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/printk.h>
  59 #include <linux/slab.h>
  60 #include <linux/mbcache.h>
  61 #include <linux/quotaops.h>
  62 #include <linux/rwsem.h>
  63 #include <linux/security.h>
  64 #include "ext2.h"
  65 #include "xattr.h"
  66 #include "acl.h"
  67
  68 #define HDR(bh) ((struct ext2_xattr_header *)((bh)->b_data))
  69 #define ENTRY(ptr) ((struct ext2_xattr_entry *)(ptr))
  70 #define FIRST_ENTRY(bh) ENTRY(HDR(bh)+1)
  71 #define IS_LAST_ENTRY(entry) (*(__u32 *)(entry) == 0)
  72
  73 #ifdef EXT2_XATTR_DEBUG
  74 # define ea_idebug(inode, f...) do { \
  75                 printk(KERN_DEBUG "inode %s:%ld: ", \
  76                         inode->i_sb->s_id, inode->i_ino); \
  77                 printk(f); \
  78                 printk("\n"); \
  79         } while (0)
  80 # define ea_bdebug(bh, f...) do { \
  81                 printk(KERN_DEBUG "block %pg:%lu: ", \
  82                         bh->b_bdev, (unsigned long) bh->b_blocknr); \
  83                 printk(f); \
  84                 printk("\n"); \
  85         } while (0)
  86 #else
  87 # define ea_idebug(inode, f...) no_printk(f)
  88 # define ea_bdebug(bh, f...)    no_printk(f)
  89 #endif
  90
  91 static int ext2_xattr_set2(struct inode *, struct buffer_head *,
  92                            struct ext2_xattr_header *);
  93
  94 static int ext2_xattr_cache_insert(struct mb_cache *, struct buffer_head *);
  95 static struct buffer_head *ext2_xattr_cache_find(struct inode *,
  96                                                  struct ext2_xattr_header *);
  97 static void ext2_xattr_rehash(struct ext2_xattr_header *,
  98                               struct ext2_xattr_entry *);
  99
 100 static const struct xattr_handler *ext2_xattr_handler_map[] = {
 101         [EXT2_XATTR_INDEX_USER]              = &ext2_xattr_user_handler,
 102 #ifdef CONFIG_EXT2_FS_POSIX_ACL
 103         [EXT2_XATTR_INDEX_POSIX_ACL_ACCESS]  = &posix_acl_access_xattr_handler,
 104         [EXT2_XATTR_INDEX_POSIX_ACL_DEFAULT] = &posix_acl_default_xattr_handler,
 105 #endif
 106         [EXT2_XATTR_INDEX_TRUSTED]           = &ext2_xattr_trusted_handler,
 107 #ifdef CONFIG_EXT2_FS_SECURITY
 108         [EXT2_XATTR_INDEX_SECURITY]          = &ext2_xattr_security_handler,
 109 #endif
 110 };
 111
 112 const struct xattr_handler *ext2_xattr_handlers[] = {
 113         &ext2_xattr_user_handler,
 114         &ext2_xattr_trusted_handler,
 115 #ifdef CONFIG_EXT2_FS_POSIX_ACL
 116         &posix_acl_access_xattr_handler,
 117         &posix_acl_default_xattr_handler,
 118 #endif
 119 #ifdef CONFIG_EXT2_FS_SECURITY
 120         &ext2_xattr_security_handler,
 121 #endif
 122         NULL
 123 };
 124
 125 static inline const struct xattr_handler *
 126 ext2_xattr_handler(int name_index)
 127 {
 128         const struct xattr_handler *handler = NULL;
 129
 130         if (name_index > 0 && name_index < ARRAY_SIZE(ext2_xattr_handler_map))
 131                 handler = ext2_xattr_handler_map[name_index];
 132         return handler;
 133 }
 134
 135 /*
 136  * ext2_xattr_get()
 137  *
 138  * Copy an extended attribute into the buffer
 139  * provided, or compute the buffer size required.
 140  * Buffer is NULL to compute the size of the buffer required.
 141  *
 142  * Returns a negative error number on failure, or the number of bytes
 143  * used / required on success.
 144  */
 145 int
 146 ext2_xattr_get(struct inode *inode, int name_index, const char *name,
 147                void *buffer, size_t buffer_size)
 148 {
 149         struct buffer_head *bh = NULL;
 150         struct ext2_xattr_entry *entry;
 151         size_t name_len, size;
 152         char *end;
 153         int error;
 154         struct mb_cache *ext2_mb_cache = EXT2_SB(inode->i_sb)->s_mb_cache;
 155
 156         ea_idebug(inode, "name=%d.%s, buffer=%p, buffer_size=%ld",
 157                   name_index, name, buffer, (long)buffer_size);
 158
 159         if (name == NULL)
 160                 return -EINVAL;
 161         name_len = strlen(name);
 162         if (name_len > 255)
 163                 return -ERANGE;
 164
 165         down_read(&EXT2_I(inode)->xattr_sem);
 166         error = -ENODATA;
 167         if (!EXT2_I(inode)->i_file_acl)
 168                 goto cleanup;
 169         ea_idebug(inode, "reading block %d", EXT2_I(inode)->i_file_acl);
 170         bh = sb_bread(inode->i_sb, EXT2_I(inode)->i_file_acl);
 171         error = -EIO;
 172         if (!bh)
 173                 goto cleanup;
 174         ea_bdebug(bh, "b_count=%d, refcount=%d",
 175                 atomic_read(&(bh->b_count)), le32_to_cpu(HDR(bh)->h_refcount));
 176         end = bh->b_data + bh->b_size;
 177         if (HDR(bh)->h_magic != cpu_to_le32(EXT2_XATTR_MAGIC) ||
 178             HDR(bh)->h_blocks != cpu_to_le32(1)) {
 179 bad_block:      ext2_error(inode->i_sb, "ext2_xattr_get",
 180                         "inode %ld: bad block %d", inode->i_ino,
 181                         EXT2_I(inode)->i_file_acl);
 182                 error = -EIO;
 183                 goto cleanup;
 184         }
 185
 186         /* find named attribute */
 187         entry = FIRST_ENTRY(bh);
 188         while (!IS_LAST_ENTRY(entry)) {
 189                 struct ext2_xattr_entry *next =
 190                         EXT2_XATTR_NEXT(entry);
 191                 if ((char *)next >= end)
 192                         goto bad_block;
 193                 if (name_index == entry->e_name_index &&
 194                     name_len == entry->e_name_len &&
 195                     memcmp(name, entry->e_name, name_len) == 0)
 196                         goto found;
 197                 entry = next;
 198         }
 199         if (ext2_xattr_cache_insert(ext2_mb_cache, bh))
 200                 ea_idebug(inode, "cache insert failed");
 201         error = -ENODATA;
 202         goto cleanup;
 203 found:
 204         /* check the buffer size */
 205         if (entry->e_value_block != 0)
 206                 goto bad_block;
 207         size = le32_to_cpu(entry->e_value_size);
 208         if (size > inode->i_sb->s_blocksize ||
 209             le16_to_cpu(entry->e_value_offs) + size > inode->i_sb->s_blocksize)
 210                 goto bad_block;
 211
 212         if (ext2_xattr_cache_insert(ext2_mb_cache, bh))
 213                 ea_idebug(inode, "cache insert failed");
 214         if (buffer) {
 215                 error = -ERANGE;
 216                 if (size > buffer_size)
 217                         goto cleanup;
 218                 /* return value of attribute */
 219                 memcpy(buffer, bh->b_data + le16_to_cpu(entry->e_value_offs),
 220                         size);
 221         }
 222         error = size;
 223
 224 cleanup:
 225         brelse(bh);
 226         up_read(&EXT2_I(inode)->xattr_sem);
 227
 228         return error;
 229 }
 230
 231 /*
 232  * ext2_xattr_list()
 233  *
 234  * Copy a list of attribute names into the buffer
 235  * provided, or compute the buffer size required.
 236  * Buffer is NULL to compute the size of the buffer required.
 237  *
 238  * Returns a negative error number on failure, or the number of bytes
 239  * used / required on success.
 240  */
 241 static int
 242 ext2_xattr_list(struct dentry *dentry, char *buffer, size_t buffer_size)
 243 {
 244         struct inode *inode = d_inode(dentry);
 245         struct buffer_head *bh = NULL;
 246         struct ext2_xattr_entry *entry;
 247         char *end;
 248         size_t rest = buffer_size;
 249         int error;
 250         struct mb_cache *ext2_mb_cache = EXT2_SB(inode->i_sb)->s_mb_cache;
 251
 252         ea_idebug(inode, "buffer=%p, buffer_size=%ld",
 253                   buffer, (long)buffer_size);
 254
 255         down_read(&EXT2_I(inode)->xattr_sem);
 256         error = 0;
 257         if (!EXT2_I(inode)->i_file_acl)
 258                 goto cleanup;
 259         ea_idebug(inode, "reading block %d", EXT2_I(inode)->i_file_acl);
 260         bh = sb_bread(inode->i_sb, EXT2_I(inode)->i_file_acl);
 261         error = -EIO;
 262         if (!bh)
 263                 goto cleanup;
 264         ea_bdebug(bh, "b_count=%d, refcount=%d",
 265                 atomic_read(&(bh->b_count)), le32_to_cpu(HDR(bh)->h_refcount));
 266         end = bh->b_data + bh->b_size;
 267         if (HDR(bh)->h_magic != cpu_to_le32(EXT2_XATTR_MAGIC) ||
 268             HDR(bh)->h_blocks != cpu_to_le32(1)) {
 269 bad_block:      ext2_error(inode->i_sb, "ext2_xattr_list",
 270                         "inode %ld: bad block %d", inode->i_ino,
 271                         EXT2_I(inode)->i_file_acl);
 272                 error = -EIO;
 273                 goto cleanup;
 274         }
 275
 276         /* check the on-disk data structure */
 277         entry = FIRST_ENTRY(bh);
 278         while (!IS_LAST_ENTRY(entry)) {
 279                 struct ext2_xattr_entry *next = EXT2_XATTR_NEXT(entry);
 280
 281                 if ((char *)next >= end)
 282                         goto bad_block;
 283                 entry = next;
 284         }
 285         if (ext2_xattr_cache_insert(ext2_mb_cache, bh))
 286                 ea_idebug(inode, "cache insert failed");
 287
 288         /* list the attribute names */
 289         for (entry = FIRST_ENTRY(bh); !IS_LAST_ENTRY(entry);
 290              entry = EXT2_XATTR_NEXT(entry)) {
 291                 const struct xattr_handler *handler =
 292                         ext2_xattr_handler(entry->e_name_index);
 293
 294                 if (handler && (!handler->list || handler->list(dentry))) {
 295                         const char *prefix = handler->prefix ?: handler->name;
 296                         size_t prefix_len = strlen(prefix);
 297                         size_t size = prefix_len + entry->e_name_len + 1;
 298
 299                         if (buffer) {
 300                                 if (size > rest) {
 301                                         error = -ERANGE;
 302                                         goto cleanup;
 303                                 }
 304                                 memcpy(buffer, prefix, prefix_len);
 305                                 buffer += prefix_len;
 306                                 memcpy(buffer, entry->e_name, entry->e_name_len);
 307                                 buffer += entry->e_name_len;
 308                                 *buffer++ = 0;
 309                         }
 310                         rest -= size;
 311                 }
 312         }
 313         error = buffer_size - rest;  /* total size */
 314
 315 cleanup:
 316         brelse(bh);
 317         up_read(&EXT2_I(inode)->xattr_sem);
 318
 319         return error;
 320 }
 321
 322 /*
 323  * Inode operation listxattr()
 324  *
 325  * d_inode(dentry)->i_mutex: don't care
 326  */
 327 ssize_t
 328 ext2_listxattr(struct dentry *dentry, char *buffer, size_t size)
 329 {
 330         return ext2_xattr_list(dentry, buffer, size);
 331 }
 332
 333 /*
 334  * If the EXT2_FEATURE_COMPAT_EXT_ATTR feature of this file system is
 335  * not set, set it.
 336  */
 337 static void ext2_xattr_update_super_block(struct super_block *sb)
 338 {
 339         if (EXT2_HAS_COMPAT_FEATURE(sb, EXT2_FEATURE_COMPAT_EXT_ATTR))
 340                 return;
 341
 342         spin_lock(&EXT2_SB(sb)->s_lock);
 343         EXT2_SET_COMPAT_FEATURE(sb, EXT2_FEATURE_COMPAT_EXT_ATTR);
 344         spin_unlock(&EXT2_SB(sb)->s_lock);
 345         mark_buffer_dirty(EXT2_SB(sb)->s_sbh);
 346 }
 347
 348 /*
 349  * ext2_xattr_set()
 350  *
 351  * Create, replace or remove an extended attribute for this inode.  Value
 352  * is NULL to remove an existing extended attribute, and non-NULL to
 353  * either replace an existing extended attribute, or create a new extended
 354  * attribute. The flags XATTR_REPLACE and XATTR_CREATE
 355  * specify that an extended attribute must exist and must not exist
 356  * previous to the call, respectively.
 357  *
 358  * Returns 0, or a negative error number on failure.
 359  */
 360 int
 361 ext2_xattr_set(struct inode *inode, int name_index, const char *name,
 362                const void *value, size_t value_len, int flags)
 363 {
 364         struct super_block *sb = inode->i_sb;
 365         struct buffer_head *bh = NULL;
 366         struct ext2_xattr_header *header = NULL;
 367         struct ext2_xattr_entry *here, *last;
 368         size_t name_len, free, min_offs = sb->s_blocksize;
 369         int not_found = 1, error;
 370         char *end;
 371
 372         /*
 373          * header -- Points either into bh, or to a temporarily
 374          *           allocated buffer.
 375          * here -- The named entry found, or the place for inserting, within
 376          *         the block pointed to by header.
 377          * last -- Points right after the last named entry within the block
 378          *         pointed to by header.
 379          * min_offs -- The offset of the first value (values are aligned
 380          *             towards the end of the block).
 381          * end -- Points right after the block pointed to by header.
 382          */
 383
 384         ea_idebug(inode, "name=%d.%s, value=%p, value_len=%ld",
 385                   name_index, name, value, (long)value_len);
 386
 387         if (value == NULL)
 388                 value_len = 0;
 389         if (name == NULL)
 390                 return -EINVAL;
 391         name_len = strlen(name);
 392         if (name_len > 255 || value_len > sb->s_blocksize)
 393                 return -ERANGE;
 394         down_write(&EXT2_I(inode)->xattr_sem);
 395         if (EXT2_I(inode)->i_file_acl) {
 396                 /* The inode already has an extended attribute block. */
 397                 bh = sb_bread(sb, EXT2_I(inode)->i_file_acl);
 398                 error = -EIO;
 399                 if (!bh)
 400                         goto cleanup;
 401                 ea_bdebug(bh, "b_count=%d, refcount=%d",
 402                         atomic_read(&(bh->b_count)),
 403                         le32_to_cpu(HDR(bh)->h_refcount));
 404                 header = HDR(bh);
 405                 end = bh->b_data + bh->b_size;
 406                 if (header->h_magic != cpu_to_le32(EXT2_XATTR_MAGIC) ||
 407                     header->h_blocks != cpu_to_le32(1)) {
 408 bad_block:              ext2_error(sb, "ext2_xattr_set",
 409                                 "inode %ld: bad block %d", inode->i_ino,
 410                                    EXT2_I(inode)->i_file_acl);
 411                         error = -EIO;
 412                         goto cleanup;
 413                 }
 414                 /* Find the named attribute. */
 415                 here = FIRST_ENTRY(bh);
 416                 while (!IS_LAST_ENTRY(here)) {
 417                         struct ext2_xattr_entry *next = EXT2_XATTR_NEXT(here);
 418                         if ((char *)next >= end)
 419                                 goto bad_block;
 420                         if (!here->e_value_block && here->e_value_size) {
 421                                 size_t offs = le16_to_cpu(here->e_value_offs);
 422                                 if (offs < min_offs)
 423                                         min_offs = offs;
 424                         }
 425                         not_found = name_index - here->e_name_index;
 426                         if (!not_found)
 427                                 not_found = name_len - here->e_name_len;
 428                         if (!not_found)
 429                                 not_found = memcmp(name, here->e_name,name_len);
 430                         if (not_found <= 0)
 431                                 break;
 432                         here = next;
 433                 }
 434                 last = here;
 435                 /* We still need to compute min_offs and last. */
 436                 while (!IS_LAST_ENTRY(last)) {
 437                         struct ext2_xattr_entry *next = EXT2_XATTR_NEXT(last);
 438                         if ((char *)next >= end)
 439                                 goto bad_block;
 440                         if (!last->e_value_block && last->e_value_size) {
 441                                 size_t offs = le16_to_cpu(last->e_value_offs);
 442                                 if (offs < min_offs)
 443                                         min_offs = offs;
 444                         }
 445                         last = next;
 446                 }
 447
 448                 /* Check whether we have enough space left. */
 449                 free = min_offs - ((char*)last - (char*)header) - sizeof(__u32);
 450         } else {
 451                 /* We will use a new extended attribute block. */
 452                 free = sb->s_blocksize -
 453                         sizeof(struct ext2_xattr_header) - sizeof(__u32);
 454                 here = last = NULL;  /* avoid gcc uninitialized warning. */
 455         }
 456
 457         if (not_found) {
 458                 /* Request to remove a nonexistent attribute? */
 459                 error = -ENODATA;
 460                 if (flags & XATTR_REPLACE)
 461                         goto cleanup;
 462                 error = 0;
 463                 if (value == NULL)
 464                         goto cleanup;
 465         } else {
 466                 /* Request to create an existing attribute? */
 467                 error = -EEXIST;
 468                 if (flags & XATTR_CREATE)
 469                         goto cleanup;
 470                 if (!here->e_value_block && here->e_value_size) {
 471                         size_t size = le32_to_cpu(here->e_value_size);
 472
 473                         if (le16_to_cpu(here->e_value_offs) + size >
 474                             sb->s_blocksize || size > sb->s_blocksize)
 475                                 goto bad_block;
 476                         free += EXT2_XATTR_SIZE(size);
 477                 }
 478                 free += EXT2_XATTR_LEN(name_len);
 479         }
 480         error = -ENOSPC;
 481         if (free < EXT2_XATTR_LEN(name_len) + EXT2_XATTR_SIZE(value_len))
 482                 goto cleanup;
 483
 484         /* Here we know that we can set the new attribute. */
 485
 486         if (header) {
 487                 /* assert(header == HDR(bh)); */
 488                 lock_buffer(bh);
 489                 if (header->h_refcount == cpu_to_le32(1)) {
 490                         __u32 hash = le32_to_cpu(header->h_hash);
 491
 492                         ea_bdebug(bh, "modifying in-place");
 493                         /*
 494                          * This must happen under buffer lock for
 495                          * ext2_xattr_set2() to reliably detect modified block
 496                          */
 497                         mb_cache_entry_delete_block(EXT2_SB(sb)->s_mb_cache,
 498                                                     hash, bh->b_blocknr);
 499
 500                         /* keep the buffer locked while modifying it. */
 501                 } else {
 502                         int offset;
 503
 504                         unlock_buffer(bh);
 505                         ea_bdebug(bh, "cloning");
 506                         header = kmalloc(bh->b_size, GFP_KERNEL);
 507                         error = -ENOMEM;
 508                         if (header == NULL)
 509                                 goto cleanup;
 510                         memcpy(header, HDR(bh), bh->b_size);
 511                         header->h_refcount = cpu_to_le32(1);
 512
 513                         offset = (char *)here - bh->b_data;
 514                         here = ENTRY((char *)header + offset);
 515                         offset = (char *)last - bh->b_data;
 516                         last = ENTRY((char *)header + offset);
 517                 }
 518         } else {
 519                 /* Allocate a buffer where we construct the new block. */
 520                 header = kzalloc(sb->s_blocksize, GFP_KERNEL);
 521                 error = -ENOMEM;
 522                 if (header == NULL)
 523                         goto cleanup;
 524                 end = (char *)header + sb->s_blocksize;
 525                 header->h_magic = cpu_to_le32(EXT2_XATTR_MAGIC);
 526                 header->h_blocks = header->h_refcount = cpu_to_le32(1);
 527                 last = here = ENTRY(header+1);
 528         }
 529
 530         /* Iff we are modifying the block in-place, bh is locked here. */
 531
 532         if (not_found) {
 533                 /* Insert the new name. */
 534                 size_t size = EXT2_XATTR_LEN(name_len);
 535                 size_t rest = (char *)last - (char *)here;
 536                 memmove((char *)here + size, here, rest);
 537                 memset(here, 0, size);
 538                 here->e_name_index = name_index;
 539                 here->e_name_len = name_len;
 540                 memcpy(here->e_name, name, name_len);
 541         } else {
 542                 if (!here->e_value_block && here->e_value_size) {
 543                         char *first_val = (char *)header + min_offs;
 544                         size_t offs = le16_to_cpu(here->e_value_offs);
 545                         char *val = (char *)header + offs;
 546                         size_t size = EXT2_XATTR_SIZE(
 547                                 le32_to_cpu(here->e_value_size));
 548
 549                         if (size == EXT2_XATTR_SIZE(value_len)) {
 550                                 /* The old and the new value have the same
 551                                    size. Just replace. */
 552                                 here->e_value_size = cpu_to_le32(value_len);
 553                                 memset(val + size - EXT2_XATTR_PAD, 0,
 554                                        EXT2_XATTR_PAD); /* Clear pad bytes. */
 555                                 memcpy(val, value, value_len);
 556                                 goto skip_replace;
 557                         }
 558
 559                         /* Remove the old value. */
 560                         memmove(first_val + size, first_val, val - first_val);
 561                         memset(first_val, 0, size);
 562                         here->e_value_offs = 0;
 563                         min_offs += size;
 564
 565                         /* Adjust all value offsets. */
 566                         last = ENTRY(header+1);
 567                         while (!IS_LAST_ENTRY(last)) {
 568                                 size_t o = le16_to_cpu(last->e_value_offs);
 569                                 if (!last->e_value_block && o < offs)
 570                                         last->e_value_offs =
 571                                                 cpu_to_le16(o + size);
 572                                 last = EXT2_XATTR_NEXT(last);
 573                         }
 574                 }
 575                 if (value == NULL) {
 576                         /* Remove the old name. */
 577                         size_t size = EXT2_XATTR_LEN(name_len);
 578                         last = ENTRY((char *)last - size);
 579                         memmove(here, (char*)here + size,
 580                                 (char*)last - (char*)here);
 581                         memset(last, 0, size);
 582                 }
 583         }
 584
 585         if (value != NULL) {
 586                 /* Insert the new value. */
 587                 here->e_value_size = cpu_to_le32(value_len);
 588                 if (value_len) {
 589                         size_t size = EXT2_XATTR_SIZE(value_len);
 590                         char *val = (char *)header + min_offs - size;
 591                         here->e_value_offs =
 592                                 cpu_to_le16((char *)val - (char *)header);
 593                         memset(val + size - EXT2_XATTR_PAD, 0,
 594                                EXT2_XATTR_PAD); /* Clear the pad bytes. */
 595                         memcpy(val, value, value_len);
 596                 }
 597         }
 598
 599 skip_replace:
 600         if (IS_LAST_ENTRY(ENTRY(header+1))) {
 601                 /* This block is now empty. */
 602                 if (bh && header == HDR(bh))
 603                         unlock_buffer(bh);  /* we were modifying in-place. */
 604                 error = ext2_xattr_set2(inode, bh, NULL);
 605         } else {
 606                 ext2_xattr_rehash(header, here);
 607                 if (bh && header == HDR(bh))
 608                         unlock_buffer(bh);  /* we were modifying in-place. */
 609                 error = ext2_xattr_set2(inode, bh, header);
 610         }
 611
 612 cleanup:
 613         if (!(bh && header == HDR(bh)))
 614                 kfree(header);
 615         brelse(bh);
 616         up_write(&EXT2_I(inode)->xattr_sem);
 617
 618         return error;
 619 }
 620
 621 /*
 622  * Second half of ext2_xattr_set(): Update the file system.
 623  */
 624 static int
 625 ext2_xattr_set2(struct inode *inode, struct buffer_head *old_bh,
 626                 struct ext2_xattr_header *header)
 627 {
 628         struct super_block *sb = inode->i_sb;
 629         struct buffer_head *new_bh = NULL;
 630         int error;
 631         struct mb_cache *ext2_mb_cache = EXT2_SB(sb)->s_mb_cache;
 632
 633         if (header) {
 634                 new_bh = ext2_xattr_cache_find(inode, header);
 635                 if (new_bh) {
 636                         /* We found an identical block in the cache. */
 637                         if (new_bh == old_bh) {
 638                                 ea_bdebug(new_bh, "keeping this block");
 639                         } else {
 640                                 /* The old block is released after updating
 641                                    the inode.  */
 642                                 ea_bdebug(new_bh, "reusing block");
 643
 644                                 error = dquot_alloc_block(inode, 1);
 645                                 if (error) {
 646                                         unlock_buffer(new_bh);
 647                                         goto cleanup;
 648                                 }
 649                                 le32_add_cpu(&HDR(new_bh)->h_refcount, 1);
 650                                 ea_bdebug(new_bh, "refcount now=%d",
 651                                         le32_to_cpu(HDR(new_bh)->h_refcount));
 652                         }
 653                         unlock_buffer(new_bh);
 654                 } else if (old_bh && header == HDR(old_bh)) {
 655                         /* Keep this block. No need to lock the block as we
 656                            don't need to change the reference count. */
 657                         new_bh = old_bh;
 658                         get_bh(new_bh);
 659                         ext2_xattr_cache_insert(ext2_mb_cache, new_bh);
 660                 } else {
 661                         /* We need to allocate a new block */
 662                         ext2_fsblk_t goal = ext2_group_first_block_no(sb,
 663                                                 EXT2_I(inode)->i_block_group);
 664                         int block = ext2_new_block(inode, goal, &error);
 665                         if (error)
 666                                 goto cleanup;
 667                         ea_idebug(inode, "creating block %d", block);
 668
 669                         new_bh = sb_getblk(sb, block);
 670                         if (unlikely(!new_bh)) {
 671                                 ext2_free_blocks(inode, block, 1);
 672                                 mark_inode_dirty(inode);
 673                                 error = -ENOMEM;
 674                                 goto cleanup;
 675                         }
 676                         lock_buffer(new_bh);
 677                         memcpy(new_bh->b_data, header, new_bh->b_size);
 678                         set_buffer_uptodate(new_bh);
 679                         unlock_buffer(new_bh);
 680                         ext2_xattr_cache_insert(ext2_mb_cache, new_bh);
 681
 682                         ext2_xattr_update_super_block(sb);
 683                 }
 684                 mark_buffer_dirty(new_bh);
 685                 if (IS_SYNC(inode)) {
 686                         sync_dirty_buffer(new_bh);
 687                         error = -EIO;
 688                         if (buffer_req(new_bh) && !buffer_uptodate(new_bh))
 689                                 goto cleanup;
 690                 }
 691         }
 692
 693         /* Update the inode. */
 694         EXT2_I(inode)->i_file_acl = new_bh ? new_bh->b_blocknr : 0;
 695         inode->i_ctime = current_time(inode);
 696         if (IS_SYNC(inode)) {
 697                 error = sync_inode_metadata(inode, 1);
 698                 /* In case sync failed due to ENOSPC the inode was actually
 699                  * written (only some dirty data were not) so we just proceed
 700                  * as if nothing happened and cleanup the unused block */
 701                 if (error && error != -ENOSPC) {
 702                         if (new_bh && new_bh != old_bh) {
 703                                 dquot_free_block_nodirty(inode, 1);
 704                                 mark_inode_dirty(inode);
 705                         }
 706                         goto cleanup;
 707                 }
 708         } else
 709                 mark_inode_dirty(inode);
 710
 711         error = 0;
 712         if (old_bh && old_bh != new_bh) {
 713                 /*
 714                  * If there was an old block and we are no longer using it,
 715                  * release the old block.
 716                  */
 717                 lock_buffer(old_bh);
 718                 if (HDR(old_bh)->h_refcount == cpu_to_le32(1)) {
 719                         __u32 hash = le32_to_cpu(HDR(old_bh)->h_hash);
 720
 721                         /*
 722                          * This must happen under buffer lock for
 723                          * ext2_xattr_set2() to reliably detect freed block
 724                          */
 725                         mb_cache_entry_delete_block(ext2_mb_cache,
 726                                                     hash, old_bh->b_blocknr);
 727                         /* Free the old block. */
 728                         ea_bdebug(old_bh, "freeing");
 729                         ext2_free_blocks(inode, old_bh->b_blocknr, 1);
 730                         mark_inode_dirty(inode);
 731                         /* We let our caller release old_bh, so we
 732                          * need to duplicate the buffer before. */
 733                         get_bh(old_bh);
 734                         bforget(old_bh);
 735                 } else {
 736                         /* Decrement the refcount only. */
 737                         le32_add_cpu(&HDR(old_bh)->h_refcount, -1);
 738                         dquot_free_block_nodirty(inode, 1);
 739                         mark_inode_dirty(inode);
 740                         mark_buffer_dirty(old_bh);
 741                         ea_bdebug(old_bh, "refcount now=%d",
 742                                 le32_to_cpu(HDR(old_bh)->h_refcount));
 743                 }
 744                 unlock_buffer(old_bh);
 745         }
 746
 747 cleanup:
 748         brelse(new_bh);
 749
 750         return error;
 751 }
 752
 753 /*
 754  * ext2_xattr_delete_inode()
 755  *
 756  * Free extended attribute resources associated with this inode. This
 757  * is called immediately before an inode is freed.
 758  */
 759 void
 760 ext2_xattr_delete_inode(struct inode *inode)
 761 {
 762         struct buffer_head *bh = NULL;
 763         struct ext2_sb_info *sbi = EXT2_SB(inode->i_sb);
 764
 765         down_write(&EXT2_I(inode)->xattr_sem);
 766         if (!EXT2_I(inode)->i_file_acl)
 767                 goto cleanup;
 768
 769         if (!ext2_data_block_valid(sbi, EXT2_I(inode)->i_file_acl, 0)) {
 770                 ext2_error(inode->i_sb, "ext2_xattr_delete_inode",
 771                         "inode %ld: xattr block %d is out of data blocks range",
 772                         inode->i_ino, EXT2_I(inode)->i_file_acl);
 773                 goto cleanup;
 774         }
 775
 776         bh = sb_bread(inode->i_sb, EXT2_I(inode)->i_file_acl);
 777         if (!bh) {
 778                 ext2_error(inode->i_sb, "ext2_xattr_delete_inode",
 779                         "inode %ld: block %d read error", inode->i_ino,
 780                         EXT2_I(inode)->i_file_acl);
 781                 goto cleanup;
 782         }
 783         ea_bdebug(bh, "b_count=%d", atomic_read(&(bh->b_count)));
 784         if (HDR(bh)->h_magic != cpu_to_le32(EXT2_XATTR_MAGIC) ||
 785             HDR(bh)->h_blocks != cpu_to_le32(1)) {
 786                 ext2_error(inode->i_sb, "ext2_xattr_delete_inode",
 787                         "inode %ld: bad block %d", inode->i_ino,
 788                         EXT2_I(inode)->i_file_acl);
 789                 goto cleanup;
 790         }
 791         lock_buffer(bh);
 792         if (HDR(bh)->h_refcount == cpu_to_le32(1)) {
 793                 __u32 hash = le32_to_cpu(HDR(bh)->h_hash);
 794
 795                 /*
 796                  * This must happen under buffer lock for ext2_xattr_set2() to
 797                  * reliably detect freed block
 798                  */
 799                 mb_cache_entry_delete_block(EXT2_SB(inode->i_sb)->s_mb_cache,
 800                                             hash, bh->b_blocknr);
 801                 ext2_free_blocks(inode, EXT2_I(inode)->i_file_acl, 1);
 802                 get_bh(bh);
 803                 bforget(bh);
 804                 unlock_buffer(bh);
 805         } else {
 806                 le32_add_cpu(&HDR(bh)->h_refcount, -1);
 807                 ea_bdebug(bh, "refcount now=%d",
 808                         le32_to_cpu(HDR(bh)->h_refcount));
 809                 unlock_buffer(bh);
 810                 mark_buffer_dirty(bh);
 811                 if (IS_SYNC(inode))
 812                         sync_dirty_buffer(bh);
 813                 dquot_free_block_nodirty(inode, 1);
 814         }
 815         EXT2_I(inode)->i_file_acl = 0;
 816
 817 cleanup:
 818         brelse(bh);
 819         up_write(&EXT2_I(inode)->xattr_sem);
 820 }
 821
 822 /*
 823  * ext2_xattr_cache_insert()
 824  *
 825  * Create a new entry in the extended attribute cache, and insert
 826  * it unless such an entry is already in the cache.
 827  *
 828  * Returns 0, or a negative error number on failure.
 829  */
 830 static int
 831 ext2_xattr_cache_insert(struct mb_cache *cache, struct buffer_head *bh)
 832 {
 833         __u32 hash = le32_to_cpu(HDR(bh)->h_hash);
 834         int error;
 835
 836         error = mb_cache_entry_create(cache, GFP_NOFS, hash, bh->b_blocknr, 1);
 837         if (error) {
 838                 if (error == -EBUSY) {
 839                         ea_bdebug(bh, "already in cache");
 840                         error = 0;
 841                 }
 842         } else
 843                 ea_bdebug(bh, "inserting [%x]", (int)hash);
 844         return error;
 845 }
 846
 847 /*
 848  * ext2_xattr_cmp()
 849  *
 850  * Compare two extended attribute blocks for equality.
 851  *
 852  * Returns 0 if the blocks are equal, 1 if they differ, and
 853  * a negative error number on errors.
 854  */
 855 static int
 856 ext2_xattr_cmp(struct ext2_xattr_header *header1,
 857                struct ext2_xattr_header *header2)
 858 {
 859         struct ext2_xattr_entry *entry1, *entry2;
 860
 861         entry1 = ENTRY(header1+1);
 862         entry2 = ENTRY(header2+1);
 863         while (!IS_LAST_ENTRY(entry1)) {
 864                 if (IS_LAST_ENTRY(entry2))
 865                         return 1;
 866                 if (entry1->e_hash != entry2->e_hash ||
 867                     entry1->e_name_index != entry2->e_name_index ||
 868                     entry1->e_name_len != entry2->e_name_len ||
 869                     entry1->e_value_size != entry2->e_value_size ||
 870                     memcmp(entry1->e_name, entry2->e_name, entry1->e_name_len))
 871                         return 1;
 872                 if (entry1->e_value_block != 0 || entry2->e_value_block != 0)
 873                         return -EIO;
 874                 if (memcmp((char *)header1 + le16_to_cpu(entry1->e_value_offs),
 875                            (char *)header2 + le16_to_cpu(entry2->e_value_offs),
 876                            le32_to_cpu(entry1->e_value_size)))
 877                         return 1;
 878
 879                 entry1 = EXT2_XATTR_NEXT(entry1);
 880                 entry2 = EXT2_XATTR_NEXT(entry2);
 881         }
 882         if (!IS_LAST_ENTRY(entry2))
 883                 return 1;
 884         return 0;
 885 }
 886
 887 /*
 888  * ext2_xattr_cache_find()
 889  *
 890  * Find an identical extended attribute block.
 891  *
 892  * Returns a locked buffer head to the block found, or NULL if such
 893  * a block was not found or an error occurred.
 894  */
 895 static struct buffer_head *
 896 ext2_xattr_cache_find(struct inode *inode, struct ext2_xattr_header *header)
 897 {
 898         __u32 hash = le32_to_cpu(header->h_hash);
 899         struct mb_cache_entry *ce;
 900         struct mb_cache *ext2_mb_cache = EXT2_SB(inode->i_sb)->s_mb_cache;
 901
 902         if (!header->h_hash)
 903                 return NULL;  /* never share */
 904         ea_idebug(inode, "looking for cached blocks [%x]", (int)hash);
 905 again:
 906         ce = mb_cache_entry_find_first(ext2_mb_cache, hash);
 907         while (ce) {
 908                 struct buffer_head *bh;
 909
 910                 bh = sb_bread(inode->i_sb, ce->e_block);
 911                 if (!bh) {
 912                         ext2_error(inode->i_sb, "ext2_xattr_cache_find",
 913                                 "inode %ld: block %ld read error",
 914                                 inode->i_ino, (unsigned long) ce->e_block);
 915                 } else {
 916                         lock_buffer(bh);
 917                         /*
 918                          * We have to be careful about races with freeing or
 919                          * rehashing of xattr block. Once we hold buffer lock
 920                          * xattr block's state is stable so we can check
 921                          * whether the block got freed / rehashed or not.
 922                          * Since we unhash mbcache entry under buffer lock when
 923                          * freeing / rehashing xattr block, checking whether
 924                          * entry is still hashed is reliable.
 925                          */
 926                         if (hlist_bl_unhashed(&ce->e_hash_list)) {
 927                                 mb_cache_entry_put(ext2_mb_cache, ce);
 928                                 unlock_buffer(bh);
 929                                 brelse(bh);
 930                                 goto again;
 931                         } else if (le32_to_cpu(HDR(bh)->h_refcount) >
 932                                    EXT2_XATTR_REFCOUNT_MAX) {
 933                                 ea_idebug(inode, "block %ld refcount %d>%d",
 934                                           (unsigned long) ce->e_block,
 935                                           le32_to_cpu(HDR(bh)->h_refcount),
 936                                           EXT2_XATTR_REFCOUNT_MAX);
 937                         } else if (!ext2_xattr_cmp(header, HDR(bh))) {
 938                                 ea_bdebug(bh, "b_count=%d",
 939                                           atomic_read(&(bh->b_count)));
 940                                 mb_cache_entry_touch(ext2_mb_cache, ce);
 941                                 mb_cache_entry_put(ext2_mb_cache, ce);
 942                                 return bh;
 943                         }
 944                         unlock_buffer(bh);
 945                         brelse(bh);
 946                 }
 947                 ce = mb_cache_entry_find_next(ext2_mb_cache, ce);
 948         }
 949         return NULL;
 950 }
 951
 952 #define NAME_HASH_SHIFT 5
 953 #define VALUE_HASH_SHIFT 16
 954
 955 /*
 956  * ext2_xattr_hash_entry()
 957  *
 958  * Compute the hash of an extended attribute.
 959  */
 960 static inline void ext2_xattr_hash_entry(struct ext2_xattr_header *header,
 961                                          struct ext2_xattr_entry *entry)
 962 {
 963         __u32 hash = 0;
 964         char *name = entry->e_name;
 965         int n;
 966
 967         for (n=0; n < entry->e_name_len; n++) {
 968                 hash = (hash << NAME_HASH_SHIFT) ^
 969                        (hash >> (8*sizeof(hash) - NAME_HASH_SHIFT)) ^
 970                        *name++;
 971         }
 972
 973         if (entry->e_value_block == 0 && entry->e_value_size != 0) {
 974                 __le32 *value = (__le32 *)((char *)header +
 975                         le16_to_cpu(entry->e_value_offs));
 976                 for (n = (le32_to_cpu(entry->e_value_size) +
 977                      EXT2_XATTR_ROUND) >> EXT2_XATTR_PAD_BITS; n; n--) {
 978                         hash = (hash << VALUE_HASH_SHIFT) ^
 979                                (hash >> (8*sizeof(hash) - VALUE_HASH_SHIFT)) ^
 980                                le32_to_cpu(*value++);
 981                 }
 982         }
 983         entry->e_hash = cpu_to_le32(hash);
 984 }
 985
 986 #undef NAME_HASH_SHIFT
 987 #undef VALUE_HASH_SHIFT
 988
 989 #define BLOCK_HASH_SHIFT 16
 990
 991 /*
 992  * ext2_xattr_rehash()
 993  *
 994  * Re-compute the extended attribute hash value after an entry has changed.
 995  */
 996 static void ext2_xattr_rehash(struct ext2_xattr_header *header,
 997                               struct ext2_xattr_entry *entry)
 998 {
 999         struct ext2_xattr_entry *here;
1000         __u32 hash = 0;
1001
1002         ext2_xattr_hash_entry(header, entry);
1003         here = ENTRY(header+1);
1004         while (!IS_LAST_ENTRY(here)) {
1005                 if (!here->e_hash) {
1006                         /* Block is not shared if an entry's hash value == 0 */
1007                         hash = 0;
1008                         break;
1009                 }
1010                 hash = (hash << BLOCK_HASH_SHIFT) ^
1011                        (hash >> (8*sizeof(hash) - BLOCK_HASH_SHIFT)) ^
1012                        le32_to_cpu(here->e_hash);
1013                 here = EXT2_XATTR_NEXT(here);
1014         }
1015         header->h_hash = cpu_to_le32(hash);
1016 }
1017
1018 #undef BLOCK_HASH_SHIFT
1019
1020 #define HASH_BUCKET_BITS 10
1021
1022 struct mb_cache *ext2_xattr_create_cache(void)
1023 {
1024         return mb_cache_create(HASH_BUCKET_BITS);
1025 }
1026
1027 void ext2_xattr_destroy_cache(struct mb_cache *cache)
1028 {
1029         if (cache)
1030                 mb_cache_destroy(cache);
1031 }