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/ext4_jbd2.h>
  57 #include <linux/ext4_fs.h>
  58 #include <linux/mbcache.h>
  59 #include <linux/quotaops.h>
  60 #include <linux/rwsem.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
  96 static struct mb_cache *ext4_xattr_cache;
  97
  98 static struct xattr_handler *ext4_xattr_handler_map[] = {
  99         [EXT4_XATTR_INDEX_USER]              = &ext4_xattr_user_handler,
 100 #ifdef CONFIG_EXT4DEV_FS_POSIX_ACL
 101         [EXT4_XATTR_INDEX_POSIX_ACL_ACCESS]  = &ext4_xattr_acl_access_handler,
 102         [EXT4_XATTR_INDEX_POSIX_ACL_DEFAULT] = &ext4_xattr_acl_default_handler,
 103 #endif
 104         [EXT4_XATTR_INDEX_TRUSTED]           = &ext4_xattr_trusted_handler,
 105 #ifdef CONFIG_EXT4DEV_FS_SECURITY
 106         [EXT4_XATTR_INDEX_SECURITY]          = &ext4_xattr_security_handler,
 107 #endif
 108 };
 109
 110 struct xattr_handler *ext4_xattr_handlers[] = {
 111         &ext4_xattr_user_handler,
 112         &ext4_xattr_trusted_handler,
 113 #ifdef CONFIG_EXT4DEV_FS_POSIX_ACL
 114         &ext4_xattr_acl_access_handler,
 115         &ext4_xattr_acl_default_handler,
 116 #endif
 117 #ifdef CONFIG_EXT4DEV_FS_SECURITY
 118         &ext4_xattr_security_handler,
 119 #endif
 120         NULL
 121 };
 122
 123 static inline struct xattr_handler *
 124 ext4_xattr_handler(int name_index)
 125 {
 126         struct xattr_handler *handler = NULL;
 127
 128         if (name_index > 0 && name_index < ARRAY_SIZE(ext4_xattr_handler_map))
 129                 handler = ext4_xattr_handler_map[name_index];
 130         return handler;
 131 }
 132
 133 /*
 134  * Inode operation listxattr()
 135  *
 136  * dentry->d_inode->i_mutex: don't care
 137  */
 138 ssize_t
 139 ext4_listxattr(struct dentry *dentry, char *buffer, size_t size)
 140 {
 141         return ext4_xattr_list(dentry->d_inode, buffer, size);
 142 }
 143
 144 static int
 145 ext4_xattr_check_names(struct ext4_xattr_entry *entry, void *end)
 146 {
 147         while (!IS_LAST_ENTRY(entry)) {
 148                 struct ext4_xattr_entry *next = EXT4_XATTR_NEXT(entry);
 149                 if ((void *)next >= end)
 150                         return -EIO;
 151                 entry = next;
 152         }
 153         return 0;
 154 }
 155
 156 static inline int
 157 ext4_xattr_check_block(struct buffer_head *bh)
 158 {
 159         int error;
 160
 161         if (BHDR(bh)->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC) ||
 162             BHDR(bh)->h_blocks != cpu_to_le32(1))
 163                 return -EIO;
 164         error = ext4_xattr_check_names(BFIRST(bh), bh->b_data + bh->b_size);
 165         return error;
 166 }
 167
 168 static inline int
 169 ext4_xattr_check_entry(struct ext4_xattr_entry *entry, size_t size)
 170 {
 171         size_t value_size = le32_to_cpu(entry->e_value_size);
 172
 173         if (entry->e_value_block != 0 || value_size > size ||
 174             le16_to_cpu(entry->e_value_offs) + value_size > size)
 175                 return -EIO;
 176         return 0;
 177 }
 178
 179 static int
 180 ext4_xattr_find_entry(struct ext4_xattr_entry **pentry, int name_index,
 181                       const char *name, size_t size, int sorted)
 182 {
 183         struct ext4_xattr_entry *entry;
 184         size_t name_len;
 185         int cmp = 1;
 186
 187         if (name == NULL)
 188                 return -EINVAL;
 189         name_len = strlen(name);
 190         entry = *pentry;
 191         for (; !IS_LAST_ENTRY(entry); entry = EXT4_XATTR_NEXT(entry)) {
 192                 cmp = name_index - entry->e_name_index;
 193                 if (!cmp)
 194                         cmp = name_len - entry->e_name_len;
 195                 if (!cmp)
 196                         cmp = memcmp(name, entry->e_name, name_len);
 197                 if (cmp <= 0 && (sorted || cmp == 0))
 198                         break;
 199         }
 200         *pentry = entry;
 201         if (!cmp && ext4_xattr_check_entry(entry, size))
 202                         return -EIO;
 203         return cmp ? -ENODATA : 0;
 204 }
 205
 206 static int
 207 ext4_xattr_block_get(struct inode *inode, int name_index, const char *name,
 208                      void *buffer, size_t buffer_size)
 209 {
 210         struct buffer_head *bh = NULL;
 211         struct ext4_xattr_entry *entry;
 212         size_t size;
 213         int error;
 214
 215         ea_idebug(inode, "name=%d.%s, buffer=%p, buffer_size=%ld",
 216                   name_index, name, buffer, (long)buffer_size);
 217
 218         error = -ENODATA;
 219         if (!EXT4_I(inode)->i_file_acl)
 220                 goto cleanup;
 221         ea_idebug(inode, "reading block %u", EXT4_I(inode)->i_file_acl);
 222         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
 223         if (!bh)
 224                 goto cleanup;
 225         ea_bdebug(bh, "b_count=%d, refcount=%d",
 226                 atomic_read(&(bh->b_count)), le32_to_cpu(BHDR(bh)->h_refcount));
 227         if (ext4_xattr_check_block(bh)) {
 228 bad_block:      ext4_error(inode->i_sb, __FUNCTION__,
 229                            "inode %lu: bad block %llu", inode->i_ino,
 230                            EXT4_I(inode)->i_file_acl);
 231                 error = -EIO;
 232                 goto cleanup;
 233         }
 234         ext4_xattr_cache_insert(bh);
 235         entry = BFIRST(bh);
 236         error = ext4_xattr_find_entry(&entry, name_index, name, bh->b_size, 1);
 237         if (error == -EIO)
 238                 goto bad_block;
 239         if (error)
 240                 goto cleanup;
 241         size = le32_to_cpu(entry->e_value_size);
 242         if (buffer) {
 243                 error = -ERANGE;
 244                 if (size > buffer_size)
 245                         goto cleanup;
 246                 memcpy(buffer, bh->b_data + le16_to_cpu(entry->e_value_offs),
 247                        size);
 248         }
 249         error = size;
 250
 251 cleanup:
 252         brelse(bh);
 253         return error;
 254 }
 255
 256 static int
 257 ext4_xattr_ibody_get(struct inode *inode, int name_index, const char *name,
 258                      void *buffer, size_t buffer_size)
 259 {
 260         struct ext4_xattr_ibody_header *header;
 261         struct ext4_xattr_entry *entry;
 262         struct ext4_inode *raw_inode;
 263         struct ext4_iloc iloc;
 264         size_t size;
 265         void *end;
 266         int error;
 267
 268         if (!(EXT4_I(inode)->i_state & EXT4_STATE_XATTR))
 269                 return -ENODATA;
 270         error = ext4_get_inode_loc(inode, &iloc);
 271         if (error)
 272                 return error;
 273         raw_inode = ext4_raw_inode(&iloc);
 274         header = IHDR(inode, raw_inode);
 275         entry = IFIRST(header);
 276         end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
 277         error = ext4_xattr_check_names(entry, end);
 278         if (error)
 279                 goto cleanup;
 280         error = ext4_xattr_find_entry(&entry, name_index, name,
 281                                       end - (void *)entry, 0);
 282         if (error)
 283                 goto cleanup;
 284         size = le32_to_cpu(entry->e_value_size);
 285         if (buffer) {
 286                 error = -ERANGE;
 287                 if (size > buffer_size)
 288                         goto cleanup;
 289                 memcpy(buffer, (void *)IFIRST(header) +
 290                        le16_to_cpu(entry->e_value_offs), size);
 291         }
 292         error = size;
 293
 294 cleanup:
 295         brelse(iloc.bh);
 296         return error;
 297 }
 298
 299 /*
 300  * ext4_xattr_get()
 301  *
 302  * Copy an extended attribute into the buffer
 303  * provided, or compute the buffer size required.
 304  * Buffer is NULL to compute the size of the buffer required.
 305  *
 306  * Returns a negative error number on failure, or the number of bytes
 307  * used / required on success.
 308  */
 309 int
 310 ext4_xattr_get(struct inode *inode, int name_index, const char *name,
 311                void *buffer, size_t buffer_size)
 312 {
 313         int error;
 314
 315         down_read(&EXT4_I(inode)->xattr_sem);
 316         error = ext4_xattr_ibody_get(inode, name_index, name, buffer,
 317                                      buffer_size);
 318         if (error == -ENODATA)
 319                 error = ext4_xattr_block_get(inode, name_index, name, buffer,
 320                                              buffer_size);
 321         up_read(&EXT4_I(inode)->xattr_sem);
 322         return error;
 323 }
 324
 325 static int
 326 ext4_xattr_list_entries(struct inode *inode, struct ext4_xattr_entry *entry,
 327                         char *buffer, size_t buffer_size)
 328 {
 329         size_t rest = buffer_size;
 330
 331         for (; !IS_LAST_ENTRY(entry); entry = EXT4_XATTR_NEXT(entry)) {
 332                 struct xattr_handler *handler =
 333                         ext4_xattr_handler(entry->e_name_index);
 334
 335                 if (handler) {
 336                         size_t size = handler->list(inode, buffer, rest,
 337                                                     entry->e_name,
 338                                                     entry->e_name_len);
 339                         if (buffer) {
 340                                 if (size > rest)
 341                                         return -ERANGE;
 342                                 buffer += size;
 343                         }
 344                         rest -= size;
 345                 }
 346         }
 347         return buffer_size - rest;
 348 }
 349
 350 static int
 351 ext4_xattr_block_list(struct inode *inode, char *buffer, size_t buffer_size)
 352 {
 353         struct buffer_head *bh = NULL;
 354         int error;
 355
 356         ea_idebug(inode, "buffer=%p, buffer_size=%ld",
 357                   buffer, (long)buffer_size);
 358
 359         error = 0;
 360         if (!EXT4_I(inode)->i_file_acl)
 361                 goto cleanup;
 362         ea_idebug(inode, "reading block %u", EXT4_I(inode)->i_file_acl);
 363         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
 364         error = -EIO;
 365         if (!bh)
 366                 goto cleanup;
 367         ea_bdebug(bh, "b_count=%d, refcount=%d",
 368                 atomic_read(&(bh->b_count)), le32_to_cpu(BHDR(bh)->h_refcount));
 369         if (ext4_xattr_check_block(bh)) {
 370                 ext4_error(inode->i_sb, __FUNCTION__,
 371                            "inode %lu: bad block %llu", inode->i_ino,
 372                            EXT4_I(inode)->i_file_acl);
 373                 error = -EIO;
 374                 goto cleanup;
 375         }
 376         ext4_xattr_cache_insert(bh);
 377         error = ext4_xattr_list_entries(inode, BFIRST(bh), buffer, buffer_size);
 378
 379 cleanup:
 380         brelse(bh);
 381
 382         return error;
 383 }
 384
 385 static int
 386 ext4_xattr_ibody_list(struct inode *inode, char *buffer, size_t buffer_size)
 387 {
 388         struct ext4_xattr_ibody_header *header;
 389         struct ext4_inode *raw_inode;
 390         struct ext4_iloc iloc;
 391         void *end;
 392         int error;
 393
 394         if (!(EXT4_I(inode)->i_state & EXT4_STATE_XATTR))
 395                 return 0;
 396         error = ext4_get_inode_loc(inode, &iloc);
 397         if (error)
 398                 return error;
 399         raw_inode = ext4_raw_inode(&iloc);
 400         header = IHDR(inode, raw_inode);
 401         end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
 402         error = ext4_xattr_check_names(IFIRST(header), end);
 403         if (error)
 404                 goto cleanup;
 405         error = ext4_xattr_list_entries(inode, IFIRST(header),
 406                                         buffer, buffer_size);
 407
 408 cleanup:
 409         brelse(iloc.bh);
 410         return error;
 411 }
 412
 413 /*
 414  * ext4_xattr_list()
 415  *
 416  * Copy a list of attribute names into the buffer
 417  * provided, or compute the buffer size required.
 418  * Buffer is NULL to compute the size of the buffer required.
 419  *
 420  * Returns a negative error number on failure, or the number of bytes
 421  * used / required on success.
 422  */
 423 int
 424 ext4_xattr_list(struct inode *inode, char *buffer, size_t buffer_size)
 425 {
 426         int i_error, b_error;
 427
 428         down_read(&EXT4_I(inode)->xattr_sem);
 429         i_error = ext4_xattr_ibody_list(inode, buffer, buffer_size);
 430         if (i_error < 0) {
 431                 b_error = 0;
 432         } else {
 433                 if (buffer) {
 434                         buffer += i_error;
 435                         buffer_size -= i_error;
 436                 }
 437                 b_error = ext4_xattr_block_list(inode, buffer, buffer_size);
 438                 if (b_error < 0)
 439                         i_error = 0;
 440         }
 441         up_read(&EXT4_I(inode)->xattr_sem);
 442         return i_error + b_error;
 443 }
 444
 445 /*
 446  * If the EXT4_FEATURE_COMPAT_EXT_ATTR feature of this file system is
 447  * not set, set it.
 448  */
 449 static void ext4_xattr_update_super_block(handle_t *handle,
 450                                           struct super_block *sb)
 451 {
 452         if (EXT4_HAS_COMPAT_FEATURE(sb, EXT4_FEATURE_COMPAT_EXT_ATTR))
 453                 return;
 454
 455         if (ext4_journal_get_write_access(handle, EXT4_SB(sb)->s_sbh) == 0) {
 456                 EXT4_SET_COMPAT_FEATURE(sb, EXT4_FEATURE_COMPAT_EXT_ATTR);
 457                 sb->s_dirt = 1;
 458                 ext4_journal_dirty_metadata(handle, EXT4_SB(sb)->s_sbh);
 459         }
 460 }
 461
 462 /*
 463  * Release the xattr block BH: If the reference count is > 1, decrement
 464  * it; otherwise free the block.
 465  */
 466 static void
 467 ext4_xattr_release_block(handle_t *handle, struct inode *inode,
 468                          struct buffer_head *bh)
 469 {
 470         struct mb_cache_entry *ce = NULL;
 471         int error = 0;
 472
 473         ce = mb_cache_entry_get(ext4_xattr_cache, bh->b_bdev, bh->b_blocknr);
 474         error = ext4_journal_get_write_access(handle, bh);
 475         if (error)
 476                 goto out;
 477
 478         lock_buffer(bh);
 479         if (BHDR(bh)->h_refcount == cpu_to_le32(1)) {
 480                 ea_bdebug(bh, "refcount now=0; freeing");
 481                 if (ce)
 482                         mb_cache_entry_free(ce);
 483                 ext4_free_blocks(handle, inode, bh->b_blocknr, 1, 1);
 484                 get_bh(bh);
 485                 ext4_forget(handle, 1, inode, bh, bh->b_blocknr);
 486         } else {
 487                 le32_add_cpu(&BHDR(bh)->h_refcount, -1);
 488                 error = ext4_journal_dirty_metadata(handle, bh);
 489                 if (IS_SYNC(inode))
 490                         handle->h_sync = 1;
 491                 DQUOT_FREE_BLOCK(inode, 1);
 492                 ea_bdebug(bh, "refcount now=%d; releasing",
 493                           le32_to_cpu(BHDR(bh)->h_refcount));
 494                 if (ce)
 495                         mb_cache_entry_release(ce);
 496         }
 497         unlock_buffer(bh);
 498 out:
 499         ext4_std_error(inode->i_sb, error);
 500         return;
 501 }
 502
 503 /*
 504  * Find the available free space for EAs. This also returns the total number of
 505  * bytes used by EA entries.
 506  */
 507 static size_t ext4_xattr_free_space(struct ext4_xattr_entry *last,
 508                                     size_t *min_offs, void *base, int *total)
 509 {
 510         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 511                 *total += EXT4_XATTR_LEN(last->e_name_len);
 512                 if (!last->e_value_block && last->e_value_size) {
 513                         size_t offs = le16_to_cpu(last->e_value_offs);
 514                         if (offs < *min_offs)
 515                                 *min_offs = offs;
 516                 }
 517         }
 518         return (*min_offs - ((void *)last - base) - sizeof(__u32));
 519 }
 520
 521 struct ext4_xattr_info {
 522         int name_index;
 523         const char *name;
 524         const void *value;
 525         size_t value_len;
 526 };
 527
 528 struct ext4_xattr_search {
 529         struct ext4_xattr_entry *first;
 530         void *base;
 531         void *end;
 532         struct ext4_xattr_entry *here;
 533         int not_found;
 534 };
 535
 536 static int
 537 ext4_xattr_set_entry(struct ext4_xattr_info *i, struct ext4_xattr_search *s)
 538 {
 539         struct ext4_xattr_entry *last;
 540         size_t free, min_offs = s->end - s->base, name_len = strlen(i->name);
 541
 542         /* Compute min_offs and last. */
 543         last = s->first;
 544         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 545                 if (!last->e_value_block && last->e_value_size) {
 546                         size_t offs = le16_to_cpu(last->e_value_offs);
 547                         if (offs < min_offs)
 548                                 min_offs = offs;
 549                 }
 550         }
 551         free = min_offs - ((void *)last - s->base) - sizeof(__u32);
 552         if (!s->not_found) {
 553                 if (!s->here->e_value_block && s->here->e_value_size) {
 554                         size_t size = le32_to_cpu(s->here->e_value_size);
 555                         free += EXT4_XATTR_SIZE(size);
 556                 }
 557                 free += EXT4_XATTR_LEN(name_len);
 558         }
 559         if (i->value) {
 560                 if (free < EXT4_XATTR_SIZE(i->value_len) ||
 561                     free < EXT4_XATTR_LEN(name_len) +
 562                            EXT4_XATTR_SIZE(i->value_len))
 563                         return -ENOSPC;
 564         }
 565
 566         if (i->value && s->not_found) {
 567                 /* Insert the new name. */
 568                 size_t size = EXT4_XATTR_LEN(name_len);
 569                 size_t rest = (void *)last - (void *)s->here + sizeof(__u32);
 570                 memmove((void *)s->here + size, s->here, rest);
 571                 memset(s->here, 0, size);
 572                 s->here->e_name_index = i->name_index;
 573                 s->here->e_name_len = name_len;
 574                 memcpy(s->here->e_name, i->name, name_len);
 575         } else {
 576                 if (!s->here->e_value_block && s->here->e_value_size) {
 577                         void *first_val = s->base + min_offs;
 578                         size_t offs = le16_to_cpu(s->here->e_value_offs);
 579                         void *val = s->base + offs;
 580                         size_t size = EXT4_XATTR_SIZE(
 581                                 le32_to_cpu(s->here->e_value_size));
 582
 583                         if (i->value && size == EXT4_XATTR_SIZE(i->value_len)) {
 584                                 /* The old and the new value have the same
 585                                    size. Just replace. */
 586                                 s->here->e_value_size =
 587                                         cpu_to_le32(i->value_len);
 588                                 memset(val + size - EXT4_XATTR_PAD, 0,
 589                                        EXT4_XATTR_PAD); /* Clear pad bytes. */
 590                                 memcpy(val, i->value, i->value_len);
 591                                 return 0;
 592                         }
 593
 594                         /* Remove the old value. */
 595                         memmove(first_val + size, first_val, val - first_val);
 596                         memset(first_val, 0, size);
 597                         s->here->e_value_size = 0;
 598                         s->here->e_value_offs = 0;
 599                         min_offs += size;
 600
 601                         /* Adjust all value offsets. */
 602                         last = s->first;
 603                         while (!IS_LAST_ENTRY(last)) {
 604                                 size_t o = le16_to_cpu(last->e_value_offs);
 605                                 if (!last->e_value_block &&
 606                                     last->e_value_size && o < offs)
 607                                         last->e_value_offs =
 608                                                 cpu_to_le16(o + size);
 609                                 last = EXT4_XATTR_NEXT(last);
 610                         }
 611                 }
 612                 if (!i->value) {
 613                         /* Remove the old name. */
 614                         size_t size = EXT4_XATTR_LEN(name_len);
 615                         last = ENTRY((void *)last - size);
 616                         memmove(s->here, (void *)s->here + size,
 617                                 (void *)last - (void *)s->here + sizeof(__u32));
 618                         memset(last, 0, size);
 619                 }
 620         }
 621
 622         if (i->value) {
 623                 /* Insert the new value. */
 624                 s->here->e_value_size = cpu_to_le32(i->value_len);
 625                 if (i->value_len) {
 626                         size_t size = EXT4_XATTR_SIZE(i->value_len);
 627                         void *val = s->base + min_offs - size;
 628                         s->here->e_value_offs = cpu_to_le16(min_offs - size);
 629                         memset(val + size - EXT4_XATTR_PAD, 0,
 630                                EXT4_XATTR_PAD); /* Clear the pad bytes. */
 631                         memcpy(val, i->value, i->value_len);
 632                 }
 633         }
 634         return 0;
 635 }
 636
 637 struct ext4_xattr_block_find {
 638         struct ext4_xattr_search s;
 639         struct buffer_head *bh;
 640 };
 641
 642 static int
 643 ext4_xattr_block_find(struct inode *inode, struct ext4_xattr_info *i,
 644                       struct ext4_xattr_block_find *bs)
 645 {
 646         struct super_block *sb = inode->i_sb;
 647         int error;
 648
 649         ea_idebug(inode, "name=%d.%s, value=%p, value_len=%ld",
 650                   i->name_index, i->name, i->value, (long)i->value_len);
 651
 652         if (EXT4_I(inode)->i_file_acl) {
 653                 /* The inode already has an extended attribute block. */
 654                 bs->bh = sb_bread(sb, EXT4_I(inode)->i_file_acl);
 655                 error = -EIO;
 656                 if (!bs->bh)
 657                         goto cleanup;
 658                 ea_bdebug(bs->bh, "b_count=%d, refcount=%d",
 659                         atomic_read(&(bs->bh->b_count)),
 660                         le32_to_cpu(BHDR(bs->bh)->h_refcount));
 661                 if (ext4_xattr_check_block(bs->bh)) {
 662                         ext4_error(sb, __FUNCTION__,
 663                                 "inode %lu: bad block %llu", inode->i_ino,
 664                                 EXT4_I(inode)->i_file_acl);
 665                         error = -EIO;
 666                         goto cleanup;
 667                 }
 668                 /* Find the named attribute. */
 669                 bs->s.base = BHDR(bs->bh);
 670                 bs->s.first = BFIRST(bs->bh);
 671                 bs->s.end = bs->bh->b_data + bs->bh->b_size;
 672                 bs->s.here = bs->s.first;
 673                 error = ext4_xattr_find_entry(&bs->s.here, i->name_index,
 674                                               i->name, bs->bh->b_size, 1);
 675                 if (error && error != -ENODATA)
 676                         goto cleanup;
 677                 bs->s.not_found = error;
 678         }
 679         error = 0;
 680
 681 cleanup:
 682         return error;
 683 }
 684
 685 static int
 686 ext4_xattr_block_set(handle_t *handle, struct inode *inode,
 687                      struct ext4_xattr_info *i,
 688                      struct ext4_xattr_block_find *bs)
 689 {
 690         struct super_block *sb = inode->i_sb;
 691         struct buffer_head *new_bh = NULL;
 692         struct ext4_xattr_search *s = &bs->s;
 693         struct mb_cache_entry *ce = NULL;
 694         int error = 0;
 695
 696 #define header(x) ((struct ext4_xattr_header *)(x))
 697
 698         if (i->value && i->value_len > sb->s_blocksize)
 699                 return -ENOSPC;
 700         if (s->base) {
 701                 ce = mb_cache_entry_get(ext4_xattr_cache, bs->bh->b_bdev,
 702                                         bs->bh->b_blocknr);
 703                 error = ext4_journal_get_write_access(handle, bs->bh);
 704                 if (error)
 705                         goto cleanup;
 706                 lock_buffer(bs->bh);
 707
 708                 if (header(s->base)->h_refcount == cpu_to_le32(1)) {
 709                         if (ce) {
 710                                 mb_cache_entry_free(ce);
 711                                 ce = NULL;
 712                         }
 713                         ea_bdebug(bs->bh, "modifying in-place");
 714                         error = ext4_xattr_set_entry(i, s);
 715                         if (!error) {
 716                                 if (!IS_LAST_ENTRY(s->first))
 717                                         ext4_xattr_rehash(header(s->base),
 718                                                           s->here);
 719                                 ext4_xattr_cache_insert(bs->bh);
 720                         }
 721                         unlock_buffer(bs->bh);
 722                         if (error == -EIO)
 723                                 goto bad_block;
 724                         if (!error)
 725                                 error = ext4_journal_dirty_metadata(handle,
 726                                                                     bs->bh);
 727                         if (error)
 728                                 goto cleanup;
 729                         goto inserted;
 730                 } else {
 731                         int offset = (char *)s->here - bs->bh->b_data;
 732
 733                         unlock_buffer(bs->bh);
 734                         jbd2_journal_release_buffer(handle, bs->bh);
 735                         if (ce) {
 736                                 mb_cache_entry_release(ce);
 737                                 ce = NULL;
 738                         }
 739                         ea_bdebug(bs->bh, "cloning");
 740                         s->base = kmalloc(bs->bh->b_size, GFP_KERNEL);
 741                         error = -ENOMEM;
 742                         if (s->base == NULL)
 743                                 goto cleanup;
 744                         memcpy(s->base, BHDR(bs->bh), bs->bh->b_size);
 745                         s->first = ENTRY(header(s->base)+1);
 746                         header(s->base)->h_refcount = cpu_to_le32(1);
 747                         s->here = ENTRY(s->base + offset);
 748                         s->end = s->base + bs->bh->b_size;
 749                 }
 750         } else {
 751                 /* Allocate a buffer where we construct the new block. */
 752                 s->base = kzalloc(sb->s_blocksize, GFP_KERNEL);
 753                 /* assert(header == s->base) */
 754                 error = -ENOMEM;
 755                 if (s->base == NULL)
 756                         goto cleanup;
 757                 header(s->base)->h_magic = cpu_to_le32(EXT4_XATTR_MAGIC);
 758                 header(s->base)->h_blocks = cpu_to_le32(1);
 759                 header(s->base)->h_refcount = cpu_to_le32(1);
 760                 s->first = ENTRY(header(s->base)+1);
 761                 s->here = ENTRY(header(s->base)+1);
 762                 s->end = s->base + sb->s_blocksize;
 763         }
 764
 765         error = ext4_xattr_set_entry(i, s);
 766         if (error == -EIO)
 767                 goto bad_block;
 768         if (error)
 769                 goto cleanup;
 770         if (!IS_LAST_ENTRY(s->first))
 771                 ext4_xattr_rehash(header(s->base), s->here);
 772
 773 inserted:
 774         if (!IS_LAST_ENTRY(s->first)) {
 775                 new_bh = ext4_xattr_cache_find(inode, header(s->base), &ce);
 776                 if (new_bh) {
 777                         /* We found an identical block in the cache. */
 778                         if (new_bh == bs->bh)
 779                                 ea_bdebug(new_bh, "keeping");
 780                         else {
 781                                 /* The old block is released after updating
 782                                    the inode. */
 783                                 error = -EDQUOT;
 784                                 if (DQUOT_ALLOC_BLOCK(inode, 1))
 785                                         goto cleanup;
 786                                 error = ext4_journal_get_write_access(handle,
 787                                                                       new_bh);
 788                                 if (error)
 789                                         goto cleanup_dquot;
 790                                 lock_buffer(new_bh);
 791                                 le32_add_cpu(&BHDR(new_bh)->h_refcount, 1);
 792                                 ea_bdebug(new_bh, "reusing; refcount now=%d",
 793                                         le32_to_cpu(BHDR(new_bh)->h_refcount));
 794                                 unlock_buffer(new_bh);
 795                                 error = ext4_journal_dirty_metadata(handle,
 796                                                                     new_bh);
 797                                 if (error)
 798                                         goto cleanup_dquot;
 799                         }
 800                         mb_cache_entry_release(ce);
 801                         ce = NULL;
 802                 } else if (bs->bh && s->base == bs->bh->b_data) {
 803                         /* We were modifying this block in-place. */
 804                         ea_bdebug(bs->bh, "keeping this block");
 805                         new_bh = bs->bh;
 806                         get_bh(new_bh);
 807                 } else {
 808                         /* We need to allocate a new block */
 809                         ext4_fsblk_t goal = le32_to_cpu(
 810                                         EXT4_SB(sb)->s_es->s_first_data_block) +
 811                                 (ext4_fsblk_t)EXT4_I(inode)->i_block_group *
 812                                 EXT4_BLOCKS_PER_GROUP(sb);
 813                         ext4_fsblk_t block = ext4_new_block(handle, inode,
 814                                                         goal, &error);
 815                         if (error)
 816                                 goto cleanup;
 817                         ea_idebug(inode, "creating block %d", block);
 818
 819                         new_bh = sb_getblk(sb, block);
 820                         if (!new_bh) {
 821 getblk_failed:
 822                                 ext4_free_blocks(handle, inode, block, 1, 1);
 823                                 error = -EIO;
 824                                 goto cleanup;
 825                         }
 826                         lock_buffer(new_bh);
 827                         error = ext4_journal_get_create_access(handle, new_bh);
 828                         if (error) {
 829                                 unlock_buffer(new_bh);
 830                                 goto getblk_failed;
 831                         }
 832                         memcpy(new_bh->b_data, s->base, new_bh->b_size);
 833                         set_buffer_uptodate(new_bh);
 834                         unlock_buffer(new_bh);
 835                         ext4_xattr_cache_insert(new_bh);
 836                         error = ext4_journal_dirty_metadata(handle, new_bh);
 837                         if (error)
 838                                 goto cleanup;
 839                 }
 840         }
 841
 842         /* Update the inode. */
 843         EXT4_I(inode)->i_file_acl = new_bh ? new_bh->b_blocknr : 0;
 844
 845         /* Drop the previous xattr block. */
 846         if (bs->bh && bs->bh != new_bh)
 847                 ext4_xattr_release_block(handle, inode, bs->bh);
 848         error = 0;
 849
 850 cleanup:
 851         if (ce)
 852                 mb_cache_entry_release(ce);
 853         brelse(new_bh);
 854         if (!(bs->bh && s->base == bs->bh->b_data))
 855                 kfree(s->base);
 856
 857         return error;
 858
 859 cleanup_dquot:
 860         DQUOT_FREE_BLOCK(inode, 1);
 861         goto cleanup;
 862
 863 bad_block:
 864         ext4_error(inode->i_sb, __FUNCTION__,
 865                    "inode %lu: bad block %llu", inode->i_ino,
 866                    EXT4_I(inode)->i_file_acl);
 867         goto cleanup;
 868
 869 #undef header
 870 }
 871
 872 struct ext4_xattr_ibody_find {
 873         struct ext4_xattr_search s;
 874         struct ext4_iloc iloc;
 875 };
 876
 877 static int
 878 ext4_xattr_ibody_find(struct inode *inode, struct ext4_xattr_info *i,
 879                       struct ext4_xattr_ibody_find *is)
 880 {
 881         struct ext4_xattr_ibody_header *header;
 882         struct ext4_inode *raw_inode;
 883         int error;
 884
 885         if (EXT4_I(inode)->i_extra_isize == 0)
 886                 return 0;
 887         raw_inode = ext4_raw_inode(&is->iloc);
 888         header = IHDR(inode, raw_inode);
 889         is->s.base = is->s.first = IFIRST(header);
 890         is->s.here = is->s.first;
 891         is->s.end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
 892         if (EXT4_I(inode)->i_state & EXT4_STATE_XATTR) {
 893                 error = ext4_xattr_check_names(IFIRST(header), is->s.end);
 894                 if (error)
 895                         return error;
 896                 /* Find the named attribute. */
 897                 error = ext4_xattr_find_entry(&is->s.here, i->name_index,
 898                                               i->name, is->s.end -
 899                                               (void *)is->s.base, 0);
 900                 if (error && error != -ENODATA)
 901                         return error;
 902                 is->s.not_found = error;
 903         }
 904         return 0;
 905 }
 906
 907 static int
 908 ext4_xattr_ibody_set(handle_t *handle, struct inode *inode,
 909                      struct ext4_xattr_info *i,
 910                      struct ext4_xattr_ibody_find *is)
 911 {
 912         struct ext4_xattr_ibody_header *header;
 913         struct ext4_xattr_search *s = &is->s;
 914         int error;
 915
 916         if (EXT4_I(inode)->i_extra_isize == 0)
 917                 return -ENOSPC;
 918         error = ext4_xattr_set_entry(i, s);
 919         if (error)
 920                 return error;
 921         header = IHDR(inode, ext4_raw_inode(&is->iloc));
 922         if (!IS_LAST_ENTRY(s->first)) {
 923                 header->h_magic = cpu_to_le32(EXT4_XATTR_MAGIC);
 924                 EXT4_I(inode)->i_state |= EXT4_STATE_XATTR;
 925         } else {
 926                 header->h_magic = cpu_to_le32(0);
 927                 EXT4_I(inode)->i_state &= ~EXT4_STATE_XATTR;
 928         }
 929         return 0;
 930 }
 931
 932 /*
 933  * ext4_xattr_set_handle()
 934  *
 935  * Create, replace or remove an extended attribute for this inode. Buffer
 936  * is NULL to remove an existing extended attribute, and non-NULL to
 937  * either replace an existing extended attribute, or create a new extended
 938  * attribute. The flags XATTR_REPLACE and XATTR_CREATE
 939  * specify that an extended attribute must exist and must not exist
 940  * previous to the call, respectively.
 941  *
 942  * Returns 0, or a negative error number on failure.
 943  */
 944 int
 945 ext4_xattr_set_handle(handle_t *handle, struct inode *inode, int name_index,
 946                       const char *name, const void *value, size_t value_len,
 947                       int flags)
 948 {
 949         struct ext4_xattr_info i = {
 950                 .name_index = name_index,
 951                 .name = name,
 952                 .value = value,
 953                 .value_len = value_len,
 954
 955         };
 956         struct ext4_xattr_ibody_find is = {
 957                 .s = { .not_found = -ENODATA, },
 958         };
 959         struct ext4_xattr_block_find bs = {
 960                 .s = { .not_found = -ENODATA, },
 961         };
 962         int error;
 963
 964         if (!name)
 965                 return -EINVAL;
 966         if (strlen(name) > 255)
 967                 return -ERANGE;
 968         down_write(&EXT4_I(inode)->xattr_sem);
 969         error = ext4_get_inode_loc(inode, &is.iloc);
 970         if (error)
 971                 goto cleanup;
 972
 973         if (EXT4_I(inode)->i_state & EXT4_STATE_NEW) {
 974                 struct ext4_inode *raw_inode = ext4_raw_inode(&is.iloc);
 975                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
 976                 EXT4_I(inode)->i_state &= ~EXT4_STATE_NEW;
 977         }
 978
 979         error = ext4_xattr_ibody_find(inode, &i, &is);
 980         if (error)
 981                 goto cleanup;
 982         if (is.s.not_found)
 983                 error = ext4_xattr_block_find(inode, &i, &bs);
 984         if (error)
 985                 goto cleanup;
 986         if (is.s.not_found && bs.s.not_found) {
 987                 error = -ENODATA;
 988                 if (flags & XATTR_REPLACE)
 989                         goto cleanup;
 990                 error = 0;
 991                 if (!value)
 992                         goto cleanup;
 993         } else {
 994                 error = -EEXIST;
 995                 if (flags & XATTR_CREATE)
 996                         goto cleanup;
 997         }
 998         error = ext4_journal_get_write_access(handle, is.iloc.bh);
 999         if (error)
1000                 goto cleanup;
1001         if (!value) {
1002                 if (!is.s.not_found)
1003                         error = ext4_xattr_ibody_set(handle, inode, &i, &is);
1004                 else if (!bs.s.not_found)
1005                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
1006         } else {
1007                 error = ext4_xattr_ibody_set(handle, inode, &i, &is);
1008                 if (!error && !bs.s.not_found) {
1009                         i.value = NULL;
1010                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
1011                 } else if (error == -ENOSPC) {
1012                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
1013                         if (error)
1014                                 goto cleanup;
1015                         if (!is.s.not_found) {
1016                                 i.value = NULL;
1017                                 error = ext4_xattr_ibody_set(handle, inode, &i,
1018                                                              &is);
1019                         }
1020                 }
1021         }
1022         if (!error) {
1023                 ext4_xattr_update_super_block(handle, inode->i_sb);
1024                 inode->i_ctime = ext4_current_time(inode);
1025                 if (!value)
1026                         EXT4_I(inode)->i_state &= ~EXT4_STATE_NO_EXPAND;
1027                 error = ext4_mark_iloc_dirty(handle, inode, &is.iloc);
1028                 /*
1029                  * The bh is consumed by ext4_mark_iloc_dirty, even with
1030                  * error != 0.
1031                  */
1032                 is.iloc.bh = NULL;
1033                 if (IS_SYNC(inode))
1034                         handle->h_sync = 1;
1035         }
1036
1037 cleanup:
1038         brelse(is.iloc.bh);
1039         brelse(bs.bh);
1040         up_write(&EXT4_I(inode)->xattr_sem);
1041         return error;
1042 }
1043
1044 /*
1045  * ext4_xattr_set()
1046  *
1047  * Like ext4_xattr_set_handle, but start from an inode. This extended
1048  * attribute modification is a filesystem transaction by itself.
1049  *
1050  * Returns 0, or a negative error number on failure.
1051  */
1052 int
1053 ext4_xattr_set(struct inode *inode, int name_index, const char *name,
1054                const void *value, size_t value_len, int flags)
1055 {
1056         handle_t *handle;
1057         int error, retries = 0;
1058
1059 retry:
1060         handle = ext4_journal_start(inode, EXT4_DATA_TRANS_BLOCKS(inode->i_sb));
1061         if (IS_ERR(handle)) {
1062                 error = PTR_ERR(handle);
1063         } else {
1064                 int error2;
1065
1066                 error = ext4_xattr_set_handle(handle, inode, name_index, name,
1067                                               value, value_len, flags);
1068                 error2 = ext4_journal_stop(handle);
1069                 if (error == -ENOSPC &&
1070                     ext4_should_retry_alloc(inode->i_sb, &retries))
1071                         goto retry;
1072                 if (error == 0)
1073                         error = error2;
1074         }
1075
1076         return error;
1077 }
1078
1079 /*
1080  * Shift the EA entries in the inode to create space for the increased
1081  * i_extra_isize.
1082  */
1083 static void ext4_xattr_shift_entries(struct ext4_xattr_entry *entry,
1084                                      int value_offs_shift, void *to,
1085                                      void *from, size_t n, int blocksize)
1086 {
1087         struct ext4_xattr_entry *last = entry;
1088         int new_offs;
1089
1090         /* Adjust the value offsets of the entries */
1091         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
1092                 if (!last->e_value_block && last->e_value_size) {
1093                         new_offs = le16_to_cpu(last->e_value_offs) +
1094                                                         value_offs_shift;
1095                         BUG_ON(new_offs + le32_to_cpu(last->e_value_size)
1096                                  > blocksize);
1097                         last->e_value_offs = cpu_to_le16(new_offs);
1098                 }
1099         }
1100         /* Shift the entries by n bytes */
1101         memmove(to, from, n);
1102 }
1103
1104 /*
1105  * Expand an inode by new_extra_isize bytes when EAs are present.
1106  * Returns 0 on success or negative error number on failure.
1107  */
1108 int ext4_expand_extra_isize_ea(struct inode *inode, int new_extra_isize,
1109                                struct ext4_inode *raw_inode, handle_t *handle)
1110 {
1111         struct ext4_xattr_ibody_header *header;
1112         struct ext4_xattr_entry *entry, *last, *first;
1113         struct buffer_head *bh = NULL;
1114         struct ext4_xattr_ibody_find *is = NULL;
1115         struct ext4_xattr_block_find *bs = NULL;
1116         char *buffer = NULL, *b_entry_name = NULL;
1117         size_t min_offs, free;
1118         int total_ino, total_blk;
1119         void *base, *start, *end;
1120         int extra_isize = 0, error = 0, tried_min_extra_isize = 0;
1121         int s_min_extra_isize = le16_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_min_extra_isize);
1122
1123         down_write(&EXT4_I(inode)->xattr_sem);
1124 retry:
1125         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize) {
1126                 up_write(&EXT4_I(inode)->xattr_sem);
1127                 return 0;
1128         }
1129
1130         header = IHDR(inode, raw_inode);
1131         entry = IFIRST(header);
1132
1133         /*
1134          * Check if enough free space is available in the inode to shift the
1135          * entries ahead by new_extra_isize.
1136          */
1137
1138         base = start = entry;
1139         end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
1140         min_offs = end - base;
1141         last = entry;
1142         total_ino = sizeof(struct ext4_xattr_ibody_header);
1143
1144         free = ext4_xattr_free_space(last, &min_offs, base, &total_ino);
1145         if (free >= new_extra_isize) {
1146                 entry = IFIRST(header);
1147                 ext4_xattr_shift_entries(entry, EXT4_I(inode)->i_extra_isize
1148                                 - new_extra_isize, (void *)raw_inode +
1149                                 EXT4_GOOD_OLD_INODE_SIZE + new_extra_isize,
1150                                 (void *)header, total_ino,
1151                                 inode->i_sb->s_blocksize);
1152                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
1153                 error = 0;
1154                 goto cleanup;
1155         }
1156
1157         /*
1158          * Enough free space isn't available in the inode, check if
1159          * EA block can hold new_extra_isize bytes.
1160          */
1161         if (EXT4_I(inode)->i_file_acl) {
1162                 bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
1163                 error = -EIO;
1164                 if (!bh)
1165                         goto cleanup;
1166                 if (ext4_xattr_check_block(bh)) {
1167                         ext4_error(inode->i_sb, __FUNCTION__,
1168                                 "inode %lu: bad block %llu", inode->i_ino,
1169                                 EXT4_I(inode)->i_file_acl);
1170                         error = -EIO;
1171                         goto cleanup;
1172                 }
1173                 base = BHDR(bh);
1174                 first = BFIRST(bh);
1175                 end = bh->b_data + bh->b_size;
1176                 min_offs = end - base;
1177                 free = ext4_xattr_free_space(first, &min_offs, base,
1178                                              &total_blk);
1179                 if (free < new_extra_isize) {
1180                         if (!tried_min_extra_isize && s_min_extra_isize) {
1181                                 tried_min_extra_isize++;
1182                                 new_extra_isize = s_min_extra_isize;
1183                                 brelse(bh);
1184                                 goto retry;
1185                         }
1186                         error = -1;
1187                         goto cleanup;
1188                 }
1189         } else {
1190                 free = inode->i_sb->s_blocksize;
1191         }
1192
1193         while (new_extra_isize > 0) {
1194                 size_t offs, size, entry_size;
1195                 struct ext4_xattr_entry *small_entry = NULL;
1196                 struct ext4_xattr_info i = {
1197                         .value = NULL,
1198                         .value_len = 0,
1199                 };
1200                 unsigned int total_size;  /* EA entry size + value size */
1201                 unsigned int shift_bytes; /* No. of bytes to shift EAs by? */
1202                 unsigned int min_total_size = ~0U;
1203
1204                 is = kzalloc(sizeof(struct ext4_xattr_ibody_find), GFP_NOFS);
1205                 bs = kzalloc(sizeof(struct ext4_xattr_block_find), GFP_NOFS);
1206                 if (!is || !bs) {
1207                         error = -ENOMEM;
1208                         goto cleanup;
1209                 }
1210
1211                 is->s.not_found = -ENODATA;
1212                 bs->s.not_found = -ENODATA;
1213                 is->iloc.bh = NULL;
1214                 bs->bh = NULL;
1215
1216                 last = IFIRST(header);
1217                 /* Find the entry best suited to be pushed into EA block */
1218                 entry = NULL;
1219                 for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
1220                         total_size =
1221                         EXT4_XATTR_SIZE(le32_to_cpu(last->e_value_size)) +
1222                                         EXT4_XATTR_LEN(last->e_name_len);
1223                         if (total_size <= free && total_size < min_total_size) {
1224                                 if (total_size < new_extra_isize) {
1225                                         small_entry = last;
1226                                 } else {
1227                                         entry = last;
1228                                         min_total_size = total_size;
1229                                 }
1230                         }
1231                 }
1232
1233                 if (entry == NULL) {
1234                         if (small_entry) {
1235                                 entry = small_entry;
1236                         } else {
1237                                 if (!tried_min_extra_isize &&
1238                                     s_min_extra_isize) {
1239                                         tried_min_extra_isize++;
1240                                         new_extra_isize = s_min_extra_isize;
1241                                         goto retry;
1242                                 }
1243                                 error = -1;
1244                                 goto cleanup;
1245                         }
1246                 }
1247                 offs = le16_to_cpu(entry->e_value_offs);
1248                 size = le32_to_cpu(entry->e_value_size);
1249                 entry_size = EXT4_XATTR_LEN(entry->e_name_len);
1250                 i.name_index = entry->e_name_index,
1251                 buffer = kmalloc(EXT4_XATTR_SIZE(size), GFP_NOFS);
1252                 b_entry_name = kmalloc(entry->e_name_len + 1, GFP_NOFS);
1253                 if (!buffer || !b_entry_name) {
1254                         error = -ENOMEM;
1255                         goto cleanup;
1256                 }
1257                 /* Save the entry name and the entry value */
1258                 memcpy(buffer, (void *)IFIRST(header) + offs,
1259                        EXT4_XATTR_SIZE(size));
1260                 memcpy(b_entry_name, entry->e_name, entry->e_name_len);
1261                 b_entry_name[entry->e_name_len] = '\0';
1262                 i.name = b_entry_name;
1263
1264                 error = ext4_get_inode_loc(inode, &is->iloc);
1265                 if (error)
1266                         goto cleanup;
1267
1268                 error = ext4_xattr_ibody_find(inode, &i, is);
1269                 if (error)
1270                         goto cleanup;
1271
1272                 /* Remove the chosen entry from the inode */
1273                 error = ext4_xattr_ibody_set(handle, inode, &i, is);
1274
1275                 entry = IFIRST(header);
1276                 if (entry_size + EXT4_XATTR_SIZE(size) >= new_extra_isize)
1277                         shift_bytes = new_extra_isize;
1278                 else
1279                         shift_bytes = entry_size + size;
1280                 /* Adjust the offsets and shift the remaining entries ahead */
1281                 ext4_xattr_shift_entries(entry, EXT4_I(inode)->i_extra_isize -
1282                         shift_bytes, (void *)raw_inode +
1283                         EXT4_GOOD_OLD_INODE_SIZE + extra_isize + shift_bytes,
1284                         (void *)header, total_ino - entry_size,
1285                         inode->i_sb->s_blocksize);
1286
1287                 extra_isize += shift_bytes;
1288                 new_extra_isize -= shift_bytes;
1289                 EXT4_I(inode)->i_extra_isize = extra_isize;
1290
1291                 i.name = b_entry_name;
1292                 i.value = buffer;
1293                 i.value_len = size;
1294                 error = ext4_xattr_block_find(inode, &i, bs);
1295                 if (error)
1296                         goto cleanup;
1297
1298                 /* Add entry which was removed from the inode into the block */
1299                 error = ext4_xattr_block_set(handle, inode, &i, bs);
1300                 if (error)
1301                         goto cleanup;
1302                 kfree(b_entry_name);
1303                 kfree(buffer);
1304                 brelse(is->iloc.bh);
1305                 kfree(is);
1306                 kfree(bs);
1307         }
1308         brelse(bh);
1309         up_write(&EXT4_I(inode)->xattr_sem);
1310         return 0;
1311
1312 cleanup:
1313         kfree(b_entry_name);
1314         kfree(buffer);
1315         if (is)
1316                 brelse(is->iloc.bh);
1317         kfree(is);
1318         kfree(bs);
1319         brelse(bh);
1320         up_write(&EXT4_I(inode)->xattr_sem);
1321         return error;
1322 }
1323
1324
1325
1326 /*
1327  * ext4_xattr_delete_inode()
1328  *
1329  * Free extended attribute resources associated with this inode. This
1330  * is called immediately before an inode is freed. We have exclusive
1331  * access to the inode.
1332  */
1333 void
1334 ext4_xattr_delete_inode(handle_t *handle, struct inode *inode)
1335 {
1336         struct buffer_head *bh = NULL;
1337
1338         if (!EXT4_I(inode)->i_file_acl)
1339                 goto cleanup;
1340         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
1341         if (!bh) {
1342                 ext4_error(inode->i_sb, __FUNCTION__,
1343                         "inode %lu: block %llu read error", inode->i_ino,
1344                         EXT4_I(inode)->i_file_acl);
1345                 goto cleanup;
1346         }
1347         if (BHDR(bh)->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC) ||
1348             BHDR(bh)->h_blocks != cpu_to_le32(1)) {
1349                 ext4_error(inode->i_sb, __FUNCTION__,
1350                         "inode %lu: bad block %llu", inode->i_ino,
1351                         EXT4_I(inode)->i_file_acl);
1352                 goto cleanup;
1353         }
1354         ext4_xattr_release_block(handle, inode, bh);
1355         EXT4_I(inode)->i_file_acl = 0;
1356
1357 cleanup:
1358         brelse(bh);
1359 }
1360
1361 /*
1362  * ext4_xattr_put_super()
1363  *
1364  * This is called when a file system is unmounted.
1365  */
1366 void
1367 ext4_xattr_put_super(struct super_block *sb)
1368 {
1369         mb_cache_shrink(sb->s_bdev);
1370 }
1371
1372 /*
1373  * ext4_xattr_cache_insert()
1374  *
1375  * Create a new entry in the extended attribute cache, and insert
1376  * it unless such an entry is already in the cache.
1377  *
1378  * Returns 0, or a negative error number on failure.
1379  */
1380 static void
1381 ext4_xattr_cache_insert(struct buffer_head *bh)
1382 {
1383         __u32 hash = le32_to_cpu(BHDR(bh)->h_hash);
1384         struct mb_cache_entry *ce;
1385         int error;
1386
1387         ce = mb_cache_entry_alloc(ext4_xattr_cache, GFP_NOFS);
1388         if (!ce) {
1389                 ea_bdebug(bh, "out of memory");
1390                 return;
1391         }
1392         error = mb_cache_entry_insert(ce, bh->b_bdev, bh->b_blocknr, &hash);
1393         if (error) {
1394                 mb_cache_entry_free(ce);
1395                 if (error == -EBUSY) {
1396                         ea_bdebug(bh, "already in cache");
1397                         error = 0;
1398                 }
1399         } else {
1400                 ea_bdebug(bh, "inserting [%x]", (int)hash);
1401                 mb_cache_entry_release(ce);
1402         }
1403 }
1404
1405 /*
1406  * ext4_xattr_cmp()
1407  *
1408  * Compare two extended attribute blocks for equality.
1409  *
1410  * Returns 0 if the blocks are equal, 1 if they differ, and
1411  * a negative error number on errors.
1412  */
1413 static int
1414 ext4_xattr_cmp(struct ext4_xattr_header *header1,
1415                struct ext4_xattr_header *header2)
1416 {
1417         struct ext4_xattr_entry *entry1, *entry2;
1418
1419         entry1 = ENTRY(header1+1);
1420         entry2 = ENTRY(header2+1);
1421         while (!IS_LAST_ENTRY(entry1)) {
1422                 if (IS_LAST_ENTRY(entry2))
1423                         return 1;
1424                 if (entry1->e_hash != entry2->e_hash ||
1425                     entry1->e_name_index != entry2->e_name_index ||
1426                     entry1->e_name_len != entry2->e_name_len ||
1427                     entry1->e_value_size != entry2->e_value_size ||
1428                     memcmp(entry1->e_name, entry2->e_name, entry1->e_name_len))
1429                         return 1;
1430                 if (entry1->e_value_block != 0 || entry2->e_value_block != 0)
1431                         return -EIO;
1432                 if (memcmp((char *)header1 + le16_to_cpu(entry1->e_value_offs),
1433                            (char *)header2 + le16_to_cpu(entry2->e_value_offs),
1434                            le32_to_cpu(entry1->e_value_size)))
1435                         return 1;
1436
1437                 entry1 = EXT4_XATTR_NEXT(entry1);
1438                 entry2 = EXT4_XATTR_NEXT(entry2);
1439         }
1440         if (!IS_LAST_ENTRY(entry2))
1441                 return 1;
1442         return 0;
1443 }
1444
1445 /*
1446  * ext4_xattr_cache_find()
1447  *
1448  * Find an identical extended attribute block.
1449  *
1450  * Returns a pointer to the block found, or NULL if such a block was
1451  * not found or an error occurred.
1452  */
1453 static struct buffer_head *
1454 ext4_xattr_cache_find(struct inode *inode, struct ext4_xattr_header *header,
1455                       struct mb_cache_entry **pce)
1456 {
1457         __u32 hash = le32_to_cpu(header->h_hash);
1458         struct mb_cache_entry *ce;
1459
1460         if (!header->h_hash)
1461                 return NULL;  /* never share */
1462         ea_idebug(inode, "looking for cached blocks [%x]", (int)hash);
1463 again:
1464         ce = mb_cache_entry_find_first(ext4_xattr_cache, 0,
1465                                        inode->i_sb->s_bdev, hash);
1466         while (ce) {
1467                 struct buffer_head *bh;
1468
1469                 if (IS_ERR(ce)) {
1470                         if (PTR_ERR(ce) == -EAGAIN)
1471                                 goto again;
1472                         break;
1473                 }
1474                 bh = sb_bread(inode->i_sb, ce->e_block);
1475                 if (!bh) {
1476                         ext4_error(inode->i_sb, __FUNCTION__,
1477                                 "inode %lu: block %lu read error",
1478                                 inode->i_ino, (unsigned long) ce->e_block);
1479                 } else if (le32_to_cpu(BHDR(bh)->h_refcount) >=
1480                                 EXT4_XATTR_REFCOUNT_MAX) {
1481                         ea_idebug(inode, "block %lu refcount %d>=%d",
1482                                   (unsigned long) ce->e_block,
1483                                   le32_to_cpu(BHDR(bh)->h_refcount),
1484                                           EXT4_XATTR_REFCOUNT_MAX);
1485                 } else if (ext4_xattr_cmp(header, BHDR(bh)) == 0) {
1486                         *pce = ce;
1487                         return bh;
1488                 }
1489                 brelse(bh);
1490                 ce = mb_cache_entry_find_next(ce, 0, inode->i_sb->s_bdev, hash);
1491         }
1492         return NULL;
1493 }
1494
1495 #define NAME_HASH_SHIFT 5
1496 #define VALUE_HASH_SHIFT 16
1497
1498 /*
1499  * ext4_xattr_hash_entry()
1500  *
1501  * Compute the hash of an extended attribute.
1502  */
1503 static inline void ext4_xattr_hash_entry(struct ext4_xattr_header *header,
1504                                          struct ext4_xattr_entry *entry)
1505 {
1506         __u32 hash = 0;
1507         char *name = entry->e_name;
1508         int n;
1509
1510         for (n=0; n < entry->e_name_len; n++) {
1511                 hash = (hash << NAME_HASH_SHIFT) ^
1512                        (hash >> (8*sizeof(hash) - NAME_HASH_SHIFT)) ^
1513                        *name++;
1514         }
1515
1516         if (entry->e_value_block == 0 && entry->e_value_size != 0) {
1517                 __le32 *value = (__le32 *)((char *)header +
1518                         le16_to_cpu(entry->e_value_offs));
1519                 for (n = (le32_to_cpu(entry->e_value_size) +
1520                      EXT4_XATTR_ROUND) >> EXT4_XATTR_PAD_BITS; n; n--) {
1521                         hash = (hash << VALUE_HASH_SHIFT) ^
1522                                (hash >> (8*sizeof(hash) - VALUE_HASH_SHIFT)) ^
1523                                le32_to_cpu(*value++);
1524                 }
1525         }
1526         entry->e_hash = cpu_to_le32(hash);
1527 }
1528
1529 #undef NAME_HASH_SHIFT
1530 #undef VALUE_HASH_SHIFT
1531
1532 #define BLOCK_HASH_SHIFT 16
1533
1534 /*
1535  * ext4_xattr_rehash()
1536  *
1537  * Re-compute the extended attribute hash value after an entry has changed.
1538  */
1539 static void ext4_xattr_rehash(struct ext4_xattr_header *header,
1540                               struct ext4_xattr_entry *entry)
1541 {
1542         struct ext4_xattr_entry *here;
1543         __u32 hash = 0;
1544
1545         ext4_xattr_hash_entry(header, entry);
1546         here = ENTRY(header+1);
1547         while (!IS_LAST_ENTRY(here)) {
1548                 if (!here->e_hash) {
1549                         /* Block is not shared if an entry's hash value == 0 */
1550                         hash = 0;
1551                         break;
1552                 }
1553                 hash = (hash << BLOCK_HASH_SHIFT) ^
1554                        (hash >> (8*sizeof(hash) - BLOCK_HASH_SHIFT)) ^
1555                        le32_to_cpu(here->e_hash);
1556                 here = EXT4_XATTR_NEXT(here);
1557         }
1558         header->h_hash = cpu_to_le32(hash);
1559 }
1560
1561 #undef BLOCK_HASH_SHIFT
1562
1563 int __init
1564 init_ext4_xattr(void)
1565 {
1566         ext4_xattr_cache = mb_cache_create("ext4_xattr", NULL,
1567                 sizeof(struct mb_cache_entry) +
1568                 sizeof(((struct mb_cache_entry *) 0)->e_indexes[0]), 1, 6);
1569         if (!ext4_xattr_cache)
1570                 return -ENOMEM;
1571         return 0;
1572 }
1573
1574 void
1575 exit_ext4_xattr(void)
1576 {
1577         if (ext4_xattr_cache)
1578                 mb_cache_destroy(ext4_xattr_cache);
1579         ext4_xattr_cache = NULL;
1580 }