fs/ext4/xattr.c

   1 /*
   2  * linux/fs/ext4/xattr.c
   3  *
   4  * Copyright (C) 2001-2003 Andreas Gruenbacher, <agruen@suse.de>
   5  *
   6  * Fix by Harrison Xing <harrison@mountainviewdata.com>.
   7  * Ext4 code with a lot of help from Eric Jarman <ejarman@acm.org>.
   8  * Extended attributes for symlinks and special files added per
   9  *  suggestion of Luka Renko <luka.renko@hermes.si>.
  10  * xattr consolidation Copyright (c) 2004 James Morris <jmorris@redhat.com>,
  11  *  Red Hat Inc.
  12  * ea-in-inode support by Alex Tomas <alex@clusterfs.com> aka bzzz
  13  *  and Andreas Gruenbacher <agruen@suse.de>.
  14  */
  15
  16 /*
  17  * Extended attributes are stored directly in inodes (on file systems with
  18  * inodes bigger than 128 bytes) and on additional disk blocks. The i_file_acl
  19  * field contains the block number if an inode uses an additional block. All
  20  * attributes must fit in the inode and one additional block. Blocks that
  21  * contain the identical set of attributes may be shared among several inodes.
  22  * Identical blocks are detected by keeping a cache of blocks that have
  23  * recently been accessed.
  24  *
  25  * The attributes in inodes and on blocks have a different header; the entries
  26  * are stored in the same format:
  27  *
  28  *   +------------------+
  29  *   | header           |
  30  *   | entry 1          | |
  31  *   | entry 2          | | growing downwards
  32  *   | entry 3          | v
  33  *   | four null bytes  |
  34  *   | . . .            |
  35  *   | value 1          | ^
  36  *   | value 3          | | growing upwards
  37  *   | value 2          | |
  38  *   +------------------+
  39  *
  40  * The header is followed by multiple entry descriptors. In disk blocks, the
  41  * entry descriptors are kept sorted. In inodes, they are unsorted. The
  42  * attribute values are aligned to the end of the block in no specific order.
  43  *
  44  * Locking strategy
  45  * ----------------
  46  * EXT4_I(inode)->i_file_acl is protected by EXT4_I(inode)->xattr_sem.
  47  * EA blocks are only changed if they are exclusive to an inode, so
  48  * holding xattr_sem also means that nothing but the EA block's reference
  49  * count can change. Multiple writers to the same block are synchronized
  50  * by the buffer lock.
  51  */
  52
  53 #include <linux/init.h>
  54 #include <linux/fs.h>
  55 #include <linux/slab.h>
  56 #include <linux/mbcache.h>
  57 #include <linux/quotaops.h>
  58 #include <linux/rwsem.h>
  59 #include "ext4_jbd2.h"
  60 #include "ext4.h"
  61 #include "xattr.h"
  62 #include "acl.h"
  63
  64 #ifdef EXT4_XATTR_DEBUG
  65 # define ea_idebug(inode, f...) do { \
  66                 printk(KERN_DEBUG "inode %s:%lu: ", \
  67                         inode->i_sb->s_id, inode->i_ino); \
  68                 printk(f); \
  69                 printk("\n"); \
  70         } while (0)
  71 # define ea_bdebug(bh, f...) do { \
  72                 char b[BDEVNAME_SIZE]; \
  73                 printk(KERN_DEBUG "block %s:%lu: ", \
  74                         bdevname(bh->b_bdev, b), \
  75                         (unsigned long) bh->b_blocknr); \
  76                 printk(f); \
  77                 printk("\n"); \
  78         } while (0)
  79 #else
  80 # define ea_idebug(inode, fmt, ...)     no_printk(fmt, ##__VA_ARGS__)
  81 # define ea_bdebug(bh, fmt, ...)        no_printk(fmt, ##__VA_ARGS__)
  82 #endif
  83
  84 static void ext4_xattr_cache_insert(struct buffer_head *);
  85 static struct buffer_head *ext4_xattr_cache_find(struct inode *,
  86                                                  struct ext4_xattr_header *,
  87                                                  struct mb_cache_entry **);
  88 static void ext4_xattr_rehash(struct ext4_xattr_header *,
  89                               struct ext4_xattr_entry *);
  90 static int ext4_xattr_list(struct dentry *dentry, char *buffer,
  91                            size_t buffer_size);
  92
  93 static struct mb_cache *ext4_xattr_cache;
  94
  95 static const struct xattr_handler *ext4_xattr_handler_map[] = {
  96         [EXT4_XATTR_INDEX_USER]              = &ext4_xattr_user_handler,
  97 #ifdef CONFIG_EXT4_FS_POSIX_ACL
  98         [EXT4_XATTR_INDEX_POSIX_ACL_ACCESS]  = &ext4_xattr_acl_access_handler,
  99         [EXT4_XATTR_INDEX_POSIX_ACL_DEFAULT] = &ext4_xattr_acl_default_handler,
 100 #endif
 101         [EXT4_XATTR_INDEX_TRUSTED]           = &ext4_xattr_trusted_handler,
 102 #ifdef CONFIG_EXT4_FS_SECURITY
 103         [EXT4_XATTR_INDEX_SECURITY]          = &ext4_xattr_security_handler,
 104 #endif
 105 };
 106
 107 const struct xattr_handler *ext4_xattr_handlers[] = {
 108         &ext4_xattr_user_handler,
 109         &ext4_xattr_trusted_handler,
 110 #ifdef CONFIG_EXT4_FS_POSIX_ACL
 111         &ext4_xattr_acl_access_handler,
 112         &ext4_xattr_acl_default_handler,
 113 #endif
 114 #ifdef CONFIG_EXT4_FS_SECURITY
 115         &ext4_xattr_security_handler,
 116 #endif
 117         NULL
 118 };
 119
 120 static __le32 ext4_xattr_block_csum(struct inode *inode,
 121                                     sector_t block_nr,
 122                                     struct ext4_xattr_header *hdr)
 123 {
 124         struct ext4_sb_info *sbi = EXT4_SB(inode->i_sb);
 125         __u32 csum, old;
 126
 127         old = hdr->h_checksum;
 128         hdr->h_checksum = 0;
 129         block_nr = cpu_to_le64(block_nr);
 130         csum = ext4_chksum(sbi, sbi->s_csum_seed, (__u8 *)&block_nr,
 131                            sizeof(block_nr));
 132         csum = ext4_chksum(sbi, csum, (__u8 *)hdr,
 133                            EXT4_BLOCK_SIZE(inode->i_sb));
 134
 135         hdr->h_checksum = old;
 136         return cpu_to_le32(csum);
 137 }
 138
 139 static int ext4_xattr_block_csum_verify(struct inode *inode,
 140                                         sector_t block_nr,
 141                                         struct ext4_xattr_header *hdr)
 142 {
 143         if (EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
 144                 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM) &&
 145             (hdr->h_checksum != ext4_xattr_block_csum(inode, block_nr, hdr)))
 146                 return 0;
 147         return 1;
 148 }
 149
 150 static void ext4_xattr_block_csum_set(struct inode *inode,
 151                                       sector_t block_nr,
 152                                       struct ext4_xattr_header *hdr)
 153 {
 154         if (!EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
 155                 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM))
 156                 return;
 157
 158         hdr->h_checksum = ext4_xattr_block_csum(inode, block_nr, hdr);
 159 }
 160
 161 static inline int ext4_handle_dirty_xattr_block(handle_t *handle,
 162                                                 struct inode *inode,
 163                                                 struct buffer_head *bh)
 164 {
 165         ext4_xattr_block_csum_set(inode, bh->b_blocknr, BHDR(bh));
 166         return ext4_handle_dirty_metadata(handle, inode, bh);
 167 }
 168
 169 static inline const struct xattr_handler *
 170 ext4_xattr_handler(int name_index)
 171 {
 172         const struct xattr_handler *handler = NULL;
 173
 174         if (name_index > 0 && name_index < ARRAY_SIZE(ext4_xattr_handler_map))
 175                 handler = ext4_xattr_handler_map[name_index];
 176         return handler;
 177 }
 178
 179 /*
 180  * Inode operation listxattr()
 181  *
 182  * dentry->d_inode->i_mutex: don't care
 183  */
 184 ssize_t
 185 ext4_listxattr(struct dentry *dentry, char *buffer, size_t size)
 186 {
 187         return ext4_xattr_list(dentry, buffer, size);
 188 }
 189
 190 static int
 191 ext4_xattr_check_names(struct ext4_xattr_entry *entry, void *end)
 192 {
 193         while (!IS_LAST_ENTRY(entry)) {
 194                 struct ext4_xattr_entry *next = EXT4_XATTR_NEXT(entry);
 195                 if ((void *)next >= end)
 196                         return -EIO;
 197                 entry = next;
 198         }
 199         return 0;
 200 }
 201
 202 static inline int
 203 ext4_xattr_check_block(struct inode *inode, struct buffer_head *bh)
 204 {
 205         int error;
 206
 207         if (buffer_verified(bh))
 208                 return 0;
 209
 210         if (BHDR(bh)->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC) ||
 211             BHDR(bh)->h_blocks != cpu_to_le32(1))
 212                 return -EIO;
 213         if (!ext4_xattr_block_csum_verify(inode, bh->b_blocknr, BHDR(bh)))
 214                 return -EIO;
 215         error = ext4_xattr_check_names(BFIRST(bh), bh->b_data + bh->b_size);
 216         if (!error)
 217                 set_buffer_verified(bh);
 218         return error;
 219 }
 220
 221 static inline int
 222 ext4_xattr_check_entry(struct ext4_xattr_entry *entry, size_t size)
 223 {
 224         size_t value_size = le32_to_cpu(entry->e_value_size);
 225
 226         if (entry->e_value_block != 0 || value_size > size ||
 227             le16_to_cpu(entry->e_value_offs) + value_size > size)
 228                 return -EIO;
 229         return 0;
 230 }
 231
 232 static int
 233 ext4_xattr_find_entry(struct ext4_xattr_entry **pentry, int name_index,
 234                       const char *name, size_t size, int sorted)
 235 {
 236         struct ext4_xattr_entry *entry;
 237         size_t name_len;
 238         int cmp = 1;
 239
 240         if (name == NULL)
 241                 return -EINVAL;
 242         name_len = strlen(name);
 243         entry = *pentry;
 244         for (; !IS_LAST_ENTRY(entry); entry = EXT4_XATTR_NEXT(entry)) {
 245                 cmp = name_index - entry->e_name_index;
 246                 if (!cmp)
 247                         cmp = name_len - entry->e_name_len;
 248                 if (!cmp)
 249                         cmp = memcmp(name, entry->e_name, name_len);
 250                 if (cmp <= 0 && (sorted || cmp == 0))
 251                         break;
 252         }
 253         *pentry = entry;
 254         if (!cmp && ext4_xattr_check_entry(entry, size))
 255                         return -EIO;
 256         return cmp ? -ENODATA : 0;
 257 }
 258
 259 static int
 260 ext4_xattr_block_get(struct inode *inode, int name_index, const char *name,
 261                      void *buffer, size_t buffer_size)
 262 {
 263         struct buffer_head *bh = NULL;
 264         struct ext4_xattr_entry *entry;
 265         size_t size;
 266         int error;
 267
 268         ea_idebug(inode, "name=%d.%s, buffer=%p, buffer_size=%ld",
 269                   name_index, name, buffer, (long)buffer_size);
 270
 271         error = -ENODATA;
 272         if (!EXT4_I(inode)->i_file_acl)
 273                 goto cleanup;
 274         ea_idebug(inode, "reading block %llu",
 275                   (unsigned long long)EXT4_I(inode)->i_file_acl);
 276         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
 277         if (!bh)
 278                 goto cleanup;
 279         ea_bdebug(bh, "b_count=%d, refcount=%d",
 280                 atomic_read(&(bh->b_count)), le32_to_cpu(BHDR(bh)->h_refcount));
 281         if (ext4_xattr_check_block(inode, bh)) {
 282 bad_block:
 283                 EXT4_ERROR_INODE(inode, "bad block %llu",
 284                                  EXT4_I(inode)->i_file_acl);
 285                 error = -EIO;
 286                 goto cleanup;
 287         }
 288         ext4_xattr_cache_insert(bh);
 289         entry = BFIRST(bh);
 290         error = ext4_xattr_find_entry(&entry, name_index, name, bh->b_size, 1);
 291         if (error == -EIO)
 292                 goto bad_block;
 293         if (error)
 294                 goto cleanup;
 295         size = le32_to_cpu(entry->e_value_size);
 296         if (buffer) {
 297                 error = -ERANGE;
 298                 if (size > buffer_size)
 299                         goto cleanup;
 300                 memcpy(buffer, bh->b_data + le16_to_cpu(entry->e_value_offs),
 301                        size);
 302         }
 303         error = size;
 304
 305 cleanup:
 306         brelse(bh);
 307         return error;
 308 }
 309
 310 int
 311 ext4_xattr_ibody_get(struct inode *inode, int name_index, const char *name,
 312                      void *buffer, size_t buffer_size)
 313 {
 314         struct ext4_xattr_ibody_header *header;
 315         struct ext4_xattr_entry *entry;
 316         struct ext4_inode *raw_inode;
 317         struct ext4_iloc iloc;
 318         size_t size;
 319         void *end;
 320         int error;
 321
 322         if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR))
 323                 return -ENODATA;
 324         error = ext4_get_inode_loc(inode, &iloc);
 325         if (error)
 326                 return error;
 327         raw_inode = ext4_raw_inode(&iloc);
 328         header = IHDR(inode, raw_inode);
 329         entry = IFIRST(header);
 330         end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
 331         error = ext4_xattr_check_names(entry, end);
 332         if (error)
 333                 goto cleanup;
 334         error = ext4_xattr_find_entry(&entry, name_index, name,
 335                                       end - (void *)entry, 0);
 336         if (error)
 337                 goto cleanup;
 338         size = le32_to_cpu(entry->e_value_size);
 339         if (buffer) {
 340                 error = -ERANGE;
 341                 if (size > buffer_size)
 342                         goto cleanup;
 343                 memcpy(buffer, (void *)IFIRST(header) +
 344                        le16_to_cpu(entry->e_value_offs), size);
 345         }
 346         error = size;
 347
 348 cleanup:
 349         brelse(iloc.bh);
 350         return error;
 351 }
 352
 353 /*
 354  * ext4_xattr_get()
 355  *
 356  * Copy an extended attribute into the buffer
 357  * provided, or compute the buffer size required.
 358  * Buffer is NULL to compute the size of the buffer required.
 359  *
 360  * Returns a negative error number on failure, or the number of bytes
 361  * used / required on success.
 362  */
 363 int
 364 ext4_xattr_get(struct inode *inode, int name_index, const char *name,
 365                void *buffer, size_t buffer_size)
 366 {
 367         int error;
 368
 369         down_read(&EXT4_I(inode)->xattr_sem);
 370         error = ext4_xattr_ibody_get(inode, name_index, name, buffer,
 371                                      buffer_size);
 372         if (error == -ENODATA)
 373                 error = ext4_xattr_block_get(inode, name_index, name, buffer,
 374                                              buffer_size);
 375         up_read(&EXT4_I(inode)->xattr_sem);
 376         return error;
 377 }
 378
 379 static int
 380 ext4_xattr_list_entries(struct dentry *dentry, struct ext4_xattr_entry *entry,
 381                         char *buffer, size_t buffer_size)
 382 {
 383         size_t rest = buffer_size;
 384
 385         for (; !IS_LAST_ENTRY(entry); entry = EXT4_XATTR_NEXT(entry)) {
 386                 const struct xattr_handler *handler =
 387                         ext4_xattr_handler(entry->e_name_index);
 388
 389                 if (handler) {
 390                         size_t size = handler->list(dentry, buffer, rest,
 391                                                     entry->e_name,
 392                                                     entry->e_name_len,
 393                                                     handler->flags);
 394                         if (buffer) {
 395                                 if (size > rest)
 396                                         return -ERANGE;
 397                                 buffer += size;
 398                         }
 399                         rest -= size;
 400                 }
 401         }
 402         return buffer_size - rest;
 403 }
 404
 405 static int
 406 ext4_xattr_block_list(struct dentry *dentry, char *buffer, size_t buffer_size)
 407 {
 408         struct inode *inode = dentry->d_inode;
 409         struct buffer_head *bh = NULL;
 410         int error;
 411
 412         ea_idebug(inode, "buffer=%p, buffer_size=%ld",
 413                   buffer, (long)buffer_size);
 414
 415         error = 0;
 416         if (!EXT4_I(inode)->i_file_acl)
 417                 goto cleanup;
 418         ea_idebug(inode, "reading block %llu",
 419                   (unsigned long long)EXT4_I(inode)->i_file_acl);
 420         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
 421         error = -EIO;
 422         if (!bh)
 423                 goto cleanup;
 424         ea_bdebug(bh, "b_count=%d, refcount=%d",
 425                 atomic_read(&(bh->b_count)), le32_to_cpu(BHDR(bh)->h_refcount));
 426         if (ext4_xattr_check_block(inode, bh)) {
 427                 EXT4_ERROR_INODE(inode, "bad block %llu",
 428                                  EXT4_I(inode)->i_file_acl);
 429                 error = -EIO;
 430                 goto cleanup;
 431         }
 432         ext4_xattr_cache_insert(bh);
 433         error = ext4_xattr_list_entries(dentry, BFIRST(bh), buffer, buffer_size);
 434
 435 cleanup:
 436         brelse(bh);
 437
 438         return error;
 439 }
 440
 441 static int
 442 ext4_xattr_ibody_list(struct dentry *dentry, char *buffer, size_t buffer_size)
 443 {
 444         struct inode *inode = dentry->d_inode;
 445         struct ext4_xattr_ibody_header *header;
 446         struct ext4_inode *raw_inode;
 447         struct ext4_iloc iloc;
 448         void *end;
 449         int error;
 450
 451         if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR))
 452                 return 0;
 453         error = ext4_get_inode_loc(inode, &iloc);
 454         if (error)
 455                 return error;
 456         raw_inode = ext4_raw_inode(&iloc);
 457         header = IHDR(inode, raw_inode);
 458         end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
 459         error = ext4_xattr_check_names(IFIRST(header), end);
 460         if (error)
 461                 goto cleanup;
 462         error = ext4_xattr_list_entries(dentry, IFIRST(header),
 463                                         buffer, buffer_size);
 464
 465 cleanup:
 466         brelse(iloc.bh);
 467         return error;
 468 }
 469
 470 /*
 471  * ext4_xattr_list()
 472  *
 473  * Copy a list of attribute names into the buffer
 474  * provided, or compute the buffer size required.
 475  * Buffer is NULL to compute the size of the buffer required.
 476  *
 477  * Returns a negative error number on failure, or the number of bytes
 478  * used / required on success.
 479  */
 480 static int
 481 ext4_xattr_list(struct dentry *dentry, char *buffer, size_t buffer_size)
 482 {
 483         int ret, ret2;
 484
 485         down_read(&EXT4_I(dentry->d_inode)->xattr_sem);
 486         ret = ret2 = ext4_xattr_ibody_list(dentry, buffer, buffer_size);
 487         if (ret < 0)
 488                 goto errout;
 489         if (buffer) {
 490                 buffer += ret;
 491                 buffer_size -= ret;
 492         }
 493         ret = ext4_xattr_block_list(dentry, buffer, buffer_size);
 494         if (ret < 0)
 495                 goto errout;
 496         ret += ret2;
 497 errout:
 498         up_read(&EXT4_I(dentry->d_inode)->xattr_sem);
 499         return ret;
 500 }
 501
 502 /*
 503  * If the EXT4_FEATURE_COMPAT_EXT_ATTR feature of this file system is
 504  * not set, set it.
 505  */
 506 static void ext4_xattr_update_super_block(handle_t *handle,
 507                                           struct super_block *sb)
 508 {
 509         if (EXT4_HAS_COMPAT_FEATURE(sb, EXT4_FEATURE_COMPAT_EXT_ATTR))
 510                 return;
 511
 512         if (ext4_journal_get_write_access(handle, EXT4_SB(sb)->s_sbh) == 0) {
 513                 EXT4_SET_COMPAT_FEATURE(sb, EXT4_FEATURE_COMPAT_EXT_ATTR);
 514                 ext4_handle_dirty_super(handle, sb);
 515         }
 516 }
 517
 518 /*
 519  * Release the xattr block BH: If the reference count is > 1, decrement
 520  * it; otherwise free the block.
 521  */
 522 static void
 523 ext4_xattr_release_block(handle_t *handle, struct inode *inode,
 524                          struct buffer_head *bh)
 525 {
 526         struct mb_cache_entry *ce = NULL;
 527         int error = 0;
 528
 529         ce = mb_cache_entry_get(ext4_xattr_cache, bh->b_bdev, bh->b_blocknr);
 530         error = ext4_journal_get_write_access(handle, bh);
 531         if (error)
 532                 goto out;
 533
 534         lock_buffer(bh);
 535         if (BHDR(bh)->h_refcount == cpu_to_le32(1)) {
 536                 ea_bdebug(bh, "refcount now=0; freeing");
 537                 if (ce)
 538                         mb_cache_entry_free(ce);
 539                 get_bh(bh);
 540                 ext4_free_blocks(handle, inode, bh, 0, 1,
 541                                  EXT4_FREE_BLOCKS_METADATA |
 542                                  EXT4_FREE_BLOCKS_FORGET);
 543                 unlock_buffer(bh);
 544         } else {
 545                 le32_add_cpu(&BHDR(bh)->h_refcount, -1);
 546                 if (ce)
 547                         mb_cache_entry_release(ce);
 548                 unlock_buffer(bh);
 549                 error = ext4_handle_dirty_xattr_block(handle, inode, bh);
 550                 if (IS_SYNC(inode))
 551                         ext4_handle_sync(handle);
 552                 dquot_free_block(inode, 1);
 553                 ea_bdebug(bh, "refcount now=%d; releasing",
 554                           le32_to_cpu(BHDR(bh)->h_refcount));
 555         }
 556 out:
 557         ext4_std_error(inode->i_sb, error);
 558         return;
 559 }
 560
 561 /*
 562  * Find the available free space for EAs. This also returns the total number of
 563  * bytes used by EA entries.
 564  */
 565 static size_t ext4_xattr_free_space(struct ext4_xattr_entry *last,
 566                                     size_t *min_offs, void *base, int *total)
 567 {
 568         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 569                 *total += EXT4_XATTR_LEN(last->e_name_len);
 570                 if (!last->e_value_block && last->e_value_size) {
 571                         size_t offs = le16_to_cpu(last->e_value_offs);
 572                         if (offs < *min_offs)
 573                                 *min_offs = offs;
 574                 }
 575         }
 576         return (*min_offs - ((void *)last - base) - sizeof(__u32));
 577 }
 578
 579 static int
 580 ext4_xattr_set_entry(struct ext4_xattr_info *i, struct ext4_xattr_search *s)
 581 {
 582         struct ext4_xattr_entry *last;
 583         size_t free, min_offs = s->end - s->base, name_len = strlen(i->name);
 584
 585         /* Compute min_offs and last. */
 586         last = s->first;
 587         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 588                 if (!last->e_value_block && last->e_value_size) {
 589                         size_t offs = le16_to_cpu(last->e_value_offs);
 590                         if (offs < min_offs)
 591                                 min_offs = offs;
 592                 }
 593         }
 594         free = min_offs - ((void *)last - s->base) - sizeof(__u32);
 595         if (!s->not_found) {
 596                 if (!s->here->e_value_block && s->here->e_value_size) {
 597                         size_t size = le32_to_cpu(s->here->e_value_size);
 598                         free += EXT4_XATTR_SIZE(size);
 599                 }
 600                 free += EXT4_XATTR_LEN(name_len);
 601         }
 602         if (i->value) {
 603                 if (free < EXT4_XATTR_SIZE(i->value_len) ||
 604                     free < EXT4_XATTR_LEN(name_len) +
 605                            EXT4_XATTR_SIZE(i->value_len))
 606                         return -ENOSPC;
 607         }
 608
 609         if (i->value && s->not_found) {
 610                 /* Insert the new name. */
 611                 size_t size = EXT4_XATTR_LEN(name_len);
 612                 size_t rest = (void *)last - (void *)s->here + sizeof(__u32);
 613                 memmove((void *)s->here + size, s->here, rest);
 614                 memset(s->here, 0, size);
 615                 s->here->e_name_index = i->name_index;
 616                 s->here->e_name_len = name_len;
 617                 memcpy(s->here->e_name, i->name, name_len);
 618         } else {
 619                 if (!s->here->e_value_block && s->here->e_value_size) {
 620                         void *first_val = s->base + min_offs;
 621                         size_t offs = le16_to_cpu(s->here->e_value_offs);
 622                         void *val = s->base + offs;
 623                         size_t size = EXT4_XATTR_SIZE(
 624                                 le32_to_cpu(s->here->e_value_size));
 625
 626                         if (i->value && size == EXT4_XATTR_SIZE(i->value_len)) {
 627                                 /* The old and the new value have the same
 628                                    size. Just replace. */
 629                                 s->here->e_value_size =
 630                                         cpu_to_le32(i->value_len);
 631                                 if (i->value == EXT4_ZERO_XATTR_VALUE) {
 632                                         memset(val, 0, size);
 633                                 } else {
 634                                         /* Clear pad bytes first. */
 635                                         memset(val + size - EXT4_XATTR_PAD, 0,
 636                                                EXT4_XATTR_PAD);
 637                                         memcpy(val, i->value, i->value_len);
 638                                 }
 639                                 return 0;
 640                         }
 641
 642                         /* Remove the old value. */
 643                         memmove(first_val + size, first_val, val - first_val);
 644                         memset(first_val, 0, size);
 645                         s->here->e_value_size = 0;
 646                         s->here->e_value_offs = 0;
 647                         min_offs += size;
 648
 649                         /* Adjust all value offsets. */
 650                         last = s->first;
 651                         while (!IS_LAST_ENTRY(last)) {
 652                                 size_t o = le16_to_cpu(last->e_value_offs);
 653                                 if (!last->e_value_block &&
 654                                     last->e_value_size && o < offs)
 655                                         last->e_value_offs =
 656                                                 cpu_to_le16(o + size);
 657                                 last = EXT4_XATTR_NEXT(last);
 658                         }
 659                 }
 660                 if (!i->value) {
 661                         /* Remove the old name. */
 662                         size_t size = EXT4_XATTR_LEN(name_len);
 663                         last = ENTRY((void *)last - size);
 664                         memmove(s->here, (void *)s->here + size,
 665                                 (void *)last - (void *)s->here + sizeof(__u32));
 666                         memset(last, 0, size);
 667                 }
 668         }
 669
 670         if (i->value) {
 671                 /* Insert the new value. */
 672                 s->here->e_value_size = cpu_to_le32(i->value_len);
 673                 if (i->value_len) {
 674                         size_t size = EXT4_XATTR_SIZE(i->value_len);
 675                         void *val = s->base + min_offs - size;
 676                         s->here->e_value_offs = cpu_to_le16(min_offs - size);
 677                         if (i->value == EXT4_ZERO_XATTR_VALUE) {
 678                                 memset(val, 0, size);
 679                         } else {
 680                                 /* Clear the pad bytes first. */
 681                                 memset(val + size - EXT4_XATTR_PAD, 0,
 682                                        EXT4_XATTR_PAD);
 683                                 memcpy(val, i->value, i->value_len);
 684                         }
 685                 }
 686         }
 687         return 0;
 688 }
 689
 690 struct ext4_xattr_block_find {
 691         struct ext4_xattr_search s;
 692         struct buffer_head *bh;
 693 };
 694
 695 static int
 696 ext4_xattr_block_find(struct inode *inode, struct ext4_xattr_info *i,
 697                       struct ext4_xattr_block_find *bs)
 698 {
 699         struct super_block *sb = inode->i_sb;
 700         int error;
 701
 702         ea_idebug(inode, "name=%d.%s, value=%p, value_len=%ld",
 703                   i->name_index, i->name, i->value, (long)i->value_len);
 704
 705         if (EXT4_I(inode)->i_file_acl) {
 706                 /* The inode already has an extended attribute block. */
 707                 bs->bh = sb_bread(sb, EXT4_I(inode)->i_file_acl);
 708                 error = -EIO;
 709                 if (!bs->bh)
 710                         goto cleanup;
 711                 ea_bdebug(bs->bh, "b_count=%d, refcount=%d",
 712                         atomic_read(&(bs->bh->b_count)),
 713                         le32_to_cpu(BHDR(bs->bh)->h_refcount));
 714                 if (ext4_xattr_check_block(inode, bs->bh)) {
 715                         EXT4_ERROR_INODE(inode, "bad block %llu",
 716                                          EXT4_I(inode)->i_file_acl);
 717                         error = -EIO;
 718                         goto cleanup;
 719                 }
 720                 /* Find the named attribute. */
 721                 bs->s.base = BHDR(bs->bh);
 722                 bs->s.first = BFIRST(bs->bh);
 723                 bs->s.end = bs->bh->b_data + bs->bh->b_size;
 724                 bs->s.here = bs->s.first;
 725                 error = ext4_xattr_find_entry(&bs->s.here, i->name_index,
 726                                               i->name, bs->bh->b_size, 1);
 727                 if (error && error != -ENODATA)
 728                         goto cleanup;
 729                 bs->s.not_found = error;
 730         }
 731         error = 0;
 732
 733 cleanup:
 734         return error;
 735 }
 736
 737 static int
 738 ext4_xattr_block_set(handle_t *handle, struct inode *inode,
 739                      struct ext4_xattr_info *i,
 740                      struct ext4_xattr_block_find *bs)
 741 {
 742         struct super_block *sb = inode->i_sb;
 743         struct buffer_head *new_bh = NULL;
 744         struct ext4_xattr_search *s = &bs->s;
 745         struct mb_cache_entry *ce = NULL;
 746         int error = 0;
 747
 748 #define header(x) ((struct ext4_xattr_header *)(x))
 749
 750         if (i->value && i->value_len > sb->s_blocksize)
 751                 return -ENOSPC;
 752         if (s->base) {
 753                 ce = mb_cache_entry_get(ext4_xattr_cache, bs->bh->b_bdev,
 754                                         bs->bh->b_blocknr);
 755                 error = ext4_journal_get_write_access(handle, bs->bh);
 756                 if (error)
 757                         goto cleanup;
 758                 lock_buffer(bs->bh);
 759
 760                 if (header(s->base)->h_refcount == cpu_to_le32(1)) {
 761                         if (ce) {
 762                                 mb_cache_entry_free(ce);
 763                                 ce = NULL;
 764                         }
 765                         ea_bdebug(bs->bh, "modifying in-place");
 766                         error = ext4_xattr_set_entry(i, s);
 767                         if (!error) {
 768                                 if (!IS_LAST_ENTRY(s->first))
 769                                         ext4_xattr_rehash(header(s->base),
 770                                                           s->here);
 771                                 ext4_xattr_cache_insert(bs->bh);
 772                         }
 773                         unlock_buffer(bs->bh);
 774                         if (error == -EIO)
 775                                 goto bad_block;
 776                         if (!error)
 777                                 error = ext4_handle_dirty_xattr_block(handle,
 778                                                                       inode,
 779                                                                       bs->bh);
 780                         if (error)
 781                                 goto cleanup;
 782                         goto inserted;
 783                 } else {
 784                         int offset = (char *)s->here - bs->bh->b_data;
 785
 786                         unlock_buffer(bs->bh);
 787                         if (ce) {
 788                                 mb_cache_entry_release(ce);
 789                                 ce = NULL;
 790                         }
 791                         ea_bdebug(bs->bh, "cloning");
 792                         s->base = kmalloc(bs->bh->b_size, GFP_NOFS);
 793                         error = -ENOMEM;
 794                         if (s->base == NULL)
 795                                 goto cleanup;
 796                         memcpy(s->base, BHDR(bs->bh), bs->bh->b_size);
 797                         s->first = ENTRY(header(s->base)+1);
 798                         header(s->base)->h_refcount = cpu_to_le32(1);
 799                         s->here = ENTRY(s->base + offset);
 800                         s->end = s->base + bs->bh->b_size;
 801                 }
 802         } else {
 803                 /* Allocate a buffer where we construct the new block. */
 804                 s->base = kzalloc(sb->s_blocksize, GFP_NOFS);
 805                 /* assert(header == s->base) */
 806                 error = -ENOMEM;
 807                 if (s->base == NULL)
 808                         goto cleanup;
 809                 header(s->base)->h_magic = cpu_to_le32(EXT4_XATTR_MAGIC);
 810                 header(s->base)->h_blocks = cpu_to_le32(1);
 811                 header(s->base)->h_refcount = cpu_to_le32(1);
 812                 s->first = ENTRY(header(s->base)+1);
 813                 s->here = ENTRY(header(s->base)+1);
 814                 s->end = s->base + sb->s_blocksize;
 815         }
 816
 817         error = ext4_xattr_set_entry(i, s);
 818         if (error == -EIO)
 819                 goto bad_block;
 820         if (error)
 821                 goto cleanup;
 822         if (!IS_LAST_ENTRY(s->first))
 823                 ext4_xattr_rehash(header(s->base), s->here);
 824
 825 inserted:
 826         if (!IS_LAST_ENTRY(s->first)) {
 827                 new_bh = ext4_xattr_cache_find(inode, header(s->base), &ce);
 828                 if (new_bh) {
 829                         /* We found an identical block in the cache. */
 830                         if (new_bh == bs->bh)
 831                                 ea_bdebug(new_bh, "keeping");
 832                         else {
 833                                 /* The old block is released after updating
 834                                    the inode. */
 835                                 error = dquot_alloc_block(inode, 1);
 836                                 if (error)
 837                                         goto cleanup;
 838                                 error = ext4_journal_get_write_access(handle,
 839                                                                       new_bh);
 840                                 if (error)
 841                                         goto cleanup_dquot;
 842                                 lock_buffer(new_bh);
 843                                 le32_add_cpu(&BHDR(new_bh)->h_refcount, 1);
 844                                 ea_bdebug(new_bh, "reusing; refcount now=%d",
 845                                         le32_to_cpu(BHDR(new_bh)->h_refcount));
 846                                 unlock_buffer(new_bh);
 847                                 error = ext4_handle_dirty_xattr_block(handle,
 848                                                                       inode,
 849                                                                       new_bh);
 850                                 if (error)
 851                                         goto cleanup_dquot;
 852                         }
 853                         mb_cache_entry_release(ce);
 854                         ce = NULL;
 855                 } else if (bs->bh && s->base == bs->bh->b_data) {
 856                         /* We were modifying this block in-place. */
 857                         ea_bdebug(bs->bh, "keeping this block");
 858                         new_bh = bs->bh;
 859                         get_bh(new_bh);
 860                 } else {
 861                         /* We need to allocate a new block */
 862                         ext4_fsblk_t goal, block;
 863
 864                         goal = ext4_group_first_block_no(sb,
 865                                                 EXT4_I(inode)->i_block_group);
 866
 867                         /* non-extent files can't have physical blocks past 2^32 */
 868                         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
 869                                 goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
 870
 871                         /*
 872                          * take i_data_sem because we will test
 873                          * i_delalloc_reserved_flag in ext4_mb_new_blocks
 874                          */
 875                         down_read((&EXT4_I(inode)->i_data_sem));
 876                         block = ext4_new_meta_blocks(handle, inode, goal, 0,
 877                                                      NULL, &error);
 878                         up_read((&EXT4_I(inode)->i_data_sem));
 879                         if (error)
 880                                 goto cleanup;
 881
 882                         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
 883                                 BUG_ON(block > EXT4_MAX_BLOCK_FILE_PHYS);
 884
 885                         ea_idebug(inode, "creating block %llu",
 886                                   (unsigned long long)block);
 887
 888                         new_bh = sb_getblk(sb, block);
 889                         if (!new_bh) {
 890 getblk_failed:
 891                                 ext4_free_blocks(handle, inode, NULL, block, 1,
 892                                                  EXT4_FREE_BLOCKS_METADATA);
 893                                 error = -EIO;
 894                                 goto cleanup;
 895                         }
 896                         lock_buffer(new_bh);
 897                         error = ext4_journal_get_create_access(handle, new_bh);
 898                         if (error) {
 899                                 unlock_buffer(new_bh);
 900                                 goto getblk_failed;
 901                         }
 902                         memcpy(new_bh->b_data, s->base, new_bh->b_size);
 903                         set_buffer_uptodate(new_bh);
 904                         unlock_buffer(new_bh);
 905                         ext4_xattr_cache_insert(new_bh);
 906                         error = ext4_handle_dirty_xattr_block(handle,
 907                                                               inode, new_bh);
 908                         if (error)
 909                                 goto cleanup;
 910                 }
 911         }
 912
 913         /* Update the inode. */
 914         EXT4_I(inode)->i_file_acl = new_bh ? new_bh->b_blocknr : 0;
 915
 916         /* Drop the previous xattr block. */
 917         if (bs->bh && bs->bh != new_bh)
 918                 ext4_xattr_release_block(handle, inode, bs->bh);
 919         error = 0;
 920
 921 cleanup:
 922         if (ce)
 923                 mb_cache_entry_release(ce);
 924         brelse(new_bh);
 925         if (!(bs->bh && s->base == bs->bh->b_data))
 926                 kfree(s->base);
 927
 928         return error;
 929
 930 cleanup_dquot:
 931         dquot_free_block(inode, 1);
 932         goto cleanup;
 933
 934 bad_block:
 935         EXT4_ERROR_INODE(inode, "bad block %llu",
 936                          EXT4_I(inode)->i_file_acl);
 937         goto cleanup;
 938
 939 #undef header
 940 }
 941
 942 int ext4_xattr_ibody_find(struct inode *inode, struct ext4_xattr_info *i,
 943                           struct ext4_xattr_ibody_find *is)
 944 {
 945         struct ext4_xattr_ibody_header *header;
 946         struct ext4_inode *raw_inode;
 947         int error;
 948
 949         if (EXT4_I(inode)->i_extra_isize == 0)
 950                 return 0;
 951         raw_inode = ext4_raw_inode(&is->iloc);
 952         header = IHDR(inode, raw_inode);
 953         is->s.base = is->s.first = IFIRST(header);
 954         is->s.here = is->s.first;
 955         is->s.end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
 956         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
 957                 error = ext4_xattr_check_names(IFIRST(header), is->s.end);
 958                 if (error)
 959                         return error;
 960                 /* Find the named attribute. */
 961                 error = ext4_xattr_find_entry(&is->s.here, i->name_index,
 962                                               i->name, is->s.end -
 963                                               (void *)is->s.base, 0);
 964                 if (error && error != -ENODATA)
 965                         return error;
 966                 is->s.not_found = error;
 967         }
 968         return 0;
 969 }
 970
 971 int ext4_xattr_ibody_inline_set(handle_t *handle, struct inode *inode,
 972                                 struct ext4_xattr_info *i,
 973                                 struct ext4_xattr_ibody_find *is)
 974 {
 975         struct ext4_xattr_ibody_header *header;
 976         struct ext4_xattr_search *s = &is->s;
 977         int error;
 978
 979         if (EXT4_I(inode)->i_extra_isize == 0)
 980                 return -ENOSPC;
 981         error = ext4_xattr_set_entry(i, s);
 982         if (error) {
 983                 if (error == -ENOSPC &&
 984                     ext4_has_inline_data(inode)) {
 985                         error = ext4_try_to_evict_inline_data(handle, inode,
 986                                         EXT4_XATTR_LEN(strlen(i->name) +
 987                                         EXT4_XATTR_SIZE(i->value_len)));
 988                         if (error)
 989                                 return error;
 990                         error = ext4_xattr_ibody_find(inode, i, is);
 991                         if (error)
 992                                 return error;
 993                         error = ext4_xattr_set_entry(i, s);
 994                 }
 995                 if (error)
 996                         return error;
 997         }
 998         header = IHDR(inode, ext4_raw_inode(&is->iloc));
 999         if (!IS_LAST_ENTRY(s->first)) {
1000                 header->h_magic = cpu_to_le32(EXT4_XATTR_MAGIC);
1001                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
1002         } else {
1003                 header->h_magic = cpu_to_le32(0);
1004                 ext4_clear_inode_state(inode, EXT4_STATE_XATTR);
1005         }
1006         return 0;
1007 }
1008
1009 static int ext4_xattr_ibody_set(handle_t *handle, struct inode *inode,
1010                                 struct ext4_xattr_info *i,
1011                                 struct ext4_xattr_ibody_find *is)
1012 {
1013         struct ext4_xattr_ibody_header *header;
1014         struct ext4_xattr_search *s = &is->s;
1015         int error;
1016
1017         if (EXT4_I(inode)->i_extra_isize == 0)
1018                 return -ENOSPC;
1019         error = ext4_xattr_set_entry(i, s);
1020         if (error)
1021                 return error;
1022         header = IHDR(inode, ext4_raw_inode(&is->iloc));
1023         if (!IS_LAST_ENTRY(s->first)) {
1024                 header->h_magic = cpu_to_le32(EXT4_XATTR_MAGIC);
1025                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
1026         } else {
1027                 header->h_magic = cpu_to_le32(0);
1028                 ext4_clear_inode_state(inode, EXT4_STATE_XATTR);
1029         }
1030         return 0;
1031 }
1032
1033 /*
1034  * ext4_xattr_set_handle()
1035  *
1036  * Create, replace or remove an extended attribute for this inode.  Value
1037  * is NULL to remove an existing extended attribute, and non-NULL to
1038  * either replace an existing extended attribute, or create a new extended
1039  * attribute. The flags XATTR_REPLACE and XATTR_CREATE
1040  * specify that an extended attribute must exist and must not exist
1041  * previous to the call, respectively.
1042  *
1043  * Returns 0, or a negative error number on failure.
1044  */
1045 int
1046 ext4_xattr_set_handle(handle_t *handle, struct inode *inode, int name_index,
1047                       const char *name, const void *value, size_t value_len,
1048                       int flags)
1049 {
1050         struct ext4_xattr_info i = {
1051                 .name_index = name_index,
1052                 .name = name,
1053                 .value = value,
1054                 .value_len = value_len,
1055
1056         };
1057         struct ext4_xattr_ibody_find is = {
1058                 .s = { .not_found = -ENODATA, },
1059         };
1060         struct ext4_xattr_block_find bs = {
1061                 .s = { .not_found = -ENODATA, },
1062         };
1063         unsigned long no_expand;
1064         int error;
1065
1066         if (!name)
1067                 return -EINVAL;
1068         if (strlen(name) > 255)
1069                 return -ERANGE;
1070         down_write(&EXT4_I(inode)->xattr_sem);
1071         no_expand = ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND);
1072         ext4_set_inode_state(inode, EXT4_STATE_NO_EXPAND);
1073
1074         error = ext4_reserve_inode_write(handle, inode, &is.iloc);
1075         if (error)
1076                 goto cleanup;
1077
1078         if (ext4_test_inode_state(inode, EXT4_STATE_NEW)) {
1079                 struct ext4_inode *raw_inode = ext4_raw_inode(&is.iloc);
1080                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
1081                 ext4_clear_inode_state(inode, EXT4_STATE_NEW);
1082         }
1083
1084         error = ext4_xattr_ibody_find(inode, &i, &is);
1085         if (error)
1086                 goto cleanup;
1087         if (is.s.not_found)
1088                 error = ext4_xattr_block_find(inode, &i, &bs);
1089         if (error)
1090                 goto cleanup;
1091         if (is.s.not_found && bs.s.not_found) {
1092                 error = -ENODATA;
1093                 if (flags & XATTR_REPLACE)
1094                         goto cleanup;
1095                 error = 0;
1096                 if (!value)
1097                         goto cleanup;
1098         } else {
1099                 error = -EEXIST;
1100                 if (flags & XATTR_CREATE)
1101                         goto cleanup;
1102         }
1103         if (!value) {
1104                 if (!is.s.not_found)
1105                         error = ext4_xattr_ibody_set(handle, inode, &i, &is);
1106                 else if (!bs.s.not_found)
1107                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
1108         } else {
1109                 error = ext4_xattr_ibody_set(handle, inode, &i, &is);
1110                 if (!error && !bs.s.not_found) {
1111                         i.value = NULL;
1112                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
1113                 } else if (error == -ENOSPC) {
1114                         if (EXT4_I(inode)->i_file_acl && !bs.s.base) {
1115                                 error = ext4_xattr_block_find(inode, &i, &bs);
1116                                 if (error)
1117                                         goto cleanup;
1118                         }
1119                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
1120                         if (error)
1121                                 goto cleanup;
1122                         if (!is.s.not_found) {
1123                                 i.value = NULL;
1124                                 error = ext4_xattr_ibody_set(handle, inode, &i,
1125                                                              &is);
1126                         }
1127                 }
1128         }
1129         if (!error) {
1130                 ext4_xattr_update_super_block(handle, inode->i_sb);
1131                 inode->i_ctime = ext4_current_time(inode);
1132                 if (!value)
1133                         ext4_clear_inode_state(inode, EXT4_STATE_NO_EXPAND);
1134                 error = ext4_mark_iloc_dirty(handle, inode, &is.iloc);
1135                 /*
1136                  * The bh is consumed by ext4_mark_iloc_dirty, even with
1137                  * error != 0.
1138                  */
1139                 is.iloc.bh = NULL;
1140                 if (IS_SYNC(inode))
1141                         ext4_handle_sync(handle);
1142         }
1143
1144 cleanup:
1145         brelse(is.iloc.bh);
1146         brelse(bs.bh);
1147         if (no_expand == 0)
1148                 ext4_clear_inode_state(inode, EXT4_STATE_NO_EXPAND);
1149         up_write(&EXT4_I(inode)->xattr_sem);
1150         return error;
1151 }
1152
1153 /*
1154  * ext4_xattr_set()
1155  *
1156  * Like ext4_xattr_set_handle, but start from an inode. This extended
1157  * attribute modification is a filesystem transaction by itself.
1158  *
1159  * Returns 0, or a negative error number on failure.
1160  */
1161 int
1162 ext4_xattr_set(struct inode *inode, int name_index, const char *name,
1163                const void *value, size_t value_len, int flags)
1164 {
1165         handle_t *handle;
1166         int error, retries = 0;
1167         int credits = EXT4_DATA_TRANS_BLOCKS(inode->i_sb);
1168
1169 retry:
1170         /*
1171          * In case of inline data, we may push out the data to a block,
1172          * So reserve the journal space first.
1173          */
1174         if (ext4_has_inline_data(inode))
1175                 credits += ext4_writepage_trans_blocks(inode) + 1;
1176
1177         handle = ext4_journal_start(inode, credits);
1178         if (IS_ERR(handle)) {
1179                 error = PTR_ERR(handle);
1180         } else {
1181                 int error2;
1182
1183                 error = ext4_xattr_set_handle(handle, inode, name_index, name,
1184                                               value, value_len, flags);
1185                 error2 = ext4_journal_stop(handle);
1186                 if (error == -ENOSPC &&
1187                     ext4_should_retry_alloc(inode->i_sb, &retries))
1188                         goto retry;
1189                 if (error == 0)
1190                         error = error2;
1191         }
1192
1193         return error;
1194 }
1195
1196 /*
1197  * Shift the EA entries in the inode to create space for the increased
1198  * i_extra_isize.
1199  */
1200 static void ext4_xattr_shift_entries(struct ext4_xattr_entry *entry,
1201                                      int value_offs_shift, void *to,
1202                                      void *from, size_t n, int blocksize)
1203 {
1204         struct ext4_xattr_entry *last = entry;
1205         int new_offs;
1206
1207         /* Adjust the value offsets of the entries */
1208         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
1209                 if (!last->e_value_block && last->e_value_size) {
1210                         new_offs = le16_to_cpu(last->e_value_offs) +
1211                                                         value_offs_shift;
1212                         BUG_ON(new_offs + le32_to_cpu(last->e_value_size)
1213                                  > blocksize);
1214                         last->e_value_offs = cpu_to_le16(new_offs);
1215                 }
1216         }
1217         /* Shift the entries by n bytes */
1218         memmove(to, from, n);
1219 }
1220
1221 /*
1222  * Expand an inode by new_extra_isize bytes when EAs are present.
1223  * Returns 0 on success or negative error number on failure.
1224  */
1225 int ext4_expand_extra_isize_ea(struct inode *inode, int new_extra_isize,
1226                                struct ext4_inode *raw_inode, handle_t *handle)
1227 {
1228         struct ext4_xattr_ibody_header *header;
1229         struct ext4_xattr_entry *entry, *last, *first;
1230         struct buffer_head *bh = NULL;
1231         struct ext4_xattr_ibody_find *is = NULL;
1232         struct ext4_xattr_block_find *bs = NULL;
1233         char *buffer = NULL, *b_entry_name = NULL;
1234         size_t min_offs, free;
1235         int total_ino, total_blk;
1236         void *base, *start, *end;
1237         int extra_isize = 0, error = 0, tried_min_extra_isize = 0;
1238         int s_min_extra_isize = le16_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_min_extra_isize);
1239
1240         down_write(&EXT4_I(inode)->xattr_sem);
1241 retry:
1242         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize) {
1243                 up_write(&EXT4_I(inode)->xattr_sem);
1244                 return 0;
1245         }
1246
1247         header = IHDR(inode, raw_inode);
1248         entry = IFIRST(header);
1249
1250         /*
1251          * Check if enough free space is available in the inode to shift the
1252          * entries ahead by new_extra_isize.
1253          */
1254
1255         base = start = entry;
1256         end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
1257         min_offs = end - base;
1258         last = entry;
1259         total_ino = sizeof(struct ext4_xattr_ibody_header);
1260
1261         free = ext4_xattr_free_space(last, &min_offs, base, &total_ino);
1262         if (free >= new_extra_isize) {
1263                 entry = IFIRST(header);
1264                 ext4_xattr_shift_entries(entry, EXT4_I(inode)->i_extra_isize
1265                                 - new_extra_isize, (void *)raw_inode +
1266                                 EXT4_GOOD_OLD_INODE_SIZE + new_extra_isize,
1267                                 (void *)header, total_ino,
1268                                 inode->i_sb->s_blocksize);
1269                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
1270                 error = 0;
1271                 goto cleanup;
1272         }
1273
1274         /*
1275          * Enough free space isn't available in the inode, check if
1276          * EA block can hold new_extra_isize bytes.
1277          */
1278         if (EXT4_I(inode)->i_file_acl) {
1279                 bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
1280                 error = -EIO;
1281                 if (!bh)
1282                         goto cleanup;
1283                 if (ext4_xattr_check_block(inode, bh)) {
1284                         EXT4_ERROR_INODE(inode, "bad block %llu",
1285                                          EXT4_I(inode)->i_file_acl);
1286                         error = -EIO;
1287                         goto cleanup;
1288                 }
1289                 base = BHDR(bh);
1290                 first = BFIRST(bh);
1291                 end = bh->b_data + bh->b_size;
1292                 min_offs = end - base;
1293                 free = ext4_xattr_free_space(first, &min_offs, base,
1294                                              &total_blk);
1295                 if (free < new_extra_isize) {
1296                         if (!tried_min_extra_isize && s_min_extra_isize) {
1297                                 tried_min_extra_isize++;
1298                                 new_extra_isize = s_min_extra_isize;
1299                                 brelse(bh);
1300                                 goto retry;
1301                         }
1302                         error = -1;
1303                         goto cleanup;
1304                 }
1305         } else {
1306                 free = inode->i_sb->s_blocksize;
1307         }
1308
1309         while (new_extra_isize > 0) {
1310                 size_t offs, size, entry_size;
1311                 struct ext4_xattr_entry *small_entry = NULL;
1312                 struct ext4_xattr_info i = {
1313                         .value = NULL,
1314                         .value_len = 0,
1315                 };
1316                 unsigned int total_size;  /* EA entry size + value size */
1317                 unsigned int shift_bytes; /* No. of bytes to shift EAs by? */
1318                 unsigned int min_total_size = ~0U;
1319
1320                 is = kzalloc(sizeof(struct ext4_xattr_ibody_find), GFP_NOFS);
1321                 bs = kzalloc(sizeof(struct ext4_xattr_block_find), GFP_NOFS);
1322                 if (!is || !bs) {
1323                         error = -ENOMEM;
1324                         goto cleanup;
1325                 }
1326
1327                 is->s.not_found = -ENODATA;
1328                 bs->s.not_found = -ENODATA;
1329                 is->iloc.bh = NULL;
1330                 bs->bh = NULL;
1331
1332                 last = IFIRST(header);
1333                 /* Find the entry best suited to be pushed into EA block */
1334                 entry = NULL;
1335                 for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
1336                         total_size =
1337                         EXT4_XATTR_SIZE(le32_to_cpu(last->e_value_size)) +
1338                                         EXT4_XATTR_LEN(last->e_name_len);
1339                         if (total_size <= free && total_size < min_total_size) {
1340                                 if (total_size < new_extra_isize) {
1341                                         small_entry = last;
1342                                 } else {
1343                                         entry = last;
1344                                         min_total_size = total_size;
1345                                 }
1346                         }
1347                 }
1348
1349                 if (entry == NULL) {
1350                         if (small_entry) {
1351                                 entry = small_entry;
1352                         } else {
1353                                 if (!tried_min_extra_isize &&
1354                                     s_min_extra_isize) {
1355                                         tried_min_extra_isize++;
1356                                         new_extra_isize = s_min_extra_isize;
1357                                         goto retry;
1358                                 }
1359                                 error = -1;
1360                                 goto cleanup;
1361                         }
1362                 }
1363                 offs = le16_to_cpu(entry->e_value_offs);
1364                 size = le32_to_cpu(entry->e_value_size);
1365                 entry_size = EXT4_XATTR_LEN(entry->e_name_len);
1366                 i.name_index = entry->e_name_index,
1367                 buffer = kmalloc(EXT4_XATTR_SIZE(size), GFP_NOFS);
1368                 b_entry_name = kmalloc(entry->e_name_len + 1, GFP_NOFS);
1369                 if (!buffer || !b_entry_name) {
1370                         error = -ENOMEM;
1371                         goto cleanup;
1372                 }
1373                 /* Save the entry name and the entry value */
1374                 memcpy(buffer, (void *)IFIRST(header) + offs,
1375                        EXT4_XATTR_SIZE(size));
1376                 memcpy(b_entry_name, entry->e_name, entry->e_name_len);
1377                 b_entry_name[entry->e_name_len] = '\0';
1378                 i.name = b_entry_name;
1379
1380                 error = ext4_get_inode_loc(inode, &is->iloc);
1381                 if (error)
1382                         goto cleanup;
1383
1384                 error = ext4_xattr_ibody_find(inode, &i, is);
1385                 if (error)
1386                         goto cleanup;
1387
1388                 /* Remove the chosen entry from the inode */
1389                 error = ext4_xattr_ibody_set(handle, inode, &i, is);
1390                 if (error)
1391                         goto cleanup;
1392
1393                 entry = IFIRST(header);
1394                 if (entry_size + EXT4_XATTR_SIZE(size) >= new_extra_isize)
1395                         shift_bytes = new_extra_isize;
1396                 else
1397                         shift_bytes = entry_size + size;
1398                 /* Adjust the offsets and shift the remaining entries ahead */
1399                 ext4_xattr_shift_entries(entry, EXT4_I(inode)->i_extra_isize -
1400                         shift_bytes, (void *)raw_inode +
1401                         EXT4_GOOD_OLD_INODE_SIZE + extra_isize + shift_bytes,
1402                         (void *)header, total_ino - entry_size,
1403                         inode->i_sb->s_blocksize);
1404
1405                 extra_isize += shift_bytes;
1406                 new_extra_isize -= shift_bytes;
1407                 EXT4_I(inode)->i_extra_isize = extra_isize;
1408
1409                 i.name = b_entry_name;
1410                 i.value = buffer;
1411                 i.value_len = size;
1412                 error = ext4_xattr_block_find(inode, &i, bs);
1413                 if (error)
1414                         goto cleanup;
1415
1416                 /* Add entry which was removed from the inode into the block */
1417                 error = ext4_xattr_block_set(handle, inode, &i, bs);
1418                 if (error)
1419                         goto cleanup;
1420                 kfree(b_entry_name);
1421                 kfree(buffer);
1422                 b_entry_name = NULL;
1423                 buffer = NULL;
1424                 brelse(is->iloc.bh);
1425                 kfree(is);
1426                 kfree(bs);
1427         }
1428         brelse(bh);
1429         up_write(&EXT4_I(inode)->xattr_sem);
1430         return 0;
1431
1432 cleanup:
1433         kfree(b_entry_name);
1434         kfree(buffer);
1435         if (is)
1436                 brelse(is->iloc.bh);
1437         kfree(is);
1438         kfree(bs);
1439         brelse(bh);
1440         up_write(&EXT4_I(inode)->xattr_sem);
1441         return error;
1442 }
1443
1444
1445
1446 /*
1447  * ext4_xattr_delete_inode()
1448  *
1449  * Free extended attribute resources associated with this inode. This
1450  * is called immediately before an inode is freed. We have exclusive
1451  * access to the inode.
1452  */
1453 void
1454 ext4_xattr_delete_inode(handle_t *handle, struct inode *inode)
1455 {
1456         struct buffer_head *bh = NULL;
1457
1458         if (!EXT4_I(inode)->i_file_acl)
1459                 goto cleanup;
1460         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
1461         if (!bh) {
1462                 EXT4_ERROR_INODE(inode, "block %llu read error",
1463                                  EXT4_I(inode)->i_file_acl);
1464                 goto cleanup;
1465         }
1466         if (BHDR(bh)->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC) ||
1467             BHDR(bh)->h_blocks != cpu_to_le32(1)) {
1468                 EXT4_ERROR_INODE(inode, "bad block %llu",
1469                                  EXT4_I(inode)->i_file_acl);
1470                 goto cleanup;
1471         }
1472         ext4_xattr_release_block(handle, inode, bh);
1473         EXT4_I(inode)->i_file_acl = 0;
1474
1475 cleanup:
1476         brelse(bh);
1477 }
1478
1479 /*
1480  * ext4_xattr_put_super()
1481  *
1482  * This is called when a file system is unmounted.
1483  */
1484 void
1485 ext4_xattr_put_super(struct super_block *sb)
1486 {
1487         mb_cache_shrink(sb->s_bdev);
1488 }
1489
1490 /*
1491  * ext4_xattr_cache_insert()
1492  *
1493  * Create a new entry in the extended attribute cache, and insert
1494  * it unless such an entry is already in the cache.
1495  *
1496  * Returns 0, or a negative error number on failure.
1497  */
1498 static void
1499 ext4_xattr_cache_insert(struct buffer_head *bh)
1500 {
1501         __u32 hash = le32_to_cpu(BHDR(bh)->h_hash);
1502         struct mb_cache_entry *ce;
1503         int error;
1504
1505         ce = mb_cache_entry_alloc(ext4_xattr_cache, GFP_NOFS);
1506         if (!ce) {
1507                 ea_bdebug(bh, "out of memory");
1508                 return;
1509         }
1510         error = mb_cache_entry_insert(ce, bh->b_bdev, bh->b_blocknr, hash);
1511         if (error) {
1512                 mb_cache_entry_free(ce);
1513                 if (error == -EBUSY) {
1514                         ea_bdebug(bh, "already in cache");
1515                         error = 0;
1516                 }
1517         } else {
1518                 ea_bdebug(bh, "inserting [%x]", (int)hash);
1519                 mb_cache_entry_release(ce);
1520         }
1521 }
1522
1523 /*
1524  * ext4_xattr_cmp()
1525  *
1526  * Compare two extended attribute blocks for equality.
1527  *
1528  * Returns 0 if the blocks are equal, 1 if they differ, and
1529  * a negative error number on errors.
1530  */
1531 static int
1532 ext4_xattr_cmp(struct ext4_xattr_header *header1,
1533                struct ext4_xattr_header *header2)
1534 {
1535         struct ext4_xattr_entry *entry1, *entry2;
1536
1537         entry1 = ENTRY(header1+1);
1538         entry2 = ENTRY(header2+1);
1539         while (!IS_LAST_ENTRY(entry1)) {
1540                 if (IS_LAST_ENTRY(entry2))
1541                         return 1;
1542                 if (entry1->e_hash != entry2->e_hash ||
1543                     entry1->e_name_index != entry2->e_name_index ||
1544                     entry1->e_name_len != entry2->e_name_len ||
1545                     entry1->e_value_size != entry2->e_value_size ||
1546                     memcmp(entry1->e_name, entry2->e_name, entry1->e_name_len))
1547                         return 1;
1548                 if (entry1->e_value_block != 0 || entry2->e_value_block != 0)
1549                         return -EIO;
1550                 if (memcmp((char *)header1 + le16_to_cpu(entry1->e_value_offs),
1551                            (char *)header2 + le16_to_cpu(entry2->e_value_offs),
1552                            le32_to_cpu(entry1->e_value_size)))
1553                         return 1;
1554
1555                 entry1 = EXT4_XATTR_NEXT(entry1);
1556                 entry2 = EXT4_XATTR_NEXT(entry2);
1557         }
1558         if (!IS_LAST_ENTRY(entry2))
1559                 return 1;
1560         return 0;
1561 }
1562
1563 /*
1564  * ext4_xattr_cache_find()
1565  *
1566  * Find an identical extended attribute block.
1567  *
1568  * Returns a pointer to the block found, or NULL if such a block was
1569  * not found or an error occurred.
1570  */
1571 static struct buffer_head *
1572 ext4_xattr_cache_find(struct inode *inode, struct ext4_xattr_header *header,
1573                       struct mb_cache_entry **pce)
1574 {
1575         __u32 hash = le32_to_cpu(header->h_hash);
1576         struct mb_cache_entry *ce;
1577
1578         if (!header->h_hash)
1579                 return NULL;  /* never share */
1580         ea_idebug(inode, "looking for cached blocks [%x]", (int)hash);
1581 again:
1582         ce = mb_cache_entry_find_first(ext4_xattr_cache, inode->i_sb->s_bdev,
1583                                        hash);
1584         while (ce) {
1585                 struct buffer_head *bh;
1586
1587                 if (IS_ERR(ce)) {
1588                         if (PTR_ERR(ce) == -EAGAIN)
1589                                 goto again;
1590                         break;
1591                 }
1592                 bh = sb_bread(inode->i_sb, ce->e_block);
1593                 if (!bh) {
1594                         EXT4_ERROR_INODE(inode, "block %lu read error",
1595                                          (unsigned long) ce->e_block);
1596                 } else if (le32_to_cpu(BHDR(bh)->h_refcount) >=
1597                                 EXT4_XATTR_REFCOUNT_MAX) {
1598                         ea_idebug(inode, "block %lu refcount %d>=%d",
1599                                   (unsigned long) ce->e_block,
1600                                   le32_to_cpu(BHDR(bh)->h_refcount),
1601                                           EXT4_XATTR_REFCOUNT_MAX);
1602                 } else if (ext4_xattr_cmp(header, BHDR(bh)) == 0) {
1603                         *pce = ce;
1604                         return bh;
1605                 }
1606                 brelse(bh);
1607                 ce = mb_cache_entry_find_next(ce, inode->i_sb->s_bdev, hash);
1608         }
1609         return NULL;
1610 }
1611
1612 #define NAME_HASH_SHIFT 5
1613 #define VALUE_HASH_SHIFT 16
1614
1615 /*
1616  * ext4_xattr_hash_entry()
1617  *
1618  * Compute the hash of an extended attribute.
1619  */
1620 static inline void ext4_xattr_hash_entry(struct ext4_xattr_header *header,
1621                                          struct ext4_xattr_entry *entry)
1622 {
1623         __u32 hash = 0;
1624         char *name = entry->e_name;
1625         int n;
1626
1627         for (n = 0; n < entry->e_name_len; n++) {
1628                 hash = (hash << NAME_HASH_SHIFT) ^
1629                        (hash >> (8*sizeof(hash) - NAME_HASH_SHIFT)) ^
1630                        *name++;
1631         }
1632
1633         if (entry->e_value_block == 0 && entry->e_value_size != 0) {
1634                 __le32 *value = (__le32 *)((char *)header +
1635                         le16_to_cpu(entry->e_value_offs));
1636                 for (n = (le32_to_cpu(entry->e_value_size) +
1637                      EXT4_XATTR_ROUND) >> EXT4_XATTR_PAD_BITS; n; n--) {
1638                         hash = (hash << VALUE_HASH_SHIFT) ^
1639                                (hash >> (8*sizeof(hash) - VALUE_HASH_SHIFT)) ^
1640                                le32_to_cpu(*value++);
1641                 }
1642         }
1643         entry->e_hash = cpu_to_le32(hash);
1644 }
1645
1646 #undef NAME_HASH_SHIFT
1647 #undef VALUE_HASH_SHIFT
1648
1649 #define BLOCK_HASH_SHIFT 16
1650
1651 /*
1652  * ext4_xattr_rehash()
1653  *
1654  * Re-compute the extended attribute hash value after an entry has changed.
1655  */
1656 static void ext4_xattr_rehash(struct ext4_xattr_header *header,
1657                               struct ext4_xattr_entry *entry)
1658 {
1659         struct ext4_xattr_entry *here;
1660         __u32 hash = 0;
1661
1662         ext4_xattr_hash_entry(header, entry);
1663         here = ENTRY(header+1);
1664         while (!IS_LAST_ENTRY(here)) {
1665                 if (!here->e_hash) {
1666                         /* Block is not shared if an entry's hash value == 0 */
1667                         hash = 0;
1668                         break;
1669                 }
1670                 hash = (hash << BLOCK_HASH_SHIFT) ^
1671                        (hash >> (8*sizeof(hash) - BLOCK_HASH_SHIFT)) ^
1672                        le32_to_cpu(here->e_hash);
1673                 here = EXT4_XATTR_NEXT(here);
1674         }
1675         header->h_hash = cpu_to_le32(hash);
1676 }
1677
1678 #undef BLOCK_HASH_SHIFT
1679
1680 int __init
1681 ext4_init_xattr(void)
1682 {
1683         ext4_xattr_cache = mb_cache_create("ext4_xattr", 6);
1684         if (!ext4_xattr_cache)
1685                 return -ENOMEM;
1686         return 0;
1687 }
1688
1689 void
1690 ext4_exit_xattr(void)
1691 {
1692         if (ext4_xattr_cache)
1693                 mb_cache_destroy(ext4_xattr_cache);
1694         ext4_xattr_cache = NULL;
1695 }