fs/ext2/xattr.c

   1 /*
   2  * linux/fs/ext2/xattr.c
   3  *
   4  * Copyright (C) 2001-2003 Andreas Gruenbacher <agruen@suse.de>
   5  *
   6  * Fix by Harrison Xing <harrison@mountainviewdata.com>.
   7  * Extended attributes for symlinks and special files added per
   8  *  suggestion of Luka Renko <luka.renko@hermes.si>.
   9  * xattr consolidation Copyright (c) 2004 James Morris <jmorris@redhat.com>,
  10  *  Red Hat Inc.
  11  *
  12  */
  13
  14 /*
  15  * Extended attributes are stored on disk blocks allocated outside of
  16  * any inode. The i_file_acl field is then made to point to this allocated
  17  * block. If all extended attributes of an inode are identical, these
  18  * inodes may share the same extended attribute block. Such situations
  19  * are automatically detected by keeping a cache of recent attribute block
  20  * numbers and hashes over the block's contents in memory.
  21  *
  22  *
  23  * Extended attribute block layout:
  24  *
  25  *   +------------------+
  26  *   | header           |
  27  *   | entry 1          | |
  28  *   | entry 2          | | growing downwards
  29  *   | entry 3          | v
  30  *   | four null bytes  |
  31  *   | . . .            |
  32  *   | value 1          | ^
  33  *   | value 3          | | growing upwards
  34  *   | value 2          | |
  35  *   +------------------+
  36  *
  37  * The block header is followed by multiple entry descriptors. These entry
  38  * descriptors are variable in size, and alligned to EXT2_XATTR_PAD
  39  * byte boundaries. The entry descriptors are sorted by attribute name,
  40  * so that two extended attribute blocks can be compared efficiently.
  41  *
  42  * Attribute values are aligned to the end of the block, stored in
  43  * no specific order. They are also padded to EXT2_XATTR_PAD byte
  44  * boundaries. No additional gaps are left between them.
  45  *
  46  * Locking strategy
  47  * ----------------
  48  * EXT2_I(inode)->i_file_acl is protected by EXT2_I(inode)->xattr_sem.
  49  * EA blocks are only changed if they are exclusive to an inode, so
  50  * holding xattr_sem also means that nothing but the EA block's reference
  51  * count will change. Multiple writers to an EA block are synchronized
  52  * by the bh lock. No more than a single bh lock is held at any time
  53  * to avoid deadlocks.
  54  */
  55
  56 #include <linux/buffer_head.h>
  57 #include <linux/module.h>
  58 #include <linux/init.h>
  59 #include <linux/slab.h>
  60 #include <linux/mbcache.h>
  61 #include <linux/quotaops.h>
  62 #include <linux/rwsem.h>
  63 #include <linux/vs_dlimit.h>
  64 #include "ext2.h"
  65 #include "xattr.h"
  66 #include "acl.h"
  67
  68 #define HDR(bh) ((struct ext2_xattr_header *)((bh)->b_data))
  69 #define ENTRY(ptr) ((struct ext2_xattr_entry *)(ptr))
  70 #define FIRST_ENTRY(bh) ENTRY(HDR(bh)+1)
  71 #define IS_LAST_ENTRY(entry) (*(__u32 *)(entry) == 0)
  72
  73 #ifdef EXT2_XATTR_DEBUG
  74 # define ea_idebug(inode, f...) do { \
  75                 printk(KERN_DEBUG "inode %s:%ld: ", \
  76                         inode->i_sb->s_id, inode->i_ino); \
  77                 printk(f); \
  78                 printk("\n"); \
  79         } while (0)
  80 # define ea_bdebug(bh, f...) do { \
  81                 char b[BDEVNAME_SIZE]; \
  82                 printk(KERN_DEBUG "block %s:%lu: ", \
  83                         bdevname(bh->b_bdev, b), \
  84                         (unsigned long) bh->b_blocknr); \
  85                 printk(f); \
  86                 printk("\n"); \
  87         } while (0)
  88 #else
  89 # define ea_idebug(f...)
  90 # define ea_bdebug(f...)
  91 #endif
  92
  93 static int ext2_xattr_set2(struct inode *, struct buffer_head *,
  94                            struct ext2_xattr_header *);
  95
  96 static int ext2_xattr_cache_insert(struct buffer_head *);
  97 static struct buffer_head *ext2_xattr_cache_find(struct inode *,
  98                                                  struct ext2_xattr_header *);
  99 static void ext2_xattr_rehash(struct ext2_xattr_header *,
 100                               struct ext2_xattr_entry *);
 101
 102 static struct mb_cache *ext2_xattr_cache;
 103
 104 static struct xattr_handler *ext2_xattr_handler_map[] = {
 105         [EXT2_XATTR_INDEX_USER]              = &ext2_xattr_user_handler,
 106 #ifdef CONFIG_EXT2_FS_POSIX_ACL
 107         [EXT2_XATTR_INDEX_POSIX_ACL_ACCESS]  = &ext2_xattr_acl_access_handler,
 108         [EXT2_XATTR_INDEX_POSIX_ACL_DEFAULT] = &ext2_xattr_acl_default_handler,
 109 #endif
 110         [EXT2_XATTR_INDEX_TRUSTED]           = &ext2_xattr_trusted_handler,
 111 #ifdef CONFIG_EXT2_FS_SECURITY
 112         [EXT2_XATTR_INDEX_SECURITY]          = &ext2_xattr_security_handler,
 113 #endif
 114 };
 115
 116 struct xattr_handler *ext2_xattr_handlers[] = {
 117         &ext2_xattr_user_handler,
 118         &ext2_xattr_trusted_handler,
 119 #ifdef CONFIG_EXT2_FS_POSIX_ACL
 120         &ext2_xattr_acl_access_handler,
 121         &ext2_xattr_acl_default_handler,
 122 #endif
 123 #ifdef CONFIG_EXT2_FS_SECURITY
 124         &ext2_xattr_security_handler,
 125 #endif
 126         NULL
 127 };
 128
 129 static inline struct xattr_handler *
 130 ext2_xattr_handler(int name_index)
 131 {
 132         struct xattr_handler *handler = NULL;
 133
 134         if (name_index > 0 && name_index < ARRAY_SIZE(ext2_xattr_handler_map))
 135                 handler = ext2_xattr_handler_map[name_index];
 136         return handler;
 137 }
 138
 139 /*
 140  * ext2_xattr_get()
 141  *
 142  * Copy an extended attribute into the buffer
 143  * provided, or compute the buffer size required.
 144  * Buffer is NULL to compute the size of the buffer required.
 145  *
 146  * Returns a negative error number on failure, or the number of bytes
 147  * used / required on success.
 148  */
 149 int
 150 ext2_xattr_get(struct inode *inode, int name_index, const char *name,
 151                void *buffer, size_t buffer_size)
 152 {
 153         struct buffer_head *bh = NULL;
 154         struct ext2_xattr_entry *entry;
 155         size_t name_len, size;
 156         char *end;
 157         int error;
 158
 159         ea_idebug(inode, "name=%d.%s, buffer=%p, buffer_size=%ld",
 160                   name_index, name, buffer, (long)buffer_size);
 161
 162         if (name == NULL)
 163                 return -EINVAL;
 164         down_read(&EXT2_I(inode)->xattr_sem);
 165         error = -ENODATA;
 166         if (!EXT2_I(inode)->i_file_acl)
 167                 goto cleanup;
 168         ea_idebug(inode, "reading block %d", EXT2_I(inode)->i_file_acl);
 169         bh = sb_bread(inode->i_sb, EXT2_I(inode)->i_file_acl);
 170         error = -EIO;
 171         if (!bh)
 172                 goto cleanup;
 173         ea_bdebug(bh, "b_count=%d, refcount=%d",
 174                 atomic_read(&(bh->b_count)), le32_to_cpu(HDR(bh)->h_refcount));
 175         end = bh->b_data + bh->b_size;
 176         if (HDR(bh)->h_magic != cpu_to_le32(EXT2_XATTR_MAGIC) ||
 177             HDR(bh)->h_blocks != cpu_to_le32(1)) {
 178 bad_block:      ext2_error(inode->i_sb, "ext2_xattr_get",
 179                         "inode %ld: bad block %d", inode->i_ino,
 180                         EXT2_I(inode)->i_file_acl);
 181                 error = -EIO;
 182                 goto cleanup;
 183         }
 184         /* find named attribute */
 185         name_len = strlen(name);
 186
 187         error = -ERANGE;
 188         if (name_len > 255)
 189                 goto cleanup;
 190         entry = FIRST_ENTRY(bh);
 191         while (!IS_LAST_ENTRY(entry)) {
 192                 struct ext2_xattr_entry *next =
 193                         EXT2_XATTR_NEXT(entry);
 194                 if ((char *)next >= end)
 195                         goto bad_block;
 196                 if (name_index == entry->e_name_index &&
 197                     name_len == entry->e_name_len &&
 198                     memcmp(name, entry->e_name, name_len) == 0)
 199                         goto found;
 200                 entry = next;
 201         }
 202         /* Check the remaining name entries */
 203         while (!IS_LAST_ENTRY(entry)) {
 204                 struct ext2_xattr_entry *next =
 205                         EXT2_XATTR_NEXT(entry);
 206                 if ((char *)next >= end)
 207                         goto bad_block;
 208                 entry = next;
 209         }
 210         if (ext2_xattr_cache_insert(bh))
 211                 ea_idebug(inode, "cache insert failed");
 212         error = -ENODATA;
 213         goto cleanup;
 214 found:
 215         /* check the buffer size */
 216         if (entry->e_value_block != 0)
 217                 goto bad_block;
 218         size = le32_to_cpu(entry->e_value_size);
 219         if (size > inode->i_sb->s_blocksize ||
 220             le16_to_cpu(entry->e_value_offs) + size > inode->i_sb->s_blocksize)
 221                 goto bad_block;
 222
 223         if (ext2_xattr_cache_insert(bh))
 224                 ea_idebug(inode, "cache insert failed");
 225         if (buffer) {
 226                 error = -ERANGE;
 227                 if (size > buffer_size)
 228                         goto cleanup;
 229                 /* return value of attribute */
 230                 memcpy(buffer, bh->b_data + le16_to_cpu(entry->e_value_offs),
 231                         size);
 232         }
 233         error = size;
 234
 235 cleanup:
 236         brelse(bh);
 237         up_read(&EXT2_I(inode)->xattr_sem);
 238
 239         return error;
 240 }
 241
 242 /*
 243  * ext2_xattr_list()
 244  *
 245  * Copy a list of attribute names into the buffer
 246  * provided, or compute the buffer size required.
 247  * Buffer is NULL to compute the size of the buffer required.
 248  *
 249  * Returns a negative error number on failure, or the number of bytes
 250  * used / required on success.
 251  */
 252 static int
 253 ext2_xattr_list(struct inode *inode, char *buffer, size_t buffer_size)
 254 {
 255         struct buffer_head *bh = NULL;
 256         struct ext2_xattr_entry *entry;
 257         char *end;
 258         size_t rest = buffer_size;
 259         int error;
 260
 261         ea_idebug(inode, "buffer=%p, buffer_size=%ld",
 262                   buffer, (long)buffer_size);
 263
 264         down_read(&EXT2_I(inode)->xattr_sem);
 265         error = 0;
 266         if (!EXT2_I(inode)->i_file_acl)
 267                 goto cleanup;
 268         ea_idebug(inode, "reading block %d", EXT2_I(inode)->i_file_acl);
 269         bh = sb_bread(inode->i_sb, EXT2_I(inode)->i_file_acl);
 270         error = -EIO;
 271         if (!bh)
 272                 goto cleanup;
 273         ea_bdebug(bh, "b_count=%d, refcount=%d",
 274                 atomic_read(&(bh->b_count)), le32_to_cpu(HDR(bh)->h_refcount));
 275         end = bh->b_data + bh->b_size;
 276         if (HDR(bh)->h_magic != cpu_to_le32(EXT2_XATTR_MAGIC) ||
 277             HDR(bh)->h_blocks != cpu_to_le32(1)) {
 278 bad_block:      ext2_error(inode->i_sb, "ext2_xattr_list",
 279                         "inode %ld: bad block %d", inode->i_ino,
 280                         EXT2_I(inode)->i_file_acl);
 281                 error = -EIO;
 282                 goto cleanup;
 283         }
 284
 285         /* check the on-disk data structure */
 286         entry = FIRST_ENTRY(bh);
 287         while (!IS_LAST_ENTRY(entry)) {
 288                 struct ext2_xattr_entry *next = EXT2_XATTR_NEXT(entry);
 289
 290                 if ((char *)next >= end)
 291                         goto bad_block;
 292                 entry = next;
 293         }
 294         if (ext2_xattr_cache_insert(bh))
 295                 ea_idebug(inode, "cache insert failed");
 296
 297         /* list the attribute names */
 298         for (entry = FIRST_ENTRY(bh); !IS_LAST_ENTRY(entry);
 299              entry = EXT2_XATTR_NEXT(entry)) {
 300                 struct xattr_handler *handler =
 301                         ext2_xattr_handler(entry->e_name_index);
 302
 303                 if (handler) {
 304                         size_t size = handler->list(inode, buffer, rest,
 305                                                     entry->e_name,
 306                                                     entry->e_name_len);
 307                         if (buffer) {
 308                                 if (size > rest) {
 309                                         error = -ERANGE;
 310                                         goto cleanup;
 311                                 }
 312                                 buffer += size;
 313                         }
 314                         rest -= size;
 315                 }
 316         }
 317         error = buffer_size - rest;  /* total size */
 318
 319 cleanup:
 320         brelse(bh);
 321         up_read(&EXT2_I(inode)->xattr_sem);
 322
 323         return error;
 324 }
 325
 326 /*
 327  * Inode operation listxattr()
 328  *
 329  * dentry->d_inode->i_sem: don't care
 330  */
 331 ssize_t
 332 ext2_listxattr(struct dentry *dentry, char *buffer, size_t size)
 333 {
 334         return ext2_xattr_list(dentry->d_inode, buffer, size);
 335 }
 336
 337 /*
 338  * If the EXT2_FEATURE_COMPAT_EXT_ATTR feature of this file system is
 339  * not set, set it.
 340  */
 341 static void ext2_xattr_update_super_block(struct super_block *sb)
 342 {
 343         if (EXT2_HAS_COMPAT_FEATURE(sb, EXT2_FEATURE_COMPAT_EXT_ATTR))
 344                 return;
 345
 346         lock_super(sb);
 347         EXT2_SB(sb)->s_es->s_feature_compat |=
 348                 cpu_to_le32(EXT2_FEATURE_COMPAT_EXT_ATTR);
 349         sb->s_dirt = 1;
 350         mark_buffer_dirty(EXT2_SB(sb)->s_sbh);
 351         unlock_super(sb);
 352 }
 353
 354 /*
 355  * ext2_xattr_set()
 356  *
 357  * Create, replace or remove an extended attribute for this inode. Buffer
 358  * is NULL to remove an existing extended attribute, and non-NULL to
 359  * either replace an existing extended attribute, or create a new extended
 360  * attribute. The flags XATTR_REPLACE and XATTR_CREATE
 361  * specify that an extended attribute must exist and must not exist
 362  * previous to the call, respectively.
 363  *
 364  * Returns 0, or a negative error number on failure.
 365  */
 366 int
 367 ext2_xattr_set(struct inode *inode, int name_index, const char *name,
 368                const void *value, size_t value_len, int flags)
 369 {
 370         struct super_block *sb = inode->i_sb;
 371         struct buffer_head *bh = NULL;
 372         struct ext2_xattr_header *header = NULL;
 373         struct ext2_xattr_entry *here, *last;
 374         size_t name_len, free, min_offs = sb->s_blocksize;
 375         int not_found = 1, error;
 376         char *end;
 377
 378         /*
 379          * header -- Points either into bh, or to a temporarily
 380          *           allocated buffer.
 381          * here -- The named entry found, or the place for inserting, within
 382          *         the block pointed to by header.
 383          * last -- Points right after the last named entry within the block
 384          *         pointed to by header.
 385          * min_offs -- The offset of the first value (values are aligned
 386          *             towards the end of the block).
 387          * end -- Points right after the block pointed to by header.
 388          */
 389
 390         ea_idebug(inode, "name=%d.%s, value=%p, value_len=%ld",
 391                   name_index, name, value, (long)value_len);
 392
 393         if (IS_RDONLY(inode))
 394                 return -EROFS;
 395         if (IS_IMMUTABLE(inode) || IS_APPEND(inode))
 396                 return -EPERM;
 397         if (value == NULL)
 398                 value_len = 0;
 399         if (name == NULL)
 400                 return -EINVAL;
 401         name_len = strlen(name);
 402         if (name_len > 255 || value_len > sb->s_blocksize)
 403                 return -ERANGE;
 404         down_write(&EXT2_I(inode)->xattr_sem);
 405         if (EXT2_I(inode)->i_file_acl) {
 406                 /* The inode already has an extended attribute block. */
 407                 bh = sb_bread(sb, EXT2_I(inode)->i_file_acl);
 408                 error = -EIO;
 409                 if (!bh)
 410                         goto cleanup;
 411                 ea_bdebug(bh, "b_count=%d, refcount=%d",
 412                         atomic_read(&(bh->b_count)),
 413                         le32_to_cpu(HDR(bh)->h_refcount));
 414                 header = HDR(bh);
 415                 end = bh->b_data + bh->b_size;
 416                 if (header->h_magic != cpu_to_le32(EXT2_XATTR_MAGIC) ||
 417                     header->h_blocks != cpu_to_le32(1)) {
 418 bad_block:              ext2_error(sb, "ext2_xattr_set",
 419                                 "inode %ld: bad block %d", inode->i_ino,
 420                                    EXT2_I(inode)->i_file_acl);
 421                         error = -EIO;
 422                         goto cleanup;
 423                 }
 424                 /* Find the named attribute. */
 425                 here = FIRST_ENTRY(bh);
 426                 while (!IS_LAST_ENTRY(here)) {
 427                         struct ext2_xattr_entry *next = EXT2_XATTR_NEXT(here);
 428                         if ((char *)next >= end)
 429                                 goto bad_block;
 430                         if (!here->e_value_block && here->e_value_size) {
 431                                 size_t offs = le16_to_cpu(here->e_value_offs);
 432                                 if (offs < min_offs)
 433                                         min_offs = offs;
 434                         }
 435                         not_found = name_index - here->e_name_index;
 436                         if (!not_found)
 437                                 not_found = name_len - here->e_name_len;
 438                         if (!not_found)
 439                                 not_found = memcmp(name, here->e_name,name_len);
 440                         if (not_found <= 0)
 441                                 break;
 442                         here = next;
 443                 }
 444                 last = here;
 445                 /* We still need to compute min_offs and last. */
 446                 while (!IS_LAST_ENTRY(last)) {
 447                         struct ext2_xattr_entry *next = EXT2_XATTR_NEXT(last);
 448                         if ((char *)next >= end)
 449                                 goto bad_block;
 450                         if (!last->e_value_block && last->e_value_size) {
 451                                 size_t offs = le16_to_cpu(last->e_value_offs);
 452                                 if (offs < min_offs)
 453                                         min_offs = offs;
 454                         }
 455                         last = next;
 456                 }
 457
 458                 /* Check whether we have enough space left. */
 459                 free = min_offs - ((char*)last - (char*)header) - sizeof(__u32);
 460         } else {
 461                 /* We will use a new extended attribute block. */
 462                 free = sb->s_blocksize -
 463                         sizeof(struct ext2_xattr_header) - sizeof(__u32);
 464                 here = last = NULL;  /* avoid gcc uninitialized warning. */
 465         }
 466
 467         if (not_found) {
 468                 /* Request to remove a nonexistent attribute? */
 469                 error = -ENODATA;
 470                 if (flags & XATTR_REPLACE)
 471                         goto cleanup;
 472                 error = 0;
 473                 if (value == NULL)
 474                         goto cleanup;
 475         } else {
 476                 /* Request to create an existing attribute? */
 477                 error = -EEXIST;
 478                 if (flags & XATTR_CREATE)
 479                         goto cleanup;
 480                 if (!here->e_value_block && here->e_value_size) {
 481                         size_t size = le32_to_cpu(here->e_value_size);
 482
 483                         if (le16_to_cpu(here->e_value_offs) + size >
 484                             sb->s_blocksize || size > sb->s_blocksize)
 485                                 goto bad_block;
 486                         free += EXT2_XATTR_SIZE(size);
 487                 }
 488                 free += EXT2_XATTR_LEN(name_len);
 489         }
 490         error = -ENOSPC;
 491         if (free < EXT2_XATTR_LEN(name_len) + EXT2_XATTR_SIZE(value_len))
 492                 goto cleanup;
 493
 494         /* Here we know that we can set the new attribute. */
 495
 496         if (header) {
 497                 struct mb_cache_entry *ce;
 498
 499                 /* assert(header == HDR(bh)); */
 500                 ce = mb_cache_entry_get(ext2_xattr_cache, bh->b_bdev,
 501                                         bh->b_blocknr);
 502                 lock_buffer(bh);
 503                 if (header->h_refcount == cpu_to_le32(1)) {
 504                         ea_bdebug(bh, "modifying in-place");
 505                         if (ce)
 506                                 mb_cache_entry_free(ce);
 507                         /* keep the buffer locked while modifying it. */
 508                 } else {
 509                         int offset;
 510
 511                         if (ce)
 512                                 mb_cache_entry_release(ce);
 513                         unlock_buffer(bh);
 514                         ea_bdebug(bh, "cloning");
 515                         header = kmalloc(bh->b_size, GFP_KERNEL);
 516                         error = -ENOMEM;
 517                         if (header == NULL)
 518                                 goto cleanup;
 519                         memcpy(header, HDR(bh), bh->b_size);
 520                         header->h_refcount = cpu_to_le32(1);
 521
 522                         offset = (char *)here - bh->b_data;
 523                         here = ENTRY((char *)header + offset);
 524                         offset = (char *)last - bh->b_data;
 525                         last = ENTRY((char *)header + offset);
 526                 }
 527         } else {
 528                 /* Allocate a buffer where we construct the new block. */
 529                 header = kmalloc(sb->s_blocksize, GFP_KERNEL);
 530                 error = -ENOMEM;
 531                 if (header == NULL)
 532                         goto cleanup;
 533                 memset(header, 0, sb->s_blocksize);
 534                 end = (char *)header + sb->s_blocksize;
 535                 header->h_magic = cpu_to_le32(EXT2_XATTR_MAGIC);
 536                 header->h_blocks = header->h_refcount = cpu_to_le32(1);
 537                 last = here = ENTRY(header+1);
 538         }
 539
 540         /* Iff we are modifying the block in-place, bh is locked here. */
 541
 542         if (not_found) {
 543                 /* Insert the new name. */
 544                 size_t size = EXT2_XATTR_LEN(name_len);
 545                 size_t rest = (char *)last - (char *)here;
 546                 memmove((char *)here + size, here, rest);
 547                 memset(here, 0, size);
 548                 here->e_name_index = name_index;
 549                 here->e_name_len = name_len;
 550                 memcpy(here->e_name, name, name_len);
 551         } else {
 552                 if (!here->e_value_block && here->e_value_size) {
 553                         char *first_val = (char *)header + min_offs;
 554                         size_t offs = le16_to_cpu(here->e_value_offs);
 555                         char *val = (char *)header + offs;
 556                         size_t size = EXT2_XATTR_SIZE(
 557                                 le32_to_cpu(here->e_value_size));
 558
 559                         if (size == EXT2_XATTR_SIZE(value_len)) {
 560                                 /* The old and the new value have the same
 561                                    size. Just replace. */
 562                                 here->e_value_size = cpu_to_le32(value_len);
 563                                 memset(val + size - EXT2_XATTR_PAD, 0,
 564                                        EXT2_XATTR_PAD); /* Clear pad bytes. */
 565                                 memcpy(val, value, value_len);
 566                                 goto skip_replace;
 567                         }
 568
 569                         /* Remove the old value. */
 570                         memmove(first_val + size, first_val, val - first_val);
 571                         memset(first_val, 0, size);
 572                         here->e_value_offs = 0;
 573                         min_offs += size;
 574
 575                         /* Adjust all value offsets. */
 576                         last = ENTRY(header+1);
 577                         while (!IS_LAST_ENTRY(last)) {
 578                                 size_t o = le16_to_cpu(last->e_value_offs);
 579                                 if (!last->e_value_block && o < offs)
 580                                         last->e_value_offs =
 581                                                 cpu_to_le16(o + size);
 582                                 last = EXT2_XATTR_NEXT(last);
 583                         }
 584                 }
 585                 if (value == NULL) {
 586                         /* Remove the old name. */
 587                         size_t size = EXT2_XATTR_LEN(name_len);
 588                         last = ENTRY((char *)last - size);
 589                         memmove(here, (char*)here + size,
 590                                 (char*)last - (char*)here);
 591                         memset(last, 0, size);
 592                 }
 593         }
 594
 595         if (value != NULL) {
 596                 /* Insert the new value. */
 597                 here->e_value_size = cpu_to_le32(value_len);
 598                 if (value_len) {
 599                         size_t size = EXT2_XATTR_SIZE(value_len);
 600                         char *val = (char *)header + min_offs - size;
 601                         here->e_value_offs =
 602                                 cpu_to_le16((char *)val - (char *)header);
 603                         memset(val + size - EXT2_XATTR_PAD, 0,
 604                                EXT2_XATTR_PAD); /* Clear the pad bytes. */
 605                         memcpy(val, value, value_len);
 606                 }
 607         }
 608
 609 skip_replace:
 610         if (IS_LAST_ENTRY(ENTRY(header+1))) {
 611                 /* This block is now empty. */
 612                 if (bh && header == HDR(bh))
 613                         unlock_buffer(bh);  /* we were modifying in-place. */
 614                 error = ext2_xattr_set2(inode, bh, NULL);
 615         } else {
 616                 ext2_xattr_rehash(header, here);
 617                 if (bh && header == HDR(bh))
 618                         unlock_buffer(bh);  /* we were modifying in-place. */
 619                 error = ext2_xattr_set2(inode, bh, header);
 620         }
 621
 622 cleanup:
 623         brelse(bh);
 624         if (!(bh && header == HDR(bh)))
 625                 kfree(header);
 626         up_write(&EXT2_I(inode)->xattr_sem);
 627
 628         return error;
 629 }
 630
 631 /*
 632  * Second half of ext2_xattr_set(): Update the file system.
 633  */
 634 static int
 635 ext2_xattr_set2(struct inode *inode, struct buffer_head *old_bh,
 636                 struct ext2_xattr_header *header)
 637 {
 638         struct super_block *sb = inode->i_sb;
 639         struct buffer_head *new_bh = NULL;
 640         int error;
 641
 642         if (header) {
 643                 new_bh = ext2_xattr_cache_find(inode, header);
 644                 if (new_bh) {
 645                         /* We found an identical block in the cache. */
 646                         if (new_bh == old_bh) {
 647                                 ea_bdebug(new_bh, "keeping this block");
 648                         } else {
 649                                 /* The old block is released after updating
 650                                    the inode.  */
 651                                 ea_bdebug(new_bh, "reusing block");
 652
 653                                 error = -ENOSPC;
 654                                 if (DLIMIT_ALLOC_BLOCK(sb, inode->i_xid, 1))
 655                                         goto cleanup;
 656                                 error = -EDQUOT;
 657                                 if (DQUOT_ALLOC_BLOCK(inode, 1)) {
 658                                         DLIMIT_FREE_BLOCK(sb, inode->i_xid, 1);
 659                                         unlock_buffer(new_bh);
 660                                         goto cleanup;
 661                                 }
 662                                 HDR(new_bh)->h_refcount = cpu_to_le32(1 +
 663                                         le32_to_cpu(HDR(new_bh)->h_refcount));
 664                                 ea_bdebug(new_bh, "refcount now=%d",
 665                                         le32_to_cpu(HDR(new_bh)->h_refcount));
 666                         }
 667                         unlock_buffer(new_bh);
 668                 } else if (old_bh && header == HDR(old_bh)) {
 669                         /* Keep this block. No need to lock the block as we
 670                            don't need to change the reference count. */
 671                         new_bh = old_bh;
 672                         get_bh(new_bh);
 673                         ext2_xattr_cache_insert(new_bh);
 674                 } else {
 675                         /* We need to allocate a new block */
 676                         int goal = le32_to_cpu(EXT2_SB(sb)->s_es->
 677                                                            s_first_data_block) +
 678                                    EXT2_I(inode)->i_block_group *
 679                                    EXT2_BLOCKS_PER_GROUP(sb);
 680                         int block = ext2_new_block(inode, goal,
 681                                                    NULL, NULL, &error);
 682                         if (error)
 683                                 goto cleanup;
 684                         ea_idebug(inode, "creating block %d", block);
 685
 686                         new_bh = sb_getblk(sb, block);
 687                         if (!new_bh) {
 688                                 ext2_free_blocks(inode, block, 1);
 689                                 error = -EIO;
 690                                 goto cleanup;
 691                         }
 692                         lock_buffer(new_bh);
 693                         memcpy(new_bh->b_data, header, new_bh->b_size);
 694                         set_buffer_uptodate(new_bh);
 695                         unlock_buffer(new_bh);
 696                         ext2_xattr_cache_insert(new_bh);
 697
 698                         ext2_xattr_update_super_block(sb);
 699                 }
 700                 mark_buffer_dirty(new_bh);
 701                 if (IS_SYNC(inode)) {
 702                         sync_dirty_buffer(new_bh);
 703                         error = -EIO;
 704                         if (buffer_req(new_bh) && !buffer_uptodate(new_bh))
 705                                 goto cleanup;
 706                 }
 707         }
 708
 709         /* Update the inode. */
 710         EXT2_I(inode)->i_file_acl = new_bh ? new_bh->b_blocknr : 0;
 711         inode->i_ctime = CURRENT_TIME_SEC;
 712         if (IS_SYNC(inode)) {
 713                 error = ext2_sync_inode (inode);
 714                 /* In case sync failed due to ENOSPC the inode was actually
 715                  * written (only some dirty data were not) so we just proceed
 716                  * as if nothing happened and cleanup the unused block */
 717                 if (error && error != -ENOSPC) {
 718                         if (new_bh && new_bh != old_bh)
 719                                 DQUOT_FREE_BLOCK(inode, 1);
 720                         goto cleanup;
 721                 }
 722         } else
 723                 mark_inode_dirty(inode);
 724
 725         error = 0;
 726         if (old_bh && old_bh != new_bh) {
 727                 struct mb_cache_entry *ce;
 728
 729                 /*
 730                  * If there was an old block and we are no longer using it,
 731                  * release the old block.
 732                  */
 733                 ce = mb_cache_entry_get(ext2_xattr_cache, old_bh->b_bdev,
 734                                         old_bh->b_blocknr);
 735                 lock_buffer(old_bh);
 736                 if (HDR(old_bh)->h_refcount == cpu_to_le32(1)) {
 737                         /* Free the old block. */
 738                         if (ce)
 739                                 mb_cache_entry_free(ce);
 740                         ea_bdebug(old_bh, "freeing");
 741                         ext2_free_blocks(inode, old_bh->b_blocknr, 1);
 742                         /* We let our caller release old_bh, so we
 743                          * need to duplicate the buffer before. */
 744                         get_bh(old_bh);
 745                         bforget(old_bh);
 746                 } else {
 747                         /* Decrement the refcount only. */
 748                         HDR(old_bh)->h_refcount = cpu_to_le32(
 749                                 le32_to_cpu(HDR(old_bh)->h_refcount) - 1);
 750                         if (ce)
 751                                 mb_cache_entry_release(ce);
 752                         DLIMIT_FREE_BLOCK(sb, inode->i_xid, 1);
 753                         DQUOT_FREE_BLOCK(inode, 1);
 754                         mark_buffer_dirty(old_bh);
 755                         ea_bdebug(old_bh, "refcount now=%d",
 756                                 le32_to_cpu(HDR(old_bh)->h_refcount));
 757                 }
 758                 unlock_buffer(old_bh);
 759         }
 760
 761 cleanup:
 762         brelse(new_bh);
 763
 764         return error;
 765 }
 766
 767 /*
 768  * ext2_xattr_delete_inode()
 769  *
 770  * Free extended attribute resources associated with this inode. This
 771  * is called immediately before an inode is freed.
 772  */
 773 void
 774 ext2_xattr_delete_inode(struct inode *inode)
 775 {
 776         struct buffer_head *bh = NULL;
 777         struct mb_cache_entry *ce;
 778
 779         down_write(&EXT2_I(inode)->xattr_sem);
 780         if (!EXT2_I(inode)->i_file_acl)
 781                 goto cleanup;
 782         bh = sb_bread(inode->i_sb, EXT2_I(inode)->i_file_acl);
 783         if (!bh) {
 784                 ext2_error(inode->i_sb, "ext2_xattr_delete_inode",
 785                         "inode %ld: block %d read error", inode->i_ino,
 786                         EXT2_I(inode)->i_file_acl);
 787                 goto cleanup;
 788         }
 789         ea_bdebug(bh, "b_count=%d", atomic_read(&(bh->b_count)));
 790         if (HDR(bh)->h_magic != cpu_to_le32(EXT2_XATTR_MAGIC) ||
 791             HDR(bh)->h_blocks != cpu_to_le32(1)) {
 792                 ext2_error(inode->i_sb, "ext2_xattr_delete_inode",
 793                         "inode %ld: bad block %d", inode->i_ino,
 794                         EXT2_I(inode)->i_file_acl);
 795                 goto cleanup;
 796         }
 797         ce = mb_cache_entry_get(ext2_xattr_cache, bh->b_bdev, bh->b_blocknr);
 798         lock_buffer(bh);
 799         if (HDR(bh)->h_refcount == cpu_to_le32(1)) {
 800                 if (ce)
 801                         mb_cache_entry_free(ce);
 802                 ext2_free_blocks(inode, EXT2_I(inode)->i_file_acl, 1);
 803                 get_bh(bh);
 804                 bforget(bh);
 805         } else {
 806                 HDR(bh)->h_refcount = cpu_to_le32(
 807                         le32_to_cpu(HDR(bh)->h_refcount) - 1);
 808                 if (ce)
 809                         mb_cache_entry_release(ce);
 810                 mark_buffer_dirty(bh);
 811                 if (IS_SYNC(inode))
 812                         sync_dirty_buffer(bh);
 813                 DLIMIT_FREE_BLOCK(inode->i_sb, inode->i_xid, 1);
 814                 DQUOT_FREE_BLOCK(inode, 1);
 815         }
 816         ea_bdebug(bh, "refcount now=%d", le32_to_cpu(HDR(bh)->h_refcount) - 1);
 817         unlock_buffer(bh);
 818         EXT2_I(inode)->i_file_acl = 0;
 819
 820 cleanup:
 821         brelse(bh);
 822         up_write(&EXT2_I(inode)->xattr_sem);
 823 }
 824
 825 /*
 826  * ext2_xattr_put_super()
 827  *
 828  * This is called when a file system is unmounted.
 829  */
 830 void
 831 ext2_xattr_put_super(struct super_block *sb)
 832 {
 833         mb_cache_shrink(ext2_xattr_cache, sb->s_bdev);
 834 }
 835
 836
 837 /*
 838  * ext2_xattr_cache_insert()
 839  *
 840  * Create a new entry in the extended attribute cache, and insert
 841  * it unless such an entry is already in the cache.
 842  *
 843  * Returns 0, or a negative error number on failure.
 844  */
 845 static int
 846 ext2_xattr_cache_insert(struct buffer_head *bh)
 847 {
 848         __u32 hash = le32_to_cpu(HDR(bh)->h_hash);
 849         struct mb_cache_entry *ce;
 850         int error;
 851
 852         ce = mb_cache_entry_alloc(ext2_xattr_cache);
 853         if (!ce)
 854                 return -ENOMEM;
 855         error = mb_cache_entry_insert(ce, bh->b_bdev, bh->b_blocknr, &hash);
 856         if (error) {
 857                 mb_cache_entry_free(ce);
 858                 if (error == -EBUSY) {
 859                         ea_bdebug(bh, "already in cache (%d cache entries)",
 860                                 atomic_read(&ext2_xattr_cache->c_entry_count));
 861                         error = 0;
 862                 }
 863         } else {
 864                 ea_bdebug(bh, "inserting [%x] (%d cache entries)", (int)hash,
 865                           atomic_read(&ext2_xattr_cache->c_entry_count));
 866                 mb_cache_entry_release(ce);
 867         }
 868         return error;
 869 }
 870
 871 /*
 872  * ext2_xattr_cmp()
 873  *
 874  * Compare two extended attribute blocks for equality.
 875  *
 876  * Returns 0 if the blocks are equal, 1 if they differ, and
 877  * a negative error number on errors.
 878  */
 879 static int
 880 ext2_xattr_cmp(struct ext2_xattr_header *header1,
 881                struct ext2_xattr_header *header2)
 882 {
 883         struct ext2_xattr_entry *entry1, *entry2;
 884
 885         entry1 = ENTRY(header1+1);
 886         entry2 = ENTRY(header2+1);
 887         while (!IS_LAST_ENTRY(entry1)) {
 888                 if (IS_LAST_ENTRY(entry2))
 889                         return 1;
 890                 if (entry1->e_hash != entry2->e_hash ||
 891                     entry1->e_name_index != entry2->e_name_index ||
 892                     entry1->e_name_len != entry2->e_name_len ||
 893                     entry1->e_value_size != entry2->e_value_size ||
 894                     memcmp(entry1->e_name, entry2->e_name, entry1->e_name_len))
 895                         return 1;
 896                 if (entry1->e_value_block != 0 || entry2->e_value_block != 0)
 897                         return -EIO;
 898                 if (memcmp((char *)header1 + le16_to_cpu(entry1->e_value_offs),
 899                            (char *)header2 + le16_to_cpu(entry2->e_value_offs),
 900                            le32_to_cpu(entry1->e_value_size)))
 901                         return 1;
 902
 903                 entry1 = EXT2_XATTR_NEXT(entry1);
 904                 entry2 = EXT2_XATTR_NEXT(entry2);
 905         }
 906         if (!IS_LAST_ENTRY(entry2))
 907                 return 1;
 908         return 0;
 909 }
 910
 911 /*
 912  * ext2_xattr_cache_find()
 913  *
 914  * Find an identical extended attribute block.
 915  *
 916  * Returns a locked buffer head to the block found, or NULL if such
 917  * a block was not found or an error occurred.
 918  */
 919 static struct buffer_head *
 920 ext2_xattr_cache_find(struct inode *inode, struct ext2_xattr_header *header)
 921 {
 922         __u32 hash = le32_to_cpu(header->h_hash);
 923         struct mb_cache_entry *ce;
 924
 925         if (!header->h_hash)
 926                 return NULL;  /* never share */
 927         ea_idebug(inode, "looking for cached blocks [%x]", (int)hash);
 928 again:
 929         ce = mb_cache_entry_find_first(ext2_xattr_cache, 0,
 930                                        inode->i_sb->s_bdev, hash);
 931         while (ce) {
 932                 struct buffer_head *bh;
 933
 934                 if (IS_ERR(ce)) {
 935                         if (PTR_ERR(ce) == -EAGAIN)
 936                                 goto again;
 937                         break;
 938                 }
 939
 940                 bh = sb_bread(inode->i_sb, ce->e_block);
 941                 if (!bh) {
 942                         ext2_error(inode->i_sb, "ext2_xattr_cache_find",
 943                                 "inode %ld: block %ld read error",
 944                                 inode->i_ino, (unsigned long) ce->e_block);
 945                 } else {
 946                         lock_buffer(bh);
 947                         if (le32_to_cpu(HDR(bh)->h_refcount) >
 948                                    EXT2_XATTR_REFCOUNT_MAX) {
 949                                 ea_idebug(inode, "block %ld refcount %d>%d",
 950                                           (unsigned long) ce->e_block,
 951                                           le32_to_cpu(HDR(bh)->h_refcount),
 952                                           EXT2_XATTR_REFCOUNT_MAX);
 953                         } else if (!ext2_xattr_cmp(header, HDR(bh))) {
 954                                 ea_bdebug(bh, "b_count=%d",
 955                                           atomic_read(&(bh->b_count)));
 956                                 mb_cache_entry_release(ce);
 957                                 return bh;
 958                         }
 959                         unlock_buffer(bh);
 960                         brelse(bh);
 961                 }
 962                 ce = mb_cache_entry_find_next(ce, 0, inode->i_sb->s_bdev, hash);
 963         }
 964         return NULL;
 965 }
 966
 967 #define NAME_HASH_SHIFT 5
 968 #define VALUE_HASH_SHIFT 16
 969
 970 /*
 971  * ext2_xattr_hash_entry()
 972  *
 973  * Compute the hash of an extended attribute.
 974  */
 975 static inline void ext2_xattr_hash_entry(struct ext2_xattr_header *header,
 976                                          struct ext2_xattr_entry *entry)
 977 {
 978         __u32 hash = 0;
 979         char *name = entry->e_name;
 980         int n;
 981
 982         for (n=0; n < entry->e_name_len; n++) {
 983                 hash = (hash << NAME_HASH_SHIFT) ^
 984                        (hash >> (8*sizeof(hash) - NAME_HASH_SHIFT)) ^
 985                        *name++;
 986         }
 987
 988         if (entry->e_value_block == 0 && entry->e_value_size != 0) {
 989                 __le32 *value = (__le32 *)((char *)header +
 990                         le16_to_cpu(entry->e_value_offs));
 991                 for (n = (le32_to_cpu(entry->e_value_size) +
 992                      EXT2_XATTR_ROUND) >> EXT2_XATTR_PAD_BITS; n; n--) {
 993                         hash = (hash << VALUE_HASH_SHIFT) ^
 994                                (hash >> (8*sizeof(hash) - VALUE_HASH_SHIFT)) ^
 995                                le32_to_cpu(*value++);
 996                 }
 997         }
 998         entry->e_hash = cpu_to_le32(hash);
 999 }
1000
1001 #undef NAME_HASH_SHIFT
1002 #undef VALUE_HASH_SHIFT
1003
1004 #define BLOCK_HASH_SHIFT 16
1005
1006 /*
1007  * ext2_xattr_rehash()
1008  *
1009  * Re-compute the extended attribute hash value after an entry has changed.
1010  */
1011 static void ext2_xattr_rehash(struct ext2_xattr_header *header,
1012                               struct ext2_xattr_entry *entry)
1013 {
1014         struct ext2_xattr_entry *here;
1015         __u32 hash = 0;
1016
1017         ext2_xattr_hash_entry(header, entry);
1018         here = ENTRY(header+1);
1019         while (!IS_LAST_ENTRY(here)) {
1020                 if (!here->e_hash) {
1021                         /* Block is not shared if an entry's hash value == 0 */
1022                         hash = 0;
1023                         break;
1024                 }
1025                 hash = (hash << BLOCK_HASH_SHIFT) ^
1026                        (hash >> (8*sizeof(hash) - BLOCK_HASH_SHIFT)) ^
1027                        le32_to_cpu(here->e_hash);
1028                 here = EXT2_XATTR_NEXT(here);
1029         }
1030         header->h_hash = cpu_to_le32(hash);
1031 }
1032
1033 #undef BLOCK_HASH_SHIFT
1034
1035 int __init
1036 init_ext2_xattr(void)
1037 {
1038         ext2_xattr_cache = mb_cache_create("ext2_xattr", NULL,
1039                 sizeof(struct mb_cache_entry) +
1040                 sizeof(((struct mb_cache_entry *) 0)->e_indexes[0]), 1, 6);
1041         if (!ext2_xattr_cache)
1042                 return -ENOMEM;
1043         return 0;
1044 }
1045
1046 void
1047 exit_ext2_xattr(void)
1048 {
1049         mb_cache_destroy(ext2_xattr_cache);
1050 }