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;
 126         __le32 save_csum;
 127         __le64 dsk_block_nr = cpu_to_le64(block_nr);
 128
 129         save_csum = hdr->h_checksum;
 130         hdr->h_checksum = 0;
 131         csum = ext4_chksum(sbi, sbi->s_csum_seed, (__u8 *)&dsk_block_nr,
 132                            sizeof(dsk_block_nr));
 133         csum = ext4_chksum(sbi, csum, (__u8 *)hdr,
 134                            EXT4_BLOCK_SIZE(inode->i_sb));
 135
 136         hdr->h_checksum = save_csum;
 137         return cpu_to_le32(csum);
 138 }
 139
 140 static int ext4_xattr_block_csum_verify(struct inode *inode,
 141                                         sector_t block_nr,
 142                                         struct ext4_xattr_header *hdr)
 143 {
 144         if (EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
 145                 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM) &&
 146             (hdr->h_checksum != ext4_xattr_block_csum(inode, block_nr, hdr)))
 147                 return 0;
 148         return 1;
 149 }
 150
 151 static void ext4_xattr_block_csum_set(struct inode *inode,
 152                                       sector_t block_nr,
 153                                       struct ext4_xattr_header *hdr)
 154 {
 155         if (!EXT4_HAS_RO_COMPAT_FEATURE(inode->i_sb,
 156                 EXT4_FEATURE_RO_COMPAT_METADATA_CSUM))
 157                 return;
 158
 159         hdr->h_checksum = ext4_xattr_block_csum(inode, block_nr, hdr);
 160 }
 161
 162 static inline int ext4_handle_dirty_xattr_block(handle_t *handle,
 163                                                 struct inode *inode,
 164                                                 struct buffer_head *bh)
 165 {
 166         ext4_xattr_block_csum_set(inode, bh->b_blocknr, BHDR(bh));
 167         return ext4_handle_dirty_metadata(handle, inode, bh);
 168 }
 169
 170 static inline const struct xattr_handler *
 171 ext4_xattr_handler(int name_index)
 172 {
 173         const struct xattr_handler *handler = NULL;
 174
 175         if (name_index > 0 && name_index < ARRAY_SIZE(ext4_xattr_handler_map))
 176                 handler = ext4_xattr_handler_map[name_index];
 177         return handler;
 178 }
 179
 180 /*
 181  * Inode operation listxattr()
 182  *
 183  * dentry->d_inode->i_mutex: don't care
 184  */
 185 ssize_t
 186 ext4_listxattr(struct dentry *dentry, char *buffer, size_t size)
 187 {
 188         return ext4_xattr_list(dentry, buffer, size);
 189 }
 190
 191 static int
 192 ext4_xattr_check_names(struct ext4_xattr_entry *entry, void *end,
 193                        void *value_start)
 194 {
 195         struct ext4_xattr_entry *e = entry;
 196
 197         while (!IS_LAST_ENTRY(e)) {
 198                 struct ext4_xattr_entry *next = EXT4_XATTR_NEXT(e);
 199                 if ((void *)next >= end)
 200                         return -EIO;
 201                 e = next;
 202         }
 203
 204         while (!IS_LAST_ENTRY(entry)) {
 205                 if (entry->e_value_size != 0 &&
 206                     (value_start + le16_to_cpu(entry->e_value_offs) <
 207                      (void *)e + sizeof(__u32) ||
 208                      value_start + le16_to_cpu(entry->e_value_offs) +
 209                     le32_to_cpu(entry->e_value_size) > end))
 210                         return -EIO;
 211                 entry = EXT4_XATTR_NEXT(entry);
 212         }
 213
 214         return 0;
 215 }
 216
 217 static inline int
 218 ext4_xattr_check_block(struct inode *inode, struct buffer_head *bh)
 219 {
 220         int error;
 221
 222         if (buffer_verified(bh))
 223                 return 0;
 224
 225         if (BHDR(bh)->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC) ||
 226             BHDR(bh)->h_blocks != cpu_to_le32(1))
 227                 return -EIO;
 228         if (!ext4_xattr_block_csum_verify(inode, bh->b_blocknr, BHDR(bh)))
 229                 return -EIO;
 230         error = ext4_xattr_check_names(BFIRST(bh), bh->b_data + bh->b_size,
 231                                        bh->b_data);
 232         if (!error)
 233                 set_buffer_verified(bh);
 234         return error;
 235 }
 236
 237 static inline int
 238 ext4_xattr_check_entry(struct ext4_xattr_entry *entry, size_t size)
 239 {
 240         size_t value_size = le32_to_cpu(entry->e_value_size);
 241
 242         if (entry->e_value_block != 0 || value_size > size ||
 243             le16_to_cpu(entry->e_value_offs) + value_size > size)
 244                 return -EIO;
 245         return 0;
 246 }
 247
 248 static int
 249 ext4_xattr_find_entry(struct ext4_xattr_entry **pentry, int name_index,
 250                       const char *name, size_t size, int sorted)
 251 {
 252         struct ext4_xattr_entry *entry;
 253         size_t name_len;
 254         int cmp = 1;
 255
 256         if (name == NULL)
 257                 return -EINVAL;
 258         name_len = strlen(name);
 259         entry = *pentry;
 260         for (; !IS_LAST_ENTRY(entry); entry = EXT4_XATTR_NEXT(entry)) {
 261                 cmp = name_index - entry->e_name_index;
 262                 if (!cmp)
 263                         cmp = name_len - entry->e_name_len;
 264                 if (!cmp)
 265                         cmp = memcmp(name, entry->e_name, name_len);
 266                 if (cmp <= 0 && (sorted || cmp == 0))
 267                         break;
 268         }
 269         *pentry = entry;
 270         if (!cmp && ext4_xattr_check_entry(entry, size))
 271                         return -EIO;
 272         return cmp ? -ENODATA : 0;
 273 }
 274
 275 static int
 276 ext4_xattr_block_get(struct inode *inode, int name_index, const char *name,
 277                      void *buffer, size_t buffer_size)
 278 {
 279         struct buffer_head *bh = NULL;
 280         struct ext4_xattr_entry *entry;
 281         size_t size;
 282         int error;
 283
 284         ea_idebug(inode, "name=%d.%s, buffer=%p, buffer_size=%ld",
 285                   name_index, name, buffer, (long)buffer_size);
 286
 287         error = -ENODATA;
 288         if (!EXT4_I(inode)->i_file_acl)
 289                 goto cleanup;
 290         ea_idebug(inode, "reading block %llu",
 291                   (unsigned long long)EXT4_I(inode)->i_file_acl);
 292         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
 293         if (!bh)
 294                 goto cleanup;
 295         ea_bdebug(bh, "b_count=%d, refcount=%d",
 296                 atomic_read(&(bh->b_count)), le32_to_cpu(BHDR(bh)->h_refcount));
 297         if (ext4_xattr_check_block(inode, bh)) {
 298 bad_block:
 299                 EXT4_ERROR_INODE(inode, "bad block %llu",
 300                                  EXT4_I(inode)->i_file_acl);
 301                 error = -EIO;
 302                 goto cleanup;
 303         }
 304         ext4_xattr_cache_insert(bh);
 305         entry = BFIRST(bh);
 306         error = ext4_xattr_find_entry(&entry, name_index, name, bh->b_size, 1);
 307         if (error == -EIO)
 308                 goto bad_block;
 309         if (error)
 310                 goto cleanup;
 311         size = le32_to_cpu(entry->e_value_size);
 312         if (buffer) {
 313                 error = -ERANGE;
 314                 if (size > buffer_size)
 315                         goto cleanup;
 316                 memcpy(buffer, bh->b_data + le16_to_cpu(entry->e_value_offs),
 317                        size);
 318         }
 319         error = size;
 320
 321 cleanup:
 322         brelse(bh);
 323         return error;
 324 }
 325
 326 int
 327 ext4_xattr_ibody_get(struct inode *inode, int name_index, const char *name,
 328                      void *buffer, size_t buffer_size)
 329 {
 330         struct ext4_xattr_ibody_header *header;
 331         struct ext4_xattr_entry *entry;
 332         struct ext4_inode *raw_inode;
 333         struct ext4_iloc iloc;
 334         size_t size;
 335         void *end;
 336         int error;
 337
 338         if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR))
 339                 return -ENODATA;
 340         error = ext4_get_inode_loc(inode, &iloc);
 341         if (error)
 342                 return error;
 343         raw_inode = ext4_raw_inode(&iloc);
 344         header = IHDR(inode, raw_inode);
 345         entry = IFIRST(header);
 346         end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
 347         error = ext4_xattr_check_names(entry, end, entry);
 348         if (error)
 349                 goto cleanup;
 350         error = ext4_xattr_find_entry(&entry, name_index, name,
 351                                       end - (void *)entry, 0);
 352         if (error)
 353                 goto cleanup;
 354         size = le32_to_cpu(entry->e_value_size);
 355         if (buffer) {
 356                 error = -ERANGE;
 357                 if (size > buffer_size)
 358                         goto cleanup;
 359                 memcpy(buffer, (void *)IFIRST(header) +
 360                        le16_to_cpu(entry->e_value_offs), size);
 361         }
 362         error = size;
 363
 364 cleanup:
 365         brelse(iloc.bh);
 366         return error;
 367 }
 368
 369 /*
 370  * ext4_xattr_get()
 371  *
 372  * Copy an extended attribute into the buffer
 373  * provided, or compute the buffer size required.
 374  * Buffer is NULL to compute the size of the buffer required.
 375  *
 376  * Returns a negative error number on failure, or the number of bytes
 377  * used / required on success.
 378  */
 379 int
 380 ext4_xattr_get(struct inode *inode, int name_index, const char *name,
 381                void *buffer, size_t buffer_size)
 382 {
 383         int error;
 384
 385         down_read(&EXT4_I(inode)->xattr_sem);
 386         error = ext4_xattr_ibody_get(inode, name_index, name, buffer,
 387                                      buffer_size);
 388         if (error == -ENODATA)
 389                 error = ext4_xattr_block_get(inode, name_index, name, buffer,
 390                                              buffer_size);
 391         up_read(&EXT4_I(inode)->xattr_sem);
 392         return error;
 393 }
 394
 395 static int
 396 ext4_xattr_list_entries(struct dentry *dentry, struct ext4_xattr_entry *entry,
 397                         char *buffer, size_t buffer_size)
 398 {
 399         size_t rest = buffer_size;
 400
 401         for (; !IS_LAST_ENTRY(entry); entry = EXT4_XATTR_NEXT(entry)) {
 402                 const struct xattr_handler *handler =
 403                         ext4_xattr_handler(entry->e_name_index);
 404
 405                 if (handler) {
 406                         size_t size = handler->list(dentry, buffer, rest,
 407                                                     entry->e_name,
 408                                                     entry->e_name_len,
 409                                                     handler->flags);
 410                         if (buffer) {
 411                                 if (size > rest)
 412                                         return -ERANGE;
 413                                 buffer += size;
 414                         }
 415                         rest -= size;
 416                 }
 417         }
 418         return buffer_size - rest;
 419 }
 420
 421 static int
 422 ext4_xattr_block_list(struct dentry *dentry, char *buffer, size_t buffer_size)
 423 {
 424         struct inode *inode = dentry->d_inode;
 425         struct buffer_head *bh = NULL;
 426         int error;
 427
 428         ea_idebug(inode, "buffer=%p, buffer_size=%ld",
 429                   buffer, (long)buffer_size);
 430
 431         error = 0;
 432         if (!EXT4_I(inode)->i_file_acl)
 433                 goto cleanup;
 434         ea_idebug(inode, "reading block %llu",
 435                   (unsigned long long)EXT4_I(inode)->i_file_acl);
 436         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
 437         error = -EIO;
 438         if (!bh)
 439                 goto cleanup;
 440         ea_bdebug(bh, "b_count=%d, refcount=%d",
 441                 atomic_read(&(bh->b_count)), le32_to_cpu(BHDR(bh)->h_refcount));
 442         if (ext4_xattr_check_block(inode, bh)) {
 443                 EXT4_ERROR_INODE(inode, "bad block %llu",
 444                                  EXT4_I(inode)->i_file_acl);
 445                 error = -EIO;
 446                 goto cleanup;
 447         }
 448         ext4_xattr_cache_insert(bh);
 449         error = ext4_xattr_list_entries(dentry, BFIRST(bh), buffer, buffer_size);
 450
 451 cleanup:
 452         brelse(bh);
 453
 454         return error;
 455 }
 456
 457 static int
 458 ext4_xattr_ibody_list(struct dentry *dentry, char *buffer, size_t buffer_size)
 459 {
 460         struct inode *inode = dentry->d_inode;
 461         struct ext4_xattr_ibody_header *header;
 462         struct ext4_inode *raw_inode;
 463         struct ext4_iloc iloc;
 464         void *end;
 465         int error;
 466
 467         if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR))
 468                 return 0;
 469         error = ext4_get_inode_loc(inode, &iloc);
 470         if (error)
 471                 return error;
 472         raw_inode = ext4_raw_inode(&iloc);
 473         header = IHDR(inode, raw_inode);
 474         end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
 475         error = ext4_xattr_check_names(IFIRST(header), end, IFIRST(header));
 476         if (error)
 477                 goto cleanup;
 478         error = ext4_xattr_list_entries(dentry, IFIRST(header),
 479                                         buffer, buffer_size);
 480
 481 cleanup:
 482         brelse(iloc.bh);
 483         return error;
 484 }
 485
 486 /*
 487  * ext4_xattr_list()
 488  *
 489  * Copy a list of attribute names into the buffer
 490  * provided, or compute the buffer size required.
 491  * Buffer is NULL to compute the size of the buffer required.
 492  *
 493  * Returns a negative error number on failure, or the number of bytes
 494  * used / required on success.
 495  */
 496 static int
 497 ext4_xattr_list(struct dentry *dentry, char *buffer, size_t buffer_size)
 498 {
 499         int ret, ret2;
 500
 501         down_read(&EXT4_I(dentry->d_inode)->xattr_sem);
 502         ret = ret2 = ext4_xattr_ibody_list(dentry, buffer, buffer_size);
 503         if (ret < 0)
 504                 goto errout;
 505         if (buffer) {
 506                 buffer += ret;
 507                 buffer_size -= ret;
 508         }
 509         ret = ext4_xattr_block_list(dentry, buffer, buffer_size);
 510         if (ret < 0)
 511                 goto errout;
 512         ret += ret2;
 513 errout:
 514         up_read(&EXT4_I(dentry->d_inode)->xattr_sem);
 515         return ret;
 516 }
 517
 518 /*
 519  * If the EXT4_FEATURE_COMPAT_EXT_ATTR feature of this file system is
 520  * not set, set it.
 521  */
 522 static void ext4_xattr_update_super_block(handle_t *handle,
 523                                           struct super_block *sb)
 524 {
 525         if (EXT4_HAS_COMPAT_FEATURE(sb, EXT4_FEATURE_COMPAT_EXT_ATTR))
 526                 return;
 527
 528         if (ext4_journal_get_write_access(handle, EXT4_SB(sb)->s_sbh) == 0) {
 529                 EXT4_SET_COMPAT_FEATURE(sb, EXT4_FEATURE_COMPAT_EXT_ATTR);
 530                 ext4_handle_dirty_super(handle, sb);
 531         }
 532 }
 533
 534 /*
 535  * Release the xattr block BH: If the reference count is > 1, decrement it;
 536  * otherwise free the block.
 537  */
 538 static void
 539 ext4_xattr_release_block(handle_t *handle, struct inode *inode,
 540                          struct buffer_head *bh)
 541 {
 542         struct mb_cache_entry *ce = NULL;
 543         int error = 0;
 544
 545         ce = mb_cache_entry_get(ext4_xattr_cache, bh->b_bdev, bh->b_blocknr);
 546         error = ext4_journal_get_write_access(handle, bh);
 547         if (error)
 548                 goto out;
 549
 550         lock_buffer(bh);
 551         if (BHDR(bh)->h_refcount == cpu_to_le32(1)) {
 552                 ea_bdebug(bh, "refcount now=0; freeing");
 553                 if (ce)
 554                         mb_cache_entry_free(ce);
 555                 get_bh(bh);
 556                 unlock_buffer(bh);
 557                 ext4_free_blocks(handle, inode, bh, 0, 1,
 558                                  EXT4_FREE_BLOCKS_METADATA |
 559                                  EXT4_FREE_BLOCKS_FORGET);
 560         } else {
 561                 le32_add_cpu(&BHDR(bh)->h_refcount, -1);
 562                 if (ce)
 563                         mb_cache_entry_release(ce);
 564                 /*
 565                  * Beware of this ugliness: Releasing of xattr block references
 566                  * from different inodes can race and so we have to protect
 567                  * from a race where someone else frees the block (and releases
 568                  * its journal_head) before we are done dirtying the buffer. In
 569                  * nojournal mode this race is harmless and we actually cannot
 570                  * call ext4_handle_dirty_xattr_block() with locked buffer as
 571                  * that function can call sync_dirty_buffer() so for that case
 572                  * we handle the dirtying after unlocking the buffer.
 573                  */
 574                 if (ext4_handle_valid(handle))
 575                         error = ext4_handle_dirty_xattr_block(handle, inode,
 576                                                               bh);
 577                 unlock_buffer(bh);
 578                 if (!ext4_handle_valid(handle))
 579                         error = ext4_handle_dirty_xattr_block(handle, inode,
 580                                                               bh);
 581                 if (IS_SYNC(inode))
 582                         ext4_handle_sync(handle);
 583                 dquot_free_block(inode, EXT4_C2B(EXT4_SB(inode->i_sb), 1));
 584                 ea_bdebug(bh, "refcount now=%d; releasing",
 585                           le32_to_cpu(BHDR(bh)->h_refcount));
 586         }
 587 out:
 588         ext4_std_error(inode->i_sb, error);
 589         return;
 590 }
 591
 592 /*
 593  * Find the available free space for EAs. This also returns the total number of
 594  * bytes used by EA entries.
 595  */
 596 static size_t ext4_xattr_free_space(struct ext4_xattr_entry *last,
 597                                     size_t *min_offs, void *base, int *total)
 598 {
 599         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 600                 *total += EXT4_XATTR_LEN(last->e_name_len);
 601                 if (!last->e_value_block && last->e_value_size) {
 602                         size_t offs = le16_to_cpu(last->e_value_offs);
 603                         if (offs < *min_offs)
 604                                 *min_offs = offs;
 605                 }
 606         }
 607         return (*min_offs - ((void *)last - base) - sizeof(__u32));
 608 }
 609
 610 static int
 611 ext4_xattr_set_entry(struct ext4_xattr_info *i, struct ext4_xattr_search *s)
 612 {
 613         struct ext4_xattr_entry *last;
 614         size_t free, min_offs = s->end - s->base, name_len = strlen(i->name);
 615
 616         /* Compute min_offs and last. */
 617         last = s->first;
 618         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 619                 if (!last->e_value_block && last->e_value_size) {
 620                         size_t offs = le16_to_cpu(last->e_value_offs);
 621                         if (offs < min_offs)
 622                                 min_offs = offs;
 623                 }
 624         }
 625         free = min_offs - ((void *)last - s->base) - sizeof(__u32);
 626         if (!s->not_found) {
 627                 if (!s->here->e_value_block && s->here->e_value_size) {
 628                         size_t size = le32_to_cpu(s->here->e_value_size);
 629                         free += EXT4_XATTR_SIZE(size);
 630                 }
 631                 free += EXT4_XATTR_LEN(name_len);
 632         }
 633         if (i->value) {
 634                 if (free < EXT4_XATTR_SIZE(i->value_len) ||
 635                     free < EXT4_XATTR_LEN(name_len) +
 636                            EXT4_XATTR_SIZE(i->value_len))
 637                         return -ENOSPC;
 638         }
 639
 640         if (i->value && s->not_found) {
 641                 /* Insert the new name. */
 642                 size_t size = EXT4_XATTR_LEN(name_len);
 643                 size_t rest = (void *)last - (void *)s->here + sizeof(__u32);
 644                 memmove((void *)s->here + size, s->here, rest);
 645                 memset(s->here, 0, size);
 646                 s->here->e_name_index = i->name_index;
 647                 s->here->e_name_len = name_len;
 648                 memcpy(s->here->e_name, i->name, name_len);
 649         } else {
 650                 if (!s->here->e_value_block && s->here->e_value_size) {
 651                         void *first_val = s->base + min_offs;
 652                         size_t offs = le16_to_cpu(s->here->e_value_offs);
 653                         void *val = s->base + offs;
 654                         size_t size = EXT4_XATTR_SIZE(
 655                                 le32_to_cpu(s->here->e_value_size));
 656
 657                         if (i->value && size == EXT4_XATTR_SIZE(i->value_len)) {
 658                                 /* The old and the new value have the same
 659                                    size. Just replace. */
 660                                 s->here->e_value_size =
 661                                         cpu_to_le32(i->value_len);
 662                                 if (i->value == EXT4_ZERO_XATTR_VALUE) {
 663                                         memset(val, 0, size);
 664                                 } else {
 665                                         /* Clear pad bytes first. */
 666                                         memset(val + size - EXT4_XATTR_PAD, 0,
 667                                                EXT4_XATTR_PAD);
 668                                         memcpy(val, i->value, i->value_len);
 669                                 }
 670                                 return 0;
 671                         }
 672
 673                         /* Remove the old value. */
 674                         memmove(first_val + size, first_val, val - first_val);
 675                         memset(first_val, 0, size);
 676                         s->here->e_value_size = 0;
 677                         s->here->e_value_offs = 0;
 678                         min_offs += size;
 679
 680                         /* Adjust all value offsets. */
 681                         last = s->first;
 682                         while (!IS_LAST_ENTRY(last)) {
 683                                 size_t o = le16_to_cpu(last->e_value_offs);
 684                                 if (!last->e_value_block &&
 685                                     last->e_value_size && o < offs)
 686                                         last->e_value_offs =
 687                                                 cpu_to_le16(o + size);
 688                                 last = EXT4_XATTR_NEXT(last);
 689                         }
 690                 }
 691                 if (!i->value) {
 692                         /* Remove the old name. */
 693                         size_t size = EXT4_XATTR_LEN(name_len);
 694                         last = ENTRY((void *)last - size);
 695                         memmove(s->here, (void *)s->here + size,
 696                                 (void *)last - (void *)s->here + sizeof(__u32));
 697                         memset(last, 0, size);
 698                 }
 699         }
 700
 701         if (i->value) {
 702                 /* Insert the new value. */
 703                 s->here->e_value_size = cpu_to_le32(i->value_len);
 704                 if (i->value_len) {
 705                         size_t size = EXT4_XATTR_SIZE(i->value_len);
 706                         void *val = s->base + min_offs - size;
 707                         s->here->e_value_offs = cpu_to_le16(min_offs - size);
 708                         if (i->value == EXT4_ZERO_XATTR_VALUE) {
 709                                 memset(val, 0, size);
 710                         } else {
 711                                 /* Clear the pad bytes first. */
 712                                 memset(val + size - EXT4_XATTR_PAD, 0,
 713                                        EXT4_XATTR_PAD);
 714                                 memcpy(val, i->value, i->value_len);
 715                         }
 716                 }
 717         }
 718         return 0;
 719 }
 720
 721 struct ext4_xattr_block_find {
 722         struct ext4_xattr_search s;
 723         struct buffer_head *bh;
 724 };
 725
 726 static int
 727 ext4_xattr_block_find(struct inode *inode, struct ext4_xattr_info *i,
 728                       struct ext4_xattr_block_find *bs)
 729 {
 730         struct super_block *sb = inode->i_sb;
 731         int error;
 732
 733         ea_idebug(inode, "name=%d.%s, value=%p, value_len=%ld",
 734                   i->name_index, i->name, i->value, (long)i->value_len);
 735
 736         if (EXT4_I(inode)->i_file_acl) {
 737                 /* The inode already has an extended attribute block. */
 738                 bs->bh = sb_bread(sb, EXT4_I(inode)->i_file_acl);
 739                 error = -EIO;
 740                 if (!bs->bh)
 741                         goto cleanup;
 742                 ea_bdebug(bs->bh, "b_count=%d, refcount=%d",
 743                         atomic_read(&(bs->bh->b_count)),
 744                         le32_to_cpu(BHDR(bs->bh)->h_refcount));
 745                 if (ext4_xattr_check_block(inode, bs->bh)) {
 746                         EXT4_ERROR_INODE(inode, "bad block %llu",
 747                                          EXT4_I(inode)->i_file_acl);
 748                         error = -EIO;
 749                         goto cleanup;
 750                 }
 751                 /* Find the named attribute. */
 752                 bs->s.base = BHDR(bs->bh);
 753                 bs->s.first = BFIRST(bs->bh);
 754                 bs->s.end = bs->bh->b_data + bs->bh->b_size;
 755                 bs->s.here = bs->s.first;
 756                 error = ext4_xattr_find_entry(&bs->s.here, i->name_index,
 757                                               i->name, bs->bh->b_size, 1);
 758                 if (error && error != -ENODATA)
 759                         goto cleanup;
 760                 bs->s.not_found = error;
 761         }
 762         error = 0;
 763
 764 cleanup:
 765         return error;
 766 }
 767
 768 static int
 769 ext4_xattr_block_set(handle_t *handle, struct inode *inode,
 770                      struct ext4_xattr_info *i,
 771                      struct ext4_xattr_block_find *bs)
 772 {
 773         struct super_block *sb = inode->i_sb;
 774         struct buffer_head *new_bh = NULL;
 775         struct ext4_xattr_search *s = &bs->s;
 776         struct mb_cache_entry *ce = NULL;
 777         int error = 0;
 778
 779 #define header(x) ((struct ext4_xattr_header *)(x))
 780
 781         if (i->value && i->value_len > sb->s_blocksize)
 782                 return -ENOSPC;
 783         if (s->base) {
 784                 ce = mb_cache_entry_get(ext4_xattr_cache, bs->bh->b_bdev,
 785                                         bs->bh->b_blocknr);
 786                 error = ext4_journal_get_write_access(handle, bs->bh);
 787                 if (error)
 788                         goto cleanup;
 789                 lock_buffer(bs->bh);
 790
 791                 if (header(s->base)->h_refcount == cpu_to_le32(1)) {
 792                         if (ce) {
 793                                 mb_cache_entry_free(ce);
 794                                 ce = NULL;
 795                         }
 796                         ea_bdebug(bs->bh, "modifying in-place");
 797                         error = ext4_xattr_set_entry(i, s);
 798                         if (!error) {
 799                                 if (!IS_LAST_ENTRY(s->first))
 800                                         ext4_xattr_rehash(header(s->base),
 801                                                           s->here);
 802                                 ext4_xattr_cache_insert(bs->bh);
 803                         }
 804                         unlock_buffer(bs->bh);
 805                         if (error == -EIO)
 806                                 goto bad_block;
 807                         if (!error)
 808                                 error = ext4_handle_dirty_xattr_block(handle,
 809                                                                       inode,
 810                                                                       bs->bh);
 811                         if (error)
 812                                 goto cleanup;
 813                         goto inserted;
 814                 } else {
 815                         int offset = (char *)s->here - bs->bh->b_data;
 816
 817                         unlock_buffer(bs->bh);
 818                         if (ce) {
 819                                 mb_cache_entry_release(ce);
 820                                 ce = NULL;
 821                         }
 822                         ea_bdebug(bs->bh, "cloning");
 823                         s->base = kmalloc(bs->bh->b_size, GFP_NOFS);
 824                         error = -ENOMEM;
 825                         if (s->base == NULL)
 826                                 goto cleanup;
 827                         memcpy(s->base, BHDR(bs->bh), bs->bh->b_size);
 828                         s->first = ENTRY(header(s->base)+1);
 829                         header(s->base)->h_refcount = cpu_to_le32(1);
 830                         s->here = ENTRY(s->base + offset);
 831                         s->end = s->base + bs->bh->b_size;
 832                 }
 833         } else {
 834                 /* Allocate a buffer where we construct the new block. */
 835                 s->base = kzalloc(sb->s_blocksize, GFP_NOFS);
 836                 /* assert(header == s->base) */
 837                 error = -ENOMEM;
 838                 if (s->base == NULL)
 839                         goto cleanup;
 840                 header(s->base)->h_magic = cpu_to_le32(EXT4_XATTR_MAGIC);
 841                 header(s->base)->h_blocks = cpu_to_le32(1);
 842                 header(s->base)->h_refcount = cpu_to_le32(1);
 843                 s->first = ENTRY(header(s->base)+1);
 844                 s->here = ENTRY(header(s->base)+1);
 845                 s->end = s->base + sb->s_blocksize;
 846         }
 847
 848         error = ext4_xattr_set_entry(i, s);
 849         if (error == -EIO)
 850                 goto bad_block;
 851         if (error)
 852                 goto cleanup;
 853         if (!IS_LAST_ENTRY(s->first))
 854                 ext4_xattr_rehash(header(s->base), s->here);
 855
 856 inserted:
 857         if (!IS_LAST_ENTRY(s->first)) {
 858                 new_bh = ext4_xattr_cache_find(inode, header(s->base), &ce);
 859                 if (new_bh) {
 860                         /* We found an identical block in the cache. */
 861                         if (new_bh == bs->bh)
 862                                 ea_bdebug(new_bh, "keeping");
 863                         else {
 864                                 /* The old block is released after updating
 865                                    the inode. */
 866                                 error = dquot_alloc_block(inode,
 867                                                 EXT4_C2B(EXT4_SB(sb), 1));
 868                                 if (error)
 869                                         goto cleanup;
 870                                 error = ext4_journal_get_write_access(handle,
 871                                                                       new_bh);
 872                                 if (error)
 873                                         goto cleanup_dquot;
 874                                 lock_buffer(new_bh);
 875                                 le32_add_cpu(&BHDR(new_bh)->h_refcount, 1);
 876                                 ea_bdebug(new_bh, "reusing; refcount now=%d",
 877                                         le32_to_cpu(BHDR(new_bh)->h_refcount));
 878                                 unlock_buffer(new_bh);
 879                                 error = ext4_handle_dirty_xattr_block(handle,
 880                                                                       inode,
 881                                                                       new_bh);
 882                                 if (error)
 883                                         goto cleanup_dquot;
 884                         }
 885                         mb_cache_entry_release(ce);
 886                         ce = NULL;
 887                 } else if (bs->bh && s->base == bs->bh->b_data) {
 888                         /* We were modifying this block in-place. */
 889                         ea_bdebug(bs->bh, "keeping this block");
 890                         new_bh = bs->bh;
 891                         get_bh(new_bh);
 892                 } else {
 893                         /* We need to allocate a new block */
 894                         ext4_fsblk_t goal, block;
 895
 896                         goal = ext4_group_first_block_no(sb,
 897                                                 EXT4_I(inode)->i_block_group);
 898
 899                         /* non-extent files can't have physical blocks past 2^32 */
 900                         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
 901                                 goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
 902
 903                         /*
 904                          * take i_data_sem because we will test
 905                          * i_delalloc_reserved_flag in ext4_mb_new_blocks
 906                          */
 907                         down_read((&EXT4_I(inode)->i_data_sem));
 908                         block = ext4_new_meta_blocks(handle, inode, goal, 0,
 909                                                      NULL, &error);
 910                         up_read((&EXT4_I(inode)->i_data_sem));
 911                         if (error)
 912                                 goto cleanup;
 913
 914                         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
 915                                 BUG_ON(block > EXT4_MAX_BLOCK_FILE_PHYS);
 916
 917                         ea_idebug(inode, "creating block %llu",
 918                                   (unsigned long long)block);
 919
 920                         new_bh = sb_getblk(sb, block);
 921                         if (unlikely(!new_bh)) {
 922                                 error = -ENOMEM;
 923 getblk_failed:
 924                                 ext4_free_blocks(handle, inode, NULL, block, 1,
 925                                                  EXT4_FREE_BLOCKS_METADATA);
 926                                 goto cleanup;
 927                         }
 928                         lock_buffer(new_bh);
 929                         error = ext4_journal_get_create_access(handle, new_bh);
 930                         if (error) {
 931                                 unlock_buffer(new_bh);
 932                                 error = -EIO;
 933                                 goto getblk_failed;
 934                         }
 935                         memcpy(new_bh->b_data, s->base, new_bh->b_size);
 936                         set_buffer_uptodate(new_bh);
 937                         unlock_buffer(new_bh);
 938                         ext4_xattr_cache_insert(new_bh);
 939                         error = ext4_handle_dirty_xattr_block(handle,
 940                                                               inode, new_bh);
 941                         if (error)
 942                                 goto cleanup;
 943                 }
 944         }
 945
 946         /* Update the inode. */
 947         EXT4_I(inode)->i_file_acl = new_bh ? new_bh->b_blocknr : 0;
 948
 949         /* Drop the previous xattr block. */
 950         if (bs->bh && bs->bh != new_bh)
 951                 ext4_xattr_release_block(handle, inode, bs->bh);
 952         error = 0;
 953
 954 cleanup:
 955         if (ce)
 956                 mb_cache_entry_release(ce);
 957         brelse(new_bh);
 958         if (!(bs->bh && s->base == bs->bh->b_data))
 959                 kfree(s->base);
 960
 961         return error;
 962
 963 cleanup_dquot:
 964         dquot_free_block(inode, EXT4_C2B(EXT4_SB(sb), 1));
 965         goto cleanup;
 966
 967 bad_block:
 968         EXT4_ERROR_INODE(inode, "bad block %llu",
 969                          EXT4_I(inode)->i_file_acl);
 970         goto cleanup;
 971
 972 #undef header
 973 }
 974
 975 int ext4_xattr_ibody_find(struct inode *inode, struct ext4_xattr_info *i,
 976                           struct ext4_xattr_ibody_find *is)
 977 {
 978         struct ext4_xattr_ibody_header *header;
 979         struct ext4_inode *raw_inode;
 980         int error;
 981
 982         if (EXT4_I(inode)->i_extra_isize == 0)
 983                 return 0;
 984         raw_inode = ext4_raw_inode(&is->iloc);
 985         header = IHDR(inode, raw_inode);
 986         is->s.base = is->s.first = IFIRST(header);
 987         is->s.here = is->s.first;
 988         is->s.end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
 989         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
 990                 error = ext4_xattr_check_names(IFIRST(header), is->s.end,
 991                                                IFIRST(header));
 992                 if (error)
 993                         return error;
 994                 /* Find the named attribute. */
 995                 error = ext4_xattr_find_entry(&is->s.here, i->name_index,
 996                                               i->name, is->s.end -
 997                                               (void *)is->s.base, 0);
 998                 if (error && error != -ENODATA)
 999                         return error;
1000                 is->s.not_found = error;
1001         }
1002         return 0;
1003 }
1004
1005 int ext4_xattr_ibody_inline_set(handle_t *handle, struct inode *inode,
1006                                 struct ext4_xattr_info *i,
1007                                 struct ext4_xattr_ibody_find *is)
1008 {
1009         struct ext4_xattr_ibody_header *header;
1010         struct ext4_xattr_search *s = &is->s;
1011         int error;
1012
1013         if (EXT4_I(inode)->i_extra_isize == 0)
1014                 return -ENOSPC;
1015         error = ext4_xattr_set_entry(i, s);
1016         if (error) {
1017                 if (error == -ENOSPC &&
1018                     ext4_has_inline_data(inode)) {
1019                         error = ext4_try_to_evict_inline_data(handle, inode,
1020                                         EXT4_XATTR_LEN(strlen(i->name) +
1021                                         EXT4_XATTR_SIZE(i->value_len)));
1022                         if (error)
1023                                 return error;
1024                         error = ext4_xattr_ibody_find(inode, i, is);
1025                         if (error)
1026                                 return error;
1027                         error = ext4_xattr_set_entry(i, s);
1028                 }
1029                 if (error)
1030                         return error;
1031         }
1032         header = IHDR(inode, ext4_raw_inode(&is->iloc));
1033         if (!IS_LAST_ENTRY(s->first)) {
1034                 header->h_magic = cpu_to_le32(EXT4_XATTR_MAGIC);
1035                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
1036         } else {
1037                 header->h_magic = cpu_to_le32(0);
1038                 ext4_clear_inode_state(inode, EXT4_STATE_XATTR);
1039         }
1040         return 0;
1041 }
1042
1043 static int ext4_xattr_ibody_set(handle_t *handle, struct inode *inode,
1044                                 struct ext4_xattr_info *i,
1045                                 struct ext4_xattr_ibody_find *is)
1046 {
1047         struct ext4_xattr_ibody_header *header;
1048         struct ext4_xattr_search *s = &is->s;
1049         int error;
1050
1051         if (EXT4_I(inode)->i_extra_isize == 0)
1052                 return -ENOSPC;
1053         error = ext4_xattr_set_entry(i, s);
1054         if (error)
1055                 return error;
1056         header = IHDR(inode, ext4_raw_inode(&is->iloc));
1057         if (!IS_LAST_ENTRY(s->first)) {
1058                 header->h_magic = cpu_to_le32(EXT4_XATTR_MAGIC);
1059                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
1060         } else {
1061                 header->h_magic = cpu_to_le32(0);
1062                 ext4_clear_inode_state(inode, EXT4_STATE_XATTR);
1063         }
1064         return 0;
1065 }
1066
1067 /*
1068  * ext4_xattr_set_handle()
1069  *
1070  * Create, replace or remove an extended attribute for this inode.  Value
1071  * is NULL to remove an existing extended attribute, and non-NULL to
1072  * either replace an existing extended attribute, or create a new extended
1073  * attribute. The flags XATTR_REPLACE and XATTR_CREATE
1074  * specify that an extended attribute must exist and must not exist
1075  * previous to the call, respectively.
1076  *
1077  * Returns 0, or a negative error number on failure.
1078  */
1079 int
1080 ext4_xattr_set_handle(handle_t *handle, struct inode *inode, int name_index,
1081                       const char *name, const void *value, size_t value_len,
1082                       int flags)
1083 {
1084         struct ext4_xattr_info i = {
1085                 .name_index = name_index,
1086                 .name = name,
1087                 .value = value,
1088                 .value_len = value_len,
1089
1090         };
1091         struct ext4_xattr_ibody_find is = {
1092                 .s = { .not_found = -ENODATA, },
1093         };
1094         struct ext4_xattr_block_find bs = {
1095                 .s = { .not_found = -ENODATA, },
1096         };
1097         unsigned long no_expand;
1098         int error;
1099
1100         if (!name)
1101                 return -EINVAL;
1102         if (strlen(name) > 255)
1103                 return -ERANGE;
1104         down_write(&EXT4_I(inode)->xattr_sem);
1105         no_expand = ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND);
1106         ext4_set_inode_state(inode, EXT4_STATE_NO_EXPAND);
1107
1108         error = ext4_reserve_inode_write(handle, inode, &is.iloc);
1109         if (error)
1110                 goto cleanup;
1111
1112         if (ext4_test_inode_state(inode, EXT4_STATE_NEW)) {
1113                 struct ext4_inode *raw_inode = ext4_raw_inode(&is.iloc);
1114                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
1115                 ext4_clear_inode_state(inode, EXT4_STATE_NEW);
1116         }
1117
1118         error = ext4_xattr_ibody_find(inode, &i, &is);
1119         if (error)
1120                 goto cleanup;
1121         if (is.s.not_found)
1122                 error = ext4_xattr_block_find(inode, &i, &bs);
1123         if (error)
1124                 goto cleanup;
1125         if (is.s.not_found && bs.s.not_found) {
1126                 error = -ENODATA;
1127                 if (flags & XATTR_REPLACE)
1128                         goto cleanup;
1129                 error = 0;
1130                 if (!value)
1131                         goto cleanup;
1132         } else {
1133                 error = -EEXIST;
1134                 if (flags & XATTR_CREATE)
1135                         goto cleanup;
1136         }
1137         if (!value) {
1138                 if (!is.s.not_found)
1139                         error = ext4_xattr_ibody_set(handle, inode, &i, &is);
1140                 else if (!bs.s.not_found)
1141                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
1142         } else {
1143                 error = ext4_xattr_ibody_set(handle, inode, &i, &is);
1144                 if (!error && !bs.s.not_found) {
1145                         i.value = NULL;
1146                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
1147                 } else if (error == -ENOSPC) {
1148                         if (EXT4_I(inode)->i_file_acl && !bs.s.base) {
1149                                 error = ext4_xattr_block_find(inode, &i, &bs);
1150                                 if (error)
1151                                         goto cleanup;
1152                         }
1153                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
1154                         if (error)
1155                                 goto cleanup;
1156                         if (!is.s.not_found) {
1157                                 i.value = NULL;
1158                                 error = ext4_xattr_ibody_set(handle, inode, &i,
1159                                                              &is);
1160                         }
1161                 }
1162         }
1163         if (!error) {
1164                 ext4_xattr_update_super_block(handle, inode->i_sb);
1165                 inode->i_ctime = ext4_current_time(inode);
1166                 if (!value)
1167                         ext4_clear_inode_state(inode, EXT4_STATE_NO_EXPAND);
1168                 error = ext4_mark_iloc_dirty(handle, inode, &is.iloc);
1169                 /*
1170                  * The bh is consumed by ext4_mark_iloc_dirty, even with
1171                  * error != 0.
1172                  */
1173                 is.iloc.bh = NULL;
1174                 if (IS_SYNC(inode))
1175                         ext4_handle_sync(handle);
1176         }
1177
1178 cleanup:
1179         brelse(is.iloc.bh);
1180         brelse(bs.bh);
1181         if (no_expand == 0)
1182                 ext4_clear_inode_state(inode, EXT4_STATE_NO_EXPAND);
1183         up_write(&EXT4_I(inode)->xattr_sem);
1184         return error;
1185 }
1186
1187 /*
1188  * ext4_xattr_set()
1189  *
1190  * Like ext4_xattr_set_handle, but start from an inode. This extended
1191  * attribute modification is a filesystem transaction by itself.
1192  *
1193  * Returns 0, or a negative error number on failure.
1194  */
1195 int
1196 ext4_xattr_set(struct inode *inode, int name_index, const char *name,
1197                const void *value, size_t value_len, int flags)
1198 {
1199         handle_t *handle;
1200         int error, retries = 0;
1201         int credits = ext4_jbd2_credits_xattr(inode);
1202
1203 retry:
1204         handle = ext4_journal_start(inode, EXT4_HT_XATTR, credits);
1205         if (IS_ERR(handle)) {
1206                 error = PTR_ERR(handle);
1207         } else {
1208                 int error2;
1209
1210                 error = ext4_xattr_set_handle(handle, inode, name_index, name,
1211                                               value, value_len, flags);
1212                 error2 = ext4_journal_stop(handle);
1213                 if (error == -ENOSPC &&
1214                     ext4_should_retry_alloc(inode->i_sb, &retries))
1215                         goto retry;
1216                 if (error == 0)
1217                         error = error2;
1218         }
1219
1220         return error;
1221 }
1222
1223 /*
1224  * Shift the EA entries in the inode to create space for the increased
1225  * i_extra_isize.
1226  */
1227 static void ext4_xattr_shift_entries(struct ext4_xattr_entry *entry,
1228                                      int value_offs_shift, void *to,
1229                                      void *from, size_t n, int blocksize)
1230 {
1231         struct ext4_xattr_entry *last = entry;
1232         int new_offs;
1233
1234         /* Adjust the value offsets of the entries */
1235         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
1236                 if (!last->e_value_block && last->e_value_size) {
1237                         new_offs = le16_to_cpu(last->e_value_offs) +
1238                                                         value_offs_shift;
1239                         BUG_ON(new_offs + le32_to_cpu(last->e_value_size)
1240                                  > blocksize);
1241                         last->e_value_offs = cpu_to_le16(new_offs);
1242                 }
1243         }
1244         /* Shift the entries by n bytes */
1245         memmove(to, from, n);
1246 }
1247
1248 /*
1249  * Expand an inode by new_extra_isize bytes when EAs are present.
1250  * Returns 0 on success or negative error number on failure.
1251  */
1252 int ext4_expand_extra_isize_ea(struct inode *inode, int new_extra_isize,
1253                                struct ext4_inode *raw_inode, handle_t *handle)
1254 {
1255         struct ext4_xattr_ibody_header *header;
1256         struct ext4_xattr_entry *entry, *last, *first;
1257         struct buffer_head *bh = NULL;
1258         struct ext4_xattr_ibody_find *is = NULL;
1259         struct ext4_xattr_block_find *bs = NULL;
1260         char *buffer = NULL, *b_entry_name = NULL;
1261         size_t min_offs, free;
1262         int total_ino, total_blk;
1263         void *base, *start, *end;
1264         int extra_isize = 0, error = 0, tried_min_extra_isize = 0;
1265         int s_min_extra_isize = le16_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_min_extra_isize);
1266
1267         down_write(&EXT4_I(inode)->xattr_sem);
1268 retry:
1269         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize) {
1270                 up_write(&EXT4_I(inode)->xattr_sem);
1271                 return 0;
1272         }
1273
1274         header = IHDR(inode, raw_inode);
1275         entry = IFIRST(header);
1276
1277         /*
1278          * Check if enough free space is available in the inode to shift the
1279          * entries ahead by new_extra_isize.
1280          */
1281
1282         base = start = entry;
1283         end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
1284         min_offs = end - base;
1285         last = entry;
1286         total_ino = sizeof(struct ext4_xattr_ibody_header);
1287
1288         free = ext4_xattr_free_space(last, &min_offs, base, &total_ino);
1289         if (free >= new_extra_isize) {
1290                 entry = IFIRST(header);
1291                 ext4_xattr_shift_entries(entry, EXT4_I(inode)->i_extra_isize
1292                                 - new_extra_isize, (void *)raw_inode +
1293                                 EXT4_GOOD_OLD_INODE_SIZE + new_extra_isize,
1294                                 (void *)header, total_ino,
1295                                 inode->i_sb->s_blocksize);
1296                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
1297                 error = 0;
1298                 goto cleanup;
1299         }
1300
1301         /*
1302          * Enough free space isn't available in the inode, check if
1303          * EA block can hold new_extra_isize bytes.
1304          */
1305         if (EXT4_I(inode)->i_file_acl) {
1306                 bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
1307                 error = -EIO;
1308                 if (!bh)
1309                         goto cleanup;
1310                 if (ext4_xattr_check_block(inode, bh)) {
1311                         EXT4_ERROR_INODE(inode, "bad block %llu",
1312                                          EXT4_I(inode)->i_file_acl);
1313                         error = -EIO;
1314                         goto cleanup;
1315                 }
1316                 base = BHDR(bh);
1317                 first = BFIRST(bh);
1318                 end = bh->b_data + bh->b_size;
1319                 min_offs = end - base;
1320                 free = ext4_xattr_free_space(first, &min_offs, base,
1321                                              &total_blk);
1322                 if (free < new_extra_isize) {
1323                         if (!tried_min_extra_isize && s_min_extra_isize) {
1324                                 tried_min_extra_isize++;
1325                                 new_extra_isize = s_min_extra_isize;
1326                                 brelse(bh);
1327                                 goto retry;
1328                         }
1329                         error = -1;
1330                         goto cleanup;
1331                 }
1332         } else {
1333                 free = inode->i_sb->s_blocksize;
1334         }
1335
1336         while (new_extra_isize > 0) {
1337                 size_t offs, size, entry_size;
1338                 struct ext4_xattr_entry *small_entry = NULL;
1339                 struct ext4_xattr_info i = {
1340                         .value = NULL,
1341                         .value_len = 0,
1342                 };
1343                 unsigned int total_size;  /* EA entry size + value size */
1344                 unsigned int shift_bytes; /* No. of bytes to shift EAs by? */
1345                 unsigned int min_total_size = ~0U;
1346
1347                 is = kzalloc(sizeof(struct ext4_xattr_ibody_find), GFP_NOFS);
1348                 bs = kzalloc(sizeof(struct ext4_xattr_block_find), GFP_NOFS);
1349                 if (!is || !bs) {
1350                         error = -ENOMEM;
1351                         goto cleanup;
1352                 }
1353
1354                 is->s.not_found = -ENODATA;
1355                 bs->s.not_found = -ENODATA;
1356                 is->iloc.bh = NULL;
1357                 bs->bh = NULL;
1358
1359                 last = IFIRST(header);
1360                 /* Find the entry best suited to be pushed into EA block */
1361                 entry = NULL;
1362                 for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
1363                         total_size =
1364                         EXT4_XATTR_SIZE(le32_to_cpu(last->e_value_size)) +
1365                                         EXT4_XATTR_LEN(last->e_name_len);
1366                         if (total_size <= free && total_size < min_total_size) {
1367                                 if (total_size < new_extra_isize) {
1368                                         small_entry = last;
1369                                 } else {
1370                                         entry = last;
1371                                         min_total_size = total_size;
1372                                 }
1373                         }
1374                 }
1375
1376                 if (entry == NULL) {
1377                         if (small_entry) {
1378                                 entry = small_entry;
1379                         } else {
1380                                 if (!tried_min_extra_isize &&
1381                                     s_min_extra_isize) {
1382                                         tried_min_extra_isize++;
1383                                         new_extra_isize = s_min_extra_isize;
1384                                         kfree(is); is = NULL;
1385                                         kfree(bs); bs = NULL;
1386                                         brelse(bh);
1387                                         goto retry;
1388                                 }
1389                                 error = -1;
1390                                 goto cleanup;
1391                         }
1392                 }
1393                 offs = le16_to_cpu(entry->e_value_offs);
1394                 size = le32_to_cpu(entry->e_value_size);
1395                 entry_size = EXT4_XATTR_LEN(entry->e_name_len);
1396                 i.name_index = entry->e_name_index,
1397                 buffer = kmalloc(EXT4_XATTR_SIZE(size), GFP_NOFS);
1398                 b_entry_name = kmalloc(entry->e_name_len + 1, GFP_NOFS);
1399                 if (!buffer || !b_entry_name) {
1400                         error = -ENOMEM;
1401                         goto cleanup;
1402                 }
1403                 /* Save the entry name and the entry value */
1404                 memcpy(buffer, (void *)IFIRST(header) + offs,
1405                        EXT4_XATTR_SIZE(size));
1406                 memcpy(b_entry_name, entry->e_name, entry->e_name_len);
1407                 b_entry_name[entry->e_name_len] = '\0';
1408                 i.name = b_entry_name;
1409
1410                 error = ext4_get_inode_loc(inode, &is->iloc);
1411                 if (error)
1412                         goto cleanup;
1413
1414                 error = ext4_xattr_ibody_find(inode, &i, is);
1415                 if (error)
1416                         goto cleanup;
1417
1418                 /* Remove the chosen entry from the inode */
1419                 error = ext4_xattr_ibody_set(handle, inode, &i, is);
1420                 if (error)
1421                         goto cleanup;
1422
1423                 entry = IFIRST(header);
1424                 if (entry_size + EXT4_XATTR_SIZE(size) >= new_extra_isize)
1425                         shift_bytes = new_extra_isize;
1426                 else
1427                         shift_bytes = entry_size + size;
1428                 /* Adjust the offsets and shift the remaining entries ahead */
1429                 ext4_xattr_shift_entries(entry, EXT4_I(inode)->i_extra_isize -
1430                         shift_bytes, (void *)raw_inode +
1431                         EXT4_GOOD_OLD_INODE_SIZE + extra_isize + shift_bytes,
1432                         (void *)header, total_ino - entry_size,
1433                         inode->i_sb->s_blocksize);
1434
1435                 extra_isize += shift_bytes;
1436                 new_extra_isize -= shift_bytes;
1437                 EXT4_I(inode)->i_extra_isize = extra_isize;
1438
1439                 i.name = b_entry_name;
1440                 i.value = buffer;
1441                 i.value_len = size;
1442                 error = ext4_xattr_block_find(inode, &i, bs);
1443                 if (error)
1444                         goto cleanup;
1445
1446                 /* Add entry which was removed from the inode into the block */
1447                 error = ext4_xattr_block_set(handle, inode, &i, bs);
1448                 if (error)
1449                         goto cleanup;
1450                 kfree(b_entry_name);
1451                 kfree(buffer);
1452                 b_entry_name = NULL;
1453                 buffer = NULL;
1454                 brelse(is->iloc.bh);
1455                 kfree(is);
1456                 kfree(bs);
1457         }
1458         brelse(bh);
1459         up_write(&EXT4_I(inode)->xattr_sem);
1460         return 0;
1461
1462 cleanup:
1463         kfree(b_entry_name);
1464         kfree(buffer);
1465         if (is)
1466                 brelse(is->iloc.bh);
1467         kfree(is);
1468         kfree(bs);
1469         brelse(bh);
1470         up_write(&EXT4_I(inode)->xattr_sem);
1471         return error;
1472 }
1473
1474
1475
1476 /*
1477  * ext4_xattr_delete_inode()
1478  *
1479  * Free extended attribute resources associated with this inode. This
1480  * is called immediately before an inode is freed. We have exclusive
1481  * access to the inode.
1482  */
1483 void
1484 ext4_xattr_delete_inode(handle_t *handle, struct inode *inode)
1485 {
1486         struct buffer_head *bh = NULL;
1487
1488         if (!EXT4_I(inode)->i_file_acl)
1489                 goto cleanup;
1490         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
1491         if (!bh) {
1492                 EXT4_ERROR_INODE(inode, "block %llu read error",
1493                                  EXT4_I(inode)->i_file_acl);
1494                 goto cleanup;
1495         }
1496         if (BHDR(bh)->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC) ||
1497             BHDR(bh)->h_blocks != cpu_to_le32(1)) {
1498                 EXT4_ERROR_INODE(inode, "bad block %llu",
1499                                  EXT4_I(inode)->i_file_acl);
1500                 goto cleanup;
1501         }
1502         ext4_xattr_release_block(handle, inode, bh);
1503         EXT4_I(inode)->i_file_acl = 0;
1504
1505 cleanup:
1506         brelse(bh);
1507 }
1508
1509 /*
1510  * ext4_xattr_put_super()
1511  *
1512  * This is called when a file system is unmounted.
1513  */
1514 void
1515 ext4_xattr_put_super(struct super_block *sb)
1516 {
1517         mb_cache_shrink(sb->s_bdev);
1518 }
1519
1520 /*
1521  * ext4_xattr_cache_insert()
1522  *
1523  * Create a new entry in the extended attribute cache, and insert
1524  * it unless such an entry is already in the cache.
1525  *
1526  * Returns 0, or a negative error number on failure.
1527  */
1528 static void
1529 ext4_xattr_cache_insert(struct buffer_head *bh)
1530 {
1531         __u32 hash = le32_to_cpu(BHDR(bh)->h_hash);
1532         struct mb_cache_entry *ce;
1533         int error;
1534
1535         ce = mb_cache_entry_alloc(ext4_xattr_cache, GFP_NOFS);
1536         if (!ce) {
1537                 ea_bdebug(bh, "out of memory");
1538                 return;
1539         }
1540         error = mb_cache_entry_insert(ce, bh->b_bdev, bh->b_blocknr, hash);
1541         if (error) {
1542                 mb_cache_entry_free(ce);
1543                 if (error == -EBUSY) {
1544                         ea_bdebug(bh, "already in cache");
1545                         error = 0;
1546                 }
1547         } else {
1548                 ea_bdebug(bh, "inserting [%x]", (int)hash);
1549                 mb_cache_entry_release(ce);
1550         }
1551 }
1552
1553 /*
1554  * ext4_xattr_cmp()
1555  *
1556  * Compare two extended attribute blocks for equality.
1557  *
1558  * Returns 0 if the blocks are equal, 1 if they differ, and
1559  * a negative error number on errors.
1560  */
1561 static int
1562 ext4_xattr_cmp(struct ext4_xattr_header *header1,
1563                struct ext4_xattr_header *header2)
1564 {
1565         struct ext4_xattr_entry *entry1, *entry2;
1566
1567         entry1 = ENTRY(header1+1);
1568         entry2 = ENTRY(header2+1);
1569         while (!IS_LAST_ENTRY(entry1)) {
1570                 if (IS_LAST_ENTRY(entry2))
1571                         return 1;
1572                 if (entry1->e_hash != entry2->e_hash ||
1573                     entry1->e_name_index != entry2->e_name_index ||
1574                     entry1->e_name_len != entry2->e_name_len ||
1575                     entry1->e_value_size != entry2->e_value_size ||
1576                     memcmp(entry1->e_name, entry2->e_name, entry1->e_name_len))
1577                         return 1;
1578                 if (entry1->e_value_block != 0 || entry2->e_value_block != 0)
1579                         return -EIO;
1580                 if (memcmp((char *)header1 + le16_to_cpu(entry1->e_value_offs),
1581                            (char *)header2 + le16_to_cpu(entry2->e_value_offs),
1582                            le32_to_cpu(entry1->e_value_size)))
1583                         return 1;
1584
1585                 entry1 = EXT4_XATTR_NEXT(entry1);
1586                 entry2 = EXT4_XATTR_NEXT(entry2);
1587         }
1588         if (!IS_LAST_ENTRY(entry2))
1589                 return 1;
1590         return 0;
1591 }
1592
1593 /*
1594  * ext4_xattr_cache_find()
1595  *
1596  * Find an identical extended attribute block.
1597  *
1598  * Returns a pointer to the block found, or NULL if such a block was
1599  * not found or an error occurred.
1600  */
1601 static struct buffer_head *
1602 ext4_xattr_cache_find(struct inode *inode, struct ext4_xattr_header *header,
1603                       struct mb_cache_entry **pce)
1604 {
1605         __u32 hash = le32_to_cpu(header->h_hash);
1606         struct mb_cache_entry *ce;
1607
1608         if (!header->h_hash)
1609                 return NULL;  /* never share */
1610         ea_idebug(inode, "looking for cached blocks [%x]", (int)hash);
1611 again:
1612         ce = mb_cache_entry_find_first(ext4_xattr_cache, inode->i_sb->s_bdev,
1613                                        hash);
1614         while (ce) {
1615                 struct buffer_head *bh;
1616
1617                 if (IS_ERR(ce)) {
1618                         if (PTR_ERR(ce) == -EAGAIN)
1619                                 goto again;
1620                         break;
1621                 }
1622                 bh = sb_bread(inode->i_sb, ce->e_block);
1623                 if (!bh) {
1624                         EXT4_ERROR_INODE(inode, "block %lu read error",
1625                                          (unsigned long) ce->e_block);
1626                 } else if (le32_to_cpu(BHDR(bh)->h_refcount) >=
1627                                 EXT4_XATTR_REFCOUNT_MAX) {
1628                         ea_idebug(inode, "block %lu refcount %d>=%d",
1629                                   (unsigned long) ce->e_block,
1630                                   le32_to_cpu(BHDR(bh)->h_refcount),
1631                                           EXT4_XATTR_REFCOUNT_MAX);
1632                 } else if (ext4_xattr_cmp(header, BHDR(bh)) == 0) {
1633                         *pce = ce;
1634                         return bh;
1635                 }
1636                 brelse(bh);
1637                 ce = mb_cache_entry_find_next(ce, inode->i_sb->s_bdev, hash);
1638         }
1639         return NULL;
1640 }
1641
1642 #define NAME_HASH_SHIFT 5
1643 #define VALUE_HASH_SHIFT 16
1644
1645 /*
1646  * ext4_xattr_hash_entry()
1647  *
1648  * Compute the hash of an extended attribute.
1649  */
1650 static inline void ext4_xattr_hash_entry(struct ext4_xattr_header *header,
1651                                          struct ext4_xattr_entry *entry)
1652 {
1653         __u32 hash = 0;
1654         char *name = entry->e_name;
1655         int n;
1656
1657         for (n = 0; n < entry->e_name_len; n++) {
1658                 hash = (hash << NAME_HASH_SHIFT) ^
1659                        (hash >> (8*sizeof(hash) - NAME_HASH_SHIFT)) ^
1660                        *name++;
1661         }
1662
1663         if (entry->e_value_block == 0 && entry->e_value_size != 0) {
1664                 __le32 *value = (__le32 *)((char *)header +
1665                         le16_to_cpu(entry->e_value_offs));
1666                 for (n = (le32_to_cpu(entry->e_value_size) +
1667                      EXT4_XATTR_ROUND) >> EXT4_XATTR_PAD_BITS; n; n--) {
1668                         hash = (hash << VALUE_HASH_SHIFT) ^
1669                                (hash >> (8*sizeof(hash) - VALUE_HASH_SHIFT)) ^
1670                                le32_to_cpu(*value++);
1671                 }
1672         }
1673         entry->e_hash = cpu_to_le32(hash);
1674 }
1675
1676 #undef NAME_HASH_SHIFT
1677 #undef VALUE_HASH_SHIFT
1678
1679 #define BLOCK_HASH_SHIFT 16
1680
1681 /*
1682  * ext4_xattr_rehash()
1683  *
1684  * Re-compute the extended attribute hash value after an entry has changed.
1685  */
1686 static void ext4_xattr_rehash(struct ext4_xattr_header *header,
1687                               struct ext4_xattr_entry *entry)
1688 {
1689         struct ext4_xattr_entry *here;
1690         __u32 hash = 0;
1691
1692         ext4_xattr_hash_entry(header, entry);
1693         here = ENTRY(header+1);
1694         while (!IS_LAST_ENTRY(here)) {
1695                 if (!here->e_hash) {
1696                         /* Block is not shared if an entry's hash value == 0 */
1697                         hash = 0;
1698                         break;
1699                 }
1700                 hash = (hash << BLOCK_HASH_SHIFT) ^
1701                        (hash >> (8*sizeof(hash) - BLOCK_HASH_SHIFT)) ^
1702                        le32_to_cpu(here->e_hash);
1703                 here = EXT4_XATTR_NEXT(here);
1704         }
1705         header->h_hash = cpu_to_le32(hash);
1706 }
1707
1708 #undef BLOCK_HASH_SHIFT
1709
1710 int __init
1711 ext4_init_xattr(void)
1712 {
1713         ext4_xattr_cache = mb_cache_create("ext4_xattr", 6);
1714         if (!ext4_xattr_cache)
1715                 return -ENOMEM;
1716         return 0;
1717 }
1718
1719 void
1720 ext4_exit_xattr(void)
1721 {
1722         if (ext4_xattr_cache)
1723                 mb_cache_destroy(ext4_xattr_cache);
1724         ext4_xattr_cache = NULL;
1725 }