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  * xattr consolidation Copyright (c) 2004 James Morris <jmorris@redhat.com>,
  10  *  Red Hat Inc.
  11  *
  12  */
  13
  14 /*
  15  * Extended attributes are stored on disk blocks allocated outside of
  16  * any inode. The i_file_acl field is then made to point to this allocated
  17  * block. If all extended attributes of an inode are identical, these
  18  * inodes may share the same extended attribute block. Such situations
  19  * are automatically detected by keeping a cache of recent attribute block
  20  * numbers and hashes over the block's contents in memory.
  21  *
  22  *
  23  * Extended attribute block layout:
  24  *
  25  *   +------------------+
  26  *   | header           |
  27  *   | entry 1          | |
  28  *   | entry 2          | | growing downwards
  29  *   | entry 3          | v
  30  *   | four null bytes  |
  31  *   | . . .            |
  32  *   | value 1          | ^
  33  *   | value 3          | | growing upwards
  34  *   | value 2          | |
  35  *   +------------------+
  36  *
  37  * The block header is followed by multiple entry descriptors. These entry
  38  * descriptors are variable in size, and alligned to EXT2_XATTR_PAD
  39  * byte boundaries. The entry descriptors are sorted by attribute name,
  40  * so that two extended attribute blocks can be compared efficiently.
  41  *
  42  * Attribute values are aligned to the end of the block, stored in
  43  * no specific order. They are also padded to EXT2_XATTR_PAD byte
  44  * boundaries. No additional gaps are left between them.
  45  *
  46  * Locking strategy
  47  * ----------------
  48  * EXT2_I(inode)->i_file_acl is protected by EXT2_I(inode)->xattr_sem.
  49  * EA blocks are only changed if they are exclusive to an inode, so
  50  * holding xattr_sem also means that nothing but the EA block's reference
  51  * count will change. Multiple writers to an EA block are synchronized
  52  * by the bh lock. No more than a single bh lock is held at any time
  53  * to avoid deadlocks.
  54  */
  55
  56 #include <linux/buffer_head.h>
  57 #include <linux/module.h>
  58 #include <linux/init.h>
  59 #include <linux/slab.h>
  60 #include <linux/mbcache.h>
  61 #include <linux/quotaops.h>
  62 #include <linux/rwsem.h>
  63 #include <linux/security.h>
  64 #include "ext2.h"
  65 #include "xattr.h"
  66 #include "acl.h"
  67
  68 #define HDR(bh) ((struct ext2_xattr_header *)((bh)->b_data))
  69 #define ENTRY(ptr) ((struct ext2_xattr_entry *)(ptr))
  70 #define FIRST_ENTRY(bh) ENTRY(HDR(bh)+1)
  71 #define IS_LAST_ENTRY(entry) (*(__u32 *)(entry) == 0)
  72
  73 #ifdef EXT2_XATTR_DEBUG
  74 # define ea_idebug(inode, f...) do { \
  75                 printk(KERN_DEBUG "inode %s:%ld: ", \
  76                         inode->i_sb->s_id, inode->i_ino); \
  77                 printk(f); \
  78                 printk("\n"); \
  79         } while (0)
  80 # define ea_bdebug(bh, f...) do { \
  81                 char b[BDEVNAME_SIZE]; \
  82                 printk(KERN_DEBUG "block %s:%lu: ", \
  83                         bdevname(bh->b_bdev, b), \
  84                         (unsigned long) bh->b_blocknr); \
  85                 printk(f); \
  86                 printk("\n"); \
  87         } while (0)
  88 #else
  89 # define ea_idebug(f...)
  90 # define ea_bdebug(f...)
  91 #endif
  92
  93 static int ext2_xattr_set2(struct inode *, struct buffer_head *,
  94                            struct ext2_xattr_header *);
  95
  96 static int ext2_xattr_cache_insert(struct buffer_head *);
  97 static struct buffer_head *ext2_xattr_cache_find(struct inode *,
  98                                                  struct ext2_xattr_header *);
  99 static void ext2_xattr_rehash(struct ext2_xattr_header *,
 100                               struct ext2_xattr_entry *);
 101
 102 static struct mb_cache *ext2_xattr_cache;
 103
 104 static const struct xattr_handler *ext2_xattr_handler_map[] = {
 105         [EXT2_XATTR_INDEX_USER]              = &ext2_xattr_user_handler,
 106 #ifdef CONFIG_EXT2_FS_POSIX_ACL
 107         [EXT2_XATTR_INDEX_POSIX_ACL_ACCESS]  = &ext2_xattr_acl_access_handler,
 108         [EXT2_XATTR_INDEX_POSIX_ACL_DEFAULT] = &ext2_xattr_acl_default_handler,
 109 #endif
 110         [EXT2_XATTR_INDEX_TRUSTED]           = &ext2_xattr_trusted_handler,
 111 #ifdef CONFIG_EXT2_FS_SECURITY
 112         [EXT2_XATTR_INDEX_SECURITY]          = &ext2_xattr_security_handler,
 113 #endif
 114 };
 115
 116 const struct xattr_handler *ext2_xattr_handlers[] = {
 117         &ext2_xattr_user_handler,
 118         &ext2_xattr_trusted_handler,
 119 #ifdef CONFIG_EXT2_FS_POSIX_ACL
 120         &ext2_xattr_acl_access_handler,
 121         &ext2_xattr_acl_default_handler,
 122 #endif
 123 #ifdef CONFIG_EXT2_FS_SECURITY
 124         &ext2_xattr_security_handler,
 125 #endif
 126         NULL
 127 };
 128
 129 static inline const struct xattr_handler *
 130 ext2_xattr_handler(int name_index)
 131 {
 132         const struct xattr_handler *handler = NULL;
 133
 134         if (name_index > 0 && name_index < ARRAY_SIZE(ext2_xattr_handler_map))
 135                 handler = ext2_xattr_handler_map[name_index];
 136         return handler;
 137 }
 138
 139 /*
 140  * ext2_xattr_get()
 141  *
 142  * Copy an extended attribute into the buffer
 143  * provided, or compute the buffer size required.
 144  * Buffer is NULL to compute the size of the buffer required.
 145  *
 146  * Returns a negative error number on failure, or the number of bytes
 147  * used / required on success.
 148  */
 149 int
 150 ext2_xattr_get(struct inode *inode, int name_index, const char *name,
 151                void *buffer, size_t buffer_size)
 152 {
 153         struct buffer_head *bh = NULL;
 154         struct ext2_xattr_entry *entry;
 155         size_t name_len, size;
 156         char *end;
 157         int error;
 158
 159         ea_idebug(inode, "name=%d.%s, buffer=%p, buffer_size=%ld",
 160                   name_index, name, buffer, (long)buffer_size);
 161
 162         if (name == NULL)
 163                 return -EINVAL;
 164         down_read(&EXT2_I(inode)->xattr_sem);
 165         error = -ENODATA;
 166         if (!EXT2_I(inode)->i_file_acl)
 167                 goto cleanup;
 168         ea_idebug(inode, "reading block %d", EXT2_I(inode)->i_file_acl);
 169         bh = sb_bread(inode->i_sb, EXT2_I(inode)->i_file_acl);
 170         error = -EIO;
 171         if (!bh)
 172                 goto cleanup;
 173         ea_bdebug(bh, "b_count=%d, refcount=%d",
 174                 atomic_read(&(bh->b_count)), le32_to_cpu(HDR(bh)->h_refcount));
 175         end = bh->b_data + bh->b_size;
 176         if (HDR(bh)->h_magic != cpu_to_le32(EXT2_XATTR_MAGIC) ||
 177             HDR(bh)->h_blocks != cpu_to_le32(1)) {
 178 bad_block:      ext2_error(inode->i_sb, "ext2_xattr_get",
 179                         "inode %ld: bad block %d", inode->i_ino,
 180                         EXT2_I(inode)->i_file_acl);
 181                 error = -EIO;
 182                 goto cleanup;
 183         }
 184         /* find named attribute */
 185         name_len = strlen(name);
 186
 187         error = -ERANGE;
 188         if (name_len > 255)
 189                 goto cleanup;
 190         entry = FIRST_ENTRY(bh);
 191         while (!IS_LAST_ENTRY(entry)) {
 192                 struct ext2_xattr_entry *next =
 193                         EXT2_XATTR_NEXT(entry);
 194                 if ((char *)next >= end)
 195                         goto bad_block;
 196                 if (name_index == entry->e_name_index &&
 197                     name_len == entry->e_name_len &&
 198                     memcmp(name, entry->e_name, name_len) == 0)
 199                         goto found;
 200                 entry = next;
 201         }
 202         if (ext2_xattr_cache_insert(bh))
 203                 ea_idebug(inode, "cache insert failed");
 204         error = -ENODATA;
 205         goto cleanup;
 206 found:
 207         /* check the buffer size */
 208         if (entry->e_value_block != 0)
 209                 goto bad_block;
 210         size = le32_to_cpu(entry->e_value_size);
 211         if (size > inode->i_sb->s_blocksize ||
 212             le16_to_cpu(entry->e_value_offs) + size > inode->i_sb->s_blocksize)
 213                 goto bad_block;
 214
 215         if (ext2_xattr_cache_insert(bh))
 216                 ea_idebug(inode, "cache insert failed");
 217         if (buffer) {
 218                 error = -ERANGE;
 219                 if (size > buffer_size)
 220                         goto cleanup;
 221                 /* return value of attribute */
 222                 memcpy(buffer, bh->b_data + le16_to_cpu(entry->e_value_offs),
 223                         size);
 224         }
 225         error = size;
 226
 227 cleanup:
 228         brelse(bh);
 229         up_read(&EXT2_I(inode)->xattr_sem);
 230
 231         return error;
 232 }
 233
 234 /*
 235  * ext2_xattr_list()
 236  *
 237  * Copy a list of attribute names into the buffer
 238  * provided, or compute the buffer size required.
 239  * Buffer is NULL to compute the size of the buffer required.
 240  *
 241  * Returns a negative error number on failure, or the number of bytes
 242  * used / required on success.
 243  */
 244 static int
 245 ext2_xattr_list(struct dentry *dentry, char *buffer, size_t buffer_size)
 246 {
 247         struct inode *inode = dentry->d_inode;
 248         struct buffer_head *bh = NULL;
 249         struct ext2_xattr_entry *entry;
 250         char *end;
 251         size_t rest = buffer_size;
 252         int error;
 253
 254         ea_idebug(inode, "buffer=%p, buffer_size=%ld",
 255                   buffer, (long)buffer_size);
 256
 257         down_read(&EXT2_I(inode)->xattr_sem);
 258         error = 0;
 259         if (!EXT2_I(inode)->i_file_acl)
 260                 goto cleanup;
 261         ea_idebug(inode, "reading block %d", EXT2_I(inode)->i_file_acl);
 262         bh = sb_bread(inode->i_sb, EXT2_I(inode)->i_file_acl);
 263         error = -EIO;
 264         if (!bh)
 265                 goto cleanup;
 266         ea_bdebug(bh, "b_count=%d, refcount=%d",
 267                 atomic_read(&(bh->b_count)), le32_to_cpu(HDR(bh)->h_refcount));
 268         end = bh->b_data + bh->b_size;
 269         if (HDR(bh)->h_magic != cpu_to_le32(EXT2_XATTR_MAGIC) ||
 270             HDR(bh)->h_blocks != cpu_to_le32(1)) {
 271 bad_block:      ext2_error(inode->i_sb, "ext2_xattr_list",
 272                         "inode %ld: bad block %d", inode->i_ino,
 273                         EXT2_I(inode)->i_file_acl);
 274                 error = -EIO;
 275                 goto cleanup;
 276         }
 277
 278         /* check the on-disk data structure */
 279         entry = FIRST_ENTRY(bh);
 280         while (!IS_LAST_ENTRY(entry)) {
 281                 struct ext2_xattr_entry *next = EXT2_XATTR_NEXT(entry);
 282
 283                 if ((char *)next >= end)
 284                         goto bad_block;
 285                 entry = next;
 286         }
 287         if (ext2_xattr_cache_insert(bh))
 288                 ea_idebug(inode, "cache insert failed");
 289
 290         /* list the attribute names */
 291         for (entry = FIRST_ENTRY(bh); !IS_LAST_ENTRY(entry);
 292              entry = EXT2_XATTR_NEXT(entry)) {
 293                 const struct xattr_handler *handler =
 294                         ext2_xattr_handler(entry->e_name_index);
 295
 296                 if (handler) {
 297                         size_t size = handler->list(dentry, buffer, rest,
 298                                                     entry->e_name,
 299                                                     entry->e_name_len,
 300                                                     handler->flags);
 301                         if (buffer) {
 302                                 if (size > rest) {
 303                                         error = -ERANGE;
 304                                         goto cleanup;
 305                                 }
 306                                 buffer += size;
 307                         }
 308                         rest -= size;
 309                 }
 310         }
 311         error = buffer_size - rest;  /* total size */
 312
 313 cleanup:
 314         brelse(bh);
 315         up_read(&EXT2_I(inode)->xattr_sem);
 316
 317         return error;
 318 }
 319
 320 /*
 321  * Inode operation listxattr()
 322  *
 323  * dentry->d_inode->i_mutex: don't care
 324  */
 325 ssize_t
 326 ext2_listxattr(struct dentry *dentry, char *buffer, size_t size)
 327 {
 328         return ext2_xattr_list(dentry, buffer, size);
 329 }
 330
 331 /*
 332  * If the EXT2_FEATURE_COMPAT_EXT_ATTR feature of this file system is
 333  * not set, set it.
 334  */
 335 static void ext2_xattr_update_super_block(struct super_block *sb)
 336 {
 337         if (EXT2_HAS_COMPAT_FEATURE(sb, EXT2_FEATURE_COMPAT_EXT_ATTR))
 338                 return;
 339
 340         spin_lock(&EXT2_SB(sb)->s_lock);
 341         EXT2_SET_COMPAT_FEATURE(sb, EXT2_FEATURE_COMPAT_EXT_ATTR);
 342         spin_unlock(&EXT2_SB(sb)->s_lock);
 343         sb->s_dirt = 1;
 344         mark_buffer_dirty(EXT2_SB(sb)->s_sbh);
 345 }
 346
 347 /*
 348  * ext2_xattr_set()
 349  *
 350  * Create, replace or remove an extended attribute for this inode.  Value
 351  * is NULL to remove an existing extended attribute, and non-NULL to
 352  * either replace an existing extended attribute, or create a new extended
 353  * attribute. The flags XATTR_REPLACE and XATTR_CREATE
 354  * specify that an extended attribute must exist and must not exist
 355  * previous to the call, respectively.
 356  *
 357  * Returns 0, or a negative error number on failure.
 358  */
 359 int
 360 ext2_xattr_set(struct inode *inode, int name_index, const char *name,
 361                const void *value, size_t value_len, int flags)
 362 {
 363         struct super_block *sb = inode->i_sb;
 364         struct buffer_head *bh = NULL;
 365         struct ext2_xattr_header *header = NULL;
 366         struct ext2_xattr_entry *here, *last;
 367         size_t name_len, free, min_offs = sb->s_blocksize;
 368         int not_found = 1, error;
 369         char *end;
 370
 371         /*
 372          * header -- Points either into bh, or to a temporarily
 373          *           allocated buffer.
 374          * here -- The named entry found, or the place for inserting, within
 375          *         the block pointed to by header.
 376          * last -- Points right after the last named entry within the block
 377          *         pointed to by header.
 378          * min_offs -- The offset of the first value (values are aligned
 379          *             towards the end of the block).
 380          * end -- Points right after the block pointed to by header.
 381          */
 382
 383         ea_idebug(inode, "name=%d.%s, value=%p, value_len=%ld",
 384                   name_index, name, value, (long)value_len);
 385
 386         if (value == NULL)
 387                 value_len = 0;
 388         if (name == NULL)
 389                 return -EINVAL;
 390         name_len = strlen(name);
 391         if (name_len > 255 || value_len > sb->s_blocksize)
 392                 return -ERANGE;
 393         down_write(&EXT2_I(inode)->xattr_sem);
 394         if (EXT2_I(inode)->i_file_acl) {
 395                 /* The inode already has an extended attribute block. */
 396                 bh = sb_bread(sb, EXT2_I(inode)->i_file_acl);
 397                 error = -EIO;
 398                 if (!bh)
 399                         goto cleanup;
 400                 ea_bdebug(bh, "b_count=%d, refcount=%d",
 401                         atomic_read(&(bh->b_count)),
 402                         le32_to_cpu(HDR(bh)->h_refcount));
 403                 header = HDR(bh);
 404                 end = bh->b_data + bh->b_size;
 405                 if (header->h_magic != cpu_to_le32(EXT2_XATTR_MAGIC) ||
 406                     header->h_blocks != cpu_to_le32(1)) {
 407 bad_block:              ext2_error(sb, "ext2_xattr_set",
 408                                 "inode %ld: bad block %d", inode->i_ino,
 409                                    EXT2_I(inode)->i_file_acl);
 410                         error = -EIO;
 411                         goto cleanup;
 412                 }
 413                 /* Find the named attribute. */
 414                 here = FIRST_ENTRY(bh);
 415                 while (!IS_LAST_ENTRY(here)) {
 416                         struct ext2_xattr_entry *next = EXT2_XATTR_NEXT(here);
 417                         if ((char *)next >= end)
 418                                 goto bad_block;
 419                         if (!here->e_value_block && here->e_value_size) {
 420                                 size_t offs = le16_to_cpu(here->e_value_offs);
 421                                 if (offs < min_offs)
 422                                         min_offs = offs;
 423                         }
 424                         not_found = name_index - here->e_name_index;
 425                         if (!not_found)
 426                                 not_found = name_len - here->e_name_len;
 427                         if (!not_found)
 428                                 not_found = memcmp(name, here->e_name,name_len);
 429                         if (not_found <= 0)
 430                                 break;
 431                         here = next;
 432                 }
 433                 last = here;
 434                 /* We still need to compute min_offs and last. */
 435                 while (!IS_LAST_ENTRY(last)) {
 436                         struct ext2_xattr_entry *next = EXT2_XATTR_NEXT(last);
 437                         if ((char *)next >= end)
 438                                 goto bad_block;
 439                         if (!last->e_value_block && last->e_value_size) {
 440                                 size_t offs = le16_to_cpu(last->e_value_offs);
 441                                 if (offs < min_offs)
 442                                         min_offs = offs;
 443                         }
 444                         last = next;
 445                 }
 446
 447                 /* Check whether we have enough space left. */
 448                 free = min_offs - ((char*)last - (char*)header) - sizeof(__u32);
 449         } else {
 450                 /* We will use a new extended attribute block. */
 451                 free = sb->s_blocksize -
 452                         sizeof(struct ext2_xattr_header) - sizeof(__u32);
 453                 here = last = NULL;  /* avoid gcc uninitialized warning. */
 454         }
 455
 456         if (not_found) {
 457                 /* Request to remove a nonexistent attribute? */
 458                 error = -ENODATA;
 459                 if (flags & XATTR_REPLACE)
 460                         goto cleanup;
 461                 error = 0;
 462                 if (value == NULL)
 463                         goto cleanup;
 464         } else {
 465                 /* Request to create an existing attribute? */
 466                 error = -EEXIST;
 467                 if (flags & XATTR_CREATE)
 468                         goto cleanup;
 469                 if (!here->e_value_block && here->e_value_size) {
 470                         size_t size = le32_to_cpu(here->e_value_size);
 471
 472                         if (le16_to_cpu(here->e_value_offs) + size >
 473                             sb->s_blocksize || size > sb->s_blocksize)
 474                                 goto bad_block;
 475                         free += EXT2_XATTR_SIZE(size);
 476                 }
 477                 free += EXT2_XATTR_LEN(name_len);
 478         }
 479         error = -ENOSPC;
 480         if (free < EXT2_XATTR_LEN(name_len) + EXT2_XATTR_SIZE(value_len))
 481                 goto cleanup;
 482
 483         /* Here we know that we can set the new attribute. */
 484
 485         if (header) {
 486                 struct mb_cache_entry *ce;
 487
 488                 /* assert(header == HDR(bh)); */
 489                 ce = mb_cache_entry_get(ext2_xattr_cache, bh->b_bdev,
 490                                         bh->b_blocknr);
 491                 lock_buffer(bh);
 492                 if (header->h_refcount == cpu_to_le32(1)) {
 493                         ea_bdebug(bh, "modifying in-place");
 494                         if (ce)
 495                                 mb_cache_entry_free(ce);
 496                         /* keep the buffer locked while modifying it. */
 497                 } else {
 498                         int offset;
 499
 500                         if (ce)
 501                                 mb_cache_entry_release(ce);
 502                         unlock_buffer(bh);
 503                         ea_bdebug(bh, "cloning");
 504                         header = kmalloc(bh->b_size, GFP_KERNEL);
 505                         error = -ENOMEM;
 506                         if (header == NULL)
 507                                 goto cleanup;
 508                         memcpy(header, HDR(bh), bh->b_size);
 509                         header->h_refcount = cpu_to_le32(1);
 510
 511                         offset = (char *)here - bh->b_data;
 512                         here = ENTRY((char *)header + offset);
 513                         offset = (char *)last - bh->b_data;
 514                         last = ENTRY((char *)header + offset);
 515                 }
 516         } else {
 517                 /* Allocate a buffer where we construct the new block. */
 518                 header = kzalloc(sb->s_blocksize, GFP_KERNEL);
 519                 error = -ENOMEM;
 520                 if (header == NULL)
 521                         goto cleanup;
 522                 end = (char *)header + sb->s_blocksize;
 523                 header->h_magic = cpu_to_le32(EXT2_XATTR_MAGIC);
 524                 header->h_blocks = header->h_refcount = cpu_to_le32(1);
 525                 last = here = ENTRY(header+1);
 526         }
 527
 528         /* Iff we are modifying the block in-place, bh is locked here. */
 529
 530         if (not_found) {
 531                 /* Insert the new name. */
 532                 size_t size = EXT2_XATTR_LEN(name_len);
 533                 size_t rest = (char *)last - (char *)here;
 534                 memmove((char *)here + size, here, rest);
 535                 memset(here, 0, size);
 536                 here->e_name_index = name_index;
 537                 here->e_name_len = name_len;
 538                 memcpy(here->e_name, name, name_len);
 539         } else {
 540                 if (!here->e_value_block && here->e_value_size) {
 541                         char *first_val = (char *)header + min_offs;
 542                         size_t offs = le16_to_cpu(here->e_value_offs);
 543                         char *val = (char *)header + offs;
 544                         size_t size = EXT2_XATTR_SIZE(
 545                                 le32_to_cpu(here->e_value_size));
 546
 547                         if (size == EXT2_XATTR_SIZE(value_len)) {
 548                                 /* The old and the new value have the same
 549                                    size. Just replace. */
 550                                 here->e_value_size = cpu_to_le32(value_len);
 551                                 memset(val + size - EXT2_XATTR_PAD, 0,
 552                                        EXT2_XATTR_PAD); /* Clear pad bytes. */
 553                                 memcpy(val, value, value_len);
 554                                 goto skip_replace;
 555                         }
 556
 557                         /* Remove the old value. */
 558                         memmove(first_val + size, first_val, val - first_val);
 559                         memset(first_val, 0, size);
 560                         here->e_value_offs = 0;
 561                         min_offs += size;
 562
 563                         /* Adjust all value offsets. */
 564                         last = ENTRY(header+1);
 565                         while (!IS_LAST_ENTRY(last)) {
 566                                 size_t o = le16_to_cpu(last->e_value_offs);
 567                                 if (!last->e_value_block && o < offs)
 568                                         last->e_value_offs =
 569                                                 cpu_to_le16(o + size);
 570                                 last = EXT2_XATTR_NEXT(last);
 571                         }
 572                 }
 573                 if (value == NULL) {
 574                         /* Remove the old name. */
 575                         size_t size = EXT2_XATTR_LEN(name_len);
 576                         last = ENTRY((char *)last - size);
 577                         memmove(here, (char*)here + size,
 578                                 (char*)last - (char*)here);
 579                         memset(last, 0, size);
 580                 }
 581         }
 582
 583         if (value != NULL) {
 584                 /* Insert the new value. */
 585                 here->e_value_size = cpu_to_le32(value_len);
 586                 if (value_len) {
 587                         size_t size = EXT2_XATTR_SIZE(value_len);
 588                         char *val = (char *)header + min_offs - size;
 589                         here->e_value_offs =
 590                                 cpu_to_le16((char *)val - (char *)header);
 591                         memset(val + size - EXT2_XATTR_PAD, 0,
 592                                EXT2_XATTR_PAD); /* Clear the pad bytes. */
 593                         memcpy(val, value, value_len);
 594                 }
 595         }
 596
 597 skip_replace:
 598         if (IS_LAST_ENTRY(ENTRY(header+1))) {
 599                 /* This block is now empty. */
 600                 if (bh && header == HDR(bh))
 601                         unlock_buffer(bh);  /* we were modifying in-place. */
 602                 error = ext2_xattr_set2(inode, bh, NULL);
 603         } else {
 604                 ext2_xattr_rehash(header, here);
 605                 if (bh && header == HDR(bh))
 606                         unlock_buffer(bh);  /* we were modifying in-place. */
 607                 error = ext2_xattr_set2(inode, bh, header);
 608         }
 609
 610 cleanup:
 611         brelse(bh);
 612         if (!(bh && header == HDR(bh)))
 613                 kfree(header);
 614         up_write(&EXT2_I(inode)->xattr_sem);
 615
 616         return error;
 617 }
 618
 619 /*
 620  * Second half of ext2_xattr_set(): Update the file system.
 621  */
 622 static int
 623 ext2_xattr_set2(struct inode *inode, struct buffer_head *old_bh,
 624                 struct ext2_xattr_header *header)
 625 {
 626         struct super_block *sb = inode->i_sb;
 627         struct buffer_head *new_bh = NULL;
 628         int error;
 629
 630         if (header) {
 631                 new_bh = ext2_xattr_cache_find(inode, header);
 632                 if (new_bh) {
 633                         /* We found an identical block in the cache. */
 634                         if (new_bh == old_bh) {
 635                                 ea_bdebug(new_bh, "keeping this block");
 636                         } else {
 637                                 /* The old block is released after updating
 638                                    the inode.  */
 639                                 ea_bdebug(new_bh, "reusing block");
 640
 641                                 error = dquot_alloc_block(inode, 1);
 642                                 if (error) {
 643                                         unlock_buffer(new_bh);
 644                                         goto cleanup;
 645                                 }
 646                                 le32_add_cpu(&HDR(new_bh)->h_refcount, 1);
 647                                 ea_bdebug(new_bh, "refcount now=%d",
 648                                         le32_to_cpu(HDR(new_bh)->h_refcount));
 649                         }
 650                         unlock_buffer(new_bh);
 651                 } else if (old_bh && header == HDR(old_bh)) {
 652                         /* Keep this block. No need to lock the block as we
 653                            don't need to change the reference count. */
 654                         new_bh = old_bh;
 655                         get_bh(new_bh);
 656                         ext2_xattr_cache_insert(new_bh);
 657                 } else {
 658                         /* We need to allocate a new block */
 659                         ext2_fsblk_t goal = ext2_group_first_block_no(sb,
 660                                                 EXT2_I(inode)->i_block_group);
 661                         int block = ext2_new_block(inode, goal, &error);
 662                         if (error)
 663                                 goto cleanup;
 664                         ea_idebug(inode, "creating block %d", block);
 665
 666                         new_bh = sb_getblk(sb, block);
 667                         if (!new_bh) {
 668                                 ext2_free_blocks(inode, block, 1);
 669                                 mark_inode_dirty(inode);
 670                                 error = -EIO;
 671                                 goto cleanup;
 672                         }
 673                         lock_buffer(new_bh);
 674                         memcpy(new_bh->b_data, header, new_bh->b_size);
 675                         set_buffer_uptodate(new_bh);
 676                         unlock_buffer(new_bh);
 677                         ext2_xattr_cache_insert(new_bh);
 678
 679                         ext2_xattr_update_super_block(sb);
 680                 }
 681                 mark_buffer_dirty(new_bh);
 682                 if (IS_SYNC(inode)) {
 683                         sync_dirty_buffer(new_bh);
 684                         error = -EIO;
 685                         if (buffer_req(new_bh) && !buffer_uptodate(new_bh))
 686                                 goto cleanup;
 687                 }
 688         }
 689
 690         /* Update the inode. */
 691         EXT2_I(inode)->i_file_acl = new_bh ? new_bh->b_blocknr : 0;
 692         inode->i_ctime = CURRENT_TIME_SEC;
 693         if (IS_SYNC(inode)) {
 694                 error = sync_inode_metadata(inode, 1);
 695                 /* In case sync failed due to ENOSPC the inode was actually
 696                  * written (only some dirty data were not) so we just proceed
 697                  * as if nothing happened and cleanup the unused block */
 698                 if (error && error != -ENOSPC) {
 699                         if (new_bh && new_bh != old_bh) {
 700                                 dquot_free_block_nodirty(inode, 1);
 701                                 mark_inode_dirty(inode);
 702                         }
 703                         goto cleanup;
 704                 }
 705         } else
 706                 mark_inode_dirty(inode);
 707
 708         error = 0;
 709         if (old_bh && old_bh != new_bh) {
 710                 struct mb_cache_entry *ce;
 711
 712                 /*
 713                  * If there was an old block and we are no longer using it,
 714                  * release the old block.
 715                  */
 716                 ce = mb_cache_entry_get(ext2_xattr_cache, old_bh->b_bdev,
 717                                         old_bh->b_blocknr);
 718                 lock_buffer(old_bh);
 719                 if (HDR(old_bh)->h_refcount == cpu_to_le32(1)) {
 720                         /* Free the old block. */
 721                         if (ce)
 722                                 mb_cache_entry_free(ce);
 723                         ea_bdebug(old_bh, "freeing");
 724                         ext2_free_blocks(inode, old_bh->b_blocknr, 1);
 725                         mark_inode_dirty(inode);
 726                         /* We let our caller release old_bh, so we
 727                          * need to duplicate the buffer before. */
 728                         get_bh(old_bh);
 729                         bforget(old_bh);
 730                 } else {
 731                         /* Decrement the refcount only. */
 732                         le32_add_cpu(&HDR(old_bh)->h_refcount, -1);
 733                         if (ce)
 734                                 mb_cache_entry_release(ce);
 735                         dquot_free_block_nodirty(inode, 1);
 736                         mark_inode_dirty(inode);
 737                         mark_buffer_dirty(old_bh);
 738                         ea_bdebug(old_bh, "refcount now=%d",
 739                                 le32_to_cpu(HDR(old_bh)->h_refcount));
 740                 }
 741                 unlock_buffer(old_bh);
 742         }
 743
 744 cleanup:
 745         brelse(new_bh);
 746
 747         return error;
 748 }
 749
 750 /*
 751  * ext2_xattr_delete_inode()
 752  *
 753  * Free extended attribute resources associated with this inode. This
 754  * is called immediately before an inode is freed.
 755  */
 756 void
 757 ext2_xattr_delete_inode(struct inode *inode)
 758 {
 759         struct buffer_head *bh = NULL;
 760         struct mb_cache_entry *ce;
 761
 762         down_write(&EXT2_I(inode)->xattr_sem);
 763         if (!EXT2_I(inode)->i_file_acl)
 764                 goto cleanup;
 765         bh = sb_bread(inode->i_sb, EXT2_I(inode)->i_file_acl);
 766         if (!bh) {
 767                 ext2_error(inode->i_sb, "ext2_xattr_delete_inode",
 768                         "inode %ld: block %d read error", inode->i_ino,
 769                         EXT2_I(inode)->i_file_acl);
 770                 goto cleanup;
 771         }
 772         ea_bdebug(bh, "b_count=%d", atomic_read(&(bh->b_count)));
 773         if (HDR(bh)->h_magic != cpu_to_le32(EXT2_XATTR_MAGIC) ||
 774             HDR(bh)->h_blocks != cpu_to_le32(1)) {
 775                 ext2_error(inode->i_sb, "ext2_xattr_delete_inode",
 776                         "inode %ld: bad block %d", inode->i_ino,
 777                         EXT2_I(inode)->i_file_acl);
 778                 goto cleanup;
 779         }
 780         ce = mb_cache_entry_get(ext2_xattr_cache, bh->b_bdev, bh->b_blocknr);
 781         lock_buffer(bh);
 782         if (HDR(bh)->h_refcount == cpu_to_le32(1)) {
 783                 if (ce)
 784                         mb_cache_entry_free(ce);
 785                 ext2_free_blocks(inode, EXT2_I(inode)->i_file_acl, 1);
 786                 get_bh(bh);
 787                 bforget(bh);
 788                 unlock_buffer(bh);
 789         } else {
 790                 le32_add_cpu(&HDR(bh)->h_refcount, -1);
 791                 if (ce)
 792                         mb_cache_entry_release(ce);
 793                 ea_bdebug(bh, "refcount now=%d",
 794                         le32_to_cpu(HDR(bh)->h_refcount));
 795                 unlock_buffer(bh);
 796                 mark_buffer_dirty(bh);
 797                 if (IS_SYNC(inode))
 798                         sync_dirty_buffer(bh);
 799                 dquot_free_block_nodirty(inode, 1);
 800         }
 801         EXT2_I(inode)->i_file_acl = 0;
 802
 803 cleanup:
 804         brelse(bh);
 805         up_write(&EXT2_I(inode)->xattr_sem);
 806 }
 807
 808 /*
 809  * ext2_xattr_put_super()
 810  *
 811  * This is called when a file system is unmounted.
 812  */
 813 void
 814 ext2_xattr_put_super(struct super_block *sb)
 815 {
 816         mb_cache_shrink(sb->s_bdev);
 817 }
 818
 819
 820 /*
 821  * ext2_xattr_cache_insert()
 822  *
 823  * Create a new entry in the extended attribute cache, and insert
 824  * it unless such an entry is already in the cache.
 825  *
 826  * Returns 0, or a negative error number on failure.
 827  */
 828 static int
 829 ext2_xattr_cache_insert(struct buffer_head *bh)
 830 {
 831         __u32 hash = le32_to_cpu(HDR(bh)->h_hash);
 832         struct mb_cache_entry *ce;
 833         int error;
 834
 835         ce = mb_cache_entry_alloc(ext2_xattr_cache, GFP_NOFS);
 836         if (!ce)
 837                 return -ENOMEM;
 838         error = mb_cache_entry_insert(ce, bh->b_bdev, bh->b_blocknr, hash);
 839         if (error) {
 840                 mb_cache_entry_free(ce);
 841                 if (error == -EBUSY) {
 842                         ea_bdebug(bh, "already in cache (%d cache entries)",
 843                                 atomic_read(&ext2_xattr_cache->c_entry_count));
 844                         error = 0;
 845                 }
 846         } else {
 847                 ea_bdebug(bh, "inserting [%x] (%d cache entries)", (int)hash,
 848                           atomic_read(&ext2_xattr_cache->c_entry_count));
 849                 mb_cache_entry_release(ce);
 850         }
 851         return error;
 852 }
 853
 854 /*
 855  * ext2_xattr_cmp()
 856  *
 857  * Compare two extended attribute blocks for equality.
 858  *
 859  * Returns 0 if the blocks are equal, 1 if they differ, and
 860  * a negative error number on errors.
 861  */
 862 static int
 863 ext2_xattr_cmp(struct ext2_xattr_header *header1,
 864                struct ext2_xattr_header *header2)
 865 {
 866         struct ext2_xattr_entry *entry1, *entry2;
 867
 868         entry1 = ENTRY(header1+1);
 869         entry2 = ENTRY(header2+1);
 870         while (!IS_LAST_ENTRY(entry1)) {
 871                 if (IS_LAST_ENTRY(entry2))
 872                         return 1;
 873                 if (entry1->e_hash != entry2->e_hash ||
 874                     entry1->e_name_index != entry2->e_name_index ||
 875                     entry1->e_name_len != entry2->e_name_len ||
 876                     entry1->e_value_size != entry2->e_value_size ||
 877                     memcmp(entry1->e_name, entry2->e_name, entry1->e_name_len))
 878                         return 1;
 879                 if (entry1->e_value_block != 0 || entry2->e_value_block != 0)
 880                         return -EIO;
 881                 if (memcmp((char *)header1 + le16_to_cpu(entry1->e_value_offs),
 882                            (char *)header2 + le16_to_cpu(entry2->e_value_offs),
 883                            le32_to_cpu(entry1->e_value_size)))
 884                         return 1;
 885
 886                 entry1 = EXT2_XATTR_NEXT(entry1);
 887                 entry2 = EXT2_XATTR_NEXT(entry2);
 888         }
 889         if (!IS_LAST_ENTRY(entry2))
 890                 return 1;
 891         return 0;
 892 }
 893
 894 /*
 895  * ext2_xattr_cache_find()
 896  *
 897  * Find an identical extended attribute block.
 898  *
 899  * Returns a locked buffer head to the block found, or NULL if such
 900  * a block was not found or an error occurred.
 901  */
 902 static struct buffer_head *
 903 ext2_xattr_cache_find(struct inode *inode, struct ext2_xattr_header *header)
 904 {
 905         __u32 hash = le32_to_cpu(header->h_hash);
 906         struct mb_cache_entry *ce;
 907
 908         if (!header->h_hash)
 909                 return NULL;  /* never share */
 910         ea_idebug(inode, "looking for cached blocks [%x]", (int)hash);
 911 again:
 912         ce = mb_cache_entry_find_first(ext2_xattr_cache, inode->i_sb->s_bdev,
 913                                        hash);
 914         while (ce) {
 915                 struct buffer_head *bh;
 916
 917                 if (IS_ERR(ce)) {
 918                         if (PTR_ERR(ce) == -EAGAIN)
 919                                 goto again;
 920                         break;
 921                 }
 922
 923                 bh = sb_bread(inode->i_sb, ce->e_block);
 924                 if (!bh) {
 925                         ext2_error(inode->i_sb, "ext2_xattr_cache_find",
 926                                 "inode %ld: block %ld read error",
 927                                 inode->i_ino, (unsigned long) ce->e_block);
 928                 } else {
 929                         lock_buffer(bh);
 930                         if (le32_to_cpu(HDR(bh)->h_refcount) >
 931                                    EXT2_XATTR_REFCOUNT_MAX) {
 932                                 ea_idebug(inode, "block %ld refcount %d>%d",
 933                                           (unsigned long) ce->e_block,
 934                                           le32_to_cpu(HDR(bh)->h_refcount),
 935                                           EXT2_XATTR_REFCOUNT_MAX);
 936                         } else if (!ext2_xattr_cmp(header, HDR(bh))) {
 937                                 ea_bdebug(bh, "b_count=%d",
 938                                           atomic_read(&(bh->b_count)));
 939                                 mb_cache_entry_release(ce);
 940                                 return bh;
 941                         }
 942                         unlock_buffer(bh);
 943                         brelse(bh);
 944                 }
 945                 ce = mb_cache_entry_find_next(ce, inode->i_sb->s_bdev, hash);
 946         }
 947         return NULL;
 948 }
 949
 950 #define NAME_HASH_SHIFT 5
 951 #define VALUE_HASH_SHIFT 16
 952
 953 /*
 954  * ext2_xattr_hash_entry()
 955  *
 956  * Compute the hash of an extended attribute.
 957  */
 958 static inline void ext2_xattr_hash_entry(struct ext2_xattr_header *header,
 959                                          struct ext2_xattr_entry *entry)
 960 {
 961         __u32 hash = 0;
 962         char *name = entry->e_name;
 963         int n;
 964
 965         for (n=0; n < entry->e_name_len; n++) {
 966                 hash = (hash << NAME_HASH_SHIFT) ^
 967                        (hash >> (8*sizeof(hash) - NAME_HASH_SHIFT)) ^
 968                        *name++;
 969         }
 970
 971         if (entry->e_value_block == 0 && entry->e_value_size != 0) {
 972                 __le32 *value = (__le32 *)((char *)header +
 973                         le16_to_cpu(entry->e_value_offs));
 974                 for (n = (le32_to_cpu(entry->e_value_size) +
 975                      EXT2_XATTR_ROUND) >> EXT2_XATTR_PAD_BITS; n; n--) {
 976                         hash = (hash << VALUE_HASH_SHIFT) ^
 977                                (hash >> (8*sizeof(hash) - VALUE_HASH_SHIFT)) ^
 978                                le32_to_cpu(*value++);
 979                 }
 980         }
 981         entry->e_hash = cpu_to_le32(hash);
 982 }
 983
 984 #undef NAME_HASH_SHIFT
 985 #undef VALUE_HASH_SHIFT
 986
 987 #define BLOCK_HASH_SHIFT 16
 988
 989 /*
 990  * ext2_xattr_rehash()
 991  *
 992  * Re-compute the extended attribute hash value after an entry has changed.
 993  */
 994 static void ext2_xattr_rehash(struct ext2_xattr_header *header,
 995                               struct ext2_xattr_entry *entry)
 996 {
 997         struct ext2_xattr_entry *here;
 998         __u32 hash = 0;
 999
1000         ext2_xattr_hash_entry(header, entry);
1001         here = ENTRY(header+1);
1002         while (!IS_LAST_ENTRY(here)) {
1003                 if (!here->e_hash) {
1004                         /* Block is not shared if an entry's hash value == 0 */
1005                         hash = 0;
1006                         break;
1007                 }
1008                 hash = (hash << BLOCK_HASH_SHIFT) ^
1009                        (hash >> (8*sizeof(hash) - BLOCK_HASH_SHIFT)) ^
1010                        le32_to_cpu(here->e_hash);
1011                 here = EXT2_XATTR_NEXT(here);
1012         }
1013         header->h_hash = cpu_to_le32(hash);
1014 }
1015
1016 #undef BLOCK_HASH_SHIFT
1017
1018 int __init
1019 init_ext2_xattr(void)
1020 {
1021         ext2_xattr_cache = mb_cache_create("ext2_xattr", 6);
1022         if (!ext2_xattr_cache)
1023                 return -ENOMEM;
1024         return 0;
1025 }
1026
1027 void
1028 exit_ext2_xattr(void)
1029 {
1030         mb_cache_destroy(ext2_xattr_cache);
1031 }