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/mbcache2.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 mb2_cache *, 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 const struct xattr_handler *ext2_xattr_handler_map[] = {
 102         [EXT2_XATTR_INDEX_USER]              = &ext2_xattr_user_handler,
 103 #ifdef CONFIG_EXT2_FS_POSIX_ACL
 104         [EXT2_XATTR_INDEX_POSIX_ACL_ACCESS]  = &posix_acl_access_xattr_handler,
 105         [EXT2_XATTR_INDEX_POSIX_ACL_DEFAULT] = &posix_acl_default_xattr_handler,
 106 #endif
 107         [EXT2_XATTR_INDEX_TRUSTED]           = &ext2_xattr_trusted_handler,
 108 #ifdef CONFIG_EXT2_FS_SECURITY
 109         [EXT2_XATTR_INDEX_SECURITY]          = &ext2_xattr_security_handler,
 110 #endif
 111 };
 112
 113 const struct xattr_handler *ext2_xattr_handlers[] = {
 114         &ext2_xattr_user_handler,
 115         &ext2_xattr_trusted_handler,
 116 #ifdef CONFIG_EXT2_FS_POSIX_ACL
 117         &posix_acl_access_xattr_handler,
 118         &posix_acl_default_xattr_handler,
 119 #endif
 120 #ifdef CONFIG_EXT2_FS_SECURITY
 121         &ext2_xattr_security_handler,
 122 #endif
 123         NULL
 124 };
 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 mb2_cache *ext2_mb_cache = EXT2_SB(inode->i_sb)->s_mb_cache;
 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(ext2_mb_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(ext2_mb_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 mb2_cache *ext2_mb_cache = EXT2_SB(inode->i_sb)->s_mb_cache;
 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(ext2_mb_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) {
 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                 /* assert(header == HDR(bh)); */
 484                 lock_buffer(bh);
 485                 if (header->h_refcount == cpu_to_le32(1)) {
 486                         __u32 hash = le32_to_cpu(header->h_hash);
 487
 488                         ea_bdebug(bh, "modifying in-place");
 489                         /*
 490                          * This must happen under buffer lock for
 491                          * ext2_xattr_set2() to reliably detect modified block
 492                          */
 493                         mb2_cache_entry_delete_block(EXT2_SB(sb)->s_mb_cache,
 494                                                      hash, bh->b_blocknr);
 495
 496                         /* keep the buffer locked while modifying it. */
 497                 } else {
 498                         int offset;
 499
 500                         unlock_buffer(bh);
 501                         ea_bdebug(bh, "cloning");
 502                         header = kmalloc(bh->b_size, GFP_KERNEL);
 503                         error = -ENOMEM;
 504                         if (header == NULL)
 505                                 goto cleanup;
 506                         memcpy(header, HDR(bh), bh->b_size);
 507                         header->h_refcount = cpu_to_le32(1);
 508
 509                         offset = (char *)here - bh->b_data;
 510                         here = ENTRY((char *)header + offset);
 511                         offset = (char *)last - bh->b_data;
 512                         last = ENTRY((char *)header + offset);
 513                 }
 514         } else {
 515                 /* Allocate a buffer where we construct the new block. */
 516                 header = kzalloc(sb->s_blocksize, GFP_KERNEL);
 517                 error = -ENOMEM;
 518                 if (header == NULL)
 519                         goto cleanup;
 520                 end = (char *)header + sb->s_blocksize;
 521                 header->h_magic = cpu_to_le32(EXT2_XATTR_MAGIC);
 522                 header->h_blocks = header->h_refcount = cpu_to_le32(1);
 523                 last = here = ENTRY(header+1);
 524         }
 525
 526         /* Iff we are modifying the block in-place, bh is locked here. */
 527
 528         if (not_found) {
 529                 /* Insert the new name. */
 530                 size_t size = EXT2_XATTR_LEN(name_len);
 531                 size_t rest = (char *)last - (char *)here;
 532                 memmove((char *)here + size, here, rest);
 533                 memset(here, 0, size);
 534                 here->e_name_index = name_index;
 535                 here->e_name_len = name_len;
 536                 memcpy(here->e_name, name, name_len);
 537         } else {
 538                 if (!here->e_value_block && here->e_value_size) {
 539                         char *first_val = (char *)header + min_offs;
 540                         size_t offs = le16_to_cpu(here->e_value_offs);
 541                         char *val = (char *)header + offs;
 542                         size_t size = EXT2_XATTR_SIZE(
 543                                 le32_to_cpu(here->e_value_size));
 544
 545                         if (size == EXT2_XATTR_SIZE(value_len)) {
 546                                 /* The old and the new value have the same
 547                                    size. Just replace. */
 548                                 here->e_value_size = cpu_to_le32(value_len);
 549                                 memset(val + size - EXT2_XATTR_PAD, 0,
 550                                        EXT2_XATTR_PAD); /* Clear pad bytes. */
 551                                 memcpy(val, value, value_len);
 552                                 goto skip_replace;
 553                         }
 554
 555                         /* Remove the old value. */
 556                         memmove(first_val + size, first_val, val - first_val);
 557                         memset(first_val, 0, size);
 558                         here->e_value_offs = 0;
 559                         min_offs += size;
 560
 561                         /* Adjust all value offsets. */
 562                         last = ENTRY(header+1);
 563                         while (!IS_LAST_ENTRY(last)) {
 564                                 size_t o = le16_to_cpu(last->e_value_offs);
 565                                 if (!last->e_value_block && o < offs)
 566                                         last->e_value_offs =
 567                                                 cpu_to_le16(o + size);
 568                                 last = EXT2_XATTR_NEXT(last);
 569                         }
 570                 }
 571                 if (value == NULL) {
 572                         /* Remove the old name. */
 573                         size_t size = EXT2_XATTR_LEN(name_len);
 574                         last = ENTRY((char *)last - size);
 575                         memmove(here, (char*)here + size,
 576                                 (char*)last - (char*)here);
 577                         memset(last, 0, size);
 578                 }
 579         }
 580
 581         if (value != NULL) {
 582                 /* Insert the new value. */
 583                 here->e_value_size = cpu_to_le32(value_len);
 584                 if (value_len) {
 585                         size_t size = EXT2_XATTR_SIZE(value_len);
 586                         char *val = (char *)header + min_offs - size;
 587                         here->e_value_offs =
 588                                 cpu_to_le16((char *)val - (char *)header);
 589                         memset(val + size - EXT2_XATTR_PAD, 0,
 590                                EXT2_XATTR_PAD); /* Clear the pad bytes. */
 591                         memcpy(val, value, value_len);
 592                 }
 593         }
 594
 595 skip_replace:
 596         if (IS_LAST_ENTRY(ENTRY(header+1))) {
 597                 /* This block is now empty. */
 598                 if (bh && header == HDR(bh))
 599                         unlock_buffer(bh);  /* we were modifying in-place. */
 600                 error = ext2_xattr_set2(inode, bh, NULL);
 601         } else {
 602                 ext2_xattr_rehash(header, here);
 603                 if (bh && header == HDR(bh))
 604                         unlock_buffer(bh);  /* we were modifying in-place. */
 605                 error = ext2_xattr_set2(inode, bh, header);
 606         }
 607
 608 cleanup:
 609         if (!(bh && header == HDR(bh)))
 610                 kfree(header);
 611         brelse(bh);
 612         up_write(&EXT2_I(inode)->xattr_sem);
 613
 614         return error;
 615 }
 616
 617 /*
 618  * Second half of ext2_xattr_set(): Update the file system.
 619  */
 620 static int
 621 ext2_xattr_set2(struct inode *inode, struct buffer_head *old_bh,
 622                 struct ext2_xattr_header *header)
 623 {
 624         struct super_block *sb = inode->i_sb;
 625         struct buffer_head *new_bh = NULL;
 626         int error;
 627         struct mb2_cache *ext2_mb_cache = EXT2_SB(sb)->s_mb_cache;
 628
 629         if (header) {
 630                 new_bh = ext2_xattr_cache_find(inode, header);
 631                 if (new_bh) {
 632                         /* We found an identical block in the cache. */
 633                         if (new_bh == old_bh) {
 634                                 ea_bdebug(new_bh, "keeping this block");
 635                         } else {
 636                                 /* The old block is released after updating
 637                                    the inode.  */
 638                                 ea_bdebug(new_bh, "reusing block");
 639
 640                                 error = dquot_alloc_block(inode, 1);
 641                                 if (error) {
 642                                         unlock_buffer(new_bh);
 643                                         goto cleanup;
 644                                 }
 645                                 le32_add_cpu(&HDR(new_bh)->h_refcount, 1);
 646                                 ea_bdebug(new_bh, "refcount now=%d",
 647                                         le32_to_cpu(HDR(new_bh)->h_refcount));
 648                         }
 649                         unlock_buffer(new_bh);
 650                 } else if (old_bh && header == HDR(old_bh)) {
 651                         /* Keep this block. No need to lock the block as we
 652                            don't need to change the reference count. */
 653                         new_bh = old_bh;
 654                         get_bh(new_bh);
 655                         ext2_xattr_cache_insert(ext2_mb_cache, new_bh);
 656                 } else {
 657                         /* We need to allocate a new block */
 658                         ext2_fsblk_t goal = ext2_group_first_block_no(sb,
 659                                                 EXT2_I(inode)->i_block_group);
 660                         int block = ext2_new_block(inode, goal, &error);
 661                         if (error)
 662                                 goto cleanup;
 663                         ea_idebug(inode, "creating block %d", block);
 664
 665                         new_bh = sb_getblk(sb, block);
 666                         if (unlikely(!new_bh)) {
 667                                 ext2_free_blocks(inode, block, 1);
 668                                 mark_inode_dirty(inode);
 669                                 error = -ENOMEM;
 670                                 goto cleanup;
 671                         }
 672                         lock_buffer(new_bh);
 673                         memcpy(new_bh->b_data, header, new_bh->b_size);
 674                         set_buffer_uptodate(new_bh);
 675                         unlock_buffer(new_bh);
 676                         ext2_xattr_cache_insert(ext2_mb_cache, new_bh);
 677
 678                         ext2_xattr_update_super_block(sb);
 679                 }
 680                 mark_buffer_dirty(new_bh);
 681                 if (IS_SYNC(inode)) {
 682                         sync_dirty_buffer(new_bh);
 683                         error = -EIO;
 684                         if (buffer_req(new_bh) && !buffer_uptodate(new_bh))
 685                                 goto cleanup;
 686                 }
 687         }
 688
 689         /* Update the inode. */
 690         EXT2_I(inode)->i_file_acl = new_bh ? new_bh->b_blocknr : 0;
 691         inode->i_ctime = CURRENT_TIME_SEC;
 692         if (IS_SYNC(inode)) {
 693                 error = sync_inode_metadata(inode, 1);
 694                 /* In case sync failed due to ENOSPC the inode was actually
 695                  * written (only some dirty data were not) so we just proceed
 696                  * as if nothing happened and cleanup the unused block */
 697                 if (error && error != -ENOSPC) {
 698                         if (new_bh && new_bh != old_bh) {
 699                                 dquot_free_block_nodirty(inode, 1);
 700                                 mark_inode_dirty(inode);
 701                         }
 702                         goto cleanup;
 703                 }
 704         } else
 705                 mark_inode_dirty(inode);
 706
 707         error = 0;
 708         if (old_bh && old_bh != new_bh) {
 709                 /*
 710                  * If there was an old block and we are no longer using it,
 711                  * release the old block.
 712                  */
 713                 lock_buffer(old_bh);
 714                 if (HDR(old_bh)->h_refcount == cpu_to_le32(1)) {
 715                         __u32 hash = le32_to_cpu(HDR(old_bh)->h_hash);
 716
 717                         /*
 718                          * This must happen under buffer lock for
 719                          * ext2_xattr_set2() to reliably detect freed block
 720                          */
 721                         mb2_cache_entry_delete_block(ext2_mb_cache,
 722                                                      hash, old_bh->b_blocknr);
 723                         /* Free the old block. */
 724                         ea_bdebug(old_bh, "freeing");
 725                         ext2_free_blocks(inode, old_bh->b_blocknr, 1);
 726                         mark_inode_dirty(inode);
 727                         /* We let our caller release old_bh, so we
 728                          * need to duplicate the buffer before. */
 729                         get_bh(old_bh);
 730                         bforget(old_bh);
 731                 } else {
 732                         /* Decrement the refcount only. */
 733                         le32_add_cpu(&HDR(old_bh)->h_refcount, -1);
 734                         dquot_free_block_nodirty(inode, 1);
 735                         mark_inode_dirty(inode);
 736                         mark_buffer_dirty(old_bh);
 737                         ea_bdebug(old_bh, "refcount now=%d",
 738                                 le32_to_cpu(HDR(old_bh)->h_refcount));
 739                 }
 740                 unlock_buffer(old_bh);
 741         }
 742
 743 cleanup:
 744         brelse(new_bh);
 745
 746         return error;
 747 }
 748
 749 /*
 750  * ext2_xattr_delete_inode()
 751  *
 752  * Free extended attribute resources associated with this inode. This
 753  * is called immediately before an inode is freed.
 754  */
 755 void
 756 ext2_xattr_delete_inode(struct inode *inode)
 757 {
 758         struct buffer_head *bh = NULL;
 759
 760         down_write(&EXT2_I(inode)->xattr_sem);
 761         if (!EXT2_I(inode)->i_file_acl)
 762                 goto cleanup;
 763         bh = sb_bread(inode->i_sb, EXT2_I(inode)->i_file_acl);
 764         if (!bh) {
 765                 ext2_error(inode->i_sb, "ext2_xattr_delete_inode",
 766                         "inode %ld: block %d read error", inode->i_ino,
 767                         EXT2_I(inode)->i_file_acl);
 768                 goto cleanup;
 769         }
 770         ea_bdebug(bh, "b_count=%d", atomic_read(&(bh->b_count)));
 771         if (HDR(bh)->h_magic != cpu_to_le32(EXT2_XATTR_MAGIC) ||
 772             HDR(bh)->h_blocks != cpu_to_le32(1)) {
 773                 ext2_error(inode->i_sb, "ext2_xattr_delete_inode",
 774                         "inode %ld: bad block %d", inode->i_ino,
 775                         EXT2_I(inode)->i_file_acl);
 776                 goto cleanup;
 777         }
 778         lock_buffer(bh);
 779         if (HDR(bh)->h_refcount == cpu_to_le32(1)) {
 780                 __u32 hash = le32_to_cpu(HDR(bh)->h_hash);
 781
 782                 /*
 783                  * This must happen under buffer lock for ext2_xattr_set2() to
 784                  * reliably detect freed block
 785                  */
 786                 mb2_cache_entry_delete_block(EXT2_SB(inode->i_sb)->s_mb_cache,
 787                                              hash, bh->b_blocknr);
 788                 ext2_free_blocks(inode, EXT2_I(inode)->i_file_acl, 1);
 789                 get_bh(bh);
 790                 bforget(bh);
 791                 unlock_buffer(bh);
 792         } else {
 793                 le32_add_cpu(&HDR(bh)->h_refcount, -1);
 794                 ea_bdebug(bh, "refcount now=%d",
 795                         le32_to_cpu(HDR(bh)->h_refcount));
 796                 unlock_buffer(bh);
 797                 mark_buffer_dirty(bh);
 798                 if (IS_SYNC(inode))
 799                         sync_dirty_buffer(bh);
 800                 dquot_free_block_nodirty(inode, 1);
 801         }
 802         EXT2_I(inode)->i_file_acl = 0;
 803
 804 cleanup:
 805         brelse(bh);
 806         up_write(&EXT2_I(inode)->xattr_sem);
 807 }
 808
 809 /*
 810  * ext2_xattr_cache_insert()
 811  *
 812  * Create a new entry in the extended attribute cache, and insert
 813  * it unless such an entry is already in the cache.
 814  *
 815  * Returns 0, or a negative error number on failure.
 816  */
 817 static int
 818 ext2_xattr_cache_insert(struct mb2_cache *cache, struct buffer_head *bh)
 819 {
 820         __u32 hash = le32_to_cpu(HDR(bh)->h_hash);
 821         int error;
 822
 823         error = mb2_cache_entry_create(cache, GFP_NOFS, hash, bh->b_blocknr);
 824         if (error) {
 825                 if (error == -EBUSY) {
 826                         ea_bdebug(bh, "already in cache (%d cache entries)",
 827                                 atomic_read(&ext2_xattr_cache->c_entry_count));
 828                         error = 0;
 829                 }
 830         } else
 831                 ea_bdebug(bh, "inserting [%x]", (int)hash);
 832         return error;
 833 }
 834
 835 /*
 836  * ext2_xattr_cmp()
 837  *
 838  * Compare two extended attribute blocks for equality.
 839  *
 840  * Returns 0 if the blocks are equal, 1 if they differ, and
 841  * a negative error number on errors.
 842  */
 843 static int
 844 ext2_xattr_cmp(struct ext2_xattr_header *header1,
 845                struct ext2_xattr_header *header2)
 846 {
 847         struct ext2_xattr_entry *entry1, *entry2;
 848
 849         entry1 = ENTRY(header1+1);
 850         entry2 = ENTRY(header2+1);
 851         while (!IS_LAST_ENTRY(entry1)) {
 852                 if (IS_LAST_ENTRY(entry2))
 853                         return 1;
 854                 if (entry1->e_hash != entry2->e_hash ||
 855                     entry1->e_name_index != entry2->e_name_index ||
 856                     entry1->e_name_len != entry2->e_name_len ||
 857                     entry1->e_value_size != entry2->e_value_size ||
 858                     memcmp(entry1->e_name, entry2->e_name, entry1->e_name_len))
 859                         return 1;
 860                 if (entry1->e_value_block != 0 || entry2->e_value_block != 0)
 861                         return -EIO;
 862                 if (memcmp((char *)header1 + le16_to_cpu(entry1->e_value_offs),
 863                            (char *)header2 + le16_to_cpu(entry2->e_value_offs),
 864                            le32_to_cpu(entry1->e_value_size)))
 865                         return 1;
 866
 867                 entry1 = EXT2_XATTR_NEXT(entry1);
 868                 entry2 = EXT2_XATTR_NEXT(entry2);
 869         }
 870         if (!IS_LAST_ENTRY(entry2))
 871                 return 1;
 872         return 0;
 873 }
 874
 875 /*
 876  * ext2_xattr_cache_find()
 877  *
 878  * Find an identical extended attribute block.
 879  *
 880  * Returns a locked buffer head to the block found, or NULL if such
 881  * a block was not found or an error occurred.
 882  */
 883 static struct buffer_head *
 884 ext2_xattr_cache_find(struct inode *inode, struct ext2_xattr_header *header)
 885 {
 886         __u32 hash = le32_to_cpu(header->h_hash);
 887         struct mb2_cache_entry *ce;
 888         struct mb2_cache *ext2_mb_cache = EXT2_SB(inode->i_sb)->s_mb_cache;
 889
 890         if (!header->h_hash)
 891                 return NULL;  /* never share */
 892         ea_idebug(inode, "looking for cached blocks [%x]", (int)hash);
 893 again:
 894         ce = mb2_cache_entry_find_first(ext2_mb_cache, hash);
 895         while (ce) {
 896                 struct buffer_head *bh;
 897
 898                 bh = sb_bread(inode->i_sb, ce->e_block);
 899                 if (!bh) {
 900                         ext2_error(inode->i_sb, "ext2_xattr_cache_find",
 901                                 "inode %ld: block %ld read error",
 902                                 inode->i_ino, (unsigned long) ce->e_block);
 903                 } else {
 904                         lock_buffer(bh);
 905                         /*
 906                          * We have to be careful about races with freeing or
 907                          * rehashing of xattr block. Once we hold buffer lock
 908                          * xattr block's state is stable so we can check
 909                          * whether the block got freed / rehashed or not.
 910                          * Since we unhash mbcache entry under buffer lock when
 911                          * freeing / rehashing xattr block, checking whether
 912                          * entry is still hashed is reliable.
 913                          */
 914                         if (hlist_bl_unhashed(&ce->e_hash_list)) {
 915                                 mb2_cache_entry_put(ext2_mb_cache, ce);
 916                                 unlock_buffer(bh);
 917                                 brelse(bh);
 918                                 goto again;
 919                         } else if (le32_to_cpu(HDR(bh)->h_refcount) >
 920                                    EXT2_XATTR_REFCOUNT_MAX) {
 921                                 ea_idebug(inode, "block %ld refcount %d>%d",
 922                                           (unsigned long) ce->e_block,
 923                                           le32_to_cpu(HDR(bh)->h_refcount),
 924                                           EXT2_XATTR_REFCOUNT_MAX);
 925                         } else if (!ext2_xattr_cmp(header, HDR(bh))) {
 926                                 ea_bdebug(bh, "b_count=%d",
 927                                           atomic_read(&(bh->b_count)));
 928                                 mb2_cache_entry_touch(ext2_mb_cache, ce);
 929                                 mb2_cache_entry_put(ext2_mb_cache, ce);
 930                                 return bh;
 931                         }
 932                         unlock_buffer(bh);
 933                         brelse(bh);
 934                 }
 935                 ce = mb2_cache_entry_find_next(ext2_mb_cache, ce);
 936         }
 937         return NULL;
 938 }
 939
 940 #define NAME_HASH_SHIFT 5
 941 #define VALUE_HASH_SHIFT 16
 942
 943 /*
 944  * ext2_xattr_hash_entry()
 945  *
 946  * Compute the hash of an extended attribute.
 947  */
 948 static inline void ext2_xattr_hash_entry(struct ext2_xattr_header *header,
 949                                          struct ext2_xattr_entry *entry)
 950 {
 951         __u32 hash = 0;
 952         char *name = entry->e_name;
 953         int n;
 954
 955         for (n=0; n < entry->e_name_len; n++) {
 956                 hash = (hash << NAME_HASH_SHIFT) ^
 957                        (hash >> (8*sizeof(hash) - NAME_HASH_SHIFT)) ^
 958                        *name++;
 959         }
 960
 961         if (entry->e_value_block == 0 && entry->e_value_size != 0) {
 962                 __le32 *value = (__le32 *)((char *)header +
 963                         le16_to_cpu(entry->e_value_offs));
 964                 for (n = (le32_to_cpu(entry->e_value_size) +
 965                      EXT2_XATTR_ROUND) >> EXT2_XATTR_PAD_BITS; n; n--) {
 966                         hash = (hash << VALUE_HASH_SHIFT) ^
 967                                (hash >> (8*sizeof(hash) - VALUE_HASH_SHIFT)) ^
 968                                le32_to_cpu(*value++);
 969                 }
 970         }
 971         entry->e_hash = cpu_to_le32(hash);
 972 }
 973
 974 #undef NAME_HASH_SHIFT
 975 #undef VALUE_HASH_SHIFT
 976
 977 #define BLOCK_HASH_SHIFT 16
 978
 979 /*
 980  * ext2_xattr_rehash()
 981  *
 982  * Re-compute the extended attribute hash value after an entry has changed.
 983  */
 984 static void ext2_xattr_rehash(struct ext2_xattr_header *header,
 985                               struct ext2_xattr_entry *entry)
 986 {
 987         struct ext2_xattr_entry *here;
 988         __u32 hash = 0;
 989
 990         ext2_xattr_hash_entry(header, entry);
 991         here = ENTRY(header+1);
 992         while (!IS_LAST_ENTRY(here)) {
 993                 if (!here->e_hash) {
 994                         /* Block is not shared if an entry's hash value == 0 */
 995                         hash = 0;
 996                         break;
 997                 }
 998                 hash = (hash << BLOCK_HASH_SHIFT) ^
 999                        (hash >> (8*sizeof(hash) - BLOCK_HASH_SHIFT)) ^
1000                        le32_to_cpu(here->e_hash);
1001                 here = EXT2_XATTR_NEXT(here);
1002         }
1003         header->h_hash = cpu_to_le32(hash);
1004 }
1005
1006 #undef BLOCK_HASH_SHIFT
1007
1008 #define HASH_BUCKET_BITS 10
1009
1010 struct mb2_cache *ext2_xattr_create_cache(void)
1011 {
1012         return mb2_cache_create(HASH_BUCKET_BITS);
1013 }
1014
1015 void ext2_xattr_destroy_cache(struct mb2_cache *cache)
1016 {
1017         if (cache)
1018                 mb2_cache_destroy(cache);
1019 }