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