static int logfs_reserve_bytes(struct inode *inode, int bytes)
1074 {
1075         struct logfs_super *super = logfs_super(inode->i_sb);
1076         u64 available = super->s_free_bytes + super->s_dirty_free_bytes
1077                         - super->s_dirty_used_bytes - super->s_dirty_pages;
1078
1079         if (!bytes)
1080                 return 0;
1081
1082         if (available < bytes)
1083                 return -ENOSPC;
1084
1085         if (available < bytes + super->s_root_reserve &&
1086                         !capable(CAP_SYS_RESOURCE))
1087                 return -ENOSPC;
1088
1089         return 0;
1090 }
1091
1092 int get_page_reserve(struct inode *inode, struct page *page)
1093 {
1094         struct logfs_super *super = logfs_super(inode->i_sb);
1095         int ret;
1096
1097         if (logfs_block(page) && logfs_block(page)->reserved_bytes)
1098                 return 0;
1099
1100         logfs_get_wblocks(inode->i_sb, page, WF_LOCK);
1101         ret = logfs_reserve_bytes(inode, 6 * LOGFS_MAX_OBJECTSIZE);
1102         if (!ret) {
1103                 alloc_data_block(inode, page);
1104                 logfs_block(page)->reserved_bytes += 6 * LOGFS_MAX_OBJECTSIZE;
1105                 super->s_dirty_pages += 6 * LOGFS_MAX_OBJECTSIZE;
1106         }
1107         logfs_put_wblocks(inode->i_sb, page, WF_LOCK);
1108         return ret;
1109 }
1110
1111 /*
1112  * We are protected by write lock.  Push victims up to superblock level
1113  * and release transaction when appropriate.
1114  */
1115 /* FIXME: This is currently called from the wrong spots. */
1116 static void logfs_handle_transaction(struct inode *inode,
1117                 struct logfs_transaction *ta)
1118 {
1119         struct logfs_super *super = logfs_super(inode->i_sb);
1120
1121         if (!ta)
1122                 return;
1123         logfs_inode(inode)->li_block->ta = NULL;
1124
1125         if (inode->i_ino != LOGFS_INO_MASTER) {
1126                 BUG(); /* FIXME: Yes, this needs more thought */
1127                 /* just remember the transaction until inode is written */
1128                 //BUG_ON(logfs_inode(inode)->li_transaction);
1129                 //logfs_inode(inode)->li_transaction = ta;
1130                 return;
1131         }
1132
1133         switch (ta->state) {
1134         case CREATE_1: /* fall through */
1135         case UNLINK_1:
1136                 BUG_ON(super->s_victim_ino);
1137                 super->s_victim_ino = ta->ino;
1138                 break;
1139         case CREATE_2: /* fall through */
1140         case UNLINK_2:
1141                 BUG_ON(super->s_victim_ino != ta->ino);
1142                 super->s_victim_ino = 0;
1143                 /* transaction ends here - free it */
1144                 kfree(ta);
1145                 break;
1146         case CROSS_RENAME_1:
1147                 BUG_ON(super->s_rename_dir);
1148                 BUG_ON(super->s_rename_pos);
1149                 super->s_rename_dir = ta->dir;
1150                 super->s_rename_pos = ta->pos;
1151                 break;
1152         case CROSS_RENAME_2:
1153                 BUG_ON(super->s_rename_dir != ta->dir);
1154                 BUG_ON(super->s_rename_pos != ta->pos);
1155                 super->s_rename_dir = 0;
1156                 super->s_rename_pos = 0;
1157                 kfree(ta);
1158                 break;
1159         case TARGET_RENAME_1:
1160                 BUG_ON(super->s_rename_dir);
1161                 BUG_ON(super->s_rename_pos);
1162                 BUG_ON(super->s_victim_ino);
1163                 super->s_rename_dir = ta->dir;
1164                 super->s_rename_pos = ta->pos;
1165                 super->s_victim_ino = ta->ino;
1166                 break;
1167         case TARGET_RENAME_2:
1168                 BUG_ON(super->s_rename_dir != ta->dir);
1169                 BUG_ON(super->s_rename_pos != ta->pos);
1170                 BUG_ON(super->s_victim_ino != ta->ino);
1171                 super->s_rename_dir = 0;
1172                 super->s_rename_pos = 0;
1173                 break;
1174         case TARGET_RENAME_3:
1175                 BUG_ON(super->s_rename_dir);
1176                 BUG_ON(super->s_rename_pos);
1177                 BUG_ON(super->s_victim_ino != ta->ino);
1178                 super->s_victim_ino = 0;
1179                 kfree(ta);
1180                 break;
1181         default:
1182                 BUG();
1183         }
1184 }
1185
1186 /*
1187  * Not strictly a reservation, but rather a check that we still have enough
1188  * space to satisfy the write.
1189  */
1190 static int logfs_reserve_blocks(struct inode *inode, int blocks)
1191 {
1192         return logfs_reserve_bytes(inode, blocks * LOGFS_MAX_OBJECTSIZE);
1193 }
1194
1195 struct write_control {
1196         u64 ofs;
1197         long flags;
1198 };
1199
1200 static struct logfs_shadow *alloc_shadow(struct inode *inode, u64 bix,
1201                 level_t level, u64 old_ofs)
1202 {
1203         struct logfs_super *super = logfs_super(inode->i_sb);
1204         struct logfs_shadow *shadow;
1205
1206         shadow = mempool_alloc(super->s_shadow_pool, GFP_NOFS);
1207         memset(shadow, 0, sizeof(*shadow));
1208         shadow->ino = inode->i_ino;
1209         shadow->bix = bix;
1210         shadow->gc_level = expand_level(inode->i_ino, level);
1211         shadow->old_ofs = old_ofs & ~LOGFS_FULLY_POPULATED;
1212         return shadow;
1213 }
1214
1215 static void free_shadow(struct inode *inode, struct logfs_shadow *shadow)
1216 {
1217         struct logfs_super *super = logfs_super(inode->i_sb);
1218
1219         mempool_free(shadow, super->s_shadow_pool);
1220 }
1221
1222 static void mark_segment(struct shadow_tree *tree, u32 segno)
1223 {
1224         int err;
1225
1226         if (!btree_lookup32(&tree->segment_map, segno)) {
1227                 err = btree_insert32(&tree->segment_map, segno, (void *)1,
1228                                 GFP_NOFS);
1229                 BUG_ON(err);
1230                 tree->no_shadowed_segments++;
1231         }
1232 }
1233
1234 /**
1235  * fill_shadow_tree - Propagate shadow tree changes due to a write
1236  * @inode:      Inode owning the page
1237  * @page:       Struct page that was written
1238  * @shadow:     Shadow for the current write
1239  *
1240  * Writes in logfs can result in two semi-valid objects.  The old object
1241  * is still valid as long as it can be reached by following pointers on
1242  * the medium.  Only when writes propagate all the way up to the journal
1243  * has the new object safely replaced the old one.
1244  *
1245  * To handle this problem, a struct logfs_shadow is used to represent
1246  * every single write.  It is attached to the indirect block, which is
1247  * marked dirty.  When the indirect block is written, its shadows are
1248  * handed up to the next indirect block (or inode).  Untimately they
1249  * will reach the master inode and be freed upon journal commit.
1250  *
1251  * This function handles a single step in the propagation.  It adds the
1252  * shadow for the current write to the tree, along with any shadows in
1253  * the page's tree, in case it was an indirect block.  If a page is
1254  * written, the inode parameter is left NULL, if an inode is written,
1255  * the page parameter is left NULL.
1256  */
1257 static void fill_shadow_tree(struct inode *inode, struct page *page,
1258                 struct logfs_shadow *shadow)
1259 {
1260         struct logfs_super *super = logfs_super(inode->i_sb);
1261         struct logfs_block *block = logfs_block(page);
1262         struct shadow_tree *tree = &super->s_shadow_tree;
1263
1264         if (PagePrivate(page)) {
1265                 if (block->alias_map)
1266                         super->s_no_object_aliases -= bitmap_weight(
1267                                         block->alias_map, LOGFS_BLOCK_FACTOR);
1268                 logfs_handle_transaction(inode, block->ta);
1269                 block->ops->free_block(inode->i_sb, block);
1270         }
1271         if (shadow) {
1272                 if (shadow->old_ofs)
1273                         btree_insert64(&tree->old, shadow->old_ofs, shadow,
1274                                         GFP_NOFS);
1275                 else
1276                         btree_insert64(&tree->new, shadow->new_ofs, shadow,
1277                                         GFP_NOFS);
1278
1279                 super->s_dirty_used_bytes += shadow->new_len;
1280                 super->s_dirty_free_bytes += shadow->old_len;
1281                 mark_segment(tree, shadow->old_ofs >> super->s_segshift);
1282                 mark_segment(tree, shadow->new_ofs >> super->s_segshift);
1283         }
1284 }
1285
1286 static void logfs_set_alias(struct super_block *sb, struct logfs_block *block,
1287                 long child_no)
1288 {
1289         struct logfs_super *super = logfs_super(sb);
1290
1291         if (block->inode && block->inode->i_ino == LOGFS_INO_MASTER) {
1292                 /* Aliases in the master inode are pointless. */
1293                 return;
1294         }
1295
1296         if (!test_bit(child_no, block->alias_map)) {
1297                 set_bit(child_no, block->alias_map);
1298                 super->s_no_object_aliases++;
1299         }
1300         list_move_tail(&block->alias_list, &super->s_object_alias);
1301 }
1302
1303 /*
1304  * Object aliases can and often do change the size and occupied space of a
1305  * file.  So not only do we have to change the pointers, we also have to
1306  * change inode->i_size and li->li_used_bytes.  Which is done by setting
1307  * another two object aliases for the inode itself.
1308  */
1309 static void set_iused(struct inode *inode, struct logfs_shadow *shadow)
1310 {
1311         struct logfs_inode *li = logfs_inode(inode);
1312
1313         if (shadow->new_len == shadow->old_len)
1314                 return;
1315
1316         alloc_inode_block(inode);
1317         li->li_used_bytes += shadow->new_len - shadow->old_len;
1318         __logfs_set_blocks(inode);
1319         logfs_set_alias(inode->i_sb, li->li_block, INODE_USED_OFS);
1320         logfs_set_alias(inode->i_sb, li->li_block, INODE_SIZE_OFS);
1321 }
1322
1323 static int logfs_write_i0(struct inode *inode, struct page *page,
1324                 struct write_control *wc)
1325 {
1326         struct logfs_shadow *shadow;
1327         u64 bix;
1328         level_t level;
1329         int full, err = 0;
1330
1331         logfs_unpack_index(page->index, &bix, &level);
1332         if (wc->ofs == 0)
1333                 if (logfs_reserve_blocks(inode, 1))
1334                         return -ENOSPC;
1335
1336         shadow = alloc_shadow(inode, bix, level, wc->ofs);
1337         if (wc->flags & WF_WRITE)
1338                 err = logfs_segment_write(inode, page, shadow);
1339         if (wc->flags & WF_DELETE)
1340                 logfs_segment_delete(inode, shadow);
1341         if (err) {
1342                 free_shadow(inode, shadow);
1343                 return err;
1344         }
1345
1346         set_iused(inode, shadow);
1347         full = 1;
1348         if (level != 0) {
1349                 alloc_indirect_block(inode, page, 0);
1350                 full = logfs_block(page)->full == LOGFS_BLOCK_FACTOR;
1351         }
1352         fill_shadow_tree(inode, page, shadow);
1353         wc->ofs = shadow->new_ofs;
1354         if (wc->ofs && full)
1355                 wc->ofs |= LOGFS_FULLY_POPULATED;
1356         return 0;
1357 }
1358
1359 static int logfs_write_direct(struct inode *inode, struct page *page,
1360                 long flags)
1361 {
1362         struct logfs_inode *li = logfs_inode(inode);
1363         struct write_control wc = {
1364                 .ofs = li->li_data[page->index],
1365                 .flags = flags,
1366         };
1367         int err;
1368
1369         alloc_inode_block(inode);
1370
1371         err = logfs_write_i0(inode, page, &wc);
1372         if (err)
1373                 return err;
1374
1375         li->li_data[page->index] = wc.ofs;
1376         logfs_set_alias(inode->i_sb, li->li_block,
1377                         page->index + INODE_POINTER_OFS);
1378         return 0;
1379 }
1380
1381 static int ptr_change(u64 ofs, struct page *page)
1382 {
1383         struct logfs_block *block = logfs_block(page);
1384         int empty0, empty1, full0, full1;
1385
1386         empty0 = ofs == 0;
1387         empty1 = block->partial == 0;
1388         if (empty0 != empty1)
1389                 return 1;
1390
1391         /* The !! is necessary to shrink result to int */
1392         full0 = !!(ofs & LOGFS_FULLY_POPULATED);
1393         full1 = block->full == LOGFS_BLOCK_FACTOR;
1394         if (full0 != full1)
1395                 return 1;
1396         return 0;
1397 }
1398
1399 static int __logfs_write_rec(struct inode *inode, struct page *page,
1400                 struct write_control *this_wc,
1401                 pgoff_t bix, level_t target_level, level_t level)
1402 {
1403         int ret, page_empty = 0;
1404         int child_no = get_bits(bix, SUBLEVEL(level));
1405         struct page *ipage;
1406         struct write_control child_wc = {
1407                 .flags = this_wc->flags,
1408         };
1409
1410         ipage = logfs_get_write_page(inode, bix, level);
1411         if (!ipage)
1412                 return -ENOMEM;
1413
1414         if (this_wc->ofs) {
1415                 ret = logfs_segment_read(inode, ipage, this_wc->ofs, bix, level);
1416                 if (ret)
1417                         goto out;
1418         } else if (!PageUptodate(ipage)) {
1419                 page_empty = 1;
1420                 logfs_read_empty(ipage);
1421         }
1422
1423         child_wc.ofs = block_get_pointer(ipage, child_no);
1424
1425         if ((__force u8)level-1 > (__force u8)target_level)
1426                 ret = __logfs_write_rec(inode, page, &child_wc, bix,
1427                                 target_level, SUBLEVEL(level));
1428         else
1429                 ret = logfs_write_i0(inode, page, &child_wc);
1430
1431         if (ret)
1432                 goto out;
1433
1434         alloc_indirect_block(inode, ipage, page_empty);
1435         block_set_pointer(ipage, child_no, child_wc.ofs);
1436         /* FIXME: first condition seems superfluous */
1437         if (child_wc.ofs || logfs_block(ipage)->partial)
1438                 this_wc->flags |= WF_WRITE;
1439         /* the condition on this_wc->ofs ensures that we won't consume extra
1440          * space for indirect blocks in the future, which we cannot reserve */
1441         if (!this_wc->ofs || ptr_change(this_wc->ofs, ipage))
1442                 ret = logfs_write_i0(inode, ipage, this_wc);
1443         else
1444                 logfs_set_alias(inode->i_sb, logfs_block(ipage), child_no);
1445 out:
1446         logfs_put_write_page(ipage);
1447         return ret;
1448 }
1449
1450 static int logfs_write_rec(struct inode *inode, struct page *page,
1451                 pgoff_t bix, level_t target_level, long flags)
1452 {
1453         struct logfs_inode *li = logfs_inode(inode);
1454         struct write_control wc = {
1455                 .ofs = li->li_data[INDIRECT_INDEX],
1456                 .flags = flags,
1457         };
1458         int ret;
1459
1460         alloc_inode_block(inode);
1461
1462         if (li->li_height > (__force u8)target_level)
1463                 ret = __logfs_write_rec(inode, page, &wc, bix, target_level,
1464                                 LEVEL(li->li_height));
1465         else
1466                 ret = logfs_write_i0(inode, page, &wc);
1467         if (ret)
1468                 return ret;
1469
1470         if (li->li_data[INDIRECT_INDEX] != wc.ofs) {
1471                 li->li_data[INDIRECT_INDEX] = wc.ofs;
1472                 logfs_set_alias(inode->i_sb, li->li_block,
1473                                 INDIRECT_INDEX + INODE_POINTER_OFS);
1474         }
1475         return ret;
1476 }
1477
1478 void logfs_add_transaction(struct inode *inode, struct logfs_transaction *ta)
1479 {
1480         alloc_inode_block(inode);
1481         logfs_inode(inode)->li_block->ta = ta;
1482 }
1483
1484 void logfs_del_transaction(struct inode *inode, struct logfs_transaction *ta)
1485 {
1486         struct logfs_block *block = logfs_inode(inode)->li_block;
1487
1488         if (block && block->ta)
1489                 block->ta = NULL;
1490 }
1491
1492 static int grow_inode(struct inode *inode, u64 bix, level_t level)
1493 {
1494         struct logfs_inode *li = logfs_inode(inode);
1495         u8 height = (__force u8)level;
1496         struct page *page;
1497         struct write_control wc = {
1498                 .flags = WF_WRITE,
1499         };
1500         int err;
1501
1502         BUG_ON(height > 5 || li->li_height > 5);
1503         while (height > li->li_height || bix >= maxbix(li->li_height)) {
1504                 page = logfs_get_write_page(inode, I0_BLOCKS + 1,
1505                                 LEVEL(li->li_height + 1));
1506                 if (!page)
1507                         return -ENOMEM;
1508                 logfs_read_empty(page);
1509                 alloc_indirect_block(inode, page, 1);
1510                 block_set_pointer(page, 0, li->li_data[INDIRECT_INDEX]);
1511                 err = logfs_write_i0(inode, page, &wc);
1512                 logfs_put_write_page(page);
1513                 if (err)
1514                         return err;
1515                 li->li_data[INDIRECT_INDEX] = wc.ofs;
1516                 wc.ofs = 0;
1517                 li->li_height++;
1518                 logfs_set_alias(inode->i_sb, li->li_block, INODE_HEIGHT_OFS);
1519         }
1520         return 0;
1521 }
1522
1523 static int __logfs_write_buf(struct inode *inode, struct page *page, long flags)
1524 {
1525         struct logfs_super *super = logfs_super(inode->i_sb);
1526         pgoff_t index = page->index;
1527         u64 bix;
1528         level_t level;
1529         int err;
1530
1531         flags |= WF_WRITE | WF_DELETE;
1532         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1533
1534         logfs_unpack_index(index, &bix, &level);
1535         if (logfs_block(page) && logfs_block(page)->reserved_bytes)
1536                 super->s_dirty_pages -= logfs_block(page)->reserved_bytes;
1537
1538         if (index < I0_BLOCKS)
1539                 return logfs_write_direct(inode, page, flags);
1540
1541         bix = adjust_bix(bix, level);
1542         err = grow_inode(inode, bix, level);
1543         if (err)
1544                 return err;
1545         return logfs_write_rec(inode, page, bix, level, flags);
1546 }
1547
1548 int logfs_write_buf(struct inode *inode, struct page *page, long flags)
1549 {
1550         struct super_block *sb = inode->i_sb;
1551         int ret;
1552
1553         logfs_get_wblocks(sb, page, flags & WF_LOCK);
1554         ret = __logfs_write_buf(inode, page, flags);
1555         logfs_put_wblocks(sb, page, flags & WF_LOCK);
1556         return ret;
1557 }
1558
1559 static int __logfs_delete(struct inode *inode, struct page *page)
1560 {
1561         long flags = WF_DELETE;
1562
1563         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1564
1565         if (page->index < I0_BLOCKS)
1566                 return logfs_write_direct(inode, page, flags);
1567         return logfs_write_rec(inode, page, page->index, 0, flags);
1568 }
1569
1570 int logfs_delete(struct inode *inode, pgoff_t index,
1571                 struct shadow_tree *shadow_tree)
1572 {
1573         struct super_block *sb = inode->i_sb;
1574         struct page *page;
1575         int ret;
1576
1577         page = logfs_get_read_page(inode, index, 0);
1578         if (!page)
1579                 return -ENOMEM;
1580
1581         logfs_get_wblocks(sb, page, 1);
1582         ret = __logfs_delete(inode, page);
1583         logfs_put_wblocks(sb, page, 1);
1584
1585         logfs_put_read_page(page);
1586
1587         return ret;
1588 }
1589
1590 int logfs_rewrite_block(struct inode *inode, u64 bix, u64 ofs,
1591                 gc_level_t gc_level, long flags)
1592 {
1593         level_t level = shrink_level(gc_level);
1594         struct page *page;
1595         int err;
1596
1597         page = logfs_get_write_page(inode, bix, level);
1598         if (!page)
1599                 return -ENOMEM;
1600
1601         err = logfs_segment_read(inode, page, ofs, bix, level);
1602         if (!err) {
1603                 if (level != 0)
1604                         alloc_indirect_block(inode, page, 0);
1605                 err = logfs_write_buf(inode, page, flags);
1606                 if (!err && shrink_level(gc_level) == 0) {
1607                         /* Rewrite cannot mark the inode dirty but has to
1608                          * write it immediatly.
1609                          * Q: Can't we just create an alias for the inode
1610                          * instead?  And if not, why not?
1611                          */
1612                         if (inode->i_ino == LOGFS_INO_MASTER)
1613                                 logfs_write_anchor(inode->i_sb);
1614                         else {
1615                                 err = __logfs_write_inode(inode, flags);
1616                         }
1617                 }
1618         }
1619         logfs_put_write_page(page);
1620         return err;
1621 }
1622
1623 static int truncate_data_block(struct inode *inode, struct page *page,
1624                 u64 ofs, struct logfs_shadow *shadow, u64 size)
1625 {
1626         loff_t pageofs = page->index << inode->i_sb->s_blocksize_bits;
1627         u64 bix;
1628         level_t level;
1629         int err;
1630
1631         /* Does truncation happen within this page? */
1632         if (size <= pageofs || size - pageofs >= PAGE_SIZE)
1633                 return 0;
1634
1635         logfs_unpack_index(page->index, &bix, &level);
1636         BUG_ON(level != 0);
1637
1638         err = logfs_segment_read(inode, page, ofs, bix, level);
1639         if (err)
1640                 return err;
1641
1642         zero_user_segment(page, size - pageofs, PAGE_CACHE_SIZE);
1643         return logfs_segment_write(inode, page, shadow);
1644 }
1645
1646 static int logfs_truncate_i0(struct inode *inode, struct page *page,
1647                 struct write_control *wc, u64 size)
1648 {
1649         struct logfs_shadow *shadow;
1650         u64 bix;
1651         level_t level;
1652         int err = 0;
1653
1654         logfs_unpack_index(page->index, &bix, &level);
1655         BUG_ON(level != 0);
1656         shadow = alloc_shadow(inode, bix, level, wc->ofs);
1657
1658         err = truncate_data_block(inode, page, wc->ofs, shadow, size);
1659         if (err) {
1660                 free_shadow(inode, shadow);
1661                 return err;
1662         }
1663
1664         logfs_segment_delete(inode, shadow);
1665         set_iused(inode, shadow);
1666         fill_shadow_tree(inode, page, shadow);
1667         wc->ofs = shadow->new_ofs;
1668         return 0;
1669 }
1670
1671 static int logfs_truncate_direct(struct inode *inode, u64 size)
1672 {
1673         struct logfs_inode *li = logfs_inode(inode);
1674         struct write_control wc;
1675         struct page *page;
1676         int e;
1677         int err;
1678
1679         alloc_inode_block(inode);
1680
1681         for (e = I0_BLOCKS - 1; e >= 0; e--) {
1682                 if (size > (e+1) * LOGFS_BLOCKSIZE)
1683                         break;
1684
1685                 wc.ofs = li->li_data[e];
1686                 if (!wc.ofs)
1687                         continue;
1688
1689                 page = logfs_get_write_page(inode, e, 0);
1690                 if (!page)
1691                         return -ENOMEM;
1692                 err = logfs_segment_read(inode, page, wc.ofs, e, 0);
1693                 if (err) {
1694                         logfs_put_write_page(page);
1695                         return err;
1696                 }
1697                 err = logfs_truncate_i0(inode, page, &wc, size);
1698                 logfs_put_write_page(page);
1699                 if (err)
1700                         return err;
1701
1702                 li->li_data[e] = wc.ofs;
1703         }
1704         return 0;
1705 }
1706
1707 /* FIXME: these need to become per-sb once we support different blocksizes */
1708 static u64 __logfs_step[] = {
1709         1,
1710         I1_BLOCKS,
1711         I2_BLOCKS,
1712         I3_BLOCKS,
1713 };
1714
1715 static u64 __logfs_start_index[] = {
1716         I0_BLOCKS,
1717         I1_BLOCKS,
1718         I2_BLOCKS,
1719         I3_BLOCKS
1720 };
1721
1722 static inline u64 logfs_step(level_t level)
1723 {
1724         return __logfs_step[(__force u8)level];
1725 }
1726
1727 static inline u64 logfs_factor(u8 level)
1728 {
1729         return __logfs_step[level] * LOGFS_BLOCKSIZE;
1730 }
1731
1732 static inline u64 logfs_start_index(level_t level)
1733 {
1734         return __logfs_start_index[(__force u8)level];
1735 }
1736
1737 static void logfs_unpack_raw_index(pgoff_t index, u64 *bix, level_t *level)
1738 {
1739         logfs_unpack_index(index, bix, level);
1740         if (*bix <= logfs_start_index(SUBLEVEL(*level)))
1741                 *bix = 0;
1742 }
1743
1744 static int __logfs_truncate_rec(struct inode *inode, struct page *ipage,
1745                 struct write_control *this_wc, u64 size)
1746 {
1747         int truncate_happened = 0;
1748         int e, err = 0;
1749         u64 bix, child_bix, next_bix;
1750         level_t level;
1751         struct page *page;
1752         struct write_control child_wc = { /* FIXME: flags */ };
1753
1754         logfs_unpack_raw_index(ipage->index, &bix, &level);
1755         err = logfs_segment_read(inode, ipage, this_wc->ofs, bix, level);
1756         if (err)
1757                 return err;
1758
1759         for (e = LOGFS_BLOCK_FACTOR - 1; e >= 0; e--) {
1760                 child_bix = bix + e * logfs_step(SUBLEVEL(level));
1761                 next_bix = child_bix + logfs_step(SUBLEVEL(level));
1762                 if (size > next_bix * LOGFS_BLOCKSIZE)
1763                         break;
1764
1765                 child_wc.ofs = pure_ofs(block_get_pointer(ipage, e));
1766                 if (!child_wc.ofs)
1767                         continue;
1768
1769                 page = logfs_get_write_page(inode, child_bix, SUBLEVEL(level));
1770                 if (!page)
1771                         return -ENOMEM;
1772
1773                 if ((__force u8)level > 1)
1774                         err = __logfs_truncate_rec(inode, page, &child_wc, size);
1775                 else
1776                         err = logfs_truncate_i0(inode, page, &child_wc, size);
1777                 logfs_put_write_page(page);
1778                 if (err)
1779                         return err;
1780
1781                 truncate_happened = 1;
1782                 alloc_indirect_block(inode, ipage, 0);
1783                 block_set_pointer(ipage, e, child_wc.ofs);
1784         }
1785
1786         if (!truncate_happened) {
1787                 printk("ineffectual truncate (%lx, %lx, %llx)\n", inode->i_ino, ipage->index, size);
1788                 return 0;
1789         }
1790
1791         this_wc->flags = WF_DELETE;
1792         if (logfs_block(ipage)->partial)
1793                 this_wc->flags |= WF_WRITE;
1794
1795         return logfs_write_i0(inode, ipage, this_wc);
1796 }
1797
1798 static int logfs_truncate_rec(struct inode *inode, u64 size)
1799 {
1800         struct logfs_inode *li = logfs_inode(inode);
1801         struct write_control wc = {
1802                 .ofs = li->li_data[INDIRECT_INDEX],
1803         };
1804         struct page *page;
1805         int err;
1806
1807         alloc_inode_block(inode);
1808
1809         if (!wc.ofs)
1810                 return 0;
1811
1812         page = logfs_get_write_page(inode, 0, LEVEL(li->li_height));
1813         if (!page)
1814                 return -ENOMEM;
1815
1816         err = __logfs_truncate_rec(inode, page, &wc, size);
1817         logfs_put_write_page(page);
1818         if (err)
1819                 return err;
1820
1821         if (li->li_data[INDIRECT_INDEX] != wc.ofs)
1822                 li->li_data[INDIRECT_INDEX] = wc.ofs;
1823         return 0;
1824 }
1825
1826 static int __logfs_truncate(struct inode *inode, u64 size)
1827 {
1828         int ret;
1829
1830         if (size >= logfs_factor(logfs_inode(inode)->li_height))
1831                 return 0;
1832
1833         ret = logfs_truncate_rec(inode, size);
1834         if (ret)
1835                 return ret;
1836
1837         return logfs_truncate_direct(inode, size);
1838 }
1839
1840 /*
1841  * Truncate, by changing the segment file, can consume a fair amount
1842  * of resources.  So back off from time to time and do some GC.
1843  * 8 or 2048 blocks should be well within safety limits even if
1844  * every single block resided in a different segment.
1845  */
1846 #define TRUNCATE_STEP   (8 * 1024 * 1024)
1847 int logfs_truncate(struct inode *inode, u64 target)
1848 {
1849         struct super_block *sb = inode->i_sb;
1850         u64 size = i_size_read(inode);
1851         int err = 0;
1852
1853         size = ALIGN(size, TRUNCATE_STEP);
1854         while (size > target) {
1855                 if (size > TRUNCATE_STEP)
1856                         size -= TRUNCATE_STEP;
1857                 else
1858                         size = 0;
1859                 if (size < target)
1860                         size = target;
1861
1862                 logfs_get_wblocks(sb, NULL, 1);
1863                 err = __logfs_truncate(inode, target);
1864                 if (!err)
1865                         err = __logfs_write_inode(inode, 0);
1866                 logfs_put_wblocks(sb, NULL, 1);
1867         }
1868
1869         if (!err)
1870                 err = vmtruncate(inode, target);
1871
1872         /* I don't trust error recovery yet. */
1873         WARN_ON(err);
1874         return err;
1875 }
1876
1877 static void move_page_to_inode(struct inode *inode, struct page *page)
1878 {
1879         struct logfs_inode *li = logfs_inode(inode);
1880         struct logfs_block *block = logfs_block(page);
1881
1882         if (!block)
1883                 return;
1884
1885         log_blockmove("move_page_to_inode(%llx, %llx, %x)\n",
1886                         block->ino, block->bix, block->level);
1887         BUG_ON(li->li_block);
1888         block->ops = &inode_block_ops;
1889         block->inode = inode;
1890         li->li_block = block;
1891
1892         block->page = NULL;
1893         page->private = 0;
1894         ClearPagePrivate(page);
1895 }
1896
1897 static void move_inode_to_page(struct page *page, struct inode *inode)
1898 {
1899         struct logfs_inode *li = logfs_inode(inode);
1900         struct logfs_block *block = li->li_block;
1901
1902         if (!block)
1903                 return;
1904
1905         log_blockmove("move_inode_to_page(%llx, %llx, %x)\n",
1906                         block->ino, block->bix, block->level);
1907         BUG_ON(PagePrivate(page));
1908         block->ops = &indirect_block_ops;
1909         block->page = page;
1910         page->private = (unsigned long)block;
1911         SetPagePrivate(page);
1912
1913         block->inode = NULL;
1914         li->li_block = NULL;
1915 }
1916
1917 int logfs_read_inode(struct inode *inode)
1918 {
1919         struct super_block *sb = inode->i_sb;
1920         struct logfs_super *super = logfs_super(sb);
1921         struct inode *master_inode = super->s_master_inode;
1922         struct page *page;
1923         struct logfs_disk_inode *di;
1924         u64 ino = inode->i_ino;
1925
1926         if (ino << sb->s_blocksize_bits > i_size_read(master_inode))
1927                 return -ENODATA;
1928         if (!logfs_exist_block(master_inode, ino))
1929                 return -ENODATA;
1930
1931         page = read_cache_page(master_inode->i_mapping, ino,
1932                         (filler_t *)logfs_readpage, NULL);
1933         if (IS_ERR(page))
1934                 return PTR_ERR(page);
1935
1936         di = kmap_atomic(page, KM_USER0);
1937         logfs_disk_to_inode(di, inode);
1938         kunmap_atomic(di, KM_USER0);
1939         move_page_to_inode(inode, page);
1940         page_cache_release(page);
1941         return 0;
1942 }
1943
1944 /* Caller must logfs_put_write_page(page); */
1945 static struct page *inode_to_page(struct inode *inode)
1946 {
1947         struct inode *master_inode = logfs_super(inode->i_sb)->s_master_inode;
1948         struct logfs_disk_inode *di;
1949         struct page *page;
1950
1951         BUG_ON(inode->i_ino == LOGFS_INO_MASTER);
1952
1953         page = logfs_get_write_page(master_inode, inode->i_ino, 0);
1954         if (!page)
1955                 return NULL;
1956
1957         di = kmap_atomic(page, KM_USER0);
1958         logfs_inode_to_disk(inode, di);
1959         kunmap_atomic(di, KM_USER0);
1960         move_inode_to_page(page, inode);
1961         return page;
1962 }
1963
1964 /* Cheaper version of write_inode.  All changes are concealed in
1965  * aliases, which are moved back.  No write to the medium happens.
1966  */
1967 void logfs_clear_inode(struct inode *inode)
1968 {
1969         struct super_block *sb = inode->i_sb;
1970         struct logfs_inode *li = logfs_inode(inode);
1971         struct logfs_block *block = li->li_block;
1972         struct page *page;
1973
1974         /* Only deleted files may be dirty at this point */
1975         BUG_ON(inode->i_state & I_DIRTY && inode->i_nlink);
1976         if (!block)
1977                 return;
1978         if ((logfs_super(sb)->s_flags & LOGFS_SB_FLAG_SHUTDOWN)) {
1979                 block->ops->free_block(inode->i_sb, block);
1980                 return;
1981         }
1982
1983         BUG_ON(inode->i_ino < LOGFS_RESERVED_INOS);
1984         page = inode_to_page(inode);
1985         BUG_ON(!page); /* FIXME: Use emergency page */
1986         logfs_put_write_page(page);
1987 }
1988
1989 static int do_write_inode(struct inode *inode)
1990 {
1991         struct super_block *sb = inode->i_sb;
1992         struct inode *master_inode = logfs_super(sb)->s_master_inode;
1993         loff_t size = (inode->i_ino + 1) << inode->i_sb->s_blocksize_bits;
1994         struct page *page;
1995         int err;
1996
1997         BUG_ON(inode->i_ino == LOGFS_INO_MASTER);
1998         /* FIXME: lock inode */
1999
2000         if (i_size_read(master_inode) < size)
2001                 i_size_write(master_inode, size);
2002
2003         /* TODO: Tell vfs this inode is clean now */
2004
2005         page = inode_to_page(inode);
2006         if (!page)
2007                 return -ENOMEM;
2008
2009         /* FIXME: transaction is part of logfs_block now.  Is that enough? */
2010         err = logfs_write_buf(master_inode, page, 0);
2011         logfs_put_write_page(page);
2012         return err;
2013 }
2014
2015 static void logfs_mod_segment_entry(struct super_block *sb, u32 segno,
2016                 int write,
2017                 void (*change_se)(struct logfs_segment_entry *, long),
2018                 long arg)
2019 {
2020         struct logfs_super *super = logfs_super(sb);
2021         struct inode *inode;
2022         struct page *page;
2023         struct logfs_segment_entry *se;
2024         pgoff_t page_no;
2025         int child_no;
2026
2027         page_no = segno >> (sb->s_blocksize_bits - 3);
2028         child_no = segno & ((sb->s_blocksize >> 3) - 1);
2029
2030         inode = super->s_segfile_inode;
2031         page = logfs_get_write_page(inode, page_no, 0);
2032         BUG_ON(!page); /* FIXME: We need some reserve page for this case */
2033         if (!PageUptodate(page))
2034                 logfs_read_block(inode, page, WRITE);
2035
2036         if (write)
2037                 alloc_indirect_block(inode, page, 0);
2038         se = kmap_atomic(page, KM_USER0);
2039         change_se(se + child_no, arg);
2040         if (write) {
2041                 logfs_set_alias(sb, logfs_block(page), child_no);
2042                 BUG_ON((int)be32_to_cpu(se[child_no].valid) > super->s_segsize);
2043         }
2044         kunmap_atomic(se, KM_USER0);
2045
2046         logfs_put_write_page(page);
2047 }
2048
2049 static void __get_segment_entry(struct logfs_segment_entry *se, long _target)
2050 {
2051         struct logfs_segment_entry *target = (void *)_target;
2052
2053         *target = *se;
2054 }
2055
2056 void logfs_get_segment_entry(struct super_block *sb, u32 segno,
2057                 struct logfs_segment_entry *se)
2058 {
2059         logfs_mod_segment_entry(sb, segno, 0, __get_segment_entry, (long)se);
2060 }
2061
2062 static void __set_segment_used(struct logfs_segment_entry *se, long increment)
2063 {
2064         u32 valid;
2065
2066         valid = be32_to_cpu(se->valid);
2067         valid += increment;
2068         se->valid = cpu_to_be32(valid);
2069 }
2070
2071 void logfs_set_segment_used(struct super_block *sb, u64 ofs, int increment)
2072 {
2073         struct logfs_super *super = logfs_super(sb);
2074         u32 segno = ofs >> super->s_segshift;
2075
2076         if (!increment)
2077                 return;
2078
2079         logfs_mod_segment_entry(sb, segno, 1, __set_segment_used, increment);
2080 }
2081
2082 static void __set_segment_erased(struct logfs_segment_entry *se, long ec_level)
2083 {
2084         se->ec_level = cpu_to_be32(ec_level);
2085 }
2086
2087 void logfs_set_segment_erased(struct super_block *sb, u32 segno, u32 ec,
2088                 gc_level_t gc_level)
2089 {
2090         u32 ec_level = ec << 4 | (__force u8)gc_level;
2091
2092         logfs_mod_segment_entry(sb, segno, 1, __set_segment_erased, ec_level);
2093 }
2094
2095 static void __set_segment_reserved(struct logfs_segment_entry *se, long ignore)
2096 {
2097         se->valid = cpu_to_be32(RESERVED);
2098 }
2099
2100 void logfs_set_segment_reserved(struct super_block *sb, u32 segno)
2101 {
2102         logfs_mod_segment_entry(sb, segno, 1, __set_segment_reserved, 0);
2103 }
2104
2105 static void __set_segment_unreserved(struct logfs_segment_entry *se,
2106                 long ec_level)
2107 {
2108         se->valid = 0;
2109         se->ec_level = cpu_to_be32(ec_level);
2110 }
2111
2112 void logfs_set_segment_unreserved(struct super_block *sb, u32 segno, u32 ec)
2113 {
2114         u32 ec_level = ec << 4;
2115
2116         logfs_mod_segment_entry(sb, segno, 1, __set_segment_unreserved,
2117                         ec_level);
2118 }
2119
2120 int __logfs_write_inode(struct inode *inode, long flags)
2121 {
2122         struct super_block *sb = inode->i_sb;
2123         int ret;
2124
2125         logfs_get_wblocks(sb, NULL, flags & WF_LOCK);
2126         ret = do_write_inode(inode);
2127         logfs_put_wblocks(sb, NULL, flags & WF_LOCK);
2128         return ret;
2129 }
2130
2131 static int do_delete_inode(struct inode *inode)
2132 {
2133         struct super_block *sb = inode->i_sb;
2134         struct inode *master_inode = logfs_super(sb)->s_master_inode;
2135         struct page *page;
2136         int ret;
2137
2138         page = logfs_get_write_page(master_inode, inode->i_ino, 0);
2139         if (!page)
2140                 return -ENOMEM;
2141
2142         move_inode_to_page(page, inode);
2143
2144         logfs_get_wblocks(sb, page, 1);
2145         ret = __logfs_delete(master_inode, page);
2146         logfs_put_wblocks(sb, page, 1);
2147
2148         logfs_put_write_page(page);
2149         return ret;
2150 }
2151
2152 /*
2153  * ZOMBIE inodes have already been deleted before and should remain dead,
2154  * if it weren't for valid checking.  No need to kill them again here.
2155  */
2156 void logfs_delete_inode(struct inode *inode)
2157 {
2158         struct logfs_inode *li = logfs_inode(inode);
2159
2160         if (!(li->li_flags & LOGFS_IF_ZOMBIE)) {
2161                 li->li_flags |= LOGFS_IF_ZOMBIE;
2162                 if (i_size_read(inode) > 0)
2163                         logfs_truncate(inode, 0);
2164                 do_delete_inode(inode);
2165         }
2166         truncate_inode_pages(&inode->i_data, 0);
2167         clear_inode(inode);
2168 }
2169
2170 void btree_write_block(struct logfs_block *block)
2171 {
2172         struct inode *inode;
2173         struct page *page;
2174         int err, cookie;
2175
2176         inode = logfs_safe_iget(block->sb, block->ino, &cookie);
2177         page = logfs_get_write_page(inode, block->bix, block->level);
2178
2179         err = logfs_readpage_nolock(page);
2180         BUG_ON(err);
2181         BUG_ON(!PagePrivate(page));
2182         BUG_ON(logfs_block(page) != block);
2183         err = __logfs_write_buf(inode, page, 0);
2184         BUG_ON(err);
2185         BUG_ON(PagePrivate(page) || page->private);
2186
2187         logfs_put_write_page(page);
2188         logfs_safe_iput(inode, cookie);
2189 }
2190
2191 /**
2192  * logfs_inode_write - write inode or dentry objects
2193  *
2194  * @inode:              parent inode (ifile or directory)
2195  * @buf:                object to write (inode or dentry)
2196  * @n:                  object size
2197  * @_pos:               object number (file position in blocks/objects)
2198  * @flags:              write flags
2199  * @lock:               0 if write lock is already taken, 1 otherwise
2200  * @shadow_tree:        shadow below this inode
2201  *
2202  * FIXME: All caller of this put a 200-300 byte variable on the stack,
2203  * only to call here and do a memcpy from that stack variable.  A good
2204  * example of wasted performance and stack space.
2205  */
2206 int logfs_inode_write(struct inode *inode, const void *buf, size_t count,
2207                 loff_t bix, long flags, struct shadow_tree *shadow_tree)
2208 {
2209         loff_t pos = bix << inode->i_sb->s_blocksize_bits;
2210         int err;
2211         struct page *page;
2212         void *pagebuf;
2213
2214         BUG_ON(pos & (LOGFS_BLOCKSIZE-1));
2215         BUG_ON(count > LOGFS_BLOCKSIZE);
2216         page = logfs_get_write_page(inode, bix, 0);
2217         if (!page)
2218                 return -ENOMEM;
2219
2220         pagebuf = kmap_atomic(page, KM_USER0);
2221         memcpy(pagebuf, buf, count);
2222         flush_dcache_page(page);
2223         kunmap_atomic(pagebuf, KM_USER0);
2224
2225         if (i_size_read(inode) < pos + LOGFS_BLOCKSIZE)
2226                 i_size_write(inode, pos + LOGFS_BLOCKSIZE);
2227
2228         err = logfs_write_buf(inode, page, flags);
2229         logfs_put_write_page(page);
2230         return err;
2231 }
2232
2233 int logfs_open_segfile(struct super_block *sb)
2234 {
2235         struct logfs_super *super = logfs_super(sb);
2236         struct inode *inode;
2237
2238         inode = logfs_read_meta_inode(sb, LOGFS_INO_SEGFILE);
2239         if (IS_ERR(inode))
2240                 return PTR_ERR(inode);
2241         super->s_segfile_inode = inode;
2242         return 0;
2243 }
2244
2245 int logfs_init_rw(struct super_block *sb)
2246 {
2247         struct logfs_super *super = logfs_super(sb);
2248         int min_fill = 3 * super->s_no_blocks;
2249
2250         INIT_LIST_HEAD(&super->s_object_alias);
2251         mutex_init(&super->s_write_mutex);
2252         super->s_block_pool = mempool_create_kmalloc_pool(min_fill,
2253                         sizeof(struct logfs_block));
2254         super->s_shadow_pool = mempool_create_kmalloc_pool(min_fill,
2255                         sizeof(struct logfs_shadow));
2256         return 0;
2257 }
2258
2259 void logfs_cleanup_rw(struct super_block *sb)
2260 {
2261         struct logfs_super *super = logfs_super(sb);
2262
2263         destroy_meta_inode(super->s_segfile_inode);
2264         logfs_mempool_destroy(super->s_block_pool);
2265         logfs_mempool_destroy(super->s_shadow_pool);
2266 }