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 #define BHDR(bh) ((struct ext4_xattr_header *)((bh)->b_data))
  65 #define ENTRY(ptr) ((struct ext4_xattr_entry *)(ptr))
  66 #define BFIRST(bh) ENTRY(BHDR(bh)+1)
  67 #define IS_LAST_ENTRY(entry) (*(__u32 *)(entry) == 0)
  68
  69 #ifdef EXT4_XATTR_DEBUG
  70 # define ea_idebug(inode, f...) do { \
  71                 printk(KERN_DEBUG "inode %s:%lu: ", \
  72                         inode->i_sb->s_id, inode->i_ino); \
  73                 printk(f); \
  74                 printk("\n"); \
  75         } while (0)
  76 # define ea_bdebug(bh, f...) do { \
  77                 char b[BDEVNAME_SIZE]; \
  78                 printk(KERN_DEBUG "block %s:%lu: ", \
  79                         bdevname(bh->b_bdev, b), \
  80                         (unsigned long) bh->b_blocknr); \
  81                 printk(f); \
  82                 printk("\n"); \
  83         } while (0)
  84 #else
  85 # define ea_idebug(f...)
  86 # define ea_bdebug(f...)
  87 #endif
  88
  89 static void ext4_xattr_cache_insert(struct buffer_head *);
  90 static struct buffer_head *ext4_xattr_cache_find(struct inode *,
  91                                                  struct ext4_xattr_header *,
  92                                                  struct mb_cache_entry **);
  93 static void ext4_xattr_rehash(struct ext4_xattr_header *,
  94                               struct ext4_xattr_entry *);
  95 static int ext4_xattr_list(struct dentry *dentry, char *buffer,
  96                            size_t buffer_size);
  97
  98 static struct mb_cache *ext4_xattr_cache;
  99
 100 static const struct xattr_handler *ext4_xattr_handler_map[] = {
 101         [EXT4_XATTR_INDEX_USER]              = &ext4_xattr_user_handler,
 102 #ifdef CONFIG_EXT4_FS_POSIX_ACL
 103         [EXT4_XATTR_INDEX_POSIX_ACL_ACCESS]  = &ext4_xattr_acl_access_handler,
 104         [EXT4_XATTR_INDEX_POSIX_ACL_DEFAULT] = &ext4_xattr_acl_default_handler,
 105 #endif
 106         [EXT4_XATTR_INDEX_TRUSTED]           = &ext4_xattr_trusted_handler,
 107 #ifdef CONFIG_EXT4_FS_SECURITY
 108         [EXT4_XATTR_INDEX_SECURITY]          = &ext4_xattr_security_handler,
 109 #endif
 110 };
 111
 112 const struct xattr_handler *ext4_xattr_handlers[] = {
 113         &ext4_xattr_user_handler,
 114         &ext4_xattr_trusted_handler,
 115 #ifdef CONFIG_EXT4_FS_POSIX_ACL
 116         &ext4_xattr_acl_access_handler,
 117         &ext4_xattr_acl_default_handler,
 118 #endif
 119 #ifdef CONFIG_EXT4_FS_SECURITY
 120         &ext4_xattr_security_handler,
 121 #endif
 122         NULL
 123 };
 124
 125 static inline const struct xattr_handler *
 126 ext4_xattr_handler(int name_index)
 127 {
 128         const struct xattr_handler *handler = NULL;
 129
 130         if (name_index > 0 && name_index < ARRAY_SIZE(ext4_xattr_handler_map))
 131                 handler = ext4_xattr_handler_map[name_index];
 132         return handler;
 133 }
 134
 135 /*
 136  * Inode operation listxattr()
 137  *
 138  * dentry->d_inode->i_mutex: don't care
 139  */
 140 ssize_t
 141 ext4_listxattr(struct dentry *dentry, char *buffer, size_t size)
 142 {
 143         return ext4_xattr_list(dentry, buffer, size);
 144 }
 145
 146 static int
 147 ext4_xattr_check_names(struct ext4_xattr_entry *entry, void *end)
 148 {
 149         while (!IS_LAST_ENTRY(entry)) {
 150                 struct ext4_xattr_entry *next = EXT4_XATTR_NEXT(entry);
 151                 if ((void *)next >= end)
 152                         return -EIO;
 153                 entry = next;
 154         }
 155         return 0;
 156 }
 157
 158 static inline int
 159 ext4_xattr_check_block(struct buffer_head *bh)
 160 {
 161         int error;
 162
 163         if (BHDR(bh)->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC) ||
 164             BHDR(bh)->h_blocks != cpu_to_le32(1))
 165                 return -EIO;
 166         error = ext4_xattr_check_names(BFIRST(bh), bh->b_data + bh->b_size);
 167         return error;
 168 }
 169
 170 static inline int
 171 ext4_xattr_check_entry(struct ext4_xattr_entry *entry, size_t size)
 172 {
 173         size_t value_size = le32_to_cpu(entry->e_value_size);
 174
 175         if (entry->e_value_block != 0 || value_size > size ||
 176             le16_to_cpu(entry->e_value_offs) + value_size > size)
 177                 return -EIO;
 178         return 0;
 179 }
 180
 181 static int
 182 ext4_xattr_find_entry(struct ext4_xattr_entry **pentry, int name_index,
 183                       const char *name, size_t size, int sorted)
 184 {
 185         struct ext4_xattr_entry *entry;
 186         size_t name_len;
 187         int cmp = 1;
 188
 189         if (name == NULL)
 190                 return -EINVAL;
 191         name_len = strlen(name);
 192         entry = *pentry;
 193         for (; !IS_LAST_ENTRY(entry); entry = EXT4_XATTR_NEXT(entry)) {
 194                 cmp = name_index - entry->e_name_index;
 195                 if (!cmp)
 196                         cmp = name_len - entry->e_name_len;
 197                 if (!cmp)
 198                         cmp = memcmp(name, entry->e_name, name_len);
 199                 if (cmp <= 0 && (sorted || cmp == 0))
 200                         break;
 201         }
 202         *pentry = entry;
 203         if (!cmp && ext4_xattr_check_entry(entry, size))
 204                         return -EIO;
 205         return cmp ? -ENODATA : 0;
 206 }
 207
 208 static int
 209 ext4_xattr_block_get(struct inode *inode, int name_index, const char *name,
 210                      void *buffer, size_t buffer_size)
 211 {
 212         struct buffer_head *bh = NULL;
 213         struct ext4_xattr_entry *entry;
 214         size_t size;
 215         int error;
 216
 217         ea_idebug(inode, "name=%d.%s, buffer=%p, buffer_size=%ld",
 218                   name_index, name, buffer, (long)buffer_size);
 219
 220         error = -ENODATA;
 221         if (!EXT4_I(inode)->i_file_acl)
 222                 goto cleanup;
 223         ea_idebug(inode, "reading block %u", EXT4_I(inode)->i_file_acl);
 224         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
 225         if (!bh)
 226                 goto cleanup;
 227         ea_bdebug(bh, "b_count=%d, refcount=%d",
 228                 atomic_read(&(bh->b_count)), le32_to_cpu(BHDR(bh)->h_refcount));
 229         if (ext4_xattr_check_block(bh)) {
 230 bad_block:
 231                 EXT4_ERROR_INODE(inode, "bad block %llu",
 232                                  EXT4_I(inode)->i_file_acl);
 233                 error = -EIO;
 234                 goto cleanup;
 235         }
 236         ext4_xattr_cache_insert(bh);
 237         entry = BFIRST(bh);
 238         error = ext4_xattr_find_entry(&entry, name_index, name, bh->b_size, 1);
 239         if (error == -EIO)
 240                 goto bad_block;
 241         if (error)
 242                 goto cleanup;
 243         size = le32_to_cpu(entry->e_value_size);
 244         if (buffer) {
 245                 error = -ERANGE;
 246                 if (size > buffer_size)
 247                         goto cleanup;
 248                 memcpy(buffer, bh->b_data + le16_to_cpu(entry->e_value_offs),
 249                        size);
 250         }
 251         error = size;
 252
 253 cleanup:
 254         brelse(bh);
 255         return error;
 256 }
 257
 258 static int
 259 ext4_xattr_ibody_get(struct inode *inode, int name_index, const char *name,
 260                      void *buffer, size_t buffer_size)
 261 {
 262         struct ext4_xattr_ibody_header *header;
 263         struct ext4_xattr_entry *entry;
 264         struct ext4_inode *raw_inode;
 265         struct ext4_iloc iloc;
 266         size_t size;
 267         void *end;
 268         int error;
 269
 270         if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR))
 271                 return -ENODATA;
 272         error = ext4_get_inode_loc(inode, &iloc);
 273         if (error)
 274                 return error;
 275         raw_inode = ext4_raw_inode(&iloc);
 276         header = IHDR(inode, raw_inode);
 277         entry = IFIRST(header);
 278         end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
 279         error = ext4_xattr_check_names(entry, end);
 280         if (error)
 281                 goto cleanup;
 282         error = ext4_xattr_find_entry(&entry, name_index, name,
 283                                       end - (void *)entry, 0);
 284         if (error)
 285                 goto cleanup;
 286         size = le32_to_cpu(entry->e_value_size);
 287         if (buffer) {
 288                 error = -ERANGE;
 289                 if (size > buffer_size)
 290                         goto cleanup;
 291                 memcpy(buffer, (void *)IFIRST(header) +
 292                        le16_to_cpu(entry->e_value_offs), size);
 293         }
 294         error = size;
 295
 296 cleanup:
 297         brelse(iloc.bh);
 298         return error;
 299 }
 300
 301 /*
 302  * ext4_xattr_get()
 303  *
 304  * Copy an extended attribute into the buffer
 305  * provided, or compute the buffer size required.
 306  * Buffer is NULL to compute the size of the buffer required.
 307  *
 308  * Returns a negative error number on failure, or the number of bytes
 309  * used / required on success.
 310  */
 311 int
 312 ext4_xattr_get(struct inode *inode, int name_index, const char *name,
 313                void *buffer, size_t buffer_size)
 314 {
 315         int error;
 316
 317         down_read(&EXT4_I(inode)->xattr_sem);
 318         error = ext4_xattr_ibody_get(inode, name_index, name, buffer,
 319                                      buffer_size);
 320         if (error == -ENODATA)
 321                 error = ext4_xattr_block_get(inode, name_index, name, buffer,
 322                                              buffer_size);
 323         up_read(&EXT4_I(inode)->xattr_sem);
 324         return error;
 325 }
 326
 327 static int
 328 ext4_xattr_list_entries(struct dentry *dentry, struct ext4_xattr_entry *entry,
 329                         char *buffer, size_t buffer_size)
 330 {
 331         size_t rest = buffer_size;
 332
 333         for (; !IS_LAST_ENTRY(entry); entry = EXT4_XATTR_NEXT(entry)) {
 334                 const struct xattr_handler *handler =
 335                         ext4_xattr_handler(entry->e_name_index);
 336
 337                 if (handler) {
 338                         size_t size = handler->list(dentry, buffer, rest,
 339                                                     entry->e_name,
 340                                                     entry->e_name_len,
 341                                                     handler->flags);
 342                         if (buffer) {
 343                                 if (size > rest)
 344                                         return -ERANGE;
 345                                 buffer += size;
 346                         }
 347                         rest -= size;
 348                 }
 349         }
 350         return buffer_size - rest;
 351 }
 352
 353 static int
 354 ext4_xattr_block_list(struct dentry *dentry, char *buffer, size_t buffer_size)
 355 {
 356         struct inode *inode = dentry->d_inode;
 357         struct buffer_head *bh = NULL;
 358         int error;
 359
 360         ea_idebug(inode, "buffer=%p, buffer_size=%ld",
 361                   buffer, (long)buffer_size);
 362
 363         error = 0;
 364         if (!EXT4_I(inode)->i_file_acl)
 365                 goto cleanup;
 366         ea_idebug(inode, "reading block %u", EXT4_I(inode)->i_file_acl);
 367         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
 368         error = -EIO;
 369         if (!bh)
 370                 goto cleanup;
 371         ea_bdebug(bh, "b_count=%d, refcount=%d",
 372                 atomic_read(&(bh->b_count)), le32_to_cpu(BHDR(bh)->h_refcount));
 373         if (ext4_xattr_check_block(bh)) {
 374                 EXT4_ERROR_INODE(inode, "bad block %llu",
 375                                  EXT4_I(inode)->i_file_acl);
 376                 error = -EIO;
 377                 goto cleanup;
 378         }
 379         ext4_xattr_cache_insert(bh);
 380         error = ext4_xattr_list_entries(dentry, BFIRST(bh), buffer, buffer_size);
 381
 382 cleanup:
 383         brelse(bh);
 384
 385         return error;
 386 }
 387
 388 static int
 389 ext4_xattr_ibody_list(struct dentry *dentry, char *buffer, size_t buffer_size)
 390 {
 391         struct inode *inode = dentry->d_inode;
 392         struct ext4_xattr_ibody_header *header;
 393         struct ext4_inode *raw_inode;
 394         struct ext4_iloc iloc;
 395         void *end;
 396         int error;
 397
 398         if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR))
 399                 return 0;
 400         error = ext4_get_inode_loc(inode, &iloc);
 401         if (error)
 402                 return error;
 403         raw_inode = ext4_raw_inode(&iloc);
 404         header = IHDR(inode, raw_inode);
 405         end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
 406         error = ext4_xattr_check_names(IFIRST(header), end);
 407         if (error)
 408                 goto cleanup;
 409         error = ext4_xattr_list_entries(dentry, IFIRST(header),
 410                                         buffer, buffer_size);
 411
 412 cleanup:
 413         brelse(iloc.bh);
 414         return error;
 415 }
 416
 417 /*
 418  * ext4_xattr_list()
 419  *
 420  * Copy a list of attribute names into the buffer
 421  * provided, or compute the buffer size required.
 422  * Buffer is NULL to compute the size of the buffer required.
 423  *
 424  * Returns a negative error number on failure, or the number of bytes
 425  * used / required on success.
 426  */
 427 static int
 428 ext4_xattr_list(struct dentry *dentry, char *buffer, size_t buffer_size)
 429 {
 430         int i_error, b_error;
 431
 432         down_read(&EXT4_I(dentry->d_inode)->xattr_sem);
 433         i_error = ext4_xattr_ibody_list(dentry, buffer, buffer_size);
 434         if (i_error < 0) {
 435                 b_error = 0;
 436         } else {
 437                 if (buffer) {
 438                         buffer += i_error;
 439                         buffer_size -= i_error;
 440                 }
 441                 b_error = ext4_xattr_block_list(dentry, buffer, buffer_size);
 442                 if (b_error < 0)
 443                         i_error = 0;
 444         }
 445         up_read(&EXT4_I(dentry->d_inode)->xattr_sem);
 446         return i_error + b_error;
 447 }
 448
 449 /*
 450  * If the EXT4_FEATURE_COMPAT_EXT_ATTR feature of this file system is
 451  * not set, set it.
 452  */
 453 static void ext4_xattr_update_super_block(handle_t *handle,
 454                                           struct super_block *sb)
 455 {
 456         if (EXT4_HAS_COMPAT_FEATURE(sb, EXT4_FEATURE_COMPAT_EXT_ATTR))
 457                 return;
 458
 459         if (ext4_journal_get_write_access(handle, EXT4_SB(sb)->s_sbh) == 0) {
 460                 EXT4_SET_COMPAT_FEATURE(sb, EXT4_FEATURE_COMPAT_EXT_ATTR);
 461                 sb->s_dirt = 1;
 462                 ext4_handle_dirty_metadata(handle, NULL, EXT4_SB(sb)->s_sbh);
 463         }
 464 }
 465
 466 /*
 467  * Release the xattr block BH: If the reference count is > 1, decrement
 468  * it; otherwise free the block.
 469  */
 470 static void
 471 ext4_xattr_release_block(handle_t *handle, struct inode *inode,
 472                          struct buffer_head *bh)
 473 {
 474         struct mb_cache_entry *ce = NULL;
 475         int error = 0;
 476
 477         ce = mb_cache_entry_get(ext4_xattr_cache, bh->b_bdev, bh->b_blocknr);
 478         error = ext4_journal_get_write_access(handle, bh);
 479         if (error)
 480                 goto out;
 481
 482         lock_buffer(bh);
 483         if (BHDR(bh)->h_refcount == cpu_to_le32(1)) {
 484                 ea_bdebug(bh, "refcount now=0; freeing");
 485                 if (ce)
 486                         mb_cache_entry_free(ce);
 487                 get_bh(bh);
 488                 ext4_free_blocks(handle, inode, bh, 0, 1,
 489                                  EXT4_FREE_BLOCKS_METADATA |
 490                                  EXT4_FREE_BLOCKS_FORGET);
 491         } else {
 492                 le32_add_cpu(&BHDR(bh)->h_refcount, -1);
 493                 error = ext4_handle_dirty_metadata(handle, inode, bh);
 494                 if (IS_SYNC(inode))
 495                         ext4_handle_sync(handle);
 496                 dquot_free_block(inode, 1);
 497                 ea_bdebug(bh, "refcount now=%d; releasing",
 498                           le32_to_cpu(BHDR(bh)->h_refcount));
 499                 if (ce)
 500                         mb_cache_entry_release(ce);
 501         }
 502         unlock_buffer(bh);
 503 out:
 504         ext4_std_error(inode->i_sb, error);
 505         return;
 506 }
 507
 508 /*
 509  * Find the available free space for EAs. This also returns the total number of
 510  * bytes used by EA entries.
 511  */
 512 static size_t ext4_xattr_free_space(struct ext4_xattr_entry *last,
 513                                     size_t *min_offs, void *base, int *total)
 514 {
 515         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 516                 *total += EXT4_XATTR_LEN(last->e_name_len);
 517                 if (!last->e_value_block && last->e_value_size) {
 518                         size_t offs = le16_to_cpu(last->e_value_offs);
 519                         if (offs < *min_offs)
 520                                 *min_offs = offs;
 521                 }
 522         }
 523         return (*min_offs - ((void *)last - base) - sizeof(__u32));
 524 }
 525
 526 struct ext4_xattr_info {
 527         int name_index;
 528         const char *name;
 529         const void *value;
 530         size_t value_len;
 531 };
 532
 533 struct ext4_xattr_search {
 534         struct ext4_xattr_entry *first;
 535         void *base;
 536         void *end;
 537         struct ext4_xattr_entry *here;
 538         int not_found;
 539 };
 540
 541 static int
 542 ext4_xattr_set_entry(struct ext4_xattr_info *i, struct ext4_xattr_search *s)
 543 {
 544         struct ext4_xattr_entry *last;
 545         size_t free, min_offs = s->end - s->base, name_len = strlen(i->name);
 546
 547         /* Compute min_offs and last. */
 548         last = s->first;
 549         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 550                 if (!last->e_value_block && last->e_value_size) {
 551                         size_t offs = le16_to_cpu(last->e_value_offs);
 552                         if (offs < min_offs)
 553                                 min_offs = offs;
 554                 }
 555         }
 556         free = min_offs - ((void *)last - s->base) - sizeof(__u32);
 557         if (!s->not_found) {
 558                 if (!s->here->e_value_block && s->here->e_value_size) {
 559                         size_t size = le32_to_cpu(s->here->e_value_size);
 560                         free += EXT4_XATTR_SIZE(size);
 561                 }
 562                 free += EXT4_XATTR_LEN(name_len);
 563         }
 564         if (i->value) {
 565                 if (free < EXT4_XATTR_SIZE(i->value_len) ||
 566                     free < EXT4_XATTR_LEN(name_len) +
 567                            EXT4_XATTR_SIZE(i->value_len))
 568                         return -ENOSPC;
 569         }
 570
 571         if (i->value && s->not_found) {
 572                 /* Insert the new name. */
 573                 size_t size = EXT4_XATTR_LEN(name_len);
 574                 size_t rest = (void *)last - (void *)s->here + sizeof(__u32);
 575                 memmove((void *)s->here + size, s->here, rest);
 576                 memset(s->here, 0, size);
 577                 s->here->e_name_index = i->name_index;
 578                 s->here->e_name_len = name_len;
 579                 memcpy(s->here->e_name, i->name, name_len);
 580         } else {
 581                 if (!s->here->e_value_block && s->here->e_value_size) {
 582                         void *first_val = s->base + min_offs;
 583                         size_t offs = le16_to_cpu(s->here->e_value_offs);
 584                         void *val = s->base + offs;
 585                         size_t size = EXT4_XATTR_SIZE(
 586                                 le32_to_cpu(s->here->e_value_size));
 587
 588                         if (i->value && size == EXT4_XATTR_SIZE(i->value_len)) {
 589                                 /* The old and the new value have the same
 590                                    size. Just replace. */
 591                                 s->here->e_value_size =
 592                                         cpu_to_le32(i->value_len);
 593                                 memset(val + size - EXT4_XATTR_PAD, 0,
 594                                        EXT4_XATTR_PAD); /* Clear pad bytes. */
 595                                 memcpy(val, i->value, i->value_len);
 596                                 return 0;
 597                         }
 598
 599                         /* Remove the old value. */
 600                         memmove(first_val + size, first_val, val - first_val);
 601                         memset(first_val, 0, size);
 602                         s->here->e_value_size = 0;
 603                         s->here->e_value_offs = 0;
 604                         min_offs += size;
 605
 606                         /* Adjust all value offsets. */
 607                         last = s->first;
 608                         while (!IS_LAST_ENTRY(last)) {
 609                                 size_t o = le16_to_cpu(last->e_value_offs);
 610                                 if (!last->e_value_block &&
 611                                     last->e_value_size && o < offs)
 612                                         last->e_value_offs =
 613                                                 cpu_to_le16(o + size);
 614                                 last = EXT4_XATTR_NEXT(last);
 615                         }
 616                 }
 617                 if (!i->value) {
 618                         /* Remove the old name. */
 619                         size_t size = EXT4_XATTR_LEN(name_len);
 620                         last = ENTRY((void *)last - size);
 621                         memmove(s->here, (void *)s->here + size,
 622                                 (void *)last - (void *)s->here + sizeof(__u32));
 623                         memset(last, 0, size);
 624                 }
 625         }
 626
 627         if (i->value) {
 628                 /* Insert the new value. */
 629                 s->here->e_value_size = cpu_to_le32(i->value_len);
 630                 if (i->value_len) {
 631                         size_t size = EXT4_XATTR_SIZE(i->value_len);
 632                         void *val = s->base + min_offs - size;
 633                         s->here->e_value_offs = cpu_to_le16(min_offs - size);
 634                         memset(val + size - EXT4_XATTR_PAD, 0,
 635                                EXT4_XATTR_PAD); /* Clear the pad bytes. */
 636                         memcpy(val, i->value, i->value_len);
 637                 }
 638         }
 639         return 0;
 640 }
 641
 642 struct ext4_xattr_block_find {
 643         struct ext4_xattr_search s;
 644         struct buffer_head *bh;
 645 };
 646
 647 static int
 648 ext4_xattr_block_find(struct inode *inode, struct ext4_xattr_info *i,
 649                       struct ext4_xattr_block_find *bs)
 650 {
 651         struct super_block *sb = inode->i_sb;
 652         int error;
 653
 654         ea_idebug(inode, "name=%d.%s, value=%p, value_len=%ld",
 655                   i->name_index, i->name, i->value, (long)i->value_len);
 656
 657         if (EXT4_I(inode)->i_file_acl) {
 658                 /* The inode already has an extended attribute block. */
 659                 bs->bh = sb_bread(sb, EXT4_I(inode)->i_file_acl);
 660                 error = -EIO;
 661                 if (!bs->bh)
 662                         goto cleanup;
 663                 ea_bdebug(bs->bh, "b_count=%d, refcount=%d",
 664                         atomic_read(&(bs->bh->b_count)),
 665                         le32_to_cpu(BHDR(bs->bh)->h_refcount));
 666                 if (ext4_xattr_check_block(bs->bh)) {
 667                         EXT4_ERROR_INODE(inode, "bad block %llu",
 668                                          EXT4_I(inode)->i_file_acl);
 669                         error = -EIO;
 670                         goto cleanup;
 671                 }
 672                 /* Find the named attribute. */
 673                 bs->s.base = BHDR(bs->bh);
 674                 bs->s.first = BFIRST(bs->bh);
 675                 bs->s.end = bs->bh->b_data + bs->bh->b_size;
 676                 bs->s.here = bs->s.first;
 677                 error = ext4_xattr_find_entry(&bs->s.here, i->name_index,
 678                                               i->name, bs->bh->b_size, 1);
 679                 if (error && error != -ENODATA)
 680                         goto cleanup;
 681                 bs->s.not_found = error;
 682         }
 683         error = 0;
 684
 685 cleanup:
 686         return error;
 687 }
 688
 689 static int
 690 ext4_xattr_block_set(handle_t *handle, struct inode *inode,
 691                      struct ext4_xattr_info *i,
 692                      struct ext4_xattr_block_find *bs)
 693 {
 694         struct super_block *sb = inode->i_sb;
 695         struct buffer_head *new_bh = NULL;
 696         struct ext4_xattr_search *s = &bs->s;
 697         struct mb_cache_entry *ce = NULL;
 698         int error = 0;
 699
 700 #define header(x) ((struct ext4_xattr_header *)(x))
 701
 702         if (i->value && i->value_len > sb->s_blocksize)
 703                 return -ENOSPC;
 704         if (s->base) {
 705                 ce = mb_cache_entry_get(ext4_xattr_cache, bs->bh->b_bdev,
 706                                         bs->bh->b_blocknr);
 707                 error = ext4_journal_get_write_access(handle, bs->bh);
 708                 if (error)
 709                         goto cleanup;
 710                 lock_buffer(bs->bh);
 711
 712                 if (header(s->base)->h_refcount == cpu_to_le32(1)) {
 713                         if (ce) {
 714                                 mb_cache_entry_free(ce);
 715                                 ce = NULL;
 716                         }
 717                         ea_bdebug(bs->bh, "modifying in-place");
 718                         error = ext4_xattr_set_entry(i, s);
 719                         if (!error) {
 720                                 if (!IS_LAST_ENTRY(s->first))
 721                                         ext4_xattr_rehash(header(s->base),
 722                                                           s->here);
 723                                 ext4_xattr_cache_insert(bs->bh);
 724                         }
 725                         unlock_buffer(bs->bh);
 726                         if (error == -EIO)
 727                                 goto bad_block;
 728                         if (!error)
 729                                 error = ext4_handle_dirty_metadata(handle,
 730                                                                    inode,
 731                                                                    bs->bh);
 732                         if (error)
 733                                 goto cleanup;
 734                         goto inserted;
 735                 } else {
 736                         int offset = (char *)s->here - bs->bh->b_data;
 737
 738                         unlock_buffer(bs->bh);
 739                         jbd2_journal_release_buffer(handle, bs->bh);
 740                         if (ce) {
 741                                 mb_cache_entry_release(ce);
 742                                 ce = NULL;
 743                         }
 744                         ea_bdebug(bs->bh, "cloning");
 745                         s->base = kmalloc(bs->bh->b_size, GFP_NOFS);
 746                         error = -ENOMEM;
 747                         if (s->base == NULL)
 748                                 goto cleanup;
 749                         memcpy(s->base, BHDR(bs->bh), bs->bh->b_size);
 750                         s->first = ENTRY(header(s->base)+1);
 751                         header(s->base)->h_refcount = cpu_to_le32(1);
 752                         s->here = ENTRY(s->base + offset);
 753                         s->end = s->base + bs->bh->b_size;
 754                 }
 755         } else {
 756                 /* Allocate a buffer where we construct the new block. */
 757                 s->base = kzalloc(sb->s_blocksize, GFP_NOFS);
 758                 /* assert(header == s->base) */
 759                 error = -ENOMEM;
 760                 if (s->base == NULL)
 761                         goto cleanup;
 762                 header(s->base)->h_magic = cpu_to_le32(EXT4_XATTR_MAGIC);
 763                 header(s->base)->h_blocks = cpu_to_le32(1);
 764                 header(s->base)->h_refcount = cpu_to_le32(1);
 765                 s->first = ENTRY(header(s->base)+1);
 766                 s->here = ENTRY(header(s->base)+1);
 767                 s->end = s->base + sb->s_blocksize;
 768         }
 769
 770         error = ext4_xattr_set_entry(i, s);
 771         if (error == -EIO)
 772                 goto bad_block;
 773         if (error)
 774                 goto cleanup;
 775         if (!IS_LAST_ENTRY(s->first))
 776                 ext4_xattr_rehash(header(s->base), s->here);
 777
 778 inserted:
 779         if (!IS_LAST_ENTRY(s->first)) {
 780                 new_bh = ext4_xattr_cache_find(inode, header(s->base), &ce);
 781                 if (new_bh) {
 782                         /* We found an identical block in the cache. */
 783                         if (new_bh == bs->bh)
 784                                 ea_bdebug(new_bh, "keeping");
 785                         else {
 786                                 /* The old block is released after updating
 787                                    the inode. */
 788                                 error = dquot_alloc_block(inode, 1);
 789                                 if (error)
 790                                         goto cleanup;
 791                                 error = ext4_journal_get_write_access(handle,
 792                                                                       new_bh);
 793                                 if (error)
 794                                         goto cleanup_dquot;
 795                                 lock_buffer(new_bh);
 796                                 le32_add_cpu(&BHDR(new_bh)->h_refcount, 1);
 797                                 ea_bdebug(new_bh, "reusing; refcount now=%d",
 798                                         le32_to_cpu(BHDR(new_bh)->h_refcount));
 799                                 unlock_buffer(new_bh);
 800                                 error = ext4_handle_dirty_metadata(handle,
 801                                                                    inode,
 802                                                                    new_bh);
 803                                 if (error)
 804                                         goto cleanup_dquot;
 805                         }
 806                         mb_cache_entry_release(ce);
 807                         ce = NULL;
 808                 } else if (bs->bh && s->base == bs->bh->b_data) {
 809                         /* We were modifying this block in-place. */
 810                         ea_bdebug(bs->bh, "keeping this block");
 811                         new_bh = bs->bh;
 812                         get_bh(new_bh);
 813                 } else {
 814                         /* We need to allocate a new block */
 815                         ext4_fsblk_t goal, block;
 816
 817                         goal = ext4_group_first_block_no(sb,
 818                                                 EXT4_I(inode)->i_block_group);
 819
 820                         /* non-extent files can't have physical blocks past 2^32 */
 821                         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
 822                                 goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
 823
 824                         block = ext4_new_meta_blocks(handle, inode,
 825                                                   goal, NULL, &error);
 826                         if (error)
 827                                 goto cleanup;
 828
 829                         if (!(ext4_test_inode_flag(inode, EXT4_INODE_EXTENTS)))
 830                                 BUG_ON(block > EXT4_MAX_BLOCK_FILE_PHYS);
 831
 832                         ea_idebug(inode, "creating block %d", block);
 833
 834                         new_bh = sb_getblk(sb, block);
 835                         if (!new_bh) {
 836 getblk_failed:
 837                                 ext4_free_blocks(handle, inode, 0, block, 1,
 838                                                  EXT4_FREE_BLOCKS_METADATA);
 839                                 error = -EIO;
 840                                 goto cleanup;
 841                         }
 842                         lock_buffer(new_bh);
 843                         error = ext4_journal_get_create_access(handle, new_bh);
 844                         if (error) {
 845                                 unlock_buffer(new_bh);
 846                                 goto getblk_failed;
 847                         }
 848                         memcpy(new_bh->b_data, s->base, new_bh->b_size);
 849                         set_buffer_uptodate(new_bh);
 850                         unlock_buffer(new_bh);
 851                         ext4_xattr_cache_insert(new_bh);
 852                         error = ext4_handle_dirty_metadata(handle,
 853                                                            inode, new_bh);
 854                         if (error)
 855                                 goto cleanup;
 856                 }
 857         }
 858
 859         /* Update the inode. */
 860         EXT4_I(inode)->i_file_acl = new_bh ? new_bh->b_blocknr : 0;
 861
 862         /* Drop the previous xattr block. */
 863         if (bs->bh && bs->bh != new_bh)
 864                 ext4_xattr_release_block(handle, inode, bs->bh);
 865         error = 0;
 866
 867 cleanup:
 868         if (ce)
 869                 mb_cache_entry_release(ce);
 870         brelse(new_bh);
 871         if (!(bs->bh && s->base == bs->bh->b_data))
 872                 kfree(s->base);
 873
 874         return error;
 875
 876 cleanup_dquot:
 877         dquot_free_block(inode, 1);
 878         goto cleanup;
 879
 880 bad_block:
 881         EXT4_ERROR_INODE(inode, "bad block %llu",
 882                          EXT4_I(inode)->i_file_acl);
 883         goto cleanup;
 884
 885 #undef header
 886 }
 887
 888 struct ext4_xattr_ibody_find {
 889         struct ext4_xattr_search s;
 890         struct ext4_iloc iloc;
 891 };
 892
 893 static int
 894 ext4_xattr_ibody_find(struct inode *inode, struct ext4_xattr_info *i,
 895                       struct ext4_xattr_ibody_find *is)
 896 {
 897         struct ext4_xattr_ibody_header *header;
 898         struct ext4_inode *raw_inode;
 899         int error;
 900
 901         if (EXT4_I(inode)->i_extra_isize == 0)
 902                 return 0;
 903         raw_inode = ext4_raw_inode(&is->iloc);
 904         header = IHDR(inode, raw_inode);
 905         is->s.base = is->s.first = IFIRST(header);
 906         is->s.here = is->s.first;
 907         is->s.end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
 908         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
 909                 error = ext4_xattr_check_names(IFIRST(header), is->s.end);
 910                 if (error)
 911                         return error;
 912                 /* Find the named attribute. */
 913                 error = ext4_xattr_find_entry(&is->s.here, i->name_index,
 914                                               i->name, is->s.end -
 915                                               (void *)is->s.base, 0);
 916                 if (error && error != -ENODATA)
 917                         return error;
 918                 is->s.not_found = error;
 919         }
 920         return 0;
 921 }
 922
 923 static int
 924 ext4_xattr_ibody_set(handle_t *handle, struct inode *inode,
 925                      struct ext4_xattr_info *i,
 926                      struct ext4_xattr_ibody_find *is)
 927 {
 928         struct ext4_xattr_ibody_header *header;
 929         struct ext4_xattr_search *s = &is->s;
 930         int error;
 931
 932         if (EXT4_I(inode)->i_extra_isize == 0)
 933                 return -ENOSPC;
 934         error = ext4_xattr_set_entry(i, s);
 935         if (error)
 936                 return error;
 937         header = IHDR(inode, ext4_raw_inode(&is->iloc));
 938         if (!IS_LAST_ENTRY(s->first)) {
 939                 header->h_magic = cpu_to_le32(EXT4_XATTR_MAGIC);
 940                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
 941         } else {
 942                 header->h_magic = cpu_to_le32(0);
 943                 ext4_clear_inode_state(inode, EXT4_STATE_XATTR);
 944         }
 945         return 0;
 946 }
 947
 948 /*
 949  * ext4_xattr_set_handle()
 950  *
 951  * Create, replace or remove an extended attribute for this inode. Buffer
 952  * is NULL to remove an existing extended attribute, and non-NULL to
 953  * either replace an existing extended attribute, or create a new extended
 954  * attribute. The flags XATTR_REPLACE and XATTR_CREATE
 955  * specify that an extended attribute must exist and must not exist
 956  * previous to the call, respectively.
 957  *
 958  * Returns 0, or a negative error number on failure.
 959  */
 960 int
 961 ext4_xattr_set_handle(handle_t *handle, struct inode *inode, int name_index,
 962                       const char *name, const void *value, size_t value_len,
 963                       int flags)
 964 {
 965         struct ext4_xattr_info i = {
 966                 .name_index = name_index,
 967                 .name = name,
 968                 .value = value,
 969                 .value_len = value_len,
 970
 971         };
 972         struct ext4_xattr_ibody_find is = {
 973                 .s = { .not_found = -ENODATA, },
 974         };
 975         struct ext4_xattr_block_find bs = {
 976                 .s = { .not_found = -ENODATA, },
 977         };
 978         unsigned long no_expand;
 979         int error;
 980
 981         if (!name)
 982                 return -EINVAL;
 983         if (strlen(name) > 255)
 984                 return -ERANGE;
 985         down_write(&EXT4_I(inode)->xattr_sem);
 986         no_expand = ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND);
 987         ext4_set_inode_state(inode, EXT4_STATE_NO_EXPAND);
 988
 989         error = ext4_get_inode_loc(inode, &is.iloc);
 990         if (error)
 991                 goto cleanup;
 992
 993         error = ext4_journal_get_write_access(handle, is.iloc.bh);
 994         if (error)
 995                 goto cleanup;
 996
 997         if (ext4_test_inode_state(inode, EXT4_STATE_NEW)) {
 998                 struct ext4_inode *raw_inode = ext4_raw_inode(&is.iloc);
 999                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
1000                 ext4_clear_inode_state(inode, EXT4_STATE_NEW);
1001         }
1002
1003         error = ext4_xattr_ibody_find(inode, &i, &is);
1004         if (error)
1005                 goto cleanup;
1006         if (is.s.not_found)
1007                 error = ext4_xattr_block_find(inode, &i, &bs);
1008         if (error)
1009                 goto cleanup;
1010         if (is.s.not_found && bs.s.not_found) {
1011                 error = -ENODATA;
1012                 if (flags & XATTR_REPLACE)
1013                         goto cleanup;
1014                 error = 0;
1015                 if (!value)
1016                         goto cleanup;
1017         } else {
1018                 error = -EEXIST;
1019                 if (flags & XATTR_CREATE)
1020                         goto cleanup;
1021         }
1022         if (!value) {
1023                 if (!is.s.not_found)
1024                         error = ext4_xattr_ibody_set(handle, inode, &i, &is);
1025                 else if (!bs.s.not_found)
1026                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
1027         } else {
1028                 error = ext4_xattr_ibody_set(handle, inode, &i, &is);
1029                 if (!error && !bs.s.not_found) {
1030                         i.value = NULL;
1031                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
1032                 } else if (error == -ENOSPC) {
1033                         if (EXT4_I(inode)->i_file_acl && !bs.s.base) {
1034                                 error = ext4_xattr_block_find(inode, &i, &bs);
1035                                 if (error)
1036                                         goto cleanup;
1037                         }
1038                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
1039                         if (error)
1040                                 goto cleanup;
1041                         if (!is.s.not_found) {
1042                                 i.value = NULL;
1043                                 error = ext4_xattr_ibody_set(handle, inode, &i,
1044                                                              &is);
1045                         }
1046                 }
1047         }
1048         if (!error) {
1049                 ext4_xattr_update_super_block(handle, inode->i_sb);
1050                 inode->i_ctime = ext4_current_time(inode);
1051                 if (!value)
1052                         ext4_clear_inode_state(inode, EXT4_STATE_NO_EXPAND);
1053                 error = ext4_mark_iloc_dirty(handle, inode, &is.iloc);
1054                 /*
1055                  * The bh is consumed by ext4_mark_iloc_dirty, even with
1056                  * error != 0.
1057                  */
1058                 is.iloc.bh = NULL;
1059                 if (IS_SYNC(inode))
1060                         ext4_handle_sync(handle);
1061         }
1062
1063 cleanup:
1064         brelse(is.iloc.bh);
1065         brelse(bs.bh);
1066         if (no_expand == 0)
1067                 ext4_clear_inode_state(inode, EXT4_STATE_NO_EXPAND);
1068         up_write(&EXT4_I(inode)->xattr_sem);
1069         return error;
1070 }
1071
1072 /*
1073  * ext4_xattr_set()
1074  *
1075  * Like ext4_xattr_set_handle, but start from an inode. This extended
1076  * attribute modification is a filesystem transaction by itself.
1077  *
1078  * Returns 0, or a negative error number on failure.
1079  */
1080 int
1081 ext4_xattr_set(struct inode *inode, int name_index, const char *name,
1082                const void *value, size_t value_len, int flags)
1083 {
1084         handle_t *handle;
1085         int error, retries = 0;
1086
1087 retry:
1088         handle = ext4_journal_start(inode, EXT4_DATA_TRANS_BLOCKS(inode->i_sb));
1089         if (IS_ERR(handle)) {
1090                 error = PTR_ERR(handle);
1091         } else {
1092                 int error2;
1093
1094                 error = ext4_xattr_set_handle(handle, inode, name_index, name,
1095                                               value, value_len, flags);
1096                 error2 = ext4_journal_stop(handle);
1097                 if (error == -ENOSPC &&
1098                     ext4_should_retry_alloc(inode->i_sb, &retries))
1099                         goto retry;
1100                 if (error == 0)
1101                         error = error2;
1102         }
1103
1104         return error;
1105 }
1106
1107 /*
1108  * Shift the EA entries in the inode to create space for the increased
1109  * i_extra_isize.
1110  */
1111 static void ext4_xattr_shift_entries(struct ext4_xattr_entry *entry,
1112                                      int value_offs_shift, void *to,
1113                                      void *from, size_t n, int blocksize)
1114 {
1115         struct ext4_xattr_entry *last = entry;
1116         int new_offs;
1117
1118         /* Adjust the value offsets of the entries */
1119         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
1120                 if (!last->e_value_block && last->e_value_size) {
1121                         new_offs = le16_to_cpu(last->e_value_offs) +
1122                                                         value_offs_shift;
1123                         BUG_ON(new_offs + le32_to_cpu(last->e_value_size)
1124                                  > blocksize);
1125                         last->e_value_offs = cpu_to_le16(new_offs);
1126                 }
1127         }
1128         /* Shift the entries by n bytes */
1129         memmove(to, from, n);
1130 }
1131
1132 /*
1133  * Expand an inode by new_extra_isize bytes when EAs are present.
1134  * Returns 0 on success or negative error number on failure.
1135  */
1136 int ext4_expand_extra_isize_ea(struct inode *inode, int new_extra_isize,
1137                                struct ext4_inode *raw_inode, handle_t *handle)
1138 {
1139         struct ext4_xattr_ibody_header *header;
1140         struct ext4_xattr_entry *entry, *last, *first;
1141         struct buffer_head *bh = NULL;
1142         struct ext4_xattr_ibody_find *is = NULL;
1143         struct ext4_xattr_block_find *bs = NULL;
1144         char *buffer = NULL, *b_entry_name = NULL;
1145         size_t min_offs, free;
1146         int total_ino, total_blk;
1147         void *base, *start, *end;
1148         int extra_isize = 0, error = 0, tried_min_extra_isize = 0;
1149         int s_min_extra_isize = le16_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_min_extra_isize);
1150
1151         down_write(&EXT4_I(inode)->xattr_sem);
1152 retry:
1153         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize) {
1154                 up_write(&EXT4_I(inode)->xattr_sem);
1155                 return 0;
1156         }
1157
1158         header = IHDR(inode, raw_inode);
1159         entry = IFIRST(header);
1160
1161         /*
1162          * Check if enough free space is available in the inode to shift the
1163          * entries ahead by new_extra_isize.
1164          */
1165
1166         base = start = entry;
1167         end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
1168         min_offs = end - base;
1169         last = entry;
1170         total_ino = sizeof(struct ext4_xattr_ibody_header);
1171
1172         free = ext4_xattr_free_space(last, &min_offs, base, &total_ino);
1173         if (free >= new_extra_isize) {
1174                 entry = IFIRST(header);
1175                 ext4_xattr_shift_entries(entry, EXT4_I(inode)->i_extra_isize
1176                                 - new_extra_isize, (void *)raw_inode +
1177                                 EXT4_GOOD_OLD_INODE_SIZE + new_extra_isize,
1178                                 (void *)header, total_ino,
1179                                 inode->i_sb->s_blocksize);
1180                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
1181                 error = 0;
1182                 goto cleanup;
1183         }
1184
1185         /*
1186          * Enough free space isn't available in the inode, check if
1187          * EA block can hold new_extra_isize bytes.
1188          */
1189         if (EXT4_I(inode)->i_file_acl) {
1190                 bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
1191                 error = -EIO;
1192                 if (!bh)
1193                         goto cleanup;
1194                 if (ext4_xattr_check_block(bh)) {
1195                         EXT4_ERROR_INODE(inode, "bad block %llu",
1196                                          EXT4_I(inode)->i_file_acl);
1197                         error = -EIO;
1198                         goto cleanup;
1199                 }
1200                 base = BHDR(bh);
1201                 first = BFIRST(bh);
1202                 end = bh->b_data + bh->b_size;
1203                 min_offs = end - base;
1204                 free = ext4_xattr_free_space(first, &min_offs, base,
1205                                              &total_blk);
1206                 if (free < new_extra_isize) {
1207                         if (!tried_min_extra_isize && s_min_extra_isize) {
1208                                 tried_min_extra_isize++;
1209                                 new_extra_isize = s_min_extra_isize;
1210                                 brelse(bh);
1211                                 goto retry;
1212                         }
1213                         error = -1;
1214                         goto cleanup;
1215                 }
1216         } else {
1217                 free = inode->i_sb->s_blocksize;
1218         }
1219
1220         while (new_extra_isize > 0) {
1221                 size_t offs, size, entry_size;
1222                 struct ext4_xattr_entry *small_entry = NULL;
1223                 struct ext4_xattr_info i = {
1224                         .value = NULL,
1225                         .value_len = 0,
1226                 };
1227                 unsigned int total_size;  /* EA entry size + value size */
1228                 unsigned int shift_bytes; /* No. of bytes to shift EAs by? */
1229                 unsigned int min_total_size = ~0U;
1230
1231                 is = kzalloc(sizeof(struct ext4_xattr_ibody_find), GFP_NOFS);
1232                 bs = kzalloc(sizeof(struct ext4_xattr_block_find), GFP_NOFS);
1233                 if (!is || !bs) {
1234                         error = -ENOMEM;
1235                         goto cleanup;
1236                 }
1237
1238                 is->s.not_found = -ENODATA;
1239                 bs->s.not_found = -ENODATA;
1240                 is->iloc.bh = NULL;
1241                 bs->bh = NULL;
1242
1243                 last = IFIRST(header);
1244                 /* Find the entry best suited to be pushed into EA block */
1245                 entry = NULL;
1246                 for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
1247                         total_size =
1248                         EXT4_XATTR_SIZE(le32_to_cpu(last->e_value_size)) +
1249                                         EXT4_XATTR_LEN(last->e_name_len);
1250                         if (total_size <= free && total_size < min_total_size) {
1251                                 if (total_size < new_extra_isize) {
1252                                         small_entry = last;
1253                                 } else {
1254                                         entry = last;
1255                                         min_total_size = total_size;
1256                                 }
1257                         }
1258                 }
1259
1260                 if (entry == NULL) {
1261                         if (small_entry) {
1262                                 entry = small_entry;
1263                         } else {
1264                                 if (!tried_min_extra_isize &&
1265                                     s_min_extra_isize) {
1266                                         tried_min_extra_isize++;
1267                                         new_extra_isize = s_min_extra_isize;
1268                                         goto retry;
1269                                 }
1270                                 error = -1;
1271                                 goto cleanup;
1272                         }
1273                 }
1274                 offs = le16_to_cpu(entry->e_value_offs);
1275                 size = le32_to_cpu(entry->e_value_size);
1276                 entry_size = EXT4_XATTR_LEN(entry->e_name_len);
1277                 i.name_index = entry->e_name_index,
1278                 buffer = kmalloc(EXT4_XATTR_SIZE(size), GFP_NOFS);
1279                 b_entry_name = kmalloc(entry->e_name_len + 1, GFP_NOFS);
1280                 if (!buffer || !b_entry_name) {
1281                         error = -ENOMEM;
1282                         goto cleanup;
1283                 }
1284                 /* Save the entry name and the entry value */
1285                 memcpy(buffer, (void *)IFIRST(header) + offs,
1286                        EXT4_XATTR_SIZE(size));
1287                 memcpy(b_entry_name, entry->e_name, entry->e_name_len);
1288                 b_entry_name[entry->e_name_len] = '\0';
1289                 i.name = b_entry_name;
1290
1291                 error = ext4_get_inode_loc(inode, &is->iloc);
1292                 if (error)
1293                         goto cleanup;
1294
1295                 error = ext4_xattr_ibody_find(inode, &i, is);
1296                 if (error)
1297                         goto cleanup;
1298
1299                 /* Remove the chosen entry from the inode */
1300                 error = ext4_xattr_ibody_set(handle, inode, &i, is);
1301                 if (error)
1302                         goto cleanup;
1303
1304                 entry = IFIRST(header);
1305                 if (entry_size + EXT4_XATTR_SIZE(size) >= new_extra_isize)
1306                         shift_bytes = new_extra_isize;
1307                 else
1308                         shift_bytes = entry_size + size;
1309                 /* Adjust the offsets and shift the remaining entries ahead */
1310                 ext4_xattr_shift_entries(entry, EXT4_I(inode)->i_extra_isize -
1311                         shift_bytes, (void *)raw_inode +
1312                         EXT4_GOOD_OLD_INODE_SIZE + extra_isize + shift_bytes,
1313                         (void *)header, total_ino - entry_size,
1314                         inode->i_sb->s_blocksize);
1315
1316                 extra_isize += shift_bytes;
1317                 new_extra_isize -= shift_bytes;
1318                 EXT4_I(inode)->i_extra_isize = extra_isize;
1319
1320                 i.name = b_entry_name;
1321                 i.value = buffer;
1322                 i.value_len = size;
1323                 error = ext4_xattr_block_find(inode, &i, bs);
1324                 if (error)
1325                         goto cleanup;
1326
1327                 /* Add entry which was removed from the inode into the block */
1328                 error = ext4_xattr_block_set(handle, inode, &i, bs);
1329                 if (error)
1330                         goto cleanup;
1331                 kfree(b_entry_name);
1332                 kfree(buffer);
1333                 b_entry_name = NULL;
1334                 buffer = NULL;
1335                 brelse(is->iloc.bh);
1336                 kfree(is);
1337                 kfree(bs);
1338         }
1339         brelse(bh);
1340         up_write(&EXT4_I(inode)->xattr_sem);
1341         return 0;
1342
1343 cleanup:
1344         kfree(b_entry_name);
1345         kfree(buffer);
1346         if (is)
1347                 brelse(is->iloc.bh);
1348         kfree(is);
1349         kfree(bs);
1350         brelse(bh);
1351         up_write(&EXT4_I(inode)->xattr_sem);
1352         return error;
1353 }
1354
1355
1356
1357 /*
1358  * ext4_xattr_delete_inode()
1359  *
1360  * Free extended attribute resources associated with this inode. This
1361  * is called immediately before an inode is freed. We have exclusive
1362  * access to the inode.
1363  */
1364 void
1365 ext4_xattr_delete_inode(handle_t *handle, struct inode *inode)
1366 {
1367         struct buffer_head *bh = NULL;
1368
1369         if (!EXT4_I(inode)->i_file_acl)
1370                 goto cleanup;
1371         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
1372         if (!bh) {
1373                 EXT4_ERROR_INODE(inode, "block %llu read error",
1374                                  EXT4_I(inode)->i_file_acl);
1375                 goto cleanup;
1376         }
1377         if (BHDR(bh)->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC) ||
1378             BHDR(bh)->h_blocks != cpu_to_le32(1)) {
1379                 EXT4_ERROR_INODE(inode, "bad block %llu",
1380                                  EXT4_I(inode)->i_file_acl);
1381                 goto cleanup;
1382         }
1383         ext4_xattr_release_block(handle, inode, bh);
1384         EXT4_I(inode)->i_file_acl = 0;
1385
1386 cleanup:
1387         brelse(bh);
1388 }
1389
1390 /*
1391  * ext4_xattr_put_super()
1392  *
1393  * This is called when a file system is unmounted.
1394  */
1395 void
1396 ext4_xattr_put_super(struct super_block *sb)
1397 {
1398         mb_cache_shrink(sb->s_bdev);
1399 }
1400
1401 /*
1402  * ext4_xattr_cache_insert()
1403  *
1404  * Create a new entry in the extended attribute cache, and insert
1405  * it unless such an entry is already in the cache.
1406  *
1407  * Returns 0, or a negative error number on failure.
1408  */
1409 static void
1410 ext4_xattr_cache_insert(struct buffer_head *bh)
1411 {
1412         __u32 hash = le32_to_cpu(BHDR(bh)->h_hash);
1413         struct mb_cache_entry *ce;
1414         int error;
1415
1416         ce = mb_cache_entry_alloc(ext4_xattr_cache, GFP_NOFS);
1417         if (!ce) {
1418                 ea_bdebug(bh, "out of memory");
1419                 return;
1420         }
1421         error = mb_cache_entry_insert(ce, bh->b_bdev, bh->b_blocknr, &hash);
1422         if (error) {
1423                 mb_cache_entry_free(ce);
1424                 if (error == -EBUSY) {
1425                         ea_bdebug(bh, "already in cache");
1426                         error = 0;
1427                 }
1428         } else {
1429                 ea_bdebug(bh, "inserting [%x]", (int)hash);
1430                 mb_cache_entry_release(ce);
1431         }
1432 }
1433
1434 /*
1435  * ext4_xattr_cmp()
1436  *
1437  * Compare two extended attribute blocks for equality.
1438  *
1439  * Returns 0 if the blocks are equal, 1 if they differ, and
1440  * a negative error number on errors.
1441  */
1442 static int
1443 ext4_xattr_cmp(struct ext4_xattr_header *header1,
1444                struct ext4_xattr_header *header2)
1445 {
1446         struct ext4_xattr_entry *entry1, *entry2;
1447
1448         entry1 = ENTRY(header1+1);
1449         entry2 = ENTRY(header2+1);
1450         while (!IS_LAST_ENTRY(entry1)) {
1451                 if (IS_LAST_ENTRY(entry2))
1452                         return 1;
1453                 if (entry1->e_hash != entry2->e_hash ||
1454                     entry1->e_name_index != entry2->e_name_index ||
1455                     entry1->e_name_len != entry2->e_name_len ||
1456                     entry1->e_value_size != entry2->e_value_size ||
1457                     memcmp(entry1->e_name, entry2->e_name, entry1->e_name_len))
1458                         return 1;
1459                 if (entry1->e_value_block != 0 || entry2->e_value_block != 0)
1460                         return -EIO;
1461                 if (memcmp((char *)header1 + le16_to_cpu(entry1->e_value_offs),
1462                            (char *)header2 + le16_to_cpu(entry2->e_value_offs),
1463                            le32_to_cpu(entry1->e_value_size)))
1464                         return 1;
1465
1466                 entry1 = EXT4_XATTR_NEXT(entry1);
1467                 entry2 = EXT4_XATTR_NEXT(entry2);
1468         }
1469         if (!IS_LAST_ENTRY(entry2))
1470                 return 1;
1471         return 0;
1472 }
1473
1474 /*
1475  * ext4_xattr_cache_find()
1476  *
1477  * Find an identical extended attribute block.
1478  *
1479  * Returns a pointer to the block found, or NULL if such a block was
1480  * not found or an error occurred.
1481  */
1482 static struct buffer_head *
1483 ext4_xattr_cache_find(struct inode *inode, struct ext4_xattr_header *header,
1484                       struct mb_cache_entry **pce)
1485 {
1486         __u32 hash = le32_to_cpu(header->h_hash);
1487         struct mb_cache_entry *ce;
1488
1489         if (!header->h_hash)
1490                 return NULL;  /* never share */
1491         ea_idebug(inode, "looking for cached blocks [%x]", (int)hash);
1492 again:
1493         ce = mb_cache_entry_find_first(ext4_xattr_cache, 0,
1494                                        inode->i_sb->s_bdev, hash);
1495         while (ce) {
1496                 struct buffer_head *bh;
1497
1498                 if (IS_ERR(ce)) {
1499                         if (PTR_ERR(ce) == -EAGAIN)
1500                                 goto again;
1501                         break;
1502                 }
1503                 bh = sb_bread(inode->i_sb, ce->e_block);
1504                 if (!bh) {
1505                         EXT4_ERROR_INODE(inode, "block %lu read error",
1506                                          (unsigned long) ce->e_block);
1507                 } else if (le32_to_cpu(BHDR(bh)->h_refcount) >=
1508                                 EXT4_XATTR_REFCOUNT_MAX) {
1509                         ea_idebug(inode, "block %lu refcount %d>=%d",
1510                                   (unsigned long) ce->e_block,
1511                                   le32_to_cpu(BHDR(bh)->h_refcount),
1512                                           EXT4_XATTR_REFCOUNT_MAX);
1513                 } else if (ext4_xattr_cmp(header, BHDR(bh)) == 0) {
1514                         *pce = ce;
1515                         return bh;
1516                 }
1517                 brelse(bh);
1518                 ce = mb_cache_entry_find_next(ce, 0, inode->i_sb->s_bdev, hash);
1519         }
1520         return NULL;
1521 }
1522
1523 #define NAME_HASH_SHIFT 5
1524 #define VALUE_HASH_SHIFT 16
1525
1526 /*
1527  * ext4_xattr_hash_entry()
1528  *
1529  * Compute the hash of an extended attribute.
1530  */
1531 static inline void ext4_xattr_hash_entry(struct ext4_xattr_header *header,
1532                                          struct ext4_xattr_entry *entry)
1533 {
1534         __u32 hash = 0;
1535         char *name = entry->e_name;
1536         int n;
1537
1538         for (n = 0; n < entry->e_name_len; n++) {
1539                 hash = (hash << NAME_HASH_SHIFT) ^
1540                        (hash >> (8*sizeof(hash) - NAME_HASH_SHIFT)) ^
1541                        *name++;
1542         }
1543
1544         if (entry->e_value_block == 0 && entry->e_value_size != 0) {
1545                 __le32 *value = (__le32 *)((char *)header +
1546                         le16_to_cpu(entry->e_value_offs));
1547                 for (n = (le32_to_cpu(entry->e_value_size) +
1548                      EXT4_XATTR_ROUND) >> EXT4_XATTR_PAD_BITS; n; n--) {
1549                         hash = (hash << VALUE_HASH_SHIFT) ^
1550                                (hash >> (8*sizeof(hash) - VALUE_HASH_SHIFT)) ^
1551                                le32_to_cpu(*value++);
1552                 }
1553         }
1554         entry->e_hash = cpu_to_le32(hash);
1555 }
1556
1557 #undef NAME_HASH_SHIFT
1558 #undef VALUE_HASH_SHIFT
1559
1560 #define BLOCK_HASH_SHIFT 16
1561
1562 /*
1563  * ext4_xattr_rehash()
1564  *
1565  * Re-compute the extended attribute hash value after an entry has changed.
1566  */
1567 static void ext4_xattr_rehash(struct ext4_xattr_header *header,
1568                               struct ext4_xattr_entry *entry)
1569 {
1570         struct ext4_xattr_entry *here;
1571         __u32 hash = 0;
1572
1573         ext4_xattr_hash_entry(header, entry);
1574         here = ENTRY(header+1);
1575         while (!IS_LAST_ENTRY(here)) {
1576                 if (!here->e_hash) {
1577                         /* Block is not shared if an entry's hash value == 0 */
1578                         hash = 0;
1579                         break;
1580                 }
1581                 hash = (hash << BLOCK_HASH_SHIFT) ^
1582                        (hash >> (8*sizeof(hash) - BLOCK_HASH_SHIFT)) ^
1583                        le32_to_cpu(here->e_hash);
1584                 here = EXT4_XATTR_NEXT(here);
1585         }
1586         header->h_hash = cpu_to_le32(hash);
1587 }
1588
1589 #undef BLOCK_HASH_SHIFT
1590
1591 int __init
1592 init_ext4_xattr(void)
1593 {
1594         ext4_xattr_cache = mb_cache_create("ext4_xattr", NULL,
1595                 sizeof(struct mb_cache_entry) +
1596                 sizeof(((struct mb_cache_entry *) 0)->e_indexes[0]), 1, 6);
1597         if (!ext4_xattr_cache)
1598                 return -ENOMEM;
1599         return 0;
1600 }
1601
1602 void
1603 exit_ext4_xattr(void)
1604 {
1605         if (ext4_xattr_cache)
1606                 mb_cache_destroy(ext4_xattr_cache);
1607         ext4_xattr_cache = NULL;
1608 }