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                 printk(KERN_DEBUG "block %pg:%lu: ", \
  81                         bh->b_bdev, (unsigned long) bh->b_blocknr); \
  82                 printk(f); \
  83                 printk("\n"); \
  84         } while (0)
  85 #else
  86 # define ea_idebug(f...)
  87 # define ea_bdebug(f...)
  88 #endif
  89
  90 static int ext2_xattr_set2(struct inode *, struct buffer_head *,
  91                            struct ext2_xattr_header *);
  92
  93 static int ext2_xattr_cache_insert(struct mb_cache *, struct buffer_head *);
  94 static struct buffer_head *ext2_xattr_cache_find(struct inode *,
  95                                                  struct ext2_xattr_header *);
  96 static void ext2_xattr_rehash(struct ext2_xattr_header *,
  97                               struct ext2_xattr_entry *);
  98
  99 static const struct xattr_handler *ext2_xattr_handler_map[] = {
 100         [EXT2_XATTR_INDEX_USER]              = &ext2_xattr_user_handler,
 101 #ifdef CONFIG_EXT2_FS_POSIX_ACL
 102         [EXT2_XATTR_INDEX_POSIX_ACL_ACCESS]  = &posix_acl_access_xattr_handler,
 103         [EXT2_XATTR_INDEX_POSIX_ACL_DEFAULT] = &posix_acl_default_xattr_handler,
 104 #endif
 105         [EXT2_XATTR_INDEX_TRUSTED]           = &ext2_xattr_trusted_handler,
 106 #ifdef CONFIG_EXT2_FS_SECURITY
 107         [EXT2_XATTR_INDEX_SECURITY]          = &ext2_xattr_security_handler,
 108 #endif
 109 };
 110
 111 const struct xattr_handler *ext2_xattr_handlers[] = {
 112         &ext2_xattr_user_handler,
 113         &ext2_xattr_trusted_handler,
 114 #ifdef CONFIG_EXT2_FS_POSIX_ACL
 115         &posix_acl_access_xattr_handler,
 116         &posix_acl_default_xattr_handler,
 117 #endif
 118 #ifdef CONFIG_EXT2_FS_SECURITY
 119         &ext2_xattr_security_handler,
 120 #endif
 121         NULL
 122 };
 123
 124 #define EA_BLOCK_CACHE(inode)   (EXT2_SB(inode->i_sb)->s_ea_block_cache)
 125
 126 static inline const struct xattr_handler *
 127 ext2_xattr_handler(int name_index)
 128 {
 129         const struct xattr_handler *handler = NULL;
 130
 131         if (name_index > 0 && name_index < ARRAY_SIZE(ext2_xattr_handler_map))
 132                 handler = ext2_xattr_handler_map[name_index];
 133         return handler;
 134 }
 135
 136 /*
 137  * ext2_xattr_get()
 138  *
 139  * Copy an extended attribute into the buffer
 140  * provided, or compute the buffer size required.
 141  * Buffer is NULL to compute the size of the buffer required.
 142  *
 143  * Returns a negative error number on failure, or the number of bytes
 144  * used / required on success.
 145  */
 146 int
 147 ext2_xattr_get(struct inode *inode, int name_index, const char *name,
 148                void *buffer, size_t buffer_size)
 149 {
 150         struct buffer_head *bh = NULL;
 151         struct ext2_xattr_entry *entry;
 152         size_t name_len, size;
 153         char *end;
 154         int error;
 155         struct mb_cache *ea_block_cache = EA_BLOCK_CACHE(inode);
 156
 157         ea_idebug(inode, "name=%d.%s, buffer=%p, buffer_size=%ld",
 158                   name_index, name, buffer, (long)buffer_size);
 159
 160         if (name == NULL)
 161                 return -EINVAL;
 162         name_len = strlen(name);
 163         if (name_len > 255)
 164                 return -ERANGE;
 165
 166         down_read(&EXT2_I(inode)->xattr_sem);
 167         error = -ENODATA;
 168         if (!EXT2_I(inode)->i_file_acl)
 169                 goto cleanup;
 170         ea_idebug(inode, "reading block %d", EXT2_I(inode)->i_file_acl);
 171         bh = sb_bread(inode->i_sb, EXT2_I(inode)->i_file_acl);
 172         error = -EIO;
 173         if (!bh)
 174                 goto cleanup;
 175         ea_bdebug(bh, "b_count=%d, refcount=%d",
 176                 atomic_read(&(bh->b_count)), le32_to_cpu(HDR(bh)->h_refcount));
 177         end = bh->b_data + bh->b_size;
 178         if (HDR(bh)->h_magic != cpu_to_le32(EXT2_XATTR_MAGIC) ||
 179             HDR(bh)->h_blocks != cpu_to_le32(1)) {
 180 bad_block:      ext2_error(inode->i_sb, "ext2_xattr_get",
 181                         "inode %ld: bad block %d", inode->i_ino,
 182                         EXT2_I(inode)->i_file_acl);
 183                 error = -EIO;
 184                 goto cleanup;
 185         }
 186
 187         /* find named attribute */
 188         entry = FIRST_ENTRY(bh);
 189         while (!IS_LAST_ENTRY(entry)) {
 190                 struct ext2_xattr_entry *next =
 191                         EXT2_XATTR_NEXT(entry);
 192                 if ((char *)next >= end)
 193                         goto bad_block;
 194                 if (name_index == entry->e_name_index &&
 195                     name_len == entry->e_name_len &&
 196                     memcmp(name, entry->e_name, name_len) == 0)
 197                         goto found;
 198                 entry = next;
 199         }
 200         if (ext2_xattr_cache_insert(ea_block_cache, bh))
 201                 ea_idebug(inode, "cache insert failed");
 202         error = -ENODATA;
 203         goto cleanup;
 204 found:
 205         /* check the buffer size */
 206         if (entry->e_value_block != 0)
 207                 goto bad_block;
 208         size = le32_to_cpu(entry->e_value_size);
 209         if (size > inode->i_sb->s_blocksize ||
 210             le16_to_cpu(entry->e_value_offs) + size > inode->i_sb->s_blocksize)
 211                 goto bad_block;
 212
 213         if (ext2_xattr_cache_insert(ea_block_cache, bh))
 214                 ea_idebug(inode, "cache insert failed");
 215         if (buffer) {
 216                 error = -ERANGE;
 217                 if (size > buffer_size)
 218                         goto cleanup;
 219                 /* return value of attribute */
 220                 memcpy(buffer, bh->b_data + le16_to_cpu(entry->e_value_offs),
 221                         size);
 222         }
 223         error = size;
 224
 225 cleanup:
 226         brelse(bh);
 227         up_read(&EXT2_I(inode)->xattr_sem);
 228
 229         return error;
 230 }
 231
 232 /*
 233  * ext2_xattr_list()
 234  *
 235  * Copy a list of attribute names into the buffer
 236  * provided, or compute the buffer size required.
 237  * Buffer is NULL to compute the size of the buffer required.
 238  *
 239  * Returns a negative error number on failure, or the number of bytes
 240  * used / required on success.
 241  */
 242 static int
 243 ext2_xattr_list(struct dentry *dentry, char *buffer, size_t buffer_size)
 244 {
 245         struct inode *inode = d_inode(dentry);
 246         struct buffer_head *bh = NULL;
 247         struct ext2_xattr_entry *entry;
 248         char *end;
 249         size_t rest = buffer_size;
 250         int error;
 251         struct mb_cache *ea_block_cache = EA_BLOCK_CACHE(inode);
 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(ea_block_cache, 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 && (!handler->list || handler->list(dentry))) {
 296                         const char *prefix = handler->prefix ?: handler->name;
 297                         size_t prefix_len = strlen(prefix);
 298                         size_t size = prefix_len + entry->e_name_len + 1;
 299
 300                         if (buffer) {
 301                                 if (size > rest) {
 302                                         error = -ERANGE;
 303                                         goto cleanup;
 304                                 }
 305                                 memcpy(buffer, prefix, prefix_len);
 306                                 buffer += prefix_len;
 307                                 memcpy(buffer, entry->e_name, entry->e_name_len);
 308                                 buffer += entry->e_name_len;
 309                                 *buffer++ = 0;
 310                         }
 311                         rest -= size;
 312                 }
 313         }
 314         error = buffer_size - rest;  /* total size */
 315
 316 cleanup:
 317         brelse(bh);
 318         up_read(&EXT2_I(inode)->xattr_sem);
 319
 320         return error;
 321 }
 322
 323 /*
 324  * Inode operation listxattr()
 325  *
 326  * d_inode(dentry)->i_mutex: don't care
 327  */
 328 ssize_t
 329 ext2_listxattr(struct dentry *dentry, char *buffer, size_t size)
 330 {
 331         return ext2_xattr_list(dentry, buffer, size);
 332 }
 333
 334 /*
 335  * If the EXT2_FEATURE_COMPAT_EXT_ATTR feature of this file system is
 336  * not set, set it.
 337  */
 338 static void ext2_xattr_update_super_block(struct super_block *sb)
 339 {
 340         if (EXT2_HAS_COMPAT_FEATURE(sb, EXT2_FEATURE_COMPAT_EXT_ATTR))
 341                 return;
 342
 343         spin_lock(&EXT2_SB(sb)->s_lock);
 344         EXT2_SET_COMPAT_FEATURE(sb, EXT2_FEATURE_COMPAT_EXT_ATTR);
 345         spin_unlock(&EXT2_SB(sb)->s_lock);
 346         mark_buffer_dirty(EXT2_SB(sb)->s_sbh);
 347 }
 348
 349 /*
 350  * ext2_xattr_set()
 351  *
 352  * Create, replace or remove an extended attribute for this inode.  Value
 353  * is NULL to remove an existing extended attribute, and non-NULL to
 354  * either replace an existing extended attribute, or create a new extended
 355  * attribute. The flags XATTR_REPLACE and XATTR_CREATE
 356  * specify that an extended attribute must exist and must not exist
 357  * previous to the call, respectively.
 358  *
 359  * Returns 0, or a negative error number on failure.
 360  */
 361 int
 362 ext2_xattr_set(struct inode *inode, int name_index, const char *name,
 363                const void *value, size_t value_len, int flags)
 364 {
 365         struct super_block *sb = inode->i_sb;
 366         struct buffer_head *bh = NULL;
 367         struct ext2_xattr_header *header = NULL;
 368         struct ext2_xattr_entry *here, *last;
 369         size_t name_len, free, min_offs = sb->s_blocksize;
 370         int not_found = 1, error;
 371         char *end;
 372
 373         /*
 374          * header -- Points either into bh, or to a temporarily
 375          *           allocated buffer.
 376          * here -- The named entry found, or the place for inserting, within
 377          *         the block pointed to by header.
 378          * last -- Points right after the last named entry within the block
 379          *         pointed to by header.
 380          * min_offs -- The offset of the first value (values are aligned
 381          *             towards the end of the block).
 382          * end -- Points right after the block pointed to by header.
 383          */
 384
 385         ea_idebug(inode, "name=%d.%s, value=%p, value_len=%ld",
 386                   name_index, name, value, (long)value_len);
 387
 388         if (value == NULL)
 389                 value_len = 0;
 390         if (name == NULL)
 391                 return -EINVAL;
 392         name_len = strlen(name);
 393         if (name_len > 255 || value_len > sb->s_blocksize)
 394                 return -ERANGE;
 395         down_write(&EXT2_I(inode)->xattr_sem);
 396         if (EXT2_I(inode)->i_file_acl) {
 397                 /* The inode already has an extended attribute block. */
 398                 bh = sb_bread(sb, EXT2_I(inode)->i_file_acl);
 399                 error = -EIO;
 400                 if (!bh)
 401                         goto cleanup;
 402                 ea_bdebug(bh, "b_count=%d, refcount=%d",
 403                         atomic_read(&(bh->b_count)),
 404                         le32_to_cpu(HDR(bh)->h_refcount));
 405                 header = HDR(bh);
 406                 end = bh->b_data + bh->b_size;
 407                 if (header->h_magic != cpu_to_le32(EXT2_XATTR_MAGIC) ||
 408                     header->h_blocks != cpu_to_le32(1)) {
 409 bad_block:              ext2_error(sb, "ext2_xattr_set",
 410                                 "inode %ld: bad block %d", inode->i_ino,
 411                                    EXT2_I(inode)->i_file_acl);
 412                         error = -EIO;
 413                         goto cleanup;
 414                 }
 415                 /* Find the named attribute. */
 416                 here = FIRST_ENTRY(bh);
 417                 while (!IS_LAST_ENTRY(here)) {
 418                         struct ext2_xattr_entry *next = EXT2_XATTR_NEXT(here);
 419                         if ((char *)next >= end)
 420                                 goto bad_block;
 421                         if (!here->e_value_block && here->e_value_size) {
 422                                 size_t offs = le16_to_cpu(here->e_value_offs);
 423                                 if (offs < min_offs)
 424                                         min_offs = offs;
 425                         }
 426                         not_found = name_index - here->e_name_index;
 427                         if (!not_found)
 428                                 not_found = name_len - here->e_name_len;
 429                         if (!not_found)
 430                                 not_found = memcmp(name, here->e_name,name_len);
 431                         if (not_found <= 0)
 432                                 break;
 433                         here = next;
 434                 }
 435                 last = here;
 436                 /* We still need to compute min_offs and last. */
 437                 while (!IS_LAST_ENTRY(last)) {
 438                         struct ext2_xattr_entry *next = EXT2_XATTR_NEXT(last);
 439                         if ((char *)next >= end)
 440                                 goto bad_block;
 441                         if (!last->e_value_block && last->e_value_size) {
 442                                 size_t offs = le16_to_cpu(last->e_value_offs);
 443                                 if (offs < min_offs)
 444                                         min_offs = offs;
 445                         }
 446                         last = next;
 447                 }
 448
 449                 /* Check whether we have enough space left. */
 450                 free = min_offs - ((char*)last - (char*)header) - sizeof(__u32);
 451         } else {
 452                 /* We will use a new extended attribute block. */
 453                 free = sb->s_blocksize -
 454                         sizeof(struct ext2_xattr_header) - sizeof(__u32);
 455                 here = last = NULL;  /* avoid gcc uninitialized warning. */
 456         }
 457
 458         if (not_found) {
 459                 /* Request to remove a nonexistent attribute? */
 460                 error = -ENODATA;
 461                 if (flags & XATTR_REPLACE)
 462                         goto cleanup;
 463                 error = 0;
 464                 if (value == NULL)
 465                         goto cleanup;
 466         } else {
 467                 /* Request to create an existing attribute? */
 468                 error = -EEXIST;
 469                 if (flags & XATTR_CREATE)
 470                         goto cleanup;
 471                 if (!here->e_value_block && here->e_value_size) {
 472                         size_t size = le32_to_cpu(here->e_value_size);
 473
 474                         if (le16_to_cpu(here->e_value_offs) + size >
 475                             sb->s_blocksize || size > sb->s_blocksize)
 476                                 goto bad_block;
 477                         free += EXT2_XATTR_SIZE(size);
 478                 }
 479                 free += EXT2_XATTR_LEN(name_len);
 480         }
 481         error = -ENOSPC;
 482         if (free < EXT2_XATTR_LEN(name_len) + EXT2_XATTR_SIZE(value_len))
 483                 goto cleanup;
 484
 485         /* Here we know that we can set the new attribute. */
 486
 487         if (header) {
 488                 /* assert(header == HDR(bh)); */
 489                 lock_buffer(bh);
 490                 if (header->h_refcount == cpu_to_le32(1)) {
 491                         __u32 hash = le32_to_cpu(header->h_hash);
 492
 493                         ea_bdebug(bh, "modifying in-place");
 494                         /*
 495                          * This must happen under buffer lock for
 496                          * ext2_xattr_set2() to reliably detect modified block
 497                          */
 498                         mb_cache_entry_delete(EA_BLOCK_CACHE(inode), hash,
 499                                               bh->b_blocknr);
 500
 501                         /* keep the buffer locked while modifying it. */
 502                 } else {
 503                         int offset;
 504
 505                         unlock_buffer(bh);
 506                         ea_bdebug(bh, "cloning");
 507                         header = kmalloc(bh->b_size, GFP_KERNEL);
 508                         error = -ENOMEM;
 509                         if (header == NULL)
 510                                 goto cleanup;
 511                         memcpy(header, HDR(bh), bh->b_size);
 512                         header->h_refcount = cpu_to_le32(1);
 513
 514                         offset = (char *)here - bh->b_data;
 515                         here = ENTRY((char *)header + offset);
 516                         offset = (char *)last - bh->b_data;
 517                         last = ENTRY((char *)header + offset);
 518                 }
 519         } else {
 520                 /* Allocate a buffer where we construct the new block. */
 521                 header = kzalloc(sb->s_blocksize, GFP_KERNEL);
 522                 error = -ENOMEM;
 523                 if (header == NULL)
 524                         goto cleanup;
 525                 end = (char *)header + sb->s_blocksize;
 526                 header->h_magic = cpu_to_le32(EXT2_XATTR_MAGIC);
 527                 header->h_blocks = header->h_refcount = cpu_to_le32(1);
 528                 last = here = ENTRY(header+1);
 529         }
 530
 531         /* Iff we are modifying the block in-place, bh is locked here. */
 532
 533         if (not_found) {
 534                 /* Insert the new name. */
 535                 size_t size = EXT2_XATTR_LEN(name_len);
 536                 size_t rest = (char *)last - (char *)here;
 537                 memmove((char *)here + size, here, rest);
 538                 memset(here, 0, size);
 539                 here->e_name_index = name_index;
 540                 here->e_name_len = name_len;
 541                 memcpy(here->e_name, name, name_len);
 542         } else {
 543                 if (!here->e_value_block && here->e_value_size) {
 544                         char *first_val = (char *)header + min_offs;
 545                         size_t offs = le16_to_cpu(here->e_value_offs);
 546                         char *val = (char *)header + offs;
 547                         size_t size = EXT2_XATTR_SIZE(
 548                                 le32_to_cpu(here->e_value_size));
 549
 550                         if (size == EXT2_XATTR_SIZE(value_len)) {
 551                                 /* The old and the new value have the same
 552                                    size. Just replace. */
 553                                 here->e_value_size = cpu_to_le32(value_len);
 554                                 memset(val + size - EXT2_XATTR_PAD, 0,
 555                                        EXT2_XATTR_PAD); /* Clear pad bytes. */
 556                                 memcpy(val, value, value_len);
 557                                 goto skip_replace;
 558                         }
 559
 560                         /* Remove the old value. */
 561                         memmove(first_val + size, first_val, val - first_val);
 562                         memset(first_val, 0, size);
 563                         here->e_value_offs = 0;
 564                         min_offs += size;
 565
 566                         /* Adjust all value offsets. */
 567                         last = ENTRY(header+1);
 568                         while (!IS_LAST_ENTRY(last)) {
 569                                 size_t o = le16_to_cpu(last->e_value_offs);
 570                                 if (!last->e_value_block && o < offs)
 571                                         last->e_value_offs =
 572                                                 cpu_to_le16(o + size);
 573                                 last = EXT2_XATTR_NEXT(last);
 574                         }
 575                 }
 576                 if (value == NULL) {
 577                         /* Remove the old name. */
 578                         size_t size = EXT2_XATTR_LEN(name_len);
 579                         last = ENTRY((char *)last - size);
 580                         memmove(here, (char*)here + size,
 581                                 (char*)last - (char*)here);
 582                         memset(last, 0, size);
 583                 }
 584         }
 585
 586         if (value != NULL) {
 587                 /* Insert the new value. */
 588                 here->e_value_size = cpu_to_le32(value_len);
 589                 if (value_len) {
 590                         size_t size = EXT2_XATTR_SIZE(value_len);
 591                         char *val = (char *)header + min_offs - size;
 592                         here->e_value_offs =
 593                                 cpu_to_le16((char *)val - (char *)header);
 594                         memset(val + size - EXT2_XATTR_PAD, 0,
 595                                EXT2_XATTR_PAD); /* Clear the pad bytes. */
 596                         memcpy(val, value, value_len);
 597                 }
 598         }
 599
 600 skip_replace:
 601         if (IS_LAST_ENTRY(ENTRY(header+1))) {
 602                 /* This block is now empty. */
 603                 if (bh && header == HDR(bh))
 604                         unlock_buffer(bh);  /* we were modifying in-place. */
 605                 error = ext2_xattr_set2(inode, bh, NULL);
 606         } else {
 607                 ext2_xattr_rehash(header, here);
 608                 if (bh && header == HDR(bh))
 609                         unlock_buffer(bh);  /* we were modifying in-place. */
 610                 error = ext2_xattr_set2(inode, bh, header);
 611         }
 612
 613 cleanup:
 614         brelse(bh);
 615         if (!(bh && header == HDR(bh)))
 616                 kfree(header);
 617         up_write(&EXT2_I(inode)->xattr_sem);
 618
 619         return error;
 620 }
 621
 622 /*
 623  * Second half of ext2_xattr_set(): Update the file system.
 624  */
 625 static int
 626 ext2_xattr_set2(struct inode *inode, struct buffer_head *old_bh,
 627                 struct ext2_xattr_header *header)
 628 {
 629         struct super_block *sb = inode->i_sb;
 630         struct buffer_head *new_bh = NULL;
 631         int error;
 632         struct mb_cache *ea_block_cache = EA_BLOCK_CACHE(inode);
 633
 634         if (header) {
 635                 new_bh = ext2_xattr_cache_find(inode, header);
 636                 if (new_bh) {
 637                         /* We found an identical block in the cache. */
 638                         if (new_bh == old_bh) {
 639                                 ea_bdebug(new_bh, "keeping this block");
 640                         } else {
 641                                 /* The old block is released after updating
 642                                    the inode.  */
 643                                 ea_bdebug(new_bh, "reusing block");
 644
 645                                 error = dquot_alloc_block(inode, 1);
 646                                 if (error) {
 647                                         unlock_buffer(new_bh);
 648                                         goto cleanup;
 649                                 }
 650                                 le32_add_cpu(&HDR(new_bh)->h_refcount, 1);
 651                                 ea_bdebug(new_bh, "refcount now=%d",
 652                                         le32_to_cpu(HDR(new_bh)->h_refcount));
 653                         }
 654                         unlock_buffer(new_bh);
 655                 } else if (old_bh && header == HDR(old_bh)) {
 656                         /* Keep this block. No need to lock the block as we
 657                            don't need to change the reference count. */
 658                         new_bh = old_bh;
 659                         get_bh(new_bh);
 660                         ext2_xattr_cache_insert(ea_block_cache, new_bh);
 661                 } else {
 662                         /* We need to allocate a new block */
 663                         ext2_fsblk_t goal = ext2_group_first_block_no(sb,
 664                                                 EXT2_I(inode)->i_block_group);
 665                         int block = ext2_new_block(inode, goal, &error);
 666                         if (error)
 667                                 goto cleanup;
 668                         ea_idebug(inode, "creating block %d", block);
 669
 670                         new_bh = sb_getblk(sb, block);
 671                         if (unlikely(!new_bh)) {
 672                                 ext2_free_blocks(inode, block, 1);
 673                                 mark_inode_dirty(inode);
 674                                 error = -ENOMEM;
 675                                 goto cleanup;
 676                         }
 677                         lock_buffer(new_bh);
 678                         memcpy(new_bh->b_data, header, new_bh->b_size);
 679                         set_buffer_uptodate(new_bh);
 680                         unlock_buffer(new_bh);
 681                         ext2_xattr_cache_insert(ea_block_cache, new_bh);
 682
 683                         ext2_xattr_update_super_block(sb);
 684                 }
 685                 mark_buffer_dirty(new_bh);
 686                 if (IS_SYNC(inode)) {
 687                         sync_dirty_buffer(new_bh);
 688                         error = -EIO;
 689                         if (buffer_req(new_bh) && !buffer_uptodate(new_bh))
 690                                 goto cleanup;
 691                 }
 692         }
 693
 694         /* Update the inode. */
 695         EXT2_I(inode)->i_file_acl = new_bh ? new_bh->b_blocknr : 0;
 696         inode->i_ctime = current_time(inode);
 697         if (IS_SYNC(inode)) {
 698                 error = sync_inode_metadata(inode, 1);
 699                 /* In case sync failed due to ENOSPC the inode was actually
 700                  * written (only some dirty data were not) so we just proceed
 701                  * as if nothing happened and cleanup the unused block */
 702                 if (error && error != -ENOSPC) {
 703                         if (new_bh && new_bh != old_bh) {
 704                                 dquot_free_block_nodirty(inode, 1);
 705                                 mark_inode_dirty(inode);
 706                         }
 707                         goto cleanup;
 708                 }
 709         } else
 710                 mark_inode_dirty(inode);
 711
 712         error = 0;
 713         if (old_bh && old_bh != new_bh) {
 714                 /*
 715                  * If there was an old block and we are no longer using it,
 716                  * release the old block.
 717                  */
 718                 lock_buffer(old_bh);
 719                 if (HDR(old_bh)->h_refcount == cpu_to_le32(1)) {
 720                         __u32 hash = le32_to_cpu(HDR(old_bh)->h_hash);
 721
 722                         /*
 723                          * This must happen under buffer lock for
 724                          * ext2_xattr_set2() to reliably detect freed block
 725                          */
 726                         mb_cache_entry_delete(ea_block_cache, hash,
 727                                               old_bh->b_blocknr);
 728                         /* Free the old block. */
 729                         ea_bdebug(old_bh, "freeing");
 730                         ext2_free_blocks(inode, old_bh->b_blocknr, 1);
 731                         mark_inode_dirty(inode);
 732                         /* We let our caller release old_bh, so we
 733                          * need to duplicate the buffer before. */
 734                         get_bh(old_bh);
 735                         bforget(old_bh);
 736                 } else {
 737                         /* Decrement the refcount only. */
 738                         le32_add_cpu(&HDR(old_bh)->h_refcount, -1);
 739                         dquot_free_block_nodirty(inode, 1);
 740                         mark_inode_dirty(inode);
 741                         mark_buffer_dirty(old_bh);
 742                         ea_bdebug(old_bh, "refcount now=%d",
 743                                 le32_to_cpu(HDR(old_bh)->h_refcount));
 744                 }
 745                 unlock_buffer(old_bh);
 746         }
 747
 748 cleanup:
 749         brelse(new_bh);
 750
 751         return error;
 752 }
 753
 754 /*
 755  * ext2_xattr_delete_inode()
 756  *
 757  * Free extended attribute resources associated with this inode. This
 758  * is called immediately before an inode is freed.
 759  */
 760 void
 761 ext2_xattr_delete_inode(struct inode *inode)
 762 {
 763         struct buffer_head *bh = NULL;
 764         struct ext2_sb_info *sbi = EXT2_SB(inode->i_sb);
 765
 766         down_write(&EXT2_I(inode)->xattr_sem);
 767         if (!EXT2_I(inode)->i_file_acl)
 768                 goto cleanup;
 769
 770         if (!ext2_data_block_valid(sbi, EXT2_I(inode)->i_file_acl, 0)) {
 771                 ext2_error(inode->i_sb, "ext2_xattr_delete_inode",
 772                         "inode %ld: xattr block %d is out of data blocks range",
 773                         inode->i_ino, EXT2_I(inode)->i_file_acl);
 774                 goto cleanup;
 775         }
 776
 777         bh = sb_bread(inode->i_sb, EXT2_I(inode)->i_file_acl);
 778         if (!bh) {
 779                 ext2_error(inode->i_sb, "ext2_xattr_delete_inode",
 780                         "inode %ld: block %d read error", inode->i_ino,
 781                         EXT2_I(inode)->i_file_acl);
 782                 goto cleanup;
 783         }
 784         ea_bdebug(bh, "b_count=%d", atomic_read(&(bh->b_count)));
 785         if (HDR(bh)->h_magic != cpu_to_le32(EXT2_XATTR_MAGIC) ||
 786             HDR(bh)->h_blocks != cpu_to_le32(1)) {
 787                 ext2_error(inode->i_sb, "ext2_xattr_delete_inode",
 788                         "inode %ld: bad block %d", inode->i_ino,
 789                         EXT2_I(inode)->i_file_acl);
 790                 goto cleanup;
 791         }
 792         lock_buffer(bh);
 793         if (HDR(bh)->h_refcount == cpu_to_le32(1)) {
 794                 __u32 hash = le32_to_cpu(HDR(bh)->h_hash);
 795
 796                 /*
 797                  * This must happen under buffer lock for ext2_xattr_set2() to
 798                  * reliably detect freed block
 799                  */
 800                 mb_cache_entry_delete(EA_BLOCK_CACHE(inode), hash,
 801                                       bh->b_blocknr);
 802                 ext2_free_blocks(inode, EXT2_I(inode)->i_file_acl, 1);
 803                 get_bh(bh);
 804                 bforget(bh);
 805                 unlock_buffer(bh);
 806         } else {
 807                 le32_add_cpu(&HDR(bh)->h_refcount, -1);
 808                 ea_bdebug(bh, "refcount now=%d",
 809                         le32_to_cpu(HDR(bh)->h_refcount));
 810                 unlock_buffer(bh);
 811                 mark_buffer_dirty(bh);
 812                 if (IS_SYNC(inode))
 813                         sync_dirty_buffer(bh);
 814                 dquot_free_block_nodirty(inode, 1);
 815         }
 816         EXT2_I(inode)->i_file_acl = 0;
 817
 818 cleanup:
 819         brelse(bh);
 820         up_write(&EXT2_I(inode)->xattr_sem);
 821 }
 822
 823 /*
 824  * ext2_xattr_cache_insert()
 825  *
 826  * Create a new entry in the extended attribute cache, and insert
 827  * it unless such an entry is already in the cache.
 828  *
 829  * Returns 0, or a negative error number on failure.
 830  */
 831 static int
 832 ext2_xattr_cache_insert(struct mb_cache *cache, struct buffer_head *bh)
 833 {
 834         __u32 hash = le32_to_cpu(HDR(bh)->h_hash);
 835         int error;
 836
 837         error = mb_cache_entry_create(cache, GFP_NOFS, hash, bh->b_blocknr, 1);
 838         if (error) {
 839                 if (error == -EBUSY) {
 840                         ea_bdebug(bh, "already in cache (%d cache entries)",
 841                                 atomic_read(&ext2_xattr_cache->c_entry_count));
 842                         error = 0;
 843                 }
 844         } else
 845                 ea_bdebug(bh, "inserting [%x]", (int)hash);
 846         return error;
 847 }
 848
 849 /*
 850  * ext2_xattr_cmp()
 851  *
 852  * Compare two extended attribute blocks for equality.
 853  *
 854  * Returns 0 if the blocks are equal, 1 if they differ, and
 855  * a negative error number on errors.
 856  */
 857 static int
 858 ext2_xattr_cmp(struct ext2_xattr_header *header1,
 859                struct ext2_xattr_header *header2)
 860 {
 861         struct ext2_xattr_entry *entry1, *entry2;
 862
 863         entry1 = ENTRY(header1+1);
 864         entry2 = ENTRY(header2+1);
 865         while (!IS_LAST_ENTRY(entry1)) {
 866                 if (IS_LAST_ENTRY(entry2))
 867                         return 1;
 868                 if (entry1->e_hash != entry2->e_hash ||
 869                     entry1->e_name_index != entry2->e_name_index ||
 870                     entry1->e_name_len != entry2->e_name_len ||
 871                     entry1->e_value_size != entry2->e_value_size ||
 872                     memcmp(entry1->e_name, entry2->e_name, entry1->e_name_len))
 873                         return 1;
 874                 if (entry1->e_value_block != 0 || entry2->e_value_block != 0)
 875                         return -EIO;
 876                 if (memcmp((char *)header1 + le16_to_cpu(entry1->e_value_offs),
 877                            (char *)header2 + le16_to_cpu(entry2->e_value_offs),
 878                            le32_to_cpu(entry1->e_value_size)))
 879                         return 1;
 880
 881                 entry1 = EXT2_XATTR_NEXT(entry1);
 882                 entry2 = EXT2_XATTR_NEXT(entry2);
 883         }
 884         if (!IS_LAST_ENTRY(entry2))
 885                 return 1;
 886         return 0;
 887 }
 888
 889 /*
 890  * ext2_xattr_cache_find()
 891  *
 892  * Find an identical extended attribute block.
 893  *
 894  * Returns a locked buffer head to the block found, or NULL if such
 895  * a block was not found or an error occurred.
 896  */
 897 static struct buffer_head *
 898 ext2_xattr_cache_find(struct inode *inode, struct ext2_xattr_header *header)
 899 {
 900         __u32 hash = le32_to_cpu(header->h_hash);
 901         struct mb_cache_entry *ce;
 902         struct mb_cache *ea_block_cache = EA_BLOCK_CACHE(inode);
 903
 904         if (!header->h_hash)
 905                 return NULL;  /* never share */
 906         ea_idebug(inode, "looking for cached blocks [%x]", (int)hash);
 907 again:
 908         ce = mb_cache_entry_find_first(ea_block_cache, hash);
 909         while (ce) {
 910                 struct buffer_head *bh;
 911
 912                 bh = sb_bread(inode->i_sb, ce->e_value);
 913                 if (!bh) {
 914                         ext2_error(inode->i_sb, "ext2_xattr_cache_find",
 915                                 "inode %ld: block %ld read error",
 916                                 inode->i_ino, (unsigned long) ce->e_value);
 917                 } else {
 918                         lock_buffer(bh);
 919                         /*
 920                          * We have to be careful about races with freeing or
 921                          * rehashing of xattr block. Once we hold buffer lock
 922                          * xattr block's state is stable so we can check
 923                          * whether the block got freed / rehashed or not.
 924                          * Since we unhash mbcache entry under buffer lock when
 925                          * freeing / rehashing xattr block, checking whether
 926                          * entry is still hashed is reliable.
 927                          */
 928                         if (hlist_bl_unhashed(&ce->e_hash_list)) {
 929                                 mb_cache_entry_put(ea_block_cache, ce);
 930                                 unlock_buffer(bh);
 931                                 brelse(bh);
 932                                 goto again;
 933                         } else if (le32_to_cpu(HDR(bh)->h_refcount) >
 934                                    EXT2_XATTR_REFCOUNT_MAX) {
 935                                 ea_idebug(inode, "block %ld refcount %d>%d",
 936                                           (unsigned long) ce->e_value,
 937                                           le32_to_cpu(HDR(bh)->h_refcount),
 938                                           EXT2_XATTR_REFCOUNT_MAX);
 939                         } else if (!ext2_xattr_cmp(header, HDR(bh))) {
 940                                 ea_bdebug(bh, "b_count=%d",
 941                                           atomic_read(&(bh->b_count)));
 942                                 mb_cache_entry_touch(ea_block_cache, ce);
 943                                 mb_cache_entry_put(ea_block_cache, ce);
 944                                 return bh;
 945                         }
 946                         unlock_buffer(bh);
 947                         brelse(bh);
 948                 }
 949                 ce = mb_cache_entry_find_next(ea_block_cache, ce);
 950         }
 951         return NULL;
 952 }
 953
 954 #define NAME_HASH_SHIFT 5
 955 #define VALUE_HASH_SHIFT 16
 956
 957 /*
 958  * ext2_xattr_hash_entry()
 959  *
 960  * Compute the hash of an extended attribute.
 961  */
 962 static inline void ext2_xattr_hash_entry(struct ext2_xattr_header *header,
 963                                          struct ext2_xattr_entry *entry)
 964 {
 965         __u32 hash = 0;
 966         char *name = entry->e_name;
 967         int n;
 968
 969         for (n=0; n < entry->e_name_len; n++) {
 970                 hash = (hash << NAME_HASH_SHIFT) ^
 971                        (hash >> (8*sizeof(hash) - NAME_HASH_SHIFT)) ^
 972                        *name++;
 973         }
 974
 975         if (entry->e_value_block == 0 && entry->e_value_size != 0) {
 976                 __le32 *value = (__le32 *)((char *)header +
 977                         le16_to_cpu(entry->e_value_offs));
 978                 for (n = (le32_to_cpu(entry->e_value_size) +
 979                      EXT2_XATTR_ROUND) >> EXT2_XATTR_PAD_BITS; n; n--) {
 980                         hash = (hash << VALUE_HASH_SHIFT) ^
 981                                (hash >> (8*sizeof(hash) - VALUE_HASH_SHIFT)) ^
 982                                le32_to_cpu(*value++);
 983                 }
 984         }
 985         entry->e_hash = cpu_to_le32(hash);
 986 }
 987
 988 #undef NAME_HASH_SHIFT
 989 #undef VALUE_HASH_SHIFT
 990
 991 #define BLOCK_HASH_SHIFT 16
 992
 993 /*
 994  * ext2_xattr_rehash()
 995  *
 996  * Re-compute the extended attribute hash value after an entry has changed.
 997  */
 998 static void ext2_xattr_rehash(struct ext2_xattr_header *header,
 999                               struct ext2_xattr_entry *entry)
1000 {
1001         struct ext2_xattr_entry *here;
1002         __u32 hash = 0;
1003
1004         ext2_xattr_hash_entry(header, entry);
1005         here = ENTRY(header+1);
1006         while (!IS_LAST_ENTRY(here)) {
1007                 if (!here->e_hash) {
1008                         /* Block is not shared if an entry's hash value == 0 */
1009                         hash = 0;
1010                         break;
1011                 }
1012                 hash = (hash << BLOCK_HASH_SHIFT) ^
1013                        (hash >> (8*sizeof(hash) - BLOCK_HASH_SHIFT)) ^
1014                        le32_to_cpu(here->e_hash);
1015                 here = EXT2_XATTR_NEXT(here);
1016         }
1017         header->h_hash = cpu_to_le32(hash);
1018 }
1019
1020 #undef BLOCK_HASH_SHIFT
1021
1022 #define HASH_BUCKET_BITS 10
1023
1024 struct mb_cache *ext2_xattr_create_cache(void)
1025 {
1026         return mb_cache_create(HASH_BUCKET_BITS);
1027 }
1028
1029 void ext2_xattr_destroy_cache(struct mb_cache *cache)
1030 {
1031         if (cache)
1032                 mb_cache_destroy(cache);
1033 }