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 #include <linux/slab.h>
  22
  23 static u64 adjust_bix(u64 bix, level_t level)
  24 {
  25         switch (level) {
  26         case 0:
  27                 return bix;
  28         case LEVEL(1):
  29                 return max_t(u64, bix, I0_BLOCKS);
  30         case LEVEL(2):
  31                 return max_t(u64, bix, I1_BLOCKS);
  32         case LEVEL(3):
  33                 return max_t(u64, bix, I2_BLOCKS);
  34         case LEVEL(4):
  35                 return max_t(u64, bix, I3_BLOCKS);
  36         case LEVEL(5):
  37                 return max_t(u64, bix, I4_BLOCKS);
  38         default:
  39                 WARN_ON(1);
  40                 return bix;
  41         }
  42 }
  43
  44 static inline u64 maxbix(u8 height)
  45 {
  46         return 1ULL << (LOGFS_BLOCK_BITS * height);
  47 }
  48
  49 /**
  50  * The inode address space is cut in two halves.  Lower half belongs to data
  51  * pages, upper half to indirect blocks.  If the high bit (INDIRECT_BIT) is
  52  * set, the actual block index (bix) and level can be derived from the page
  53  * index.
  54  *
  55  * The lowest three bits of the block index are set to 0 after packing and
  56  * unpacking.  Since the lowest n bits (9 for 4KiB blocksize) are ignored
  57  * anyway this is harmless.
  58  */
  59 #define ARCH_SHIFT      (BITS_PER_LONG - 32)
  60 #define INDIRECT_BIT    (0x80000000UL << ARCH_SHIFT)
  61 #define LEVEL_SHIFT     (28 + ARCH_SHIFT)
  62 static inline pgoff_t first_indirect_block(void)
  63 {
  64         return INDIRECT_BIT | (1ULL << LEVEL_SHIFT);
  65 }
  66
  67 pgoff_t logfs_pack_index(u64 bix, level_t level)
  68 {
  69         pgoff_t index;
  70
  71         BUG_ON(bix >= INDIRECT_BIT);
  72         if (level == 0)
  73                 return bix;
  74
  75         index  = INDIRECT_BIT;
  76         index |= (__force long)level << LEVEL_SHIFT;
  77         index |= bix >> ((__force u8)level * LOGFS_BLOCK_BITS);
  78         return index;
  79 }
  80
  81 void logfs_unpack_index(pgoff_t index, u64 *bix, level_t *level)
  82 {
  83         u8 __level;
  84
  85         if (!(index & INDIRECT_BIT)) {
  86                 *bix = index;
  87                 *level = 0;
  88                 return;
  89         }
  90
  91         __level = (index & ~INDIRECT_BIT) >> LEVEL_SHIFT;
  92         *level = LEVEL(__level);
  93         *bix = (index << (__level * LOGFS_BLOCK_BITS)) & ~INDIRECT_BIT;
  94         *bix = adjust_bix(*bix, *level);
  95         return;
  96 }
  97 #undef ARCH_SHIFT
  98 #undef INDIRECT_BIT
  99 #undef LEVEL_SHIFT
 100
 101 /*
 102  * Time is stored as nanoseconds since the epoch.
 103  */
 104 static struct timespec be64_to_timespec(__be64 betime)
 105 {
 106         return ns_to_timespec(be64_to_cpu(betime));
 107 }
 108
 109 static __be64 timespec_to_be64(struct timespec tsp)
 110 {
 111         return cpu_to_be64((u64)tsp.tv_sec * NSEC_PER_SEC + tsp.tv_nsec);
 112 }
 113
 114 static void logfs_disk_to_inode(struct logfs_disk_inode *di, struct inode*inode)
 115 {
 116         struct logfs_inode *li = logfs_inode(inode);
 117         int i;
 118
 119         inode->i_mode   = be16_to_cpu(di->di_mode);
 120         li->li_height   = di->di_height;
 121         li->li_flags    = be32_to_cpu(di->di_flags);
 122         inode->i_uid    = be32_to_cpu(di->di_uid);
 123         inode->i_gid    = be32_to_cpu(di->di_gid);
 124         inode->i_size   = be64_to_cpu(di->di_size);
 125         logfs_set_blocks(inode, be64_to_cpu(di->di_used_bytes));
 126         inode->i_atime  = be64_to_timespec(di->di_atime);
 127         inode->i_ctime  = be64_to_timespec(di->di_ctime);
 128         inode->i_mtime  = be64_to_timespec(di->di_mtime);
 129         set_nlink(inode, be32_to_cpu(di->di_refcount));
 130         inode->i_generation = be32_to_cpu(di->di_generation);
 131
 132         switch (inode->i_mode & S_IFMT) {
 133         case S_IFSOCK:  /* fall through */
 134         case S_IFBLK:   /* fall through */
 135         case S_IFCHR:   /* fall through */
 136         case S_IFIFO:
 137                 inode->i_rdev = be64_to_cpu(di->di_data[0]);
 138                 break;
 139         case S_IFDIR:   /* fall through */
 140         case S_IFREG:   /* fall through */
 141         case S_IFLNK:
 142                 for (i = 0; i < LOGFS_EMBEDDED_FIELDS; i++)
 143                         li->li_data[i] = be64_to_cpu(di->di_data[i]);
 144                 break;
 145         default:
 146                 BUG();
 147         }
 148 }
 149
 150 static void logfs_inode_to_disk(struct inode *inode, struct logfs_disk_inode*di)
 151 {
 152         struct logfs_inode *li = logfs_inode(inode);
 153         int i;
 154
 155         di->di_mode     = cpu_to_be16(inode->i_mode);
 156         di->di_height   = li->li_height;
 157         di->di_pad      = 0;
 158         di->di_flags    = cpu_to_be32(li->li_flags);
 159         di->di_uid      = cpu_to_be32(inode->i_uid);
 160         di->di_gid      = cpu_to_be32(inode->i_gid);
 161         di->di_size     = cpu_to_be64(i_size_read(inode));
 162         di->di_used_bytes = cpu_to_be64(li->li_used_bytes);
 163         di->di_atime    = timespec_to_be64(inode->i_atime);
 164         di->di_ctime    = timespec_to_be64(inode->i_ctime);
 165         di->di_mtime    = timespec_to_be64(inode->i_mtime);
 166         di->di_refcount = cpu_to_be32(inode->i_nlink);
 167         di->di_generation = cpu_to_be32(inode->i_generation);
 168
 169         switch (inode->i_mode & S_IFMT) {
 170         case S_IFSOCK:  /* fall through */
 171         case S_IFBLK:   /* fall through */
 172         case S_IFCHR:   /* fall through */
 173         case S_IFIFO:
 174                 di->di_data[0] = cpu_to_be64(inode->i_rdev);
 175                 break;
 176         case S_IFDIR:   /* fall through */
 177         case S_IFREG:   /* fall through */
 178         case S_IFLNK:
 179                 for (i = 0; i < LOGFS_EMBEDDED_FIELDS; i++)
 180                         di->di_data[i] = cpu_to_be64(li->li_data[i]);
 181                 break;
 182         default:
 183                 BUG();
 184         }
 185 }
 186
 187 static void __logfs_set_blocks(struct inode *inode)
 188 {
 189         struct super_block *sb = inode->i_sb;
 190         struct logfs_inode *li = logfs_inode(inode);
 191
 192         inode->i_blocks = ULONG_MAX;
 193         if (li->li_used_bytes >> sb->s_blocksize_bits < ULONG_MAX)
 194                 inode->i_blocks = ALIGN(li->li_used_bytes, 512) >> 9;
 195 }
 196
 197 void logfs_set_blocks(struct inode *inode, u64 bytes)
 198 {
 199         struct logfs_inode *li = logfs_inode(inode);
 200
 201         li->li_used_bytes = bytes;
 202         __logfs_set_blocks(inode);
 203 }
 204
 205 static void prelock_page(struct super_block *sb, struct page *page, int lock)
 206 {
 207         struct logfs_super *super = logfs_super(sb);
 208
 209         BUG_ON(!PageLocked(page));
 210         if (lock) {
 211                 BUG_ON(PagePreLocked(page));
 212                 SetPagePreLocked(page);
 213         } else {
 214                 /* We are in GC path. */
 215                 if (PagePreLocked(page))
 216                         super->s_lock_count++;
 217                 else
 218                         SetPagePreLocked(page);
 219         }
 220 }
 221
 222 static void preunlock_page(struct super_block *sb, struct page *page, int lock)
 223 {
 224         struct logfs_super *super = logfs_super(sb);
 225
 226         BUG_ON(!PageLocked(page));
 227         if (lock)
 228                 ClearPagePreLocked(page);
 229         else {
 230                 /* We are in GC path. */
 231                 BUG_ON(!PagePreLocked(page));
 232                 if (super->s_lock_count)
 233                         super->s_lock_count--;
 234                 else
 235                         ClearPagePreLocked(page);
 236         }
 237 }
 238
 239 /*
 240  * Logfs is prone to an AB-BA deadlock where one task tries to acquire
 241  * s_write_mutex with a locked page and GC tries to get that page while holding
 242  * s_write_mutex.
 243  * To solve this issue logfs will ignore the page lock iff the page in question
 244  * is waiting for s_write_mutex.  We annotate this fact by setting PG_pre_locked
 245  * in addition to PG_locked.
 246  */
 247 static void logfs_get_wblocks(struct super_block *sb, struct page *page,
 248                 int lock)
 249 {
 250         struct logfs_super *super = logfs_super(sb);
 251
 252         if (page)
 253                 prelock_page(sb, page, lock);
 254
 255         if (lock) {
 256                 mutex_lock(&super->s_write_mutex);
 257                 logfs_gc_pass(sb);
 258                 /* FIXME: We also have to check for shadowed space
 259                  * and mempool fill grade */
 260         }
 261 }
 262
 263 static void logfs_put_wblocks(struct super_block *sb, struct page *page,
 264                 int lock)
 265 {
 266         struct logfs_super *super = logfs_super(sb);
 267
 268         if (page)
 269                 preunlock_page(sb, page, lock);
 270         /* Order matters - we must clear PG_pre_locked before releasing
 271          * s_write_mutex or we could race against another task. */
 272         if (lock)
 273                 mutex_unlock(&super->s_write_mutex);
 274 }
 275
 276 static struct page *logfs_get_read_page(struct inode *inode, u64 bix,
 277                 level_t level)
 278 {
 279         return find_or_create_page(inode->i_mapping,
 280                         logfs_pack_index(bix, level), GFP_NOFS);
 281 }
 282
 283 static void logfs_put_read_page(struct page *page)
 284 {
 285         unlock_page(page);
 286         page_cache_release(page);
 287 }
 288
 289 static void logfs_lock_write_page(struct page *page)
 290 {
 291         int loop = 0;
 292
 293         while (unlikely(!trylock_page(page))) {
 294                 if (loop++ > 0x1000) {
 295                         /* Has been observed once so far... */
 296                         printk(KERN_ERR "stack at %p\n", &loop);
 297                         BUG();
 298                 }
 299                 if (PagePreLocked(page)) {
 300                         /* Holder of page lock is waiting for us, it
 301                          * is safe to use this page. */
 302                         break;
 303                 }
 304                 /* Some other process has this page locked and has
 305                  * nothing to do with us.  Wait for it to finish.
 306                  */
 307                 schedule();
 308         }
 309         BUG_ON(!PageLocked(page));
 310 }
 311
 312 static struct page *logfs_get_write_page(struct inode *inode, u64 bix,
 313                 level_t level)
 314 {
 315         struct address_space *mapping = inode->i_mapping;
 316         pgoff_t index = logfs_pack_index(bix, level);
 317         struct page *page;
 318         int err;
 319
 320 repeat:
 321         page = find_get_page(mapping, index);
 322         if (!page) {
 323                 page = __page_cache_alloc(GFP_NOFS);
 324                 if (!page)
 325                         return NULL;
 326                 err = add_to_page_cache_lru(page, mapping, index, GFP_NOFS);
 327                 if (unlikely(err)) {
 328                         page_cache_release(page);
 329                         if (err == -EEXIST)
 330                                 goto repeat;
 331                         return NULL;
 332                 }
 333         } else logfs_lock_write_page(page);
 334         BUG_ON(!PageLocked(page));
 335         return page;
 336 }
 337
 338 static void logfs_unlock_write_page(struct page *page)
 339 {
 340         if (!PagePreLocked(page))
 341                 unlock_page(page);
 342 }
 343
 344 static void logfs_put_write_page(struct page *page)
 345 {
 346         logfs_unlock_write_page(page);
 347         page_cache_release(page);
 348 }
 349
 350 static struct page *logfs_get_page(struct inode *inode, u64 bix, level_t level,
 351                 int rw)
 352 {
 353         if (rw == READ)
 354                 return logfs_get_read_page(inode, bix, level);
 355         else
 356                 return logfs_get_write_page(inode, bix, level);
 357 }
 358
 359 static void logfs_put_page(struct page *page, int rw)
 360 {
 361         if (rw == READ)
 362                 logfs_put_read_page(page);
 363         else
 364                 logfs_put_write_page(page);
 365 }
 366
 367 static unsigned long __get_bits(u64 val, int skip, int no)
 368 {
 369         u64 ret = val;
 370
 371         ret >>= skip * no;
 372         ret <<= 64 - no;
 373         ret >>= 64 - no;
 374         return ret;
 375 }
 376
 377 static unsigned long get_bits(u64 val, level_t skip)
 378 {
 379         return __get_bits(val, (__force int)skip, LOGFS_BLOCK_BITS);
 380 }
 381
 382 static inline void init_shadow_tree(struct super_block *sb,
 383                 struct shadow_tree *tree)
 384 {
 385         struct logfs_super *super = logfs_super(sb);
 386
 387         btree_init_mempool64(&tree->new, super->s_btree_pool);
 388         btree_init_mempool64(&tree->old, super->s_btree_pool);
 389 }
 390
 391 static void indirect_write_block(struct logfs_block *block)
 392 {
 393         struct page *page;
 394         struct inode *inode;
 395         int ret;
 396
 397         page = block->page;
 398         inode = page->mapping->host;
 399         logfs_lock_write_page(page);
 400         ret = logfs_write_buf(inode, page, 0);
 401         logfs_unlock_write_page(page);
 402         /*
 403          * This needs some rework.  Unless you want your filesystem to run
 404          * completely synchronously (you don't), the filesystem will always
 405          * report writes as 'successful' before the actual work has been
 406          * done.  The actual work gets done here and this is where any errors
 407          * will show up.  And there isn't much we can do about it, really.
 408          *
 409          * Some attempts to fix the errors (move from bad blocks, retry io,...)
 410          * have already been done, so anything left should be either a broken
 411          * device or a bug somewhere in logfs itself.  Being relatively new,
 412          * the odds currently favor a bug, so for now the line below isn't
 413          * entirely tasteles.
 414          */
 415         BUG_ON(ret);
 416 }
 417
 418 static void inode_write_block(struct logfs_block *block)
 419 {
 420         struct inode *inode;
 421         int ret;
 422
 423         inode = block->inode;
 424         if (inode->i_ino == LOGFS_INO_MASTER)
 425                 logfs_write_anchor(inode->i_sb);
 426         else {
 427                 ret = __logfs_write_inode(inode, 0);
 428                 /* see indirect_write_block comment */
 429                 BUG_ON(ret);
 430         }
 431 }
 432
 433 /*
 434  * This silences a false, yet annoying gcc warning.  I hate it when my editor
 435  * jumps into bitops.h each time I recompile this file.
 436  * TODO: Complain to gcc folks about this and upgrade compiler.
 437  */
 438 static unsigned long fnb(const unsigned long *addr,
 439                 unsigned long size, unsigned long offset)
 440 {
 441         return find_next_bit(addr, size, offset);
 442 }
 443
 444 static __be64 inode_val0(struct inode *inode)
 445 {
 446         struct logfs_inode *li = logfs_inode(inode);
 447         u64 val;
 448
 449         /*
 450          * Explicit shifting generates good code, but must match the format
 451          * of the structure.  Add some paranoia just in case.
 452          */
 453         BUILD_BUG_ON(offsetof(struct logfs_disk_inode, di_mode) != 0);
 454         BUILD_BUG_ON(offsetof(struct logfs_disk_inode, di_height) != 2);
 455         BUILD_BUG_ON(offsetof(struct logfs_disk_inode, di_flags) != 4);
 456
 457         val =   (u64)inode->i_mode << 48 |
 458                 (u64)li->li_height << 40 |
 459                 (u64)li->li_flags;
 460         return cpu_to_be64(val);
 461 }
 462
 463 static int inode_write_alias(struct super_block *sb,
 464                 struct logfs_block *block, write_alias_t *write_one_alias)
 465 {
 466         struct inode *inode = block->inode;
 467         struct logfs_inode *li = logfs_inode(inode);
 468         unsigned long pos;
 469         u64 ino , bix;
 470         __be64 val;
 471         level_t level;
 472         int err;
 473
 474         for (pos = 0; ; pos++) {
 475                 pos = fnb(block->alias_map, LOGFS_BLOCK_FACTOR, pos);
 476                 if (pos >= LOGFS_EMBEDDED_FIELDS + INODE_POINTER_OFS)
 477                         return 0;
 478
 479                 switch (pos) {
 480                 case INODE_HEIGHT_OFS:
 481                         val = inode_val0(inode);
 482                         break;
 483                 case INODE_USED_OFS:
 484                         val = cpu_to_be64(li->li_used_bytes);
 485                         break;
 486                 case INODE_SIZE_OFS:
 487                         val = cpu_to_be64(i_size_read(inode));
 488                         break;
 489                 case INODE_POINTER_OFS ... INODE_POINTER_OFS + LOGFS_EMBEDDED_FIELDS - 1:
 490                         val = cpu_to_be64(li->li_data[pos - INODE_POINTER_OFS]);
 491                         break;
 492                 default:
 493                         BUG();
 494                 }
 495
 496                 ino = LOGFS_INO_MASTER;
 497                 bix = inode->i_ino;
 498                 level = LEVEL(0);
 499                 err = write_one_alias(sb, ino, bix, level, pos, val);
 500                 if (err)
 501                         return err;
 502         }
 503 }
 504
 505 static int indirect_write_alias(struct super_block *sb,
 506                 struct logfs_block *block, write_alias_t *write_one_alias)
 507 {
 508         unsigned long pos;
 509         struct page *page = block->page;
 510         u64 ino , bix;
 511         __be64 *child, val;
 512         level_t level;
 513         int err;
 514
 515         for (pos = 0; ; pos++) {
 516                 pos = fnb(block->alias_map, LOGFS_BLOCK_FACTOR, pos);
 517                 if (pos >= LOGFS_BLOCK_FACTOR)
 518                         return 0;
 519
 520                 ino = page->mapping->host->i_ino;
 521                 logfs_unpack_index(page->index, &bix, &level);
 522                 child = kmap_atomic(page, KM_USER0);
 523                 val = child[pos];
 524                 kunmap_atomic(child, KM_USER0);
 525                 err = write_one_alias(sb, ino, bix, level, pos, val);
 526                 if (err)
 527                         return err;
 528         }
 529 }
 530
 531 int logfs_write_obj_aliases_pagecache(struct super_block *sb)
 532 {
 533         struct logfs_super *super = logfs_super(sb);
 534         struct logfs_block *block;
 535         int err;
 536
 537         list_for_each_entry(block, &super->s_object_alias, alias_list) {
 538                 err = block->ops->write_alias(sb, block, write_alias_journal);
 539                 if (err)
 540                         return err;
 541         }
 542         return 0;
 543 }
 544
 545 void __free_block(struct super_block *sb, struct logfs_block *block)
 546 {
 547         BUG_ON(!list_empty(&block->item_list));
 548         list_del(&block->alias_list);
 549         mempool_free(block, logfs_super(sb)->s_block_pool);
 550 }
 551
 552 static void inode_free_block(struct super_block *sb, struct logfs_block *block)
 553 {
 554         struct inode *inode = block->inode;
 555
 556         logfs_inode(inode)->li_block = NULL;
 557         __free_block(sb, block);
 558 }
 559
 560 static void indirect_free_block(struct super_block *sb,
 561                 struct logfs_block *block)
 562 {
 563         ClearPagePrivate(block->page);
 564         block->page->private = 0;
 565         __free_block(sb, block);
 566 }
 567
 568
 569 static struct logfs_block_ops inode_block_ops = {
 570         .write_block = inode_write_block,
 571         .free_block = inode_free_block,
 572         .write_alias = inode_write_alias,
 573 };
 574
 575 struct logfs_block_ops indirect_block_ops = {
 576         .write_block = indirect_write_block,
 577         .free_block = indirect_free_block,
 578         .write_alias = indirect_write_alias,
 579 };
 580
 581 struct logfs_block *__alloc_block(struct super_block *sb,
 582                 u64 ino, u64 bix, level_t level)
 583 {
 584         struct logfs_super *super = logfs_super(sb);
 585         struct logfs_block *block;
 586
 587         block = mempool_alloc(super->s_block_pool, GFP_NOFS);
 588         memset(block, 0, sizeof(*block));
 589         INIT_LIST_HEAD(&block->alias_list);
 590         INIT_LIST_HEAD(&block->item_list);
 591         block->sb = sb;
 592         block->ino = ino;
 593         block->bix = bix;
 594         block->level = level;
 595         return block;
 596 }
 597
 598 static void alloc_inode_block(struct inode *inode)
 599 {
 600         struct logfs_inode *li = logfs_inode(inode);
 601         struct logfs_block *block;
 602
 603         if (li->li_block)
 604                 return;
 605
 606         block = __alloc_block(inode->i_sb, LOGFS_INO_MASTER, inode->i_ino, 0);
 607         block->inode = inode;
 608         li->li_block = block;
 609         block->ops = &inode_block_ops;
 610 }
 611
 612 void initialize_block_counters(struct page *page, struct logfs_block *block,
 613                 __be64 *array, int page_is_empty)
 614 {
 615         u64 ptr;
 616         int i, start;
 617
 618         block->partial = 0;
 619         block->full = 0;
 620         start = 0;
 621         if (page->index < first_indirect_block()) {
 622                 /* Counters are pointless on level 0 */
 623                 return;
 624         }
 625         if (page->index == first_indirect_block()) {
 626                 /* Skip unused pointers */
 627                 start = I0_BLOCKS;
 628                 block->full = I0_BLOCKS;
 629         }
 630         if (!page_is_empty) {
 631                 for (i = start; i < LOGFS_BLOCK_FACTOR; i++) {
 632                         ptr = be64_to_cpu(array[i]);
 633                         if (ptr)
 634                                 block->partial++;
 635                         if (ptr & LOGFS_FULLY_POPULATED)
 636                                 block->full++;
 637                 }
 638         }
 639 }
 640
 641 static void alloc_data_block(struct inode *inode, struct page *page)
 642 {
 643         struct logfs_block *block;
 644         u64 bix;
 645         level_t level;
 646
 647         if (PagePrivate(page))
 648                 return;
 649
 650         logfs_unpack_index(page->index, &bix, &level);
 651         block = __alloc_block(inode->i_sb, inode->i_ino, bix, level);
 652         block->page = page;
 653         SetPagePrivate(page);
 654         page->private = (unsigned long)block;
 655         block->ops = &indirect_block_ops;
 656 }
 657
 658 static void alloc_indirect_block(struct inode *inode, struct page *page,
 659                 int page_is_empty)
 660 {
 661         struct logfs_block *block;
 662         __be64 *array;
 663
 664         if (PagePrivate(page))
 665                 return;
 666
 667         alloc_data_block(inode, page);
 668
 669         block = logfs_block(page);
 670         array = kmap_atomic(page, KM_USER0);
 671         initialize_block_counters(page, block, array, page_is_empty);
 672         kunmap_atomic(array, KM_USER0);
 673 }
 674
 675 static void block_set_pointer(struct page *page, int index, u64 ptr)
 676 {
 677         struct logfs_block *block = logfs_block(page);
 678         __be64 *array;
 679         u64 oldptr;
 680
 681         BUG_ON(!block);
 682         array = kmap_atomic(page, KM_USER0);
 683         oldptr = be64_to_cpu(array[index]);
 684         array[index] = cpu_to_be64(ptr);
 685         kunmap_atomic(array, KM_USER0);
 686         SetPageUptodate(page);
 687
 688         block->full += !!(ptr & LOGFS_FULLY_POPULATED)
 689                 - !!(oldptr & LOGFS_FULLY_POPULATED);
 690         block->partial += !!ptr - !!oldptr;
 691 }
 692
 693 static u64 block_get_pointer(struct page *page, int index)
 694 {
 695         __be64 *block;
 696         u64 ptr;
 697
 698         block = kmap_atomic(page, KM_USER0);
 699         ptr = be64_to_cpu(block[index]);
 700         kunmap_atomic(block, KM_USER0);
 701         return ptr;
 702 }
 703
 704 static int logfs_read_empty(struct page *page)
 705 {
 706         zero_user_segment(page, 0, PAGE_CACHE_SIZE);
 707         return 0;
 708 }
 709
 710 static int logfs_read_direct(struct inode *inode, struct page *page)
 711 {
 712         struct logfs_inode *li = logfs_inode(inode);
 713         pgoff_t index = page->index;
 714         u64 block;
 715
 716         block = li->li_data[index];
 717         if (!block)
 718                 return logfs_read_empty(page);
 719
 720         return logfs_segment_read(inode, page, block, index, 0);
 721 }
 722
 723 static int logfs_read_loop(struct inode *inode, struct page *page,
 724                 int rw_context)
 725 {
 726         struct logfs_inode *li = logfs_inode(inode);
 727         u64 bix, bofs = li->li_data[INDIRECT_INDEX];
 728         level_t level, target_level;
 729         int ret;
 730         struct page *ipage;
 731
 732         logfs_unpack_index(page->index, &bix, &target_level);
 733         if (!bofs)
 734                 return logfs_read_empty(page);
 735
 736         if (bix >= maxbix(li->li_height))
 737                 return logfs_read_empty(page);
 738
 739         for (level = LEVEL(li->li_height);
 740                         (__force u8)level > (__force u8)target_level;
 741                         level = SUBLEVEL(level)){
 742                 ipage = logfs_get_page(inode, bix, level, rw_context);
 743                 if (!ipage)
 744                         return -ENOMEM;
 745
 746                 ret = logfs_segment_read(inode, ipage, bofs, bix, level);
 747                 if (ret) {
 748                         logfs_put_read_page(ipage);
 749                         return ret;
 750                 }
 751
 752                 bofs = block_get_pointer(ipage, get_bits(bix, SUBLEVEL(level)));
 753                 logfs_put_page(ipage, rw_context);
 754                 if (!bofs)
 755                         return logfs_read_empty(page);
 756         }
 757
 758         return logfs_segment_read(inode, page, bofs, bix, 0);
 759 }
 760
 761 static int logfs_read_block(struct inode *inode, struct page *page,
 762                 int rw_context)
 763 {
 764         pgoff_t index = page->index;
 765
 766         if (index < I0_BLOCKS)
 767                 return logfs_read_direct(inode, page);
 768         return logfs_read_loop(inode, page, rw_context);
 769 }
 770
 771 static int logfs_exist_loop(struct inode *inode, u64 bix)
 772 {
 773         struct logfs_inode *li = logfs_inode(inode);
 774         u64 bofs = li->li_data[INDIRECT_INDEX];
 775         level_t level;
 776         int ret;
 777         struct page *ipage;
 778
 779         if (!bofs)
 780                 return 0;
 781         if (bix >= maxbix(li->li_height))
 782                 return 0;
 783
 784         for (level = LEVEL(li->li_height); level != 0; level = SUBLEVEL(level)) {
 785                 ipage = logfs_get_read_page(inode, bix, level);
 786                 if (!ipage)
 787                         return -ENOMEM;
 788
 789                 ret = logfs_segment_read(inode, ipage, bofs, bix, level);
 790                 if (ret) {
 791                         logfs_put_read_page(ipage);
 792                         return ret;
 793                 }
 794
 795                 bofs = block_get_pointer(ipage, get_bits(bix, SUBLEVEL(level)));
 796                 logfs_put_read_page(ipage);
 797                 if (!bofs)
 798                         return 0;
 799         }
 800
 801         return 1;
 802 }
 803
 804 int logfs_exist_block(struct inode *inode, u64 bix)
 805 {
 806         struct logfs_inode *li = logfs_inode(inode);
 807
 808         if (bix < I0_BLOCKS)
 809                 return !!li->li_data[bix];
 810         return logfs_exist_loop(inode, bix);
 811 }
 812
 813 static u64 seek_holedata_direct(struct inode *inode, u64 bix, int data)
 814 {
 815         struct logfs_inode *li = logfs_inode(inode);
 816
 817         for (; bix < I0_BLOCKS; bix++)
 818                 if (data ^ (li->li_data[bix] == 0))
 819                         return bix;
 820         return I0_BLOCKS;
 821 }
 822
 823 static u64 seek_holedata_loop(struct inode *inode, u64 bix, int data)
 824 {
 825         struct logfs_inode *li = logfs_inode(inode);
 826         __be64 *rblock;
 827         u64 increment, bofs = li->li_data[INDIRECT_INDEX];
 828         level_t level;
 829         int ret, slot;
 830         struct page *page;
 831
 832         BUG_ON(!bofs);
 833
 834         for (level = LEVEL(li->li_height); level != 0; level = SUBLEVEL(level)) {
 835                 increment = 1 << (LOGFS_BLOCK_BITS * ((__force u8)level-1));
 836                 page = logfs_get_read_page(inode, bix, level);
 837                 if (!page)
 838                         return bix;
 839
 840                 ret = logfs_segment_read(inode, page, bofs, bix, level);
 841                 if (ret) {
 842                         logfs_put_read_page(page);
 843                         return bix;
 844                 }
 845
 846                 slot = get_bits(bix, SUBLEVEL(level));
 847                 rblock = kmap_atomic(page, KM_USER0);
 848                 while (slot < LOGFS_BLOCK_FACTOR) {
 849                         if (data && (rblock[slot] != 0))
 850                                 break;
 851                         if (!data && !(be64_to_cpu(rblock[slot]) & LOGFS_FULLY_POPULATED))
 852                                 break;
 853                         slot++;
 854                         bix += increment;
 855                         bix &= ~(increment - 1);
 856                 }
 857                 if (slot >= LOGFS_BLOCK_FACTOR) {
 858                         kunmap_atomic(rblock, KM_USER0);
 859                         logfs_put_read_page(page);
 860                         return bix;
 861                 }
 862                 bofs = be64_to_cpu(rblock[slot]);
 863                 kunmap_atomic(rblock, KM_USER0);
 864                 logfs_put_read_page(page);
 865                 if (!bofs) {
 866                         BUG_ON(data);
 867                         return bix;
 868                 }
 869         }
 870         return bix;
 871 }
 872
 873 /**
 874  * logfs_seek_hole - find next hole starting at a given block index
 875  * @inode:              inode to search in
 876  * @bix:                block index to start searching
 877  *
 878  * Returns next hole.  If the file doesn't contain any further holes, the
 879  * block address next to eof is returned instead.
 880  */
 881 u64 logfs_seek_hole(struct inode *inode, u64 bix)
 882 {
 883         struct logfs_inode *li = logfs_inode(inode);
 884
 885         if (bix < I0_BLOCKS) {
 886                 bix = seek_holedata_direct(inode, bix, 0);
 887                 if (bix < I0_BLOCKS)
 888                         return bix;
 889         }
 890
 891         if (!li->li_data[INDIRECT_INDEX])
 892                 return bix;
 893         else if (li->li_data[INDIRECT_INDEX] & LOGFS_FULLY_POPULATED)
 894                 bix = maxbix(li->li_height);
 895         else if (bix >= maxbix(li->li_height))
 896                 return bix;
 897         else {
 898                 bix = seek_holedata_loop(inode, bix, 0);
 899                 if (bix < maxbix(li->li_height))
 900                         return bix;
 901                 /* Should not happen anymore.  But if some port writes semi-
 902                  * corrupt images (as this one used to) we might run into it.
 903                  */
 904                 WARN_ON_ONCE(bix == maxbix(li->li_height));
 905         }
 906
 907         return bix;
 908 }
 909
 910 static u64 __logfs_seek_data(struct inode *inode, u64 bix)
 911 {
 912         struct logfs_inode *li = logfs_inode(inode);
 913
 914         if (bix < I0_BLOCKS) {
 915                 bix = seek_holedata_direct(inode, bix, 1);
 916                 if (bix < I0_BLOCKS)
 917                         return bix;
 918         }
 919
 920         if (bix < maxbix(li->li_height)) {
 921                 if (!li->li_data[INDIRECT_INDEX])
 922                         bix = maxbix(li->li_height);
 923                 else
 924                         return seek_holedata_loop(inode, bix, 1);
 925         }
 926
 927         return bix;
 928 }
 929
 930 /**
 931  * logfs_seek_data - find next data block after a given block index
 932  * @inode:              inode to search in
 933  * @bix:                block index to start searching
 934  *
 935  * Returns next data block.  If the file doesn't contain any further data
 936  * blocks, the last block in the file is returned instead.
 937  */
 938 u64 logfs_seek_data(struct inode *inode, u64 bix)
 939 {
 940         struct super_block *sb = inode->i_sb;
 941         u64 ret, end;
 942
 943         ret = __logfs_seek_data(inode, bix);
 944         end = i_size_read(inode) >> sb->s_blocksize_bits;
 945         if (ret >= end)
 946                 ret = max(bix, end);
 947         return ret;
 948 }
 949
 950 static int logfs_is_valid_direct(struct logfs_inode *li, u64 bix, u64 ofs)
 951 {
 952         return pure_ofs(li->li_data[bix]) == ofs;
 953 }
 954
 955 static int __logfs_is_valid_loop(struct inode *inode, u64 bix,
 956                 u64 ofs, u64 bofs)
 957 {
 958         struct logfs_inode *li = logfs_inode(inode);
 959         level_t level;
 960         int ret;
 961         struct page *page;
 962
 963         for (level = LEVEL(li->li_height); level != 0; level = SUBLEVEL(level)){
 964                 page = logfs_get_write_page(inode, bix, level);
 965                 BUG_ON(!page);
 966
 967                 ret = logfs_segment_read(inode, page, bofs, bix, level);
 968                 if (ret) {
 969                         logfs_put_write_page(page);
 970                         return 0;
 971                 }
 972
 973                 bofs = block_get_pointer(page, get_bits(bix, SUBLEVEL(level)));
 974                 logfs_put_write_page(page);
 975                 if (!bofs)
 976                         return 0;
 977
 978                 if (pure_ofs(bofs) == ofs)
 979                         return 1;
 980         }
 981         return 0;
 982 }
 983
 984 static int logfs_is_valid_loop(struct inode *inode, u64 bix, u64 ofs)
 985 {
 986         struct logfs_inode *li = logfs_inode(inode);
 987         u64 bofs = li->li_data[INDIRECT_INDEX];
 988
 989         if (!bofs)
 990                 return 0;
 991
 992         if (bix >= maxbix(li->li_height))
 993                 return 0;
 994
 995         if (pure_ofs(bofs) == ofs)
 996                 return 1;
 997
 998         return __logfs_is_valid_loop(inode, bix, ofs, bofs);
 999 }
1000
1001 static int __logfs_is_valid_block(struct inode *inode, u64 bix, u64 ofs)
1002 {
1003         struct logfs_inode *li = logfs_inode(inode);
1004
1005         if ((inode->i_nlink == 0) && atomic_read(&inode->i_count) == 1)
1006                 return 0;
1007
1008         if (bix < I0_BLOCKS)
1009                 return logfs_is_valid_direct(li, bix, ofs);
1010         return logfs_is_valid_loop(inode, bix, ofs);
1011 }
1012
1013 /**
1014  * logfs_is_valid_block - check whether this block is still valid
1015  *
1016  * @sb  - superblock
1017  * @ofs - block physical offset
1018  * @ino - block inode number
1019  * @bix - block index
1020  * @level - block level
1021  *
1022  * Returns 0 if the block is invalid, 1 if it is valid and 2 if it will
1023  * become invalid once the journal is written.
1024  */
1025 int logfs_is_valid_block(struct super_block *sb, u64 ofs, u64 ino, u64 bix,
1026                 gc_level_t gc_level)
1027 {
1028         struct logfs_super *super = logfs_super(sb);
1029         struct inode *inode;
1030         int ret, cookie;
1031
1032         /* Umount closes a segment with free blocks remaining.  Those
1033          * blocks are by definition invalid. */
1034         if (ino == -1)
1035                 return 0;
1036
1037         LOGFS_BUG_ON((u64)(u_long)ino != ino, sb);
1038
1039         inode = logfs_safe_iget(sb, ino, &cookie);
1040         if (IS_ERR(inode))
1041                 goto invalid;
1042
1043         ret = __logfs_is_valid_block(inode, bix, ofs);
1044         logfs_safe_iput(inode, cookie);
1045         if (ret)
1046                 return ret;
1047
1048 invalid:
1049         /* Block is nominally invalid, but may still sit in the shadow tree,
1050          * waiting for a journal commit.
1051          */
1052         if (btree_lookup64(&super->s_shadow_tree.old, ofs))
1053                 return 2;
1054         return 0;
1055 }
1056
1057 int logfs_readpage_nolock(struct page *page)
1058 {
1059         struct inode *inode = page->mapping->host;
1060         int ret = -EIO;
1061
1062         ret = logfs_read_block(inode, page, READ);
1063
1064         if (ret) {
1065                 ClearPageUptodate(page);
1066                 SetPageError(page);
1067         } else {
1068                 SetPageUptodate(page);
1069                 ClearPageError(page);
1070         }
1071         flush_dcache_page(page);
1072
1073         return ret;
1074 }
1075
1076 static int logfs_reserve_bytes(struct inode *inode, int bytes)
1077 {
1078         struct logfs_super *super = logfs_super(inode->i_sb);
1079         u64 available = super->s_free_bytes + super->s_dirty_free_bytes
1080                         - super->s_dirty_used_bytes - super->s_dirty_pages;
1081
1082         if (!bytes)
1083                 return 0;
1084
1085         if (available < bytes)
1086                 return -ENOSPC;
1087
1088         if (available < bytes + super->s_root_reserve &&
1089                         !capable(CAP_SYS_RESOURCE))
1090                 return -ENOSPC;
1091
1092         return 0;
1093 }
1094
1095 int get_page_reserve(struct inode *inode, struct page *page)
1096 {
1097         struct logfs_super *super = logfs_super(inode->i_sb);
1098         struct logfs_block *block = logfs_block(page);
1099         int ret;
1100
1101         if (block && block->reserved_bytes)
1102                 return 0;
1103
1104         logfs_get_wblocks(inode->i_sb, page, WF_LOCK);
1105         while ((ret = logfs_reserve_bytes(inode, 6 * LOGFS_MAX_OBJECTSIZE)) &&
1106                         !list_empty(&super->s_writeback_list)) {
1107                 block = list_entry(super->s_writeback_list.next,
1108                                 struct logfs_block, alias_list);
1109                 block->ops->write_block(block);
1110         }
1111         if (!ret) {
1112                 alloc_data_block(inode, page);
1113                 block = logfs_block(page);
1114                 block->reserved_bytes += 6 * LOGFS_MAX_OBJECTSIZE;
1115                 super->s_dirty_pages += 6 * LOGFS_MAX_OBJECTSIZE;
1116                 list_move_tail(&block->alias_list, &super->s_writeback_list);
1117         }
1118         logfs_put_wblocks(inode->i_sb, page, WF_LOCK);
1119         return ret;
1120 }
1121
1122 /*
1123  * We are protected by write lock.  Push victims up to superblock level
1124  * and release transaction when appropriate.
1125  */
1126 /* FIXME: This is currently called from the wrong spots. */
1127 static void logfs_handle_transaction(struct inode *inode,
1128                 struct logfs_transaction *ta)
1129 {
1130         struct logfs_super *super = logfs_super(inode->i_sb);
1131
1132         if (!ta)
1133                 return;
1134         logfs_inode(inode)->li_block->ta = NULL;
1135
1136         if (inode->i_ino != LOGFS_INO_MASTER) {
1137                 BUG(); /* FIXME: Yes, this needs more thought */
1138                 /* just remember the transaction until inode is written */
1139                 //BUG_ON(logfs_inode(inode)->li_transaction);
1140                 //logfs_inode(inode)->li_transaction = ta;
1141                 return;
1142         }
1143
1144         switch (ta->state) {
1145         case CREATE_1: /* fall through */
1146         case UNLINK_1:
1147                 BUG_ON(super->s_victim_ino);
1148                 super->s_victim_ino = ta->ino;
1149                 break;
1150         case CREATE_2: /* fall through */
1151         case UNLINK_2:
1152                 BUG_ON(super->s_victim_ino != ta->ino);
1153                 super->s_victim_ino = 0;
1154                 /* transaction ends here - free it */
1155                 kfree(ta);
1156                 break;
1157         case CROSS_RENAME_1:
1158                 BUG_ON(super->s_rename_dir);
1159                 BUG_ON(super->s_rename_pos);
1160                 super->s_rename_dir = ta->dir;
1161                 super->s_rename_pos = ta->pos;
1162                 break;
1163         case CROSS_RENAME_2:
1164                 BUG_ON(super->s_rename_dir != ta->dir);
1165                 BUG_ON(super->s_rename_pos != ta->pos);
1166                 super->s_rename_dir = 0;
1167                 super->s_rename_pos = 0;
1168                 kfree(ta);
1169                 break;
1170         case TARGET_RENAME_1:
1171                 BUG_ON(super->s_rename_dir);
1172                 BUG_ON(super->s_rename_pos);
1173                 BUG_ON(super->s_victim_ino);
1174                 super->s_rename_dir = ta->dir;
1175                 super->s_rename_pos = ta->pos;
1176                 super->s_victim_ino = ta->ino;
1177                 break;
1178         case TARGET_RENAME_2:
1179                 BUG_ON(super->s_rename_dir != ta->dir);
1180                 BUG_ON(super->s_rename_pos != ta->pos);
1181                 BUG_ON(super->s_victim_ino != ta->ino);
1182                 super->s_rename_dir = 0;
1183                 super->s_rename_pos = 0;
1184                 break;
1185         case TARGET_RENAME_3:
1186                 BUG_ON(super->s_rename_dir);
1187                 BUG_ON(super->s_rename_pos);
1188                 BUG_ON(super->s_victim_ino != ta->ino);
1189                 super->s_victim_ino = 0;
1190                 kfree(ta);
1191                 break;
1192         default:
1193                 BUG();
1194         }
1195 }
1196
1197 /*
1198  * Not strictly a reservation, but rather a check that we still have enough
1199  * space to satisfy the write.
1200  */
1201 static int logfs_reserve_blocks(struct inode *inode, int blocks)
1202 {
1203         return logfs_reserve_bytes(inode, blocks * LOGFS_MAX_OBJECTSIZE);
1204 }
1205
1206 struct write_control {
1207         u64 ofs;
1208         long flags;
1209 };
1210
1211 static struct logfs_shadow *alloc_shadow(struct inode *inode, u64 bix,
1212                 level_t level, u64 old_ofs)
1213 {
1214         struct logfs_super *super = logfs_super(inode->i_sb);
1215         struct logfs_shadow *shadow;
1216
1217         shadow = mempool_alloc(super->s_shadow_pool, GFP_NOFS);
1218         memset(shadow, 0, sizeof(*shadow));
1219         shadow->ino = inode->i_ino;
1220         shadow->bix = bix;
1221         shadow->gc_level = expand_level(inode->i_ino, level);
1222         shadow->old_ofs = old_ofs & ~LOGFS_FULLY_POPULATED;
1223         return shadow;
1224 }
1225
1226 static void free_shadow(struct inode *inode, struct logfs_shadow *shadow)
1227 {
1228         struct logfs_super *super = logfs_super(inode->i_sb);
1229
1230         mempool_free(shadow, super->s_shadow_pool);
1231 }
1232
1233 static void mark_segment(struct shadow_tree *tree, u32 segno)
1234 {
1235         int err;
1236
1237         if (!btree_lookup32(&tree->segment_map, segno)) {
1238                 err = btree_insert32(&tree->segment_map, segno, (void *)1,
1239                                 GFP_NOFS);
1240                 BUG_ON(err);
1241                 tree->no_shadowed_segments++;
1242         }
1243 }
1244
1245 /**
1246  * fill_shadow_tree - Propagate shadow tree changes due to a write
1247  * @inode:      Inode owning the page
1248  * @page:       Struct page that was written
1249  * @shadow:     Shadow for the current write
1250  *
1251  * Writes in logfs can result in two semi-valid objects.  The old object
1252  * is still valid as long as it can be reached by following pointers on
1253  * the medium.  Only when writes propagate all the way up to the journal
1254  * has the new object safely replaced the old one.
1255  *
1256  * To handle this problem, a struct logfs_shadow is used to represent
1257  * every single write.  It is attached to the indirect block, which is
1258  * marked dirty.  When the indirect block is written, its shadows are
1259  * handed up to the next indirect block (or inode).  Untimately they
1260  * will reach the master inode and be freed upon journal commit.
1261  *
1262  * This function handles a single step in the propagation.  It adds the
1263  * shadow for the current write to the tree, along with any shadows in
1264  * the page's tree, in case it was an indirect block.  If a page is
1265  * written, the inode parameter is left NULL, if an inode is written,
1266  * the page parameter is left NULL.
1267  */
1268 static void fill_shadow_tree(struct inode *inode, struct page *page,
1269                 struct logfs_shadow *shadow)
1270 {
1271         struct logfs_super *super = logfs_super(inode->i_sb);
1272         struct logfs_block *block = logfs_block(page);
1273         struct shadow_tree *tree = &super->s_shadow_tree;
1274
1275         if (PagePrivate(page)) {
1276                 if (block->alias_map)
1277                         super->s_no_object_aliases -= bitmap_weight(
1278                                         block->alias_map, LOGFS_BLOCK_FACTOR);
1279                 logfs_handle_transaction(inode, block->ta);
1280                 block->ops->free_block(inode->i_sb, block);
1281         }
1282         if (shadow) {
1283                 if (shadow->old_ofs)
1284                         btree_insert64(&tree->old, shadow->old_ofs, shadow,
1285                                         GFP_NOFS);
1286                 else
1287                         btree_insert64(&tree->new, shadow->new_ofs, shadow,
1288                                         GFP_NOFS);
1289
1290                 super->s_dirty_used_bytes += shadow->new_len;
1291                 super->s_dirty_free_bytes += shadow->old_len;
1292                 mark_segment(tree, shadow->old_ofs >> super->s_segshift);
1293                 mark_segment(tree, shadow->new_ofs >> super->s_segshift);
1294         }
1295 }
1296
1297 static void logfs_set_alias(struct super_block *sb, struct logfs_block *block,
1298                 long child_no)
1299 {
1300         struct logfs_super *super = logfs_super(sb);
1301
1302         if (block->inode && block->inode->i_ino == LOGFS_INO_MASTER) {
1303                 /* Aliases in the master inode are pointless. */
1304                 return;
1305         }
1306
1307         if (!test_bit(child_no, block->alias_map)) {
1308                 set_bit(child_no, block->alias_map);
1309                 super->s_no_object_aliases++;
1310         }
1311         list_move_tail(&block->alias_list, &super->s_object_alias);
1312 }
1313
1314 /*
1315  * Object aliases can and often do change the size and occupied space of a
1316  * file.  So not only do we have to change the pointers, we also have to
1317  * change inode->i_size and li->li_used_bytes.  Which is done by setting
1318  * another two object aliases for the inode itself.
1319  */
1320 static void set_iused(struct inode *inode, struct logfs_shadow *shadow)
1321 {
1322         struct logfs_inode *li = logfs_inode(inode);
1323
1324         if (shadow->new_len == shadow->old_len)
1325                 return;
1326
1327         alloc_inode_block(inode);
1328         li->li_used_bytes += shadow->new_len - shadow->old_len;
1329         __logfs_set_blocks(inode);
1330         logfs_set_alias(inode->i_sb, li->li_block, INODE_USED_OFS);
1331         logfs_set_alias(inode->i_sb, li->li_block, INODE_SIZE_OFS);
1332 }
1333
1334 static int logfs_write_i0(struct inode *inode, struct page *page,
1335                 struct write_control *wc)
1336 {
1337         struct logfs_shadow *shadow;
1338         u64 bix;
1339         level_t level;
1340         int full, err = 0;
1341
1342         logfs_unpack_index(page->index, &bix, &level);
1343         if (wc->ofs == 0)
1344                 if (logfs_reserve_blocks(inode, 1))
1345                         return -ENOSPC;
1346
1347         shadow = alloc_shadow(inode, bix, level, wc->ofs);
1348         if (wc->flags & WF_WRITE)
1349                 err = logfs_segment_write(inode, page, shadow);
1350         if (wc->flags & WF_DELETE)
1351                 logfs_segment_delete(inode, shadow);
1352         if (err) {
1353                 free_shadow(inode, shadow);
1354                 return err;
1355         }
1356
1357         set_iused(inode, shadow);
1358         full = 1;
1359         if (level != 0) {
1360                 alloc_indirect_block(inode, page, 0);
1361                 full = logfs_block(page)->full == LOGFS_BLOCK_FACTOR;
1362         }
1363         fill_shadow_tree(inode, page, shadow);
1364         wc->ofs = shadow->new_ofs;
1365         if (wc->ofs && full)
1366                 wc->ofs |= LOGFS_FULLY_POPULATED;
1367         return 0;
1368 }
1369
1370 static int logfs_write_direct(struct inode *inode, struct page *page,
1371                 long flags)
1372 {
1373         struct logfs_inode *li = logfs_inode(inode);
1374         struct write_control wc = {
1375                 .ofs = li->li_data[page->index],
1376                 .flags = flags,
1377         };
1378         int err;
1379
1380         alloc_inode_block(inode);
1381
1382         err = logfs_write_i0(inode, page, &wc);
1383         if (err)
1384                 return err;
1385
1386         li->li_data[page->index] = wc.ofs;
1387         logfs_set_alias(inode->i_sb, li->li_block,
1388                         page->index + INODE_POINTER_OFS);
1389         return 0;
1390 }
1391
1392 static int ptr_change(u64 ofs, struct page *page)
1393 {
1394         struct logfs_block *block = logfs_block(page);
1395         int empty0, empty1, full0, full1;
1396
1397         empty0 = ofs == 0;
1398         empty1 = block->partial == 0;
1399         if (empty0 != empty1)
1400                 return 1;
1401
1402         /* The !! is necessary to shrink result to int */
1403         full0 = !!(ofs & LOGFS_FULLY_POPULATED);
1404         full1 = block->full == LOGFS_BLOCK_FACTOR;
1405         if (full0 != full1)
1406                 return 1;
1407         return 0;
1408 }
1409
1410 static int __logfs_write_rec(struct inode *inode, struct page *page,
1411                 struct write_control *this_wc,
1412                 pgoff_t bix, level_t target_level, level_t level)
1413 {
1414         int ret, page_empty = 0;
1415         int child_no = get_bits(bix, SUBLEVEL(level));
1416         struct page *ipage;
1417         struct write_control child_wc = {
1418                 .flags = this_wc->flags,
1419         };
1420
1421         ipage = logfs_get_write_page(inode, bix, level);
1422         if (!ipage)
1423                 return -ENOMEM;
1424
1425         if (this_wc->ofs) {
1426                 ret = logfs_segment_read(inode, ipage, this_wc->ofs, bix, level);
1427                 if (ret)
1428                         goto out;
1429         } else if (!PageUptodate(ipage)) {
1430                 page_empty = 1;
1431                 logfs_read_empty(ipage);
1432         }
1433
1434         child_wc.ofs = block_get_pointer(ipage, child_no);
1435
1436         if ((__force u8)level-1 > (__force u8)target_level)
1437                 ret = __logfs_write_rec(inode, page, &child_wc, bix,
1438                                 target_level, SUBLEVEL(level));
1439         else
1440                 ret = logfs_write_i0(inode, page, &child_wc);
1441
1442         if (ret)
1443                 goto out;
1444
1445         alloc_indirect_block(inode, ipage, page_empty);
1446         block_set_pointer(ipage, child_no, child_wc.ofs);
1447         /* FIXME: first condition seems superfluous */
1448         if (child_wc.ofs || logfs_block(ipage)->partial)
1449                 this_wc->flags |= WF_WRITE;
1450         /* the condition on this_wc->ofs ensures that we won't consume extra
1451          * space for indirect blocks in the future, which we cannot reserve */
1452         if (!this_wc->ofs || ptr_change(this_wc->ofs, ipage))
1453                 ret = logfs_write_i0(inode, ipage, this_wc);
1454         else
1455                 logfs_set_alias(inode->i_sb, logfs_block(ipage), child_no);
1456 out:
1457         logfs_put_write_page(ipage);
1458         return ret;
1459 }
1460
1461 static int logfs_write_rec(struct inode *inode, struct page *page,
1462                 pgoff_t bix, level_t target_level, long flags)
1463 {
1464         struct logfs_inode *li = logfs_inode(inode);
1465         struct write_control wc = {
1466                 .ofs = li->li_data[INDIRECT_INDEX],
1467                 .flags = flags,
1468         };
1469         int ret;
1470
1471         alloc_inode_block(inode);
1472
1473         if (li->li_height > (__force u8)target_level)
1474                 ret = __logfs_write_rec(inode, page, &wc, bix, target_level,
1475                                 LEVEL(li->li_height));
1476         else
1477                 ret = logfs_write_i0(inode, page, &wc);
1478         if (ret)
1479                 return ret;
1480
1481         if (li->li_data[INDIRECT_INDEX] != wc.ofs) {
1482                 li->li_data[INDIRECT_INDEX] = wc.ofs;
1483                 logfs_set_alias(inode->i_sb, li->li_block,
1484                                 INDIRECT_INDEX + INODE_POINTER_OFS);
1485         }
1486         return ret;
1487 }
1488
1489 void logfs_add_transaction(struct inode *inode, struct logfs_transaction *ta)
1490 {
1491         alloc_inode_block(inode);
1492         logfs_inode(inode)->li_block->ta = ta;
1493 }
1494
1495 void logfs_del_transaction(struct inode *inode, struct logfs_transaction *ta)
1496 {
1497         struct logfs_block *block = logfs_inode(inode)->li_block;
1498
1499         if (block && block->ta)
1500                 block->ta = NULL;
1501 }
1502
1503 static int grow_inode(struct inode *inode, u64 bix, level_t level)
1504 {
1505         struct logfs_inode *li = logfs_inode(inode);
1506         u8 height = (__force u8)level;
1507         struct page *page;
1508         struct write_control wc = {
1509                 .flags = WF_WRITE,
1510         };
1511         int err;
1512
1513         BUG_ON(height > 5 || li->li_height > 5);
1514         while (height > li->li_height || bix >= maxbix(li->li_height)) {
1515                 page = logfs_get_write_page(inode, I0_BLOCKS + 1,
1516                                 LEVEL(li->li_height + 1));
1517                 if (!page)
1518                         return -ENOMEM;
1519                 logfs_read_empty(page);
1520                 alloc_indirect_block(inode, page, 1);
1521                 block_set_pointer(page, 0, li->li_data[INDIRECT_INDEX]);
1522                 err = logfs_write_i0(inode, page, &wc);
1523                 logfs_put_write_page(page);
1524                 if (err)
1525                         return err;
1526                 li->li_data[INDIRECT_INDEX] = wc.ofs;
1527                 wc.ofs = 0;
1528                 li->li_height++;
1529                 logfs_set_alias(inode->i_sb, li->li_block, INODE_HEIGHT_OFS);
1530         }
1531         return 0;
1532 }
1533
1534 static int __logfs_write_buf(struct inode *inode, struct page *page, long flags)
1535 {
1536         struct logfs_super *super = logfs_super(inode->i_sb);
1537         pgoff_t index = page->index;
1538         u64 bix;
1539         level_t level;
1540         int err;
1541
1542         flags |= WF_WRITE | WF_DELETE;
1543         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1544
1545         logfs_unpack_index(index, &bix, &level);
1546         if (logfs_block(page) && logfs_block(page)->reserved_bytes)
1547                 super->s_dirty_pages -= logfs_block(page)->reserved_bytes;
1548
1549         if (index < I0_BLOCKS)
1550                 return logfs_write_direct(inode, page, flags);
1551
1552         bix = adjust_bix(bix, level);
1553         err = grow_inode(inode, bix, level);
1554         if (err)
1555                 return err;
1556         return logfs_write_rec(inode, page, bix, level, flags);
1557 }
1558
1559 int logfs_write_buf(struct inode *inode, struct page *page, long flags)
1560 {
1561         struct super_block *sb = inode->i_sb;
1562         int ret;
1563
1564         logfs_get_wblocks(sb, page, flags & WF_LOCK);
1565         ret = __logfs_write_buf(inode, page, flags);
1566         logfs_put_wblocks(sb, page, flags & WF_LOCK);
1567         return ret;
1568 }
1569
1570 static int __logfs_delete(struct inode *inode, struct page *page)
1571 {
1572         long flags = WF_DELETE;
1573
1574         inode->i_ctime = inode->i_mtime = CURRENT_TIME;
1575
1576         if (page->index < I0_BLOCKS)
1577                 return logfs_write_direct(inode, page, flags);
1578         return logfs_write_rec(inode, page, page->index, 0, flags);
1579 }
1580
1581 int logfs_delete(struct inode *inode, pgoff_t index,
1582                 struct shadow_tree *shadow_tree)
1583 {
1584         struct super_block *sb = inode->i_sb;
1585         struct page *page;
1586         int ret;
1587
1588         page = logfs_get_read_page(inode, index, 0);
1589         if (!page)
1590                 return -ENOMEM;
1591
1592         logfs_get_wblocks(sb, page, 1);
1593         ret = __logfs_delete(inode, page);
1594         logfs_put_wblocks(sb, page, 1);
1595
1596         logfs_put_read_page(page);
1597
1598         return ret;
1599 }
1600
1601 int logfs_rewrite_block(struct inode *inode, u64 bix, u64 ofs,
1602                 gc_level_t gc_level, long flags)
1603 {
1604         level_t level = shrink_level(gc_level);
1605         struct page *page;
1606         int err;
1607
1608         page = logfs_get_write_page(inode, bix, level);
1609         if (!page)
1610                 return -ENOMEM;
1611
1612         err = logfs_segment_read(inode, page, ofs, bix, level);
1613         if (!err) {
1614                 if (level != 0)
1615                         alloc_indirect_block(inode, page, 0);
1616                 err = logfs_write_buf(inode, page, flags);
1617                 if (!err && shrink_level(gc_level) == 0) {
1618                         /* Rewrite cannot mark the inode dirty but has to
1619                          * write it immediately.
1620                          * Q: Can't we just create an alias for the inode
1621                          * instead?  And if not, why not?
1622                          */
1623                         if (inode->i_ino == LOGFS_INO_MASTER)
1624                                 logfs_write_anchor(inode->i_sb);
1625                         else {
1626                                 err = __logfs_write_inode(inode, flags);
1627                         }
1628                 }
1629         }
1630         logfs_put_write_page(page);
1631         return err;
1632 }
1633
1634 static int truncate_data_block(struct inode *inode, struct page *page,
1635                 u64 ofs, struct logfs_shadow *shadow, u64 size)
1636 {
1637         loff_t pageofs = page->index << inode->i_sb->s_blocksize_bits;
1638         u64 bix;
1639         level_t level;
1640         int err;
1641
1642         /* Does truncation happen within this page? */
1643         if (size <= pageofs || size - pageofs >= PAGE_SIZE)
1644                 return 0;
1645
1646         logfs_unpack_index(page->index, &bix, &level);
1647         BUG_ON(level != 0);
1648
1649         err = logfs_segment_read(inode, page, ofs, bix, level);
1650         if (err)
1651                 return err;
1652
1653         zero_user_segment(page, size - pageofs, PAGE_CACHE_SIZE);
1654         return logfs_segment_write(inode, page, shadow);
1655 }
1656
1657 static int logfs_truncate_i0(struct inode *inode, struct page *page,
1658                 struct write_control *wc, u64 size)
1659 {
1660         struct logfs_shadow *shadow;
1661         u64 bix;
1662         level_t level;
1663         int err = 0;
1664
1665         logfs_unpack_index(page->index, &bix, &level);
1666         BUG_ON(level != 0);
1667         shadow = alloc_shadow(inode, bix, level, wc->ofs);
1668
1669         err = truncate_data_block(inode, page, wc->ofs, shadow, size);
1670         if (err) {
1671                 free_shadow(inode, shadow);
1672                 return err;
1673         }
1674
1675         logfs_segment_delete(inode, shadow);
1676         set_iused(inode, shadow);
1677         fill_shadow_tree(inode, page, shadow);
1678         wc->ofs = shadow->new_ofs;
1679         return 0;
1680 }
1681
1682 static int logfs_truncate_direct(struct inode *inode, u64 size)
1683 {
1684         struct logfs_inode *li = logfs_inode(inode);
1685         struct write_control wc;
1686         struct page *page;
1687         int e;
1688         int err;
1689
1690         alloc_inode_block(inode);
1691
1692         for (e = I0_BLOCKS - 1; e >= 0; e--) {
1693                 if (size > (e+1) * LOGFS_BLOCKSIZE)
1694                         break;
1695
1696                 wc.ofs = li->li_data[e];
1697                 if (!wc.ofs)
1698                         continue;
1699
1700                 page = logfs_get_write_page(inode, e, 0);
1701                 if (!page)
1702                         return -ENOMEM;
1703                 err = logfs_segment_read(inode, page, wc.ofs, e, 0);
1704                 if (err) {
1705                         logfs_put_write_page(page);
1706                         return err;
1707                 }
1708                 err = logfs_truncate_i0(inode, page, &wc, size);
1709                 logfs_put_write_page(page);
1710                 if (err)
1711                         return err;
1712
1713                 li->li_data[e] = wc.ofs;
1714         }
1715         return 0;
1716 }
1717
1718 /* FIXME: these need to become per-sb once we support different blocksizes */
1719 static u64 __logfs_step[] = {
1720         1,
1721         I1_BLOCKS,
1722         I2_BLOCKS,
1723         I3_BLOCKS,
1724 };
1725
1726 static u64 __logfs_start_index[] = {
1727         I0_BLOCKS,
1728         I1_BLOCKS,
1729         I2_BLOCKS,
1730         I3_BLOCKS
1731 };
1732
1733 static inline u64 logfs_step(level_t level)
1734 {
1735         return __logfs_step[(__force u8)level];
1736 }
1737
1738 static inline u64 logfs_factor(u8 level)
1739 {
1740         return __logfs_step[level] * LOGFS_BLOCKSIZE;
1741 }
1742
1743 static inline u64 logfs_start_index(level_t level)
1744 {
1745         return __logfs_start_index[(__force u8)level];
1746 }
1747
1748 static void logfs_unpack_raw_index(pgoff_t index, u64 *bix, level_t *level)
1749 {
1750         logfs_unpack_index(index, bix, level);
1751         if (*bix <= logfs_start_index(SUBLEVEL(*level)))
1752                 *bix = 0;
1753 }
1754
1755 static int __logfs_truncate_rec(struct inode *inode, struct page *ipage,
1756                 struct write_control *this_wc, u64 size)
1757 {
1758         int truncate_happened = 0;
1759         int e, err = 0;
1760         u64 bix, child_bix, next_bix;
1761         level_t level;
1762         struct page *page;
1763         struct write_control child_wc = { /* FIXME: flags */ };
1764
1765         logfs_unpack_raw_index(ipage->index, &bix, &level);
1766         err = logfs_segment_read(inode, ipage, this_wc->ofs, bix, level);
1767         if (err)
1768                 return err;
1769
1770         for (e = LOGFS_BLOCK_FACTOR - 1; e >= 0; e--) {
1771                 child_bix = bix + e * logfs_step(SUBLEVEL(level));
1772                 next_bix = child_bix + logfs_step(SUBLEVEL(level));
1773                 if (size > next_bix * LOGFS_BLOCKSIZE)
1774                         break;
1775
1776                 child_wc.ofs = pure_ofs(block_get_pointer(ipage, e));
1777                 if (!child_wc.ofs)
1778                         continue;
1779
1780                 page = logfs_get_write_page(inode, child_bix, SUBLEVEL(level));
1781                 if (!page)
1782                         return -ENOMEM;
1783
1784                 if ((__force u8)level > 1)
1785                         err = __logfs_truncate_rec(inode, page, &child_wc, size);
1786                 else
1787                         err = logfs_truncate_i0(inode, page, &child_wc, size);
1788                 logfs_put_write_page(page);
1789                 if (err)
1790                         return err;
1791
1792                 truncate_happened = 1;
1793                 alloc_indirect_block(inode, ipage, 0);
1794                 block_set_pointer(ipage, e, child_wc.ofs);
1795         }
1796
1797         if (!truncate_happened) {
1798                 printk("ineffectual truncate (%lx, %lx, %llx)\n", inode->i_ino, ipage->index, size);
1799                 return 0;
1800         }
1801
1802         this_wc->flags = WF_DELETE;
1803         if (logfs_block(ipage)->partial)
1804                 this_wc->flags |= WF_WRITE;
1805
1806         return logfs_write_i0(inode, ipage, this_wc);
1807 }
1808
1809 static int logfs_truncate_rec(struct inode *inode, u64 size)
1810 {
1811         struct logfs_inode *li = logfs_inode(inode);
1812         struct write_control wc = {
1813                 .ofs = li->li_data[INDIRECT_INDEX],
1814         };
1815         struct page *page;
1816         int err;
1817
1818         alloc_inode_block(inode);
1819
1820         if (!wc.ofs)
1821                 return 0;
1822
1823         page = logfs_get_write_page(inode, 0, LEVEL(li->li_height));
1824         if (!page)
1825                 return -ENOMEM;
1826
1827         err = __logfs_truncate_rec(inode, page, &wc, size);
1828         logfs_put_write_page(page);
1829         if (err)
1830                 return err;
1831
1832         if (li->li_data[INDIRECT_INDEX] != wc.ofs)
1833                 li->li_data[INDIRECT_INDEX] = wc.ofs;
1834         return 0;
1835 }
1836
1837 static int __logfs_truncate(struct inode *inode, u64 size)
1838 {
1839         int ret;
1840
1841         if (size >= logfs_factor(logfs_inode(inode)->li_height))
1842                 return 0;
1843
1844         ret = logfs_truncate_rec(inode, size);
1845         if (ret)
1846                 return ret;
1847
1848         return logfs_truncate_direct(inode, size);
1849 }
1850
1851 /*
1852  * Truncate, by changing the segment file, can consume a fair amount
1853  * of resources.  So back off from time to time and do some GC.
1854  * 8 or 2048 blocks should be well within safety limits even if
1855  * every single block resided in a different segment.
1856  */
1857 #define TRUNCATE_STEP   (8 * 1024 * 1024)
1858 int logfs_truncate(struct inode *inode, u64 target)
1859 {
1860         struct super_block *sb = inode->i_sb;
1861         u64 size = i_size_read(inode);
1862         int err = 0;
1863
1864         size = ALIGN(size, TRUNCATE_STEP);
1865         while (size > target) {
1866                 if (size > TRUNCATE_STEP)
1867                         size -= TRUNCATE_STEP;
1868                 else
1869                         size = 0;
1870                 if (size < target)
1871                         size = target;
1872
1873                 logfs_get_wblocks(sb, NULL, 1);
1874                 err = __logfs_truncate(inode, size);
1875                 if (!err)
1876                         err = __logfs_write_inode(inode, 0);
1877                 logfs_put_wblocks(sb, NULL, 1);
1878         }
1879
1880         if (!err)
1881                 err = vmtruncate(inode, target);
1882
1883         /* I don't trust error recovery yet. */
1884         WARN_ON(err);
1885         return err;
1886 }
1887
1888 static void move_page_to_inode(struct inode *inode, struct page *page)
1889 {
1890         struct logfs_inode *li = logfs_inode(inode);
1891         struct logfs_block *block = logfs_block(page);
1892
1893         if (!block)
1894                 return;
1895
1896         log_blockmove("move_page_to_inode(%llx, %llx, %x)\n",
1897                         block->ino, block->bix, block->level);
1898         BUG_ON(li->li_block);
1899         block->ops = &inode_block_ops;
1900         block->inode = inode;
1901         li->li_block = block;
1902
1903         block->page = NULL;
1904         page->private = 0;
1905         ClearPagePrivate(page);
1906 }
1907
1908 static void move_inode_to_page(struct page *page, struct inode *inode)
1909 {
1910         struct logfs_inode *li = logfs_inode(inode);
1911         struct logfs_block *block = li->li_block;
1912
1913         if (!block)
1914                 return;
1915
1916         log_blockmove("move_inode_to_page(%llx, %llx, %x)\n",
1917                         block->ino, block->bix, block->level);
1918         BUG_ON(PagePrivate(page));
1919         block->ops = &indirect_block_ops;
1920         block->page = page;
1921         page->private = (unsigned long)block;
1922         SetPagePrivate(page);
1923
1924         block->inode = NULL;
1925         li->li_block = NULL;
1926 }
1927
1928 int logfs_read_inode(struct inode *inode)
1929 {
1930         struct super_block *sb = inode->i_sb;
1931         struct logfs_super *super = logfs_super(sb);
1932         struct inode *master_inode = super->s_master_inode;
1933         struct page *page;
1934         struct logfs_disk_inode *di;
1935         u64 ino = inode->i_ino;
1936
1937         if (ino << sb->s_blocksize_bits > i_size_read(master_inode))
1938                 return -ENODATA;
1939         if (!logfs_exist_block(master_inode, ino))
1940                 return -ENODATA;
1941
1942         page = read_cache_page(master_inode->i_mapping, ino,
1943                         (filler_t *)logfs_readpage, NULL);
1944         if (IS_ERR(page))
1945                 return PTR_ERR(page);
1946
1947         di = kmap_atomic(page, KM_USER0);
1948         logfs_disk_to_inode(di, inode);
1949         kunmap_atomic(di, KM_USER0);
1950         move_page_to_inode(inode, page);
1951         page_cache_release(page);
1952         return 0;
1953 }
1954
1955 /* Caller must logfs_put_write_page(page); */
1956 static struct page *inode_to_page(struct inode *inode)
1957 {
1958         struct inode *master_inode = logfs_super(inode->i_sb)->s_master_inode;
1959         struct logfs_disk_inode *di;
1960         struct page *page;
1961
1962         BUG_ON(inode->i_ino == LOGFS_INO_MASTER);
1963
1964         page = logfs_get_write_page(master_inode, inode->i_ino, 0);
1965         if (!page)
1966                 return NULL;
1967
1968         di = kmap_atomic(page, KM_USER0);
1969         logfs_inode_to_disk(inode, di);
1970         kunmap_atomic(di, KM_USER0);
1971         move_inode_to_page(page, inode);
1972         return page;
1973 }
1974
1975 static int do_write_inode(struct inode *inode)
1976 {
1977         struct super_block *sb = inode->i_sb;
1978         struct inode *master_inode = logfs_super(sb)->s_master_inode;
1979         loff_t size = (inode->i_ino + 1) << inode->i_sb->s_blocksize_bits;
1980         struct page *page;
1981         int err;
1982
1983         BUG_ON(inode->i_ino == LOGFS_INO_MASTER);
1984         /* FIXME: lock inode */
1985
1986         if (i_size_read(master_inode) < size)
1987                 i_size_write(master_inode, size);
1988
1989         /* TODO: Tell vfs this inode is clean now */
1990
1991         page = inode_to_page(inode);
1992         if (!page)
1993                 return -ENOMEM;
1994
1995         /* FIXME: transaction is part of logfs_block now.  Is that enough? */
1996         err = logfs_write_buf(master_inode, page, 0);
1997         if (err)
1998                 move_page_to_inode(inode, page);
1999
2000         logfs_put_write_page(page);
2001         return err;
2002 }
2003
2004 static void logfs_mod_segment_entry(struct super_block *sb, u32 segno,
2005                 int write,
2006                 void (*change_se)(struct logfs_segment_entry *, long),
2007                 long arg)
2008 {
2009         struct logfs_super *super = logfs_super(sb);
2010         struct inode *inode;
2011         struct page *page;
2012         struct logfs_segment_entry *se;
2013         pgoff_t page_no;
2014         int child_no;
2015
2016         page_no = segno >> (sb->s_blocksize_bits - 3);
2017         child_no = segno & ((sb->s_blocksize >> 3) - 1);
2018
2019         inode = super->s_segfile_inode;
2020         page = logfs_get_write_page(inode, page_no, 0);
2021         BUG_ON(!page); /* FIXME: We need some reserve page for this case */
2022         if (!PageUptodate(page))
2023                 logfs_read_block(inode, page, WRITE);
2024
2025         if (write)
2026                 alloc_indirect_block(inode, page, 0);
2027         se = kmap_atomic(page, KM_USER0);
2028         change_se(se + child_no, arg);
2029         if (write) {
2030                 logfs_set_alias(sb, logfs_block(page), child_no);
2031                 BUG_ON((int)be32_to_cpu(se[child_no].valid) > super->s_segsize);
2032         }
2033         kunmap_atomic(se, KM_USER0);
2034
2035         logfs_put_write_page(page);
2036 }
2037
2038 static void __get_segment_entry(struct logfs_segment_entry *se, long _target)
2039 {
2040         struct logfs_segment_entry *target = (void *)_target;
2041
2042         *target = *se;
2043 }
2044
2045 void logfs_get_segment_entry(struct super_block *sb, u32 segno,
2046                 struct logfs_segment_entry *se)
2047 {
2048         logfs_mod_segment_entry(sb, segno, 0, __get_segment_entry, (long)se);
2049 }
2050
2051 static void __set_segment_used(struct logfs_segment_entry *se, long increment)
2052 {
2053         u32 valid;
2054
2055         valid = be32_to_cpu(se->valid);
2056         valid += increment;
2057         se->valid = cpu_to_be32(valid);
2058 }
2059
2060 void logfs_set_segment_used(struct super_block *sb, u64 ofs, int increment)
2061 {
2062         struct logfs_super *super = logfs_super(sb);
2063         u32 segno = ofs >> super->s_segshift;
2064
2065         if (!increment)
2066                 return;
2067
2068         logfs_mod_segment_entry(sb, segno, 1, __set_segment_used, increment);
2069 }
2070
2071 static void __set_segment_erased(struct logfs_segment_entry *se, long ec_level)
2072 {
2073         se->ec_level = cpu_to_be32(ec_level);
2074 }
2075
2076 void logfs_set_segment_erased(struct super_block *sb, u32 segno, u32 ec,
2077                 gc_level_t gc_level)
2078 {
2079         u32 ec_level = ec << 4 | (__force u8)gc_level;
2080
2081         logfs_mod_segment_entry(sb, segno, 1, __set_segment_erased, ec_level);
2082 }
2083
2084 static void __set_segment_reserved(struct logfs_segment_entry *se, long ignore)
2085 {
2086         se->valid = cpu_to_be32(RESERVED);
2087 }
2088
2089 void logfs_set_segment_reserved(struct super_block *sb, u32 segno)
2090 {
2091         logfs_mod_segment_entry(sb, segno, 1, __set_segment_reserved, 0);
2092 }
2093
2094 static void __set_segment_unreserved(struct logfs_segment_entry *se,
2095                 long ec_level)
2096 {
2097         se->valid = 0;
2098         se->ec_level = cpu_to_be32(ec_level);
2099 }
2100
2101 void logfs_set_segment_unreserved(struct super_block *sb, u32 segno, u32 ec)
2102 {
2103         u32 ec_level = ec << 4;
2104
2105         logfs_mod_segment_entry(sb, segno, 1, __set_segment_unreserved,
2106                         ec_level);
2107 }
2108
2109 int __logfs_write_inode(struct inode *inode, long flags)
2110 {
2111         struct super_block *sb = inode->i_sb;
2112         int ret;
2113
2114         logfs_get_wblocks(sb, NULL, flags & WF_LOCK);
2115         ret = do_write_inode(inode);
2116         logfs_put_wblocks(sb, NULL, flags & WF_LOCK);
2117         return ret;
2118 }
2119
2120 static int do_delete_inode(struct inode *inode)
2121 {
2122         struct super_block *sb = inode->i_sb;
2123         struct inode *master_inode = logfs_super(sb)->s_master_inode;
2124         struct page *page;
2125         int ret;
2126
2127         page = logfs_get_write_page(master_inode, inode->i_ino, 0);
2128         if (!page)
2129                 return -ENOMEM;
2130
2131         move_inode_to_page(page, inode);
2132
2133         logfs_get_wblocks(sb, page, 1);
2134         ret = __logfs_delete(master_inode, page);
2135         logfs_put_wblocks(sb, page, 1);
2136
2137         logfs_put_write_page(page);
2138         return ret;
2139 }
2140
2141 /*
2142  * ZOMBIE inodes have already been deleted before and should remain dead,
2143  * if it weren't for valid checking.  No need to kill them again here.
2144  */
2145 void logfs_evict_inode(struct inode *inode)
2146 {
2147         struct super_block *sb = inode->i_sb;
2148         struct logfs_inode *li = logfs_inode(inode);
2149         struct logfs_block *block = li->li_block;
2150         struct page *page;
2151
2152         if (!inode->i_nlink) {
2153                 if (!(li->li_flags & LOGFS_IF_ZOMBIE)) {
2154                         li->li_flags |= LOGFS_IF_ZOMBIE;
2155                         if (i_size_read(inode) > 0)
2156                                 logfs_truncate(inode, 0);
2157                         do_delete_inode(inode);
2158                 }
2159         }
2160         truncate_inode_pages(&inode->i_data, 0);
2161         end_writeback(inode);
2162
2163         /* Cheaper version of write_inode.  All changes are concealed in
2164          * aliases, which are moved back.  No write to the medium happens.
2165          */
2166         /* Only deleted files may be dirty at this point */
2167         BUG_ON(inode->i_state & I_DIRTY && inode->i_nlink);
2168         if (!block)
2169                 return;
2170         if ((logfs_super(sb)->s_flags & LOGFS_SB_FLAG_SHUTDOWN)) {
2171                 block->ops->free_block(inode->i_sb, block);
2172                 return;
2173         }
2174
2175         BUG_ON(inode->i_ino < LOGFS_RESERVED_INOS);
2176         page = inode_to_page(inode);
2177         BUG_ON(!page); /* FIXME: Use emergency page */
2178         logfs_put_write_page(page);
2179 }
2180
2181 void btree_write_block(struct logfs_block *block)
2182 {
2183         struct inode *inode;
2184         struct page *page;
2185         int err, cookie;
2186
2187         inode = logfs_safe_iget(block->sb, block->ino, &cookie);
2188         page = logfs_get_write_page(inode, block->bix, block->level);
2189
2190         err = logfs_readpage_nolock(page);
2191         BUG_ON(err);
2192         BUG_ON(!PagePrivate(page));
2193         BUG_ON(logfs_block(page) != block);
2194         err = __logfs_write_buf(inode, page, 0);
2195         BUG_ON(err);
2196         BUG_ON(PagePrivate(page) || page->private);
2197
2198         logfs_put_write_page(page);
2199         logfs_safe_iput(inode, cookie);
2200 }
2201
2202 /**
2203  * logfs_inode_write - write inode or dentry objects
2204  *
2205  * @inode:              parent inode (ifile or directory)
2206  * @buf:                object to write (inode or dentry)
2207  * @n:                  object size
2208  * @_pos:               object number (file position in blocks/objects)
2209  * @flags:              write flags
2210  * @lock:               0 if write lock is already taken, 1 otherwise
2211  * @shadow_tree:        shadow below this inode
2212  *
2213  * FIXME: All caller of this put a 200-300 byte variable on the stack,
2214  * only to call here and do a memcpy from that stack variable.  A good
2215  * example of wasted performance and stack space.
2216  */
2217 int logfs_inode_write(struct inode *inode, const void *buf, size_t count,
2218                 loff_t bix, long flags, struct shadow_tree *shadow_tree)
2219 {
2220         loff_t pos = bix << inode->i_sb->s_blocksize_bits;
2221         int err;
2222         struct page *page;
2223         void *pagebuf;
2224
2225         BUG_ON(pos & (LOGFS_BLOCKSIZE-1));
2226         BUG_ON(count > LOGFS_BLOCKSIZE);
2227         page = logfs_get_write_page(inode, bix, 0);
2228         if (!page)
2229                 return -ENOMEM;
2230
2231         pagebuf = kmap_atomic(page, KM_USER0);
2232         memcpy(pagebuf, buf, count);
2233         flush_dcache_page(page);
2234         kunmap_atomic(pagebuf, KM_USER0);
2235
2236         if (i_size_read(inode) < pos + LOGFS_BLOCKSIZE)
2237                 i_size_write(inode, pos + LOGFS_BLOCKSIZE);
2238
2239         err = logfs_write_buf(inode, page, flags);
2240         logfs_put_write_page(page);
2241         return err;
2242 }
2243
2244 int logfs_open_segfile(struct super_block *sb)
2245 {
2246         struct logfs_super *super = logfs_super(sb);
2247         struct inode *inode;
2248
2249         inode = logfs_read_meta_inode(sb, LOGFS_INO_SEGFILE);
2250         if (IS_ERR(inode))
2251                 return PTR_ERR(inode);
2252         super->s_segfile_inode = inode;
2253         return 0;
2254 }
2255
2256 int logfs_init_rw(struct super_block *sb)
2257 {
2258         struct logfs_super *super = logfs_super(sb);
2259         int min_fill = 3 * super->s_no_blocks;
2260
2261         INIT_LIST_HEAD(&super->s_object_alias);
2262         INIT_LIST_HEAD(&super->s_writeback_list);
2263         mutex_init(&super->s_write_mutex);
2264         super->s_block_pool = mempool_create_kmalloc_pool(min_fill,
2265                         sizeof(struct logfs_block));
2266         super->s_shadow_pool = mempool_create_kmalloc_pool(min_fill,
2267                         sizeof(struct logfs_shadow));
2268         return 0;
2269 }
2270
2271 void logfs_cleanup_rw(struct super_block *sb)
2272 {
2273         struct logfs_super *super = logfs_super(sb);
2274
2275         logfs_mempool_destroy(super->s_block_pool);
2276         logfs_mempool_destroy(super->s_shadow_pool);
2277 }