fs/ext2/xattr.c

   1 /*
   2  * linux/fs/ext2/xattr.c
   3  *
   4  * Copyright (C) 2001-2003 Andreas Gruenbacher <agruen@suse.de>
   5  *
   6  * Fix by Harrison Xing <harrison@mountainviewdata.com>.
   7  * Extended attributes for symlinks and special files added per
   8  *  suggestion of Luka Renko <luka.renko@hermes.si>.
   9  */
  10
  11 /*
  12  * Extended attributes are stored on disk blocks allocated outside of
  13  * any inode. The i_file_acl field is then made to point to this allocated
  14  * block. If all extended attributes of an inode are identical, these
  15  * inodes may share the same extended attribute block. Such situations
  16  * are automatically detected by keeping a cache of recent attribute block
  17  * numbers and hashes over the block's contents in memory.
  18  *
  19  *
  20  * Extended attribute block layout:
  21  *
  22  *   +------------------+
  23  *   | header           |
  24  *   ¦ entry 1          | |
  25  *   | entry 2          | | growing downwards
  26  *   | entry 3          | v
  27  *   | four null bytes  |
  28  *   | . . .            |
  29  *   | value 1          | ^
  30  *   | value 3          | | growing upwards
  31  *   | value 2          | |
  32  *   +------------------+
  33  *
  34  * The block header is followed by multiple entry descriptors. These entry
  35  * descriptors are variable in size, and alligned to EXT2_XATTR_PAD
  36  * byte boundaries. The entry descriptors are sorted by attribute name,
  37  * so that two extended attribute blocks can be compared efficiently.
  38  *
  39  * Attribute values are aligned to the end of the block, stored in
  40  * no specific order. They are also padded to EXT2_XATTR_PAD byte
  41  * boundaries. No additional gaps are left between them.
  42  *
  43  * Locking strategy
  44  * ----------------
  45  * EXT2_I(inode)->i_file_acl is protected by EXT2_I(inode)->xattr_sem.
  46  * EA blocks are only changed if they are exclusive to an inode, so
  47  * holding xattr_sem also means that nothing but the EA block's reference
  48  * count will change. Multiple writers to an EA block are synchronized
  49  * by the bh lock. No more than a single bh lock is held at any time
  50  * to avoid deadlocks.
  51  */
  52
  53 #include <linux/buffer_head.h>
  54 #include <linux/module.h>
  55 #include <linux/init.h>
  56 #include <linux/slab.h>
  57 #include <linux/mbcache.h>
  58 #include <linux/quotaops.h>
  59 #include <linux/rwsem.h>
  60 #include "ext2.h"
  61 #include "xattr.h"
  62 #include "acl.h"
  63
  64 /* These symbols may be needed by a module. */
  65 EXPORT_SYMBOL(ext2_xattr_register);
  66 EXPORT_SYMBOL(ext2_xattr_unregister);
  67 EXPORT_SYMBOL(ext2_xattr_get);
  68 EXPORT_SYMBOL(ext2_xattr_list);
  69 EXPORT_SYMBOL(ext2_xattr_set);
  70
  71 #define HDR(bh) ((struct ext2_xattr_header *)((bh)->b_data))
  72 #define ENTRY(ptr) ((struct ext2_xattr_entry *)(ptr))
  73 #define FIRST_ENTRY(bh) ENTRY(HDR(bh)+1)
  74 #define IS_LAST_ENTRY(entry) (*(__u32 *)(entry) == 0)
  75
  76 #ifdef EXT2_XATTR_DEBUG
  77 # define ea_idebug(inode, f...) do { \
  78                 printk(KERN_DEBUG "inode %s:%ld: ", \
  79                         inode->i_sb->s_id, inode->i_ino); \
  80                 printk(f); \
  81                 printk("\n"); \
  82         } while (0)
  83 # define ea_bdebug(bh, f...) do { \
  84                 char b[BDEVNAME_SIZE]; \
  85                 printk(KERN_DEBUG "block %s:%lu: ", \
  86                         bdevname(bh->b_bdev, b), \
  87                         (unsigned long) bh->b_blocknr); \
  88                 printk(f); \
  89                 printk("\n"); \
  90         } while (0)
  91 #else
  92 # define ea_idebug(f...)
  93 # define ea_bdebug(f...)
  94 #endif
  95
  96 static int ext2_xattr_set2(struct inode *, struct buffer_head *,
  97                            struct ext2_xattr_header *);
  98
  99 static int ext2_xattr_cache_insert(struct buffer_head *);
 100 static struct buffer_head *ext2_xattr_cache_find(struct inode *,
 101                                                  struct ext2_xattr_header *);
 102 static void ext2_xattr_cache_remove(struct buffer_head *);
 103 static void ext2_xattr_rehash(struct ext2_xattr_header *,
 104                               struct ext2_xattr_entry *);
 105
 106 static struct mb_cache *ext2_xattr_cache;
 107 static struct ext2_xattr_handler *ext2_xattr_handlers[EXT2_XATTR_INDEX_MAX];
 108 static rwlock_t ext2_handler_lock = RW_LOCK_UNLOCKED;
 109
 110 int
 111 ext2_xattr_register(int name_index, struct ext2_xattr_handler *handler)
 112 {
 113         int error = -EINVAL;
 114
 115         if (name_index > 0 && name_index <= EXT2_XATTR_INDEX_MAX) {
 116                 write_lock(&ext2_handler_lock);
 117                 if (!ext2_xattr_handlers[name_index-1]) {
 118                         ext2_xattr_handlers[name_index-1] = handler;
 119                         error = 0;
 120                 }
 121                 write_unlock(&ext2_handler_lock);
 122         }
 123         return error;
 124 }
 125
 126 void
 127 ext2_xattr_unregister(int name_index, struct ext2_xattr_handler *handler)
 128 {
 129         if (name_index > 0 || name_index <= EXT2_XATTR_INDEX_MAX) {
 130                 write_lock(&ext2_handler_lock);
 131                 ext2_xattr_handlers[name_index-1] = NULL;
 132                 write_unlock(&ext2_handler_lock);
 133         }
 134 }
 135
 136 static inline const char *
 137 strcmp_prefix(const char *a, const char *a_prefix)
 138 {
 139         while (*a_prefix && *a == *a_prefix) {
 140                 a++;
 141                 a_prefix++;
 142         }
 143         return *a_prefix ? NULL : a;
 144 }
 145
 146 /*
 147  * Decode the extended attribute name, and translate it into
 148  * the name_index and name suffix.
 149  */
 150 static struct ext2_xattr_handler *
 151 ext2_xattr_resolve_name(const char **name)
 152 {
 153         struct ext2_xattr_handler *handler = NULL;
 154         int i;
 155
 156         if (!*name)
 157                 return NULL;
 158         read_lock(&ext2_handler_lock);
 159         for (i=0; i<EXT2_XATTR_INDEX_MAX; i++) {
 160                 if (ext2_xattr_handlers[i]) {
 161                         const char *n = strcmp_prefix(*name,
 162                                 ext2_xattr_handlers[i]->prefix);
 163                         if (n) {
 164                                 handler = ext2_xattr_handlers[i];
 165                                 *name = n;
 166                                 break;
 167                         }
 168                 }
 169         }
 170         read_unlock(&ext2_handler_lock);
 171         return handler;
 172 }
 173
 174 static inline struct ext2_xattr_handler *
 175 ext2_xattr_handler(int name_index)
 176 {
 177         struct ext2_xattr_handler *handler = NULL;
 178         if (name_index > 0 && name_index <= EXT2_XATTR_INDEX_MAX) {
 179                 read_lock(&ext2_handler_lock);
 180                 handler = ext2_xattr_handlers[name_index-1];
 181                 read_unlock(&ext2_handler_lock);
 182         }
 183         return handler;
 184 }
 185
 186 /*
 187  * Inode operation getxattr()
 188  *
 189  * dentry->d_inode->i_sem: don't care
 190  */
 191 ssize_t
 192 ext2_getxattr(struct dentry *dentry, const char *name,
 193               void *buffer, size_t size)
 194 {
 195         struct ext2_xattr_handler *handler;
 196         struct inode *inode = dentry->d_inode;
 197
 198         handler = ext2_xattr_resolve_name(&name);
 199         if (!handler)
 200                 return -EOPNOTSUPP;
 201         return handler->get(inode, name, buffer, size);
 202 }
 203
 204 /*
 205  * Inode operation listxattr()
 206  *
 207  * dentry->d_inode->i_sem: don't care
 208  */
 209 ssize_t
 210 ext2_listxattr(struct dentry *dentry, char *buffer, size_t size)
 211 {
 212         return ext2_xattr_list(dentry->d_inode, buffer, size);
 213 }
 214
 215 /*
 216  * Inode operation setxattr()
 217  *
 218  * dentry->d_inode->i_sem: down
 219  */
 220 int
 221 ext2_setxattr(struct dentry *dentry, const char *name,
 222               const void *value, size_t size, int flags)
 223 {
 224         struct ext2_xattr_handler *handler;
 225         struct inode *inode = dentry->d_inode;
 226
 227         if (size == 0)
 228                 value = "";  /* empty EA, do not remove */
 229         handler = ext2_xattr_resolve_name(&name);
 230         if (!handler)
 231                 return -EOPNOTSUPP;
 232         return handler->set(inode, name, value, size, flags);
 233 }
 234
 235 /*
 236  * Inode operation removexattr()
 237  *
 238  * dentry->d_inode->i_sem: down
 239  */
 240 int
 241 ext2_removexattr(struct dentry *dentry, const char *name)
 242 {
 243         struct ext2_xattr_handler *handler;
 244         struct inode *inode = dentry->d_inode;
 245
 246         handler = ext2_xattr_resolve_name(&name);
 247         if (!handler)
 248                 return -EOPNOTSUPP;
 249         return handler->set(inode, name, NULL, 0, XATTR_REPLACE);
 250 }
 251
 252 /*
 253  * ext2_xattr_get()
 254  *
 255  * Copy an extended attribute into the buffer
 256  * provided, or compute the buffer size required.
 257  * Buffer is NULL to compute the size of the buffer required.
 258  *
 259  * Returns a negative error number on failure, or the number of bytes
 260  * used / required on success.
 261  */
 262 int
 263 ext2_xattr_get(struct inode *inode, int name_index, const char *name,
 264                void *buffer, size_t buffer_size)
 265 {
 266         struct buffer_head *bh = NULL;
 267         struct ext2_xattr_entry *entry;
 268         size_t name_len, size;
 269         char *end;
 270         int error;
 271
 272         ea_idebug(inode, "name=%d.%s, buffer=%p, buffer_size=%ld",
 273                   name_index, name, buffer, (long)buffer_size);
 274
 275         if (name == NULL)
 276                 return -EINVAL;
 277         down_read(&EXT2_I(inode)->xattr_sem);
 278         error = -ENODATA;
 279         if (!EXT2_I(inode)->i_file_acl)
 280                 goto cleanup;
 281         ea_idebug(inode, "reading block %d", EXT2_I(inode)->i_file_acl);
 282         bh = sb_bread(inode->i_sb, EXT2_I(inode)->i_file_acl);
 283         error = -EIO;
 284         if (!bh)
 285                 goto cleanup;
 286         ea_bdebug(bh, "b_count=%d, refcount=%d",
 287                 atomic_read(&(bh->b_count)), le32_to_cpu(HDR(bh)->h_refcount));
 288         end = bh->b_data + bh->b_size;
 289         if (HDR(bh)->h_magic != cpu_to_le32(EXT2_XATTR_MAGIC) ||
 290             HDR(bh)->h_blocks != cpu_to_le32(1)) {
 291 bad_block:      ext2_error(inode->i_sb, "ext2_xattr_get",
 292                         "inode %ld: bad block %d", inode->i_ino,
 293                         EXT2_I(inode)->i_file_acl);
 294                 error = -EIO;
 295                 goto cleanup;
 296         }
 297         /* find named attribute */
 298         name_len = strlen(name);
 299
 300         error = -ERANGE;
 301         if (name_len > 255)
 302                 goto cleanup;
 303         entry = FIRST_ENTRY(bh);
 304         while (!IS_LAST_ENTRY(entry)) {
 305                 struct ext2_xattr_entry *next =
 306                         EXT2_XATTR_NEXT(entry);
 307                 if ((char *)next >= end)
 308                         goto bad_block;
 309                 if (name_index == entry->e_name_index &&
 310                     name_len == entry->e_name_len &&
 311                     memcmp(name, entry->e_name, name_len) == 0)
 312                         goto found;
 313                 entry = next;
 314         }
 315         /* Check the remaining name entries */
 316         while (!IS_LAST_ENTRY(entry)) {
 317                 struct ext2_xattr_entry *next =
 318                         EXT2_XATTR_NEXT(entry);
 319                 if ((char *)next >= end)
 320                         goto bad_block;
 321                 entry = next;
 322         }
 323         if (ext2_xattr_cache_insert(bh))
 324                 ea_idebug(inode, "cache insert failed");
 325         error = -ENODATA;
 326         goto cleanup;
 327 found:
 328         /* check the buffer size */
 329         if (entry->e_value_block != 0)
 330                 goto bad_block;
 331         size = le32_to_cpu(entry->e_value_size);
 332         if (size > inode->i_sb->s_blocksize ||
 333             le16_to_cpu(entry->e_value_offs) + size > inode->i_sb->s_blocksize)
 334                 goto bad_block;
 335
 336         if (ext2_xattr_cache_insert(bh))
 337                 ea_idebug(inode, "cache insert failed");
 338         if (buffer) {
 339                 error = -ERANGE;
 340                 if (size > buffer_size)
 341                         goto cleanup;
 342                 /* return value of attribute */
 343                 memcpy(buffer, bh->b_data + le16_to_cpu(entry->e_value_offs),
 344                         size);
 345         }
 346         error = size;
 347
 348 cleanup:
 349         brelse(bh);
 350         up_read(&EXT2_I(inode)->xattr_sem);
 351
 352         return error;
 353 }
 354
 355 /*
 356  * ext2_xattr_list()
 357  *
 358  * Copy a list of attribute names into the buffer
 359  * provided, or compute the buffer size required.
 360  * Buffer is NULL to compute the size of the buffer required.
 361  *
 362  * Returns a negative error number on failure, or the number of bytes
 363  * used / required on success.
 364  */
 365 int
 366 ext2_xattr_list(struct inode *inode, char *buffer, size_t buffer_size)
 367 {
 368         struct buffer_head *bh = NULL;
 369         struct ext2_xattr_entry *entry;
 370         size_t size = 0;
 371         char *buf, *end;
 372         int error;
 373
 374         ea_idebug(inode, "buffer=%p, buffer_size=%ld",
 375                   buffer, (long)buffer_size);
 376
 377         down_read(&EXT2_I(inode)->xattr_sem);
 378         error = 0;
 379         if (!EXT2_I(inode)->i_file_acl)
 380                 goto cleanup;
 381         ea_idebug(inode, "reading block %d", EXT2_I(inode)->i_file_acl);
 382         bh = sb_bread(inode->i_sb, EXT2_I(inode)->i_file_acl);
 383         error = -EIO;
 384         if (!bh)
 385                 goto cleanup;
 386         ea_bdebug(bh, "b_count=%d, refcount=%d",
 387                 atomic_read(&(bh->b_count)), le32_to_cpu(HDR(bh)->h_refcount));
 388         end = bh->b_data + bh->b_size;
 389         if (HDR(bh)->h_magic != cpu_to_le32(EXT2_XATTR_MAGIC) ||
 390             HDR(bh)->h_blocks != cpu_to_le32(1)) {
 391 bad_block:      ext2_error(inode->i_sb, "ext2_xattr_list",
 392                         "inode %ld: bad block %d", inode->i_ino,
 393                         EXT2_I(inode)->i_file_acl);
 394                 error = -EIO;
 395                 goto cleanup;
 396         }
 397         /* compute the size required for the list of attribute names */
 398         for (entry = FIRST_ENTRY(bh); !IS_LAST_ENTRY(entry);
 399              entry = EXT2_XATTR_NEXT(entry)) {
 400                 struct ext2_xattr_handler *handler;
 401                 struct ext2_xattr_entry *next =
 402                         EXT2_XATTR_NEXT(entry);
 403                 if ((char *)next >= end)
 404                         goto bad_block;
 405
 406                 handler = ext2_xattr_handler(entry->e_name_index);
 407                 if (handler)
 408                         size += handler->list(NULL, inode, entry->e_name,
 409                                               entry->e_name_len);
 410         }
 411
 412         if (ext2_xattr_cache_insert(bh))
 413                 ea_idebug(inode, "cache insert failed");
 414         if (!buffer) {
 415                 error = size;
 416                 goto cleanup;
 417         } else {
 418                 error = -ERANGE;
 419                 if (size > buffer_size)
 420                         goto cleanup;
 421         }
 422
 423         /* list the attribute names */
 424         buf = buffer;
 425         for (entry = FIRST_ENTRY(bh); !IS_LAST_ENTRY(entry);
 426              entry = EXT2_XATTR_NEXT(entry)) {
 427                 struct ext2_xattr_handler *handler;
 428
 429                 handler = ext2_xattr_handler(entry->e_name_index);
 430                 if (handler)
 431                         buf += handler->list(buf, inode, entry->e_name,
 432                                              entry->e_name_len);
 433         }
 434         error = size;
 435
 436 cleanup:
 437         brelse(bh);
 438         up_read(&EXT2_I(inode)->xattr_sem);
 439
 440         return error;
 441 }
 442
 443 /*
 444  * If the EXT2_FEATURE_COMPAT_EXT_ATTR feature of this file system is
 445  * not set, set it.
 446  */
 447 static void ext2_xattr_update_super_block(struct super_block *sb)
 448 {
 449         if (EXT2_HAS_COMPAT_FEATURE(sb, EXT2_FEATURE_COMPAT_EXT_ATTR))
 450                 return;
 451
 452         lock_super(sb);
 453         EXT2_SB(sb)->s_es->s_feature_compat |=
 454                 cpu_to_le32(EXT2_FEATURE_COMPAT_EXT_ATTR);
 455         sb->s_dirt = 1;
 456         mark_buffer_dirty(EXT2_SB(sb)->s_sbh);
 457         unlock_super(sb);
 458 }
 459
 460 /*
 461  * ext2_xattr_set()
 462  *
 463  * Create, replace or remove an extended attribute for this inode. Buffer
 464  * is NULL to remove an existing extended attribute, and non-NULL to
 465  * either replace an existing extended attribute, or create a new extended
 466  * attribute. The flags XATTR_REPLACE and XATTR_CREATE
 467  * specify that an extended attribute must exist and must not exist
 468  * previous to the call, respectively.
 469  *
 470  * Returns 0, or a negative error number on failure.
 471  */
 472 int
 473 ext2_xattr_set(struct inode *inode, int name_index, const char *name,
 474                const void *value, size_t value_len, int flags)
 475 {
 476         struct super_block *sb = inode->i_sb;
 477         struct buffer_head *bh = NULL;
 478         struct ext2_xattr_header *header = NULL;
 479         struct ext2_xattr_entry *here, *last;
 480         size_t name_len, free, min_offs = sb->s_blocksize;
 481         int not_found = 1, error;
 482         char *end;
 483
 484         /*
 485          * header -- Points either into bh, or to a temporarily
 486          *           allocated buffer.
 487          * here -- The named entry found, or the place for inserting, within
 488          *         the block pointed to by header.
 489          * last -- Points right after the last named entry within the block
 490          *         pointed to by header.
 491          * min_offs -- The offset of the first value (values are aligned
 492          *             towards the end of the block).
 493          * end -- Points right after the block pointed to by header.
 494          */
 495
 496         ea_idebug(inode, "name=%d.%s, value=%p, value_len=%ld",
 497                   name_index, name, value, (long)value_len);
 498
 499         if (IS_RDONLY(inode))
 500                 return -EROFS;
 501         if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 502                 return -EPERM;
 503         if (value == NULL)
 504                 value_len = 0;
 505         if (name == NULL)
 506                 return -EINVAL;
 507         name_len = strlen(name);
 508         if (name_len > 255 || value_len > sb->s_blocksize)
 509                 return -ERANGE;
 510         down_write(&EXT2_I(inode)->xattr_sem);
 511         if (EXT2_I(inode)->i_file_acl) {
 512                 /* The inode already has an extended attribute block. */
 513                 bh = sb_bread(sb, EXT2_I(inode)->i_file_acl);
 514                 error = -EIO;
 515                 if (!bh)
 516                         goto cleanup;
 517                 ea_bdebug(bh, "b_count=%d, refcount=%d",
 518                         atomic_read(&(bh->b_count)),
 519                         le32_to_cpu(HDR(bh)->h_refcount));
 520                 header = HDR(bh);
 521                 end = bh->b_data + bh->b_size;
 522                 if (header->h_magic != cpu_to_le32(EXT2_XATTR_MAGIC) ||
 523                     header->h_blocks != cpu_to_le32(1)) {
 524 bad_block:              ext2_error(sb, "ext2_xattr_set",
 525                                 "inode %ld: bad block %d", inode->i_ino,
 526                                    EXT2_I(inode)->i_file_acl);
 527                         error = -EIO;
 528                         goto cleanup;
 529                 }
 530                 /* Find the named attribute. */
 531                 here = FIRST_ENTRY(bh);
 532                 while (!IS_LAST_ENTRY(here)) {
 533                         struct ext2_xattr_entry *next = EXT2_XATTR_NEXT(here);
 534                         if ((char *)next >= end)
 535                                 goto bad_block;
 536                         if (!here->e_value_block && here->e_value_size) {
 537                                 size_t offs = le16_to_cpu(here->e_value_offs);
 538                                 if (offs < min_offs)
 539                                         min_offs = offs;
 540                         }
 541                         not_found = name_index - here->e_name_index;
 542                         if (!not_found)
 543                                 not_found = name_len - here->e_name_len;
 544                         if (!not_found)
 545                                 not_found = memcmp(name, here->e_name,name_len);
 546                         if (not_found <= 0)
 547                                 break;
 548                         here = next;
 549                 }
 550                 last = here;
 551                 /* We still need to compute min_offs and last. */
 552                 while (!IS_LAST_ENTRY(last)) {
 553                         struct ext2_xattr_entry *next = EXT2_XATTR_NEXT(last);
 554                         if ((char *)next >= end)
 555                                 goto bad_block;
 556                         if (!last->e_value_block && last->e_value_size) {
 557                                 size_t offs = le16_to_cpu(last->e_value_offs);
 558                                 if (offs < min_offs)
 559                                         min_offs = offs;
 560                         }
 561                         last = next;
 562                 }
 563
 564                 /* Check whether we have enough space left. */
 565                 free = min_offs - ((char*)last - (char*)header) - sizeof(__u32);
 566         } else {
 567                 /* We will use a new extended attribute block. */
 568                 free = sb->s_blocksize -
 569                         sizeof(struct ext2_xattr_header) - sizeof(__u32);
 570                 here = last = NULL;  /* avoid gcc uninitialized warning. */
 571         }
 572
 573         if (not_found) {
 574                 /* Request to remove a nonexistent attribute? */
 575                 error = -ENODATA;
 576                 if (flags & XATTR_REPLACE)
 577                         goto cleanup;
 578                 error = 0;
 579                 if (value == NULL)
 580                         goto cleanup;
 581         } else {
 582                 /* Request to create an existing attribute? */
 583                 error = -EEXIST;
 584                 if (flags & XATTR_CREATE)
 585                         goto cleanup;
 586                 if (!here->e_value_block && here->e_value_size) {
 587                         size_t size = le32_to_cpu(here->e_value_size);
 588
 589                         if (le16_to_cpu(here->e_value_offs) + size >
 590                             sb->s_blocksize || size > sb->s_blocksize)
 591                                 goto bad_block;
 592                         free += EXT2_XATTR_SIZE(size);
 593                 }
 594                 free += EXT2_XATTR_LEN(name_len);
 595         }
 596         error = -ENOSPC;
 597         if (free < EXT2_XATTR_LEN(name_len) + EXT2_XATTR_SIZE(value_len))
 598                 goto cleanup;
 599
 600         /* Here we know that we can set the new attribute. */
 601
 602         if (header) {
 603                 /* assert(header == HDR(bh)); */
 604                 lock_buffer(bh);
 605                 if (header->h_refcount == cpu_to_le32(1)) {
 606                         ea_bdebug(bh, "modifying in-place");
 607                         ext2_xattr_cache_remove(bh);
 608                         /* keep the buffer locked while modifying it. */
 609                 } else {
 610                         int offset;
 611
 612                         unlock_buffer(bh);
 613                         ea_bdebug(bh, "cloning");
 614                         header = kmalloc(bh->b_size, GFP_KERNEL);
 615                         error = -ENOMEM;
 616                         if (header == NULL)
 617                                 goto cleanup;
 618                         memcpy(header, HDR(bh), bh->b_size);
 619                         header->h_refcount = cpu_to_le32(1);
 620
 621                         offset = (char *)here - bh->b_data;
 622                         here = ENTRY((char *)header + offset);
 623                         offset = (char *)last - bh->b_data;
 624                         last = ENTRY((char *)header + offset);
 625                 }
 626         } else {
 627                 /* Allocate a buffer where we construct the new block. */
 628                 header = kmalloc(sb->s_blocksize, GFP_KERNEL);
 629                 error = -ENOMEM;
 630                 if (header == NULL)
 631                         goto cleanup;
 632                 memset(header, 0, sb->s_blocksize);
 633                 end = (char *)header + sb->s_blocksize;
 634                 header->h_magic = cpu_to_le32(EXT2_XATTR_MAGIC);
 635                 header->h_blocks = header->h_refcount = cpu_to_le32(1);
 636                 last = here = ENTRY(header+1);
 637         }
 638
 639         /* Iff we are modifying the block in-place, bh is locked here. */
 640
 641         if (not_found) {
 642                 /* Insert the new name. */
 643                 size_t size = EXT2_XATTR_LEN(name_len);
 644                 size_t rest = (char *)last - (char *)here;
 645                 memmove((char *)here + size, here, rest);
 646                 memset(here, 0, size);
 647                 here->e_name_index = name_index;
 648                 here->e_name_len = name_len;
 649                 memcpy(here->e_name, name, name_len);
 650         } else {
 651                 if (!here->e_value_block && here->e_value_size) {
 652                         char *first_val = (char *)header + min_offs;
 653                         size_t offs = le16_to_cpu(here->e_value_offs);
 654                         char *val = (char *)header + offs;
 655                         size_t size = EXT2_XATTR_SIZE(
 656                                 le32_to_cpu(here->e_value_size));
 657
 658                         if (size == EXT2_XATTR_SIZE(value_len)) {
 659                                 /* The old and the new value have the same
 660                                    size. Just replace. */
 661                                 here->e_value_size = cpu_to_le32(value_len);
 662                                 memset(val + size - EXT2_XATTR_PAD, 0,
 663                                        EXT2_XATTR_PAD); /* Clear pad bytes. */
 664                                 memcpy(val, value, value_len);
 665                                 goto skip_replace;
 666                         }
 667
 668                         /* Remove the old value. */
 669                         memmove(first_val + size, first_val, val - first_val);
 670                         memset(first_val, 0, size);
 671                         here->e_value_offs = 0;
 672                         min_offs += size;
 673
 674                         /* Adjust all value offsets. */
 675                         last = ENTRY(header+1);
 676                         while (!IS_LAST_ENTRY(last)) {
 677                                 size_t o = le16_to_cpu(last->e_value_offs);
 678                                 if (!last->e_value_block && o < offs)
 679                                         last->e_value_offs =
 680                                                 cpu_to_le16(o + size);
 681                                 last = EXT2_XATTR_NEXT(last);
 682                         }
 683                 }
 684                 if (value == NULL) {
 685                         /* Remove the old name. */
 686                         size_t size = EXT2_XATTR_LEN(name_len);
 687                         last = ENTRY((char *)last - size);
 688                         memmove(here, (char*)here + size,
 689                                 (char*)last - (char*)here);
 690                         memset(last, 0, size);
 691                 }
 692         }
 693
 694         if (value != NULL) {
 695                 /* Insert the new value. */
 696                 here->e_value_size = cpu_to_le32(value_len);
 697                 if (value_len) {
 698                         size_t size = EXT2_XATTR_SIZE(value_len);
 699                         char *val = (char *)header + min_offs - size;
 700                         here->e_value_offs =
 701                                 cpu_to_le16((char *)val - (char *)header);
 702                         memset(val + size - EXT2_XATTR_PAD, 0,
 703                                EXT2_XATTR_PAD); /* Clear the pad bytes. */
 704                         memcpy(val, value, value_len);
 705                 }
 706         }
 707
 708 skip_replace:
 709         if (IS_LAST_ENTRY(ENTRY(header+1))) {
 710                 /* This block is now empty. */
 711                 if (bh && header == HDR(bh))
 712                         unlock_buffer(bh);  /* we were modifying in-place. */
 713                 error = ext2_xattr_set2(inode, bh, NULL);
 714         } else {
 715                 ext2_xattr_rehash(header, here);
 716                 if (bh && header == HDR(bh))
 717                         unlock_buffer(bh);  /* we were modifying in-place. */
 718                 error = ext2_xattr_set2(inode, bh, header);
 719         }
 720
 721 cleanup:
 722         brelse(bh);
 723         if (!(bh && header == HDR(bh)))
 724                 kfree(header);
 725         up_write(&EXT2_I(inode)->xattr_sem);
 726
 727         return error;
 728 }
 729
 730 /*
 731  * Second half of ext2_xattr_set(): Update the file system.
 732  */
 733 static int
 734 ext2_xattr_set2(struct inode *inode, struct buffer_head *old_bh,
 735                 struct ext2_xattr_header *header)
 736 {
 737         struct super_block *sb = inode->i_sb;
 738         struct buffer_head *new_bh = NULL;
 739         int error;
 740
 741         if (header) {
 742                 new_bh = ext2_xattr_cache_find(inode, header);
 743                 if (new_bh) {
 744                         /* We found an identical block in the cache. */
 745                         if (new_bh == old_bh) {
 746                                 ea_bdebug(new_bh, "keeping this block");
 747                         } else {
 748                                 /* The old block is released after updating
 749                                    the inode.  */
 750                                 ea_bdebug(new_bh, "reusing block");
 751
 752                                 error = -EDQUOT;
 753                                 if (DQUOT_ALLOC_BLOCK(inode, 1)) {
 754                                         unlock_buffer(new_bh);
 755                                         goto cleanup;
 756                                 }
 757                                 HDR(new_bh)->h_refcount = cpu_to_le32(1 +
 758                                         le32_to_cpu(HDR(new_bh)->h_refcount));
 759                                 ea_bdebug(new_bh, "refcount now=%d",
 760                                         le32_to_cpu(HDR(new_bh)->h_refcount));
 761                         }
 762                         unlock_buffer(new_bh);
 763                 } else if (old_bh && header == HDR(old_bh)) {
 764                         /* Keep this block. No need to lock the block as we
 765                            don't need to change the reference count. */
 766                         new_bh = old_bh;
 767                         get_bh(new_bh);
 768                         ext2_xattr_cache_insert(new_bh);
 769                 } else {
 770                         /* We need to allocate a new block */
 771                         int goal = le32_to_cpu(EXT2_SB(sb)->s_es->
 772                                                            s_first_data_block) +
 773                                    EXT2_I(inode)->i_block_group *
 774                                    EXT2_BLOCKS_PER_GROUP(sb);
 775                         int block = ext2_new_block(inode, goal, 0, 0, &error);
 776                         if (error)
 777                                 goto cleanup;
 778                         ea_idebug(inode, "creating block %d", block);
 779
 780                         new_bh = sb_getblk(sb, block);
 781                         if (!new_bh) {
 782                                 ext2_free_blocks(inode, block, 1);
 783                                 error = -EIO;
 784                                 goto cleanup;
 785                         }
 786                         lock_buffer(new_bh);
 787                         memcpy(new_bh->b_data, header, new_bh->b_size);
 788                         set_buffer_uptodate(new_bh);
 789                         unlock_buffer(new_bh);
 790                         ext2_xattr_cache_insert(new_bh);
 791
 792                         ext2_xattr_update_super_block(sb);
 793                 }
 794                 mark_buffer_dirty(new_bh);
 795                 if (IS_SYNC(inode)) {
 796                         sync_dirty_buffer(new_bh);
 797                         error = -EIO;
 798                         if (buffer_req(new_bh) && !buffer_uptodate(new_bh))
 799                                 goto cleanup;
 800                 }
 801         }
 802
 803         /* Update the inode. */
 804         EXT2_I(inode)->i_file_acl = new_bh ? new_bh->b_blocknr : 0;
 805         inode->i_ctime = CURRENT_TIME;
 806         if (IS_SYNC(inode)) {
 807                 error = ext2_sync_inode (inode);
 808                 if (error)
 809                         goto cleanup;
 810         } else
 811                 mark_inode_dirty(inode);
 812
 813         error = 0;
 814         if (old_bh && old_bh != new_bh) {
 815                 /*
 816                  * If there was an old block and we are no longer using it,
 817                  * release the old block.
 818                  */
 819                 lock_buffer(old_bh);
 820                 if (HDR(old_bh)->h_refcount == cpu_to_le32(1)) {
 821                         /* Free the old block. */
 822                         ea_bdebug(old_bh, "freeing");
 823                         ext2_free_blocks(inode, old_bh->b_blocknr, 1);
 824                         /* We let our caller release old_bh, so we
 825                          * need to duplicate the buffer before. */
 826                         get_bh(old_bh);
 827                         bforget(old_bh);
 828                 } else {
 829                         /* Decrement the refcount only. */
 830                         HDR(old_bh)->h_refcount = cpu_to_le32(
 831                                 le32_to_cpu(HDR(old_bh)->h_refcount) - 1);
 832                         DQUOT_FREE_BLOCK(inode, 1);
 833                         mark_buffer_dirty(old_bh);
 834                         ea_bdebug(old_bh, "refcount now=%d",
 835                                 le32_to_cpu(HDR(old_bh)->h_refcount));
 836                 }
 837                 unlock_buffer(old_bh);
 838         }
 839
 840 cleanup:
 841         brelse(new_bh);
 842
 843         return error;
 844 }
 845
 846 /*
 847  * ext2_xattr_delete_inode()
 848  *
 849  * Free extended attribute resources associated with this inode. This
 850  * is called immediately before an inode is freed.
 851  */
 852 void
 853 ext2_xattr_delete_inode(struct inode *inode)
 854 {
 855         struct buffer_head *bh = NULL;
 856
 857         down_write(&EXT2_I(inode)->xattr_sem);
 858         if (!EXT2_I(inode)->i_file_acl)
 859                 goto cleanup;
 860         bh = sb_bread(inode->i_sb, EXT2_I(inode)->i_file_acl);
 861         if (!bh) {
 862                 ext2_error(inode->i_sb, "ext2_xattr_delete_inode",
 863                         "inode %ld: block %d read error", inode->i_ino,
 864                         EXT2_I(inode)->i_file_acl);
 865                 goto cleanup;
 866         }
 867         ea_bdebug(bh, "b_count=%d", atomic_read(&(bh->b_count)));
 868         if (HDR(bh)->h_magic != cpu_to_le32(EXT2_XATTR_MAGIC) ||
 869             HDR(bh)->h_blocks != cpu_to_le32(1)) {
 870                 ext2_error(inode->i_sb, "ext2_xattr_delete_inode",
 871                         "inode %ld: bad block %d", inode->i_ino,
 872                         EXT2_I(inode)->i_file_acl);
 873                 goto cleanup;
 874         }
 875         lock_buffer(bh);
 876         if (HDR(bh)->h_refcount == cpu_to_le32(1)) {
 877                 ext2_xattr_cache_remove(bh);
 878                 ext2_free_blocks(inode, EXT2_I(inode)->i_file_acl, 1);
 879                 get_bh(bh);
 880                 bforget(bh);
 881         } else {
 882                 HDR(bh)->h_refcount = cpu_to_le32(
 883                         le32_to_cpu(HDR(bh)->h_refcount) - 1);
 884                 mark_buffer_dirty(bh);
 885                 if (IS_SYNC(inode))
 886                         sync_dirty_buffer(bh);
 887                 DQUOT_FREE_BLOCK(inode, 1);
 888         }
 889         ea_bdebug(bh, "refcount now=%d", le32_to_cpu(HDR(bh)->h_refcount) - 1);
 890         unlock_buffer(bh);
 891         EXT2_I(inode)->i_file_acl = 0;
 892
 893 cleanup:
 894         brelse(bh);
 895         up_write(&EXT2_I(inode)->xattr_sem);
 896 }
 897
 898 /*
 899  * ext2_xattr_put_super()
 900  *
 901  * This is called when a file system is unmounted.
 902  */
 903 void
 904 ext2_xattr_put_super(struct super_block *sb)
 905 {
 906         mb_cache_shrink(ext2_xattr_cache, sb->s_bdev);
 907 }
 908
 909
 910 /*
 911  * ext2_xattr_cache_insert()
 912  *
 913  * Create a new entry in the extended attribute cache, and insert
 914  * it unless such an entry is already in the cache.
 915  *
 916  * Returns 0, or a negative error number on failure.
 917  */
 918 static int
 919 ext2_xattr_cache_insert(struct buffer_head *bh)
 920 {
 921         __u32 hash = le32_to_cpu(HDR(bh)->h_hash);
 922         struct mb_cache_entry *ce;
 923         int error;
 924
 925         ce = mb_cache_entry_alloc(ext2_xattr_cache);
 926         if (!ce)
 927                 return -ENOMEM;
 928         error = mb_cache_entry_insert(ce, bh->b_bdev, bh->b_blocknr, &hash);
 929         if (error) {
 930                 mb_cache_entry_free(ce);
 931                 if (error == -EBUSY) {
 932                         ea_bdebug(bh, "already in cache (%d cache entries)",
 933                                 atomic_read(&ext2_xattr_cache->c_entry_count));
 934                         error = 0;
 935                 }
 936         } else {
 937                 ea_bdebug(bh, "inserting [%x] (%d cache entries)", (int)hash,
 938                           atomic_read(&ext2_xattr_cache->c_entry_count));
 939                 mb_cache_entry_release(ce);
 940         }
 941         return error;
 942 }
 943
 944 /*
 945  * ext2_xattr_cmp()
 946  *
 947  * Compare two extended attribute blocks for equality.
 948  *
 949  * Returns 0 if the blocks are equal, 1 if they differ, and
 950  * a negative error number on errors.
 951  */
 952 static int
 953 ext2_xattr_cmp(struct ext2_xattr_header *header1,
 954                struct ext2_xattr_header *header2)
 955 {
 956         struct ext2_xattr_entry *entry1, *entry2;
 957
 958         entry1 = ENTRY(header1+1);
 959         entry2 = ENTRY(header2+1);
 960         while (!IS_LAST_ENTRY(entry1)) {
 961                 if (IS_LAST_ENTRY(entry2))
 962                         return 1;
 963                 if (entry1->e_hash != entry2->e_hash ||
 964                     entry1->e_name_len != entry2->e_name_len ||
 965                     entry1->e_value_size != entry2->e_value_size ||
 966                     memcmp(entry1->e_name, entry2->e_name, entry1->e_name_len))
 967                         return 1;
 968                 if (entry1->e_value_block != 0 || entry2->e_value_block != 0)
 969                         return -EIO;
 970                 if (memcmp((char *)header1 + le16_to_cpu(entry1->e_value_offs),
 971                            (char *)header2 + le16_to_cpu(entry2->e_value_offs),
 972                            le32_to_cpu(entry1->e_value_size)))
 973                         return 1;
 974
 975                 entry1 = EXT2_XATTR_NEXT(entry1);
 976                 entry2 = EXT2_XATTR_NEXT(entry2);
 977         }
 978         if (!IS_LAST_ENTRY(entry2))
 979                 return 1;
 980         return 0;
 981 }
 982
 983 /*
 984  * ext2_xattr_cache_find()
 985  *
 986  * Find an identical extended attribute block.
 987  *
 988  * Returns a locked buffer head to the block found, or NULL if such
 989  * a block was not found or an error occurred.
 990  */
 991 static struct buffer_head *
 992 ext2_xattr_cache_find(struct inode *inode, struct ext2_xattr_header *header)
 993 {
 994         __u32 hash = le32_to_cpu(header->h_hash);
 995         struct mb_cache_entry *ce;
 996
 997         if (!header->h_hash)
 998                 return NULL;  /* never share */
 999         ea_idebug(inode, "looking for cached blocks [%x]", (int)hash);
1000         ce = mb_cache_entry_find_first(ext2_xattr_cache, 0,
1001                                        inode->i_sb->s_bdev, hash);
1002         while (ce) {
1003                 struct buffer_head *bh = sb_bread(inode->i_sb, ce->e_block);
1004
1005                 if (!bh) {
1006                         ext2_error(inode->i_sb, "ext2_xattr_cache_find",
1007                                 "inode %ld: block %ld read error",
1008                                 inode->i_ino, (unsigned long) ce->e_block);
1009                 } else {
1010                         lock_buffer(bh);
1011                         if (le32_to_cpu(HDR(bh)->h_refcount) >
1012                                    EXT2_XATTR_REFCOUNT_MAX) {
1013                                 ea_idebug(inode, "block %ld refcount %d>%d",
1014                                           (unsigned long) ce->e_block,
1015                                           le32_to_cpu(HDR(bh)->h_refcount),
1016                                           EXT2_XATTR_REFCOUNT_MAX);
1017                         } else if (!ext2_xattr_cmp(header, HDR(bh))) {
1018                                 ea_bdebug(bh, "b_count=%d",
1019                                           atomic_read(&(bh->b_count)));
1020                                 mb_cache_entry_release(ce);
1021                                 return bh;
1022                         }
1023                         unlock_buffer(bh);
1024                         brelse(bh);
1025                 }
1026                 ce = mb_cache_entry_find_next(ce, 0, inode->i_sb->s_bdev, hash);
1027         }
1028         return NULL;
1029 }
1030
1031 /*
1032  * ext2_xattr_cache_remove()
1033  *
1034  * Remove the cache entry of a block from the cache. Called when a
1035  * block becomes invalid.
1036  */
1037 static void
1038 ext2_xattr_cache_remove(struct buffer_head *bh)
1039 {
1040         struct mb_cache_entry *ce;
1041
1042         ce = mb_cache_entry_get(ext2_xattr_cache, bh->b_bdev, bh->b_blocknr);
1043         if (ce) {
1044                 ea_bdebug(bh, "removing (%d cache entries remaining)",
1045                           atomic_read(&ext2_xattr_cache->c_entry_count)-1);
1046                 mb_cache_entry_free(ce);
1047         } else
1048                 ea_bdebug(bh, "no cache entry");
1049 }
1050
1051 #define NAME_HASH_SHIFT 5
1052 #define VALUE_HASH_SHIFT 16
1053
1054 /*
1055  * ext2_xattr_hash_entry()
1056  *
1057  * Compute the hash of an extended attribute.
1058  */
1059 static inline void ext2_xattr_hash_entry(struct ext2_xattr_header *header,
1060                                          struct ext2_xattr_entry *entry)
1061 {
1062         __u32 hash = 0;
1063         char *name = entry->e_name;
1064         int n;
1065
1066         for (n=0; n < entry->e_name_len; n++) {
1067                 hash = (hash << NAME_HASH_SHIFT) ^
1068                        (hash >> (8*sizeof(hash) - NAME_HASH_SHIFT)) ^
1069                        *name++;
1070         }
1071
1072         if (entry->e_value_block == 0 && entry->e_value_size != 0) {
1073                 __u32 *value = (__u32 *)((char *)header +
1074                         le16_to_cpu(entry->e_value_offs));
1075                 for (n = (le32_to_cpu(entry->e_value_size) +
1076                      EXT2_XATTR_ROUND) >> EXT2_XATTR_PAD_BITS; n; n--) {
1077                         hash = (hash << VALUE_HASH_SHIFT) ^
1078                                (hash >> (8*sizeof(hash) - VALUE_HASH_SHIFT)) ^
1079                                le32_to_cpu(*value++);
1080                 }
1081         }
1082         entry->e_hash = cpu_to_le32(hash);
1083 }
1084
1085 #undef NAME_HASH_SHIFT
1086 #undef VALUE_HASH_SHIFT
1087
1088 #define BLOCK_HASH_SHIFT 16
1089
1090 /*
1091  * ext2_xattr_rehash()
1092  *
1093  * Re-compute the extended attribute hash value after an entry has changed.
1094  */
1095 static void ext2_xattr_rehash(struct ext2_xattr_header *header,
1096                               struct ext2_xattr_entry *entry)
1097 {
1098         struct ext2_xattr_entry *here;
1099         __u32 hash = 0;
1100
1101         ext2_xattr_hash_entry(header, entry);
1102         here = ENTRY(header+1);
1103         while (!IS_LAST_ENTRY(here)) {
1104                 if (!here->e_hash) {
1105                         /* Block is not shared if an entry's hash value == 0 */
1106                         hash = 0;
1107                         break;
1108                 }
1109                 hash = (hash << BLOCK_HASH_SHIFT) ^
1110                        (hash >> (8*sizeof(hash) - BLOCK_HASH_SHIFT)) ^
1111                        le32_to_cpu(here->e_hash);
1112                 here = EXT2_XATTR_NEXT(here);
1113         }
1114         header->h_hash = cpu_to_le32(hash);
1115 }
1116
1117 #undef BLOCK_HASH_SHIFT
1118
1119 int __init
1120 init_ext2_xattr(void)
1121 {
1122         int     err;
1123
1124         err = ext2_xattr_register(EXT2_XATTR_INDEX_USER,
1125                                   &ext2_xattr_user_handler);
1126         if (err)
1127                 return err;
1128         err = ext2_xattr_register(EXT2_XATTR_INDEX_TRUSTED,
1129                                   &ext2_xattr_trusted_handler);
1130         if (err)
1131                 goto out;
1132 #ifdef CONFIG_EXT2_FS_SECURITY
1133         err = ext2_xattr_register(EXT2_XATTR_INDEX_SECURITY,
1134                                   &ext2_xattr_security_handler);
1135         if (err)
1136                 goto out1;
1137 #endif
1138 #ifdef CONFIG_EXT2_FS_POSIX_ACL
1139         err = init_ext2_acl();
1140         if (err)
1141                 goto out2;
1142 #endif
1143         ext2_xattr_cache = mb_cache_create("ext2_xattr", NULL,
1144                 sizeof(struct mb_cache_entry) +
1145                 sizeof(struct mb_cache_entry_index), 1, 6);
1146         if (!ext2_xattr_cache) {
1147                 err = -ENOMEM;
1148                 goto out3;
1149         }
1150         return 0;
1151 out3:
1152 #ifdef CONFIG_EXT2_FS_POSIX_ACL
1153         exit_ext2_acl();
1154 out2:
1155 #endif
1156 #ifdef CONFIG_EXT2_FS_SECURITY
1157         ext2_xattr_unregister(EXT2_XATTR_INDEX_SECURITY,
1158                               &ext2_xattr_security_handler);
1159 out1:
1160 #endif
1161         ext2_xattr_unregister(EXT2_XATTR_INDEX_TRUSTED,
1162                               &ext2_xattr_trusted_handler);
1163 out:
1164         ext2_xattr_unregister(EXT2_XATTR_INDEX_USER,
1165                               &ext2_xattr_user_handler);
1166         return err;
1167 }
1168
1169 void
1170 exit_ext2_xattr(void)
1171 {
1172         mb_cache_destroy(ext2_xattr_cache);
1173 #ifdef CONFIG_EXT2_FS_POSIX_ACL
1174         exit_ext2_acl();
1175 #endif
1176 #ifdef CONFIG_EXT2_FS_SECURITY
1177         ext2_xattr_unregister(EXT2_XATTR_INDEX_SECURITY,
1178                               &ext2_xattr_security_handler);
1179 #endif
1180         ext2_xattr_unregister(EXT2_XATTR_INDEX_TRUSTED,
1181                               &ext2_xattr_trusted_handler);
1182         ext2_xattr_unregister(EXT2_XATTR_INDEX_USER,
1183                               &ext2_xattr_user_handler);
1184 }