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 = ext4_group_first_block_no(sb,
 810                                                 EXT4_I(inode)->i_block_group);
 811                         ext4_fsblk_t block = ext4_new_block(handle, inode,
 812                                                         goal, &error);
 813                         if (error)
 814                                 goto cleanup;
 815                         ea_idebug(inode, "creating block %d", block);
 816
 817                         new_bh = sb_getblk(sb, block);
 818                         if (!new_bh) {
 819 getblk_failed:
 820                                 ext4_free_blocks(handle, inode, block, 1, 1);
 821                                 error = -EIO;
 822                                 goto cleanup;
 823                         }
 824                         lock_buffer(new_bh);
 825                         error = ext4_journal_get_create_access(handle, new_bh);
 826                         if (error) {
 827                                 unlock_buffer(new_bh);
 828                                 goto getblk_failed;
 829                         }
 830                         memcpy(new_bh->b_data, s->base, new_bh->b_size);
 831                         set_buffer_uptodate(new_bh);
 832                         unlock_buffer(new_bh);
 833                         ext4_xattr_cache_insert(new_bh);
 834                         error = ext4_journal_dirty_metadata(handle, new_bh);
 835                         if (error)
 836                                 goto cleanup;
 837                 }
 838         }
 839
 840         /* Update the inode. */
 841         EXT4_I(inode)->i_file_acl = new_bh ? new_bh->b_blocknr : 0;
 842
 843         /* Drop the previous xattr block. */
 844         if (bs->bh && bs->bh != new_bh)
 845                 ext4_xattr_release_block(handle, inode, bs->bh);
 846         error = 0;
 847
 848 cleanup:
 849         if (ce)
 850                 mb_cache_entry_release(ce);
 851         brelse(new_bh);
 852         if (!(bs->bh && s->base == bs->bh->b_data))
 853                 kfree(s->base);
 854
 855         return error;
 856
 857 cleanup_dquot:
 858         DQUOT_FREE_BLOCK(inode, 1);
 859         goto cleanup;
 860
 861 bad_block:
 862         ext4_error(inode->i_sb, __FUNCTION__,
 863                    "inode %lu: bad block %llu", inode->i_ino,
 864                    EXT4_I(inode)->i_file_acl);
 865         goto cleanup;
 866
 867 #undef header
 868 }
 869
 870 struct ext4_xattr_ibody_find {
 871         struct ext4_xattr_search s;
 872         struct ext4_iloc iloc;
 873 };
 874
 875 static int
 876 ext4_xattr_ibody_find(struct inode *inode, struct ext4_xattr_info *i,
 877                       struct ext4_xattr_ibody_find *is)
 878 {
 879         struct ext4_xattr_ibody_header *header;
 880         struct ext4_inode *raw_inode;
 881         int error;
 882
 883         if (EXT4_I(inode)->i_extra_isize == 0)
 884                 return 0;
 885         raw_inode = ext4_raw_inode(&is->iloc);
 886         header = IHDR(inode, raw_inode);
 887         is->s.base = is->s.first = IFIRST(header);
 888         is->s.here = is->s.first;
 889         is->s.end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
 890         if (EXT4_I(inode)->i_state & EXT4_STATE_XATTR) {
 891                 error = ext4_xattr_check_names(IFIRST(header), is->s.end);
 892                 if (error)
 893                         return error;
 894                 /* Find the named attribute. */
 895                 error = ext4_xattr_find_entry(&is->s.here, i->name_index,
 896                                               i->name, is->s.end -
 897                                               (void *)is->s.base, 0);
 898                 if (error && error != -ENODATA)
 899                         return error;
 900                 is->s.not_found = error;
 901         }
 902         return 0;
 903 }
 904
 905 static int
 906 ext4_xattr_ibody_set(handle_t *handle, struct inode *inode,
 907                      struct ext4_xattr_info *i,
 908                      struct ext4_xattr_ibody_find *is)
 909 {
 910         struct ext4_xattr_ibody_header *header;
 911         struct ext4_xattr_search *s = &is->s;
 912         int error;
 913
 914         if (EXT4_I(inode)->i_extra_isize == 0)
 915                 return -ENOSPC;
 916         error = ext4_xattr_set_entry(i, s);
 917         if (error)
 918                 return error;
 919         header = IHDR(inode, ext4_raw_inode(&is->iloc));
 920         if (!IS_LAST_ENTRY(s->first)) {
 921                 header->h_magic = cpu_to_le32(EXT4_XATTR_MAGIC);
 922                 EXT4_I(inode)->i_state |= EXT4_STATE_XATTR;
 923         } else {
 924                 header->h_magic = cpu_to_le32(0);
 925                 EXT4_I(inode)->i_state &= ~EXT4_STATE_XATTR;
 926         }
 927         return 0;
 928 }
 929
 930 /*
 931  * ext4_xattr_set_handle()
 932  *
 933  * Create, replace or remove an extended attribute for this inode. Buffer
 934  * is NULL to remove an existing extended attribute, and non-NULL to
 935  * either replace an existing extended attribute, or create a new extended
 936  * attribute. The flags XATTR_REPLACE and XATTR_CREATE
 937  * specify that an extended attribute must exist and must not exist
 938  * previous to the call, respectively.
 939  *
 940  * Returns 0, or a negative error number on failure.
 941  */
 942 int
 943 ext4_xattr_set_handle(handle_t *handle, struct inode *inode, int name_index,
 944                       const char *name, const void *value, size_t value_len,
 945                       int flags)
 946 {
 947         struct ext4_xattr_info i = {
 948                 .name_index = name_index,
 949                 .name = name,
 950                 .value = value,
 951                 .value_len = value_len,
 952
 953         };
 954         struct ext4_xattr_ibody_find is = {
 955                 .s = { .not_found = -ENODATA, },
 956         };
 957         struct ext4_xattr_block_find bs = {
 958                 .s = { .not_found = -ENODATA, },
 959         };
 960         int error;
 961
 962         if (!name)
 963                 return -EINVAL;
 964         if (strlen(name) > 255)
 965                 return -ERANGE;
 966         down_write(&EXT4_I(inode)->xattr_sem);
 967         error = ext4_get_inode_loc(inode, &is.iloc);
 968         if (error)
 969                 goto cleanup;
 970
 971         if (EXT4_I(inode)->i_state & EXT4_STATE_NEW) {
 972                 struct ext4_inode *raw_inode = ext4_raw_inode(&is.iloc);
 973                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
 974                 EXT4_I(inode)->i_state &= ~EXT4_STATE_NEW;
 975         }
 976
 977         error = ext4_xattr_ibody_find(inode, &i, &is);
 978         if (error)
 979                 goto cleanup;
 980         if (is.s.not_found)
 981                 error = ext4_xattr_block_find(inode, &i, &bs);
 982         if (error)
 983                 goto cleanup;
 984         if (is.s.not_found && bs.s.not_found) {
 985                 error = -ENODATA;
 986                 if (flags & XATTR_REPLACE)
 987                         goto cleanup;
 988                 error = 0;
 989                 if (!value)
 990                         goto cleanup;
 991         } else {
 992                 error = -EEXIST;
 993                 if (flags & XATTR_CREATE)
 994                         goto cleanup;
 995         }
 996         error = ext4_journal_get_write_access(handle, is.iloc.bh);
 997         if (error)
 998                 goto cleanup;
 999         if (!value) {
1000                 if (!is.s.not_found)
1001                         error = ext4_xattr_ibody_set(handle, inode, &i, &is);
1002                 else if (!bs.s.not_found)
1003                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
1004         } else {
1005                 error = ext4_xattr_ibody_set(handle, inode, &i, &is);
1006                 if (!error && !bs.s.not_found) {
1007                         i.value = NULL;
1008                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
1009                 } else if (error == -ENOSPC) {
1010                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
1011                         if (error)
1012                                 goto cleanup;
1013                         if (!is.s.not_found) {
1014                                 i.value = NULL;
1015                                 error = ext4_xattr_ibody_set(handle, inode, &i,
1016                                                              &is);
1017                         }
1018                 }
1019         }
1020         if (!error) {
1021                 ext4_xattr_update_super_block(handle, inode->i_sb);
1022                 inode->i_ctime = ext4_current_time(inode);
1023                 if (!value)
1024                         EXT4_I(inode)->i_state &= ~EXT4_STATE_NO_EXPAND;
1025                 error = ext4_mark_iloc_dirty(handle, inode, &is.iloc);
1026                 /*
1027                  * The bh is consumed by ext4_mark_iloc_dirty, even with
1028                  * error != 0.
1029                  */
1030                 is.iloc.bh = NULL;
1031                 if (IS_SYNC(inode))
1032                         handle->h_sync = 1;
1033         }
1034
1035 cleanup:
1036         brelse(is.iloc.bh);
1037         brelse(bs.bh);
1038         up_write(&EXT4_I(inode)->xattr_sem);
1039         return error;
1040 }
1041
1042 /*
1043  * ext4_xattr_set()
1044  *
1045  * Like ext4_xattr_set_handle, but start from an inode. This extended
1046  * attribute modification is a filesystem transaction by itself.
1047  *
1048  * Returns 0, or a negative error number on failure.
1049  */
1050 int
1051 ext4_xattr_set(struct inode *inode, int name_index, const char *name,
1052                const void *value, size_t value_len, int flags)
1053 {
1054         handle_t *handle;
1055         int error, retries = 0;
1056
1057 retry:
1058         handle = ext4_journal_start(inode, EXT4_DATA_TRANS_BLOCKS(inode->i_sb));
1059         if (IS_ERR(handle)) {
1060                 error = PTR_ERR(handle);
1061         } else {
1062                 int error2;
1063
1064                 error = ext4_xattr_set_handle(handle, inode, name_index, name,
1065                                               value, value_len, flags);
1066                 error2 = ext4_journal_stop(handle);
1067                 if (error == -ENOSPC &&
1068                     ext4_should_retry_alloc(inode->i_sb, &retries))
1069                         goto retry;
1070                 if (error == 0)
1071                         error = error2;
1072         }
1073
1074         return error;
1075 }
1076
1077 /*
1078  * Shift the EA entries in the inode to create space for the increased
1079  * i_extra_isize.
1080  */
1081 static void ext4_xattr_shift_entries(struct ext4_xattr_entry *entry,
1082                                      int value_offs_shift, void *to,
1083                                      void *from, size_t n, int blocksize)
1084 {
1085         struct ext4_xattr_entry *last = entry;
1086         int new_offs;
1087
1088         /* Adjust the value offsets of the entries */
1089         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
1090                 if (!last->e_value_block && last->e_value_size) {
1091                         new_offs = le16_to_cpu(last->e_value_offs) +
1092                                                         value_offs_shift;
1093                         BUG_ON(new_offs + le32_to_cpu(last->e_value_size)
1094                                  > blocksize);
1095                         last->e_value_offs = cpu_to_le16(new_offs);
1096                 }
1097         }
1098         /* Shift the entries by n bytes */
1099         memmove(to, from, n);
1100 }
1101
1102 /*
1103  * Expand an inode by new_extra_isize bytes when EAs are present.
1104  * Returns 0 on success or negative error number on failure.
1105  */
1106 int ext4_expand_extra_isize_ea(struct inode *inode, int new_extra_isize,
1107                                struct ext4_inode *raw_inode, handle_t *handle)
1108 {
1109         struct ext4_xattr_ibody_header *header;
1110         struct ext4_xattr_entry *entry, *last, *first;
1111         struct buffer_head *bh = NULL;
1112         struct ext4_xattr_ibody_find *is = NULL;
1113         struct ext4_xattr_block_find *bs = NULL;
1114         char *buffer = NULL, *b_entry_name = NULL;
1115         size_t min_offs, free;
1116         int total_ino, total_blk;
1117         void *base, *start, *end;
1118         int extra_isize = 0, error = 0, tried_min_extra_isize = 0;
1119         int s_min_extra_isize = le16_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_min_extra_isize);
1120
1121         down_write(&EXT4_I(inode)->xattr_sem);
1122 retry:
1123         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize) {
1124                 up_write(&EXT4_I(inode)->xattr_sem);
1125                 return 0;
1126         }
1127
1128         header = IHDR(inode, raw_inode);
1129         entry = IFIRST(header);
1130
1131         /*
1132          * Check if enough free space is available in the inode to shift the
1133          * entries ahead by new_extra_isize.
1134          */
1135
1136         base = start = entry;
1137         end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
1138         min_offs = end - base;
1139         last = entry;
1140         total_ino = sizeof(struct ext4_xattr_ibody_header);
1141
1142         free = ext4_xattr_free_space(last, &min_offs, base, &total_ino);
1143         if (free >= new_extra_isize) {
1144                 entry = IFIRST(header);
1145                 ext4_xattr_shift_entries(entry, EXT4_I(inode)->i_extra_isize
1146                                 - new_extra_isize, (void *)raw_inode +
1147                                 EXT4_GOOD_OLD_INODE_SIZE + new_extra_isize,
1148                                 (void *)header, total_ino,
1149                                 inode->i_sb->s_blocksize);
1150                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
1151                 error = 0;
1152                 goto cleanup;
1153         }
1154
1155         /*
1156          * Enough free space isn't available in the inode, check if
1157          * EA block can hold new_extra_isize bytes.
1158          */
1159         if (EXT4_I(inode)->i_file_acl) {
1160                 bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
1161                 error = -EIO;
1162                 if (!bh)
1163                         goto cleanup;
1164                 if (ext4_xattr_check_block(bh)) {
1165                         ext4_error(inode->i_sb, __FUNCTION__,
1166                                 "inode %lu: bad block %llu", inode->i_ino,
1167                                 EXT4_I(inode)->i_file_acl);
1168                         error = -EIO;
1169                         goto cleanup;
1170                 }
1171                 base = BHDR(bh);
1172                 first = BFIRST(bh);
1173                 end = bh->b_data + bh->b_size;
1174                 min_offs = end - base;
1175                 free = ext4_xattr_free_space(first, &min_offs, base,
1176                                              &total_blk);
1177                 if (free < new_extra_isize) {
1178                         if (!tried_min_extra_isize && s_min_extra_isize) {
1179                                 tried_min_extra_isize++;
1180                                 new_extra_isize = s_min_extra_isize;
1181                                 brelse(bh);
1182                                 goto retry;
1183                         }
1184                         error = -1;
1185                         goto cleanup;
1186                 }
1187         } else {
1188                 free = inode->i_sb->s_blocksize;
1189         }
1190
1191         while (new_extra_isize > 0) {
1192                 size_t offs, size, entry_size;
1193                 struct ext4_xattr_entry *small_entry = NULL;
1194                 struct ext4_xattr_info i = {
1195                         .value = NULL,
1196                         .value_len = 0,
1197                 };
1198                 unsigned int total_size;  /* EA entry size + value size */
1199                 unsigned int shift_bytes; /* No. of bytes to shift EAs by? */
1200                 unsigned int min_total_size = ~0U;
1201
1202                 is = kzalloc(sizeof(struct ext4_xattr_ibody_find), GFP_NOFS);
1203                 bs = kzalloc(sizeof(struct ext4_xattr_block_find), GFP_NOFS);
1204                 if (!is || !bs) {
1205                         error = -ENOMEM;
1206                         goto cleanup;
1207                 }
1208
1209                 is->s.not_found = -ENODATA;
1210                 bs->s.not_found = -ENODATA;
1211                 is->iloc.bh = NULL;
1212                 bs->bh = NULL;
1213
1214                 last = IFIRST(header);
1215                 /* Find the entry best suited to be pushed into EA block */
1216                 entry = NULL;
1217                 for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
1218                         total_size =
1219                         EXT4_XATTR_SIZE(le32_to_cpu(last->e_value_size)) +
1220                                         EXT4_XATTR_LEN(last->e_name_len);
1221                         if (total_size <= free && total_size < min_total_size) {
1222                                 if (total_size < new_extra_isize) {
1223                                         small_entry = last;
1224                                 } else {
1225                                         entry = last;
1226                                         min_total_size = total_size;
1227                                 }
1228                         }
1229                 }
1230
1231                 if (entry == NULL) {
1232                         if (small_entry) {
1233                                 entry = small_entry;
1234                         } else {
1235                                 if (!tried_min_extra_isize &&
1236                                     s_min_extra_isize) {
1237                                         tried_min_extra_isize++;
1238                                         new_extra_isize = s_min_extra_isize;
1239                                         goto retry;
1240                                 }
1241                                 error = -1;
1242                                 goto cleanup;
1243                         }
1244                 }
1245                 offs = le16_to_cpu(entry->e_value_offs);
1246                 size = le32_to_cpu(entry->e_value_size);
1247                 entry_size = EXT4_XATTR_LEN(entry->e_name_len);
1248                 i.name_index = entry->e_name_index,
1249                 buffer = kmalloc(EXT4_XATTR_SIZE(size), GFP_NOFS);
1250                 b_entry_name = kmalloc(entry->e_name_len + 1, GFP_NOFS);
1251                 if (!buffer || !b_entry_name) {
1252                         error = -ENOMEM;
1253                         goto cleanup;
1254                 }
1255                 /* Save the entry name and the entry value */
1256                 memcpy(buffer, (void *)IFIRST(header) + offs,
1257                        EXT4_XATTR_SIZE(size));
1258                 memcpy(b_entry_name, entry->e_name, entry->e_name_len);
1259                 b_entry_name[entry->e_name_len] = '\0';
1260                 i.name = b_entry_name;
1261
1262                 error = ext4_get_inode_loc(inode, &is->iloc);
1263                 if (error)
1264                         goto cleanup;
1265
1266                 error = ext4_xattr_ibody_find(inode, &i, is);
1267                 if (error)
1268                         goto cleanup;
1269
1270                 /* Remove the chosen entry from the inode */
1271                 error = ext4_xattr_ibody_set(handle, inode, &i, is);
1272
1273                 entry = IFIRST(header);
1274                 if (entry_size + EXT4_XATTR_SIZE(size) >= new_extra_isize)
1275                         shift_bytes = new_extra_isize;
1276                 else
1277                         shift_bytes = entry_size + size;
1278                 /* Adjust the offsets and shift the remaining entries ahead */
1279                 ext4_xattr_shift_entries(entry, EXT4_I(inode)->i_extra_isize -
1280                         shift_bytes, (void *)raw_inode +
1281                         EXT4_GOOD_OLD_INODE_SIZE + extra_isize + shift_bytes,
1282                         (void *)header, total_ino - entry_size,
1283                         inode->i_sb->s_blocksize);
1284
1285                 extra_isize += shift_bytes;
1286                 new_extra_isize -= shift_bytes;
1287                 EXT4_I(inode)->i_extra_isize = extra_isize;
1288
1289                 i.name = b_entry_name;
1290                 i.value = buffer;
1291                 i.value_len = size;
1292                 error = ext4_xattr_block_find(inode, &i, bs);
1293                 if (error)
1294                         goto cleanup;
1295
1296                 /* Add entry which was removed from the inode into the block */
1297                 error = ext4_xattr_block_set(handle, inode, &i, bs);
1298                 if (error)
1299                         goto cleanup;
1300                 kfree(b_entry_name);
1301                 kfree(buffer);
1302                 brelse(is->iloc.bh);
1303                 kfree(is);
1304                 kfree(bs);
1305         }
1306         brelse(bh);
1307         up_write(&EXT4_I(inode)->xattr_sem);
1308         return 0;
1309
1310 cleanup:
1311         kfree(b_entry_name);
1312         kfree(buffer);
1313         if (is)
1314                 brelse(is->iloc.bh);
1315         kfree(is);
1316         kfree(bs);
1317         brelse(bh);
1318         up_write(&EXT4_I(inode)->xattr_sem);
1319         return error;
1320 }
1321
1322
1323
1324 /*
1325  * ext4_xattr_delete_inode()
1326  *
1327  * Free extended attribute resources associated with this inode. This
1328  * is called immediately before an inode is freed. We have exclusive
1329  * access to the inode.
1330  */
1331 void
1332 ext4_xattr_delete_inode(handle_t *handle, struct inode *inode)
1333 {
1334         struct buffer_head *bh = NULL;
1335
1336         if (!EXT4_I(inode)->i_file_acl)
1337                 goto cleanup;
1338         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
1339         if (!bh) {
1340                 ext4_error(inode->i_sb, __FUNCTION__,
1341                         "inode %lu: block %llu read error", inode->i_ino,
1342                         EXT4_I(inode)->i_file_acl);
1343                 goto cleanup;
1344         }
1345         if (BHDR(bh)->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC) ||
1346             BHDR(bh)->h_blocks != cpu_to_le32(1)) {
1347                 ext4_error(inode->i_sb, __FUNCTION__,
1348                         "inode %lu: bad block %llu", inode->i_ino,
1349                         EXT4_I(inode)->i_file_acl);
1350                 goto cleanup;
1351         }
1352         ext4_xattr_release_block(handle, inode, bh);
1353         EXT4_I(inode)->i_file_acl = 0;
1354
1355 cleanup:
1356         brelse(bh);
1357 }
1358
1359 /*
1360  * ext4_xattr_put_super()
1361  *
1362  * This is called when a file system is unmounted.
1363  */
1364 void
1365 ext4_xattr_put_super(struct super_block *sb)
1366 {
1367         mb_cache_shrink(sb->s_bdev);
1368 }
1369
1370 /*
1371  * ext4_xattr_cache_insert()
1372  *
1373  * Create a new entry in the extended attribute cache, and insert
1374  * it unless such an entry is already in the cache.
1375  *
1376  * Returns 0, or a negative error number on failure.
1377  */
1378 static void
1379 ext4_xattr_cache_insert(struct buffer_head *bh)
1380 {
1381         __u32 hash = le32_to_cpu(BHDR(bh)->h_hash);
1382         struct mb_cache_entry *ce;
1383         int error;
1384
1385         ce = mb_cache_entry_alloc(ext4_xattr_cache, GFP_NOFS);
1386         if (!ce) {
1387                 ea_bdebug(bh, "out of memory");
1388                 return;
1389         }
1390         error = mb_cache_entry_insert(ce, bh->b_bdev, bh->b_blocknr, &hash);
1391         if (error) {
1392                 mb_cache_entry_free(ce);
1393                 if (error == -EBUSY) {
1394                         ea_bdebug(bh, "already in cache");
1395                         error = 0;
1396                 }
1397         } else {
1398                 ea_bdebug(bh, "inserting [%x]", (int)hash);
1399                 mb_cache_entry_release(ce);
1400         }
1401 }
1402
1403 /*
1404  * ext4_xattr_cmp()
1405  *
1406  * Compare two extended attribute blocks for equality.
1407  *
1408  * Returns 0 if the blocks are equal, 1 if they differ, and
1409  * a negative error number on errors.
1410  */
1411 static int
1412 ext4_xattr_cmp(struct ext4_xattr_header *header1,
1413                struct ext4_xattr_header *header2)
1414 {
1415         struct ext4_xattr_entry *entry1, *entry2;
1416
1417         entry1 = ENTRY(header1+1);
1418         entry2 = ENTRY(header2+1);
1419         while (!IS_LAST_ENTRY(entry1)) {
1420                 if (IS_LAST_ENTRY(entry2))
1421                         return 1;
1422                 if (entry1->e_hash != entry2->e_hash ||
1423                     entry1->e_name_index != entry2->e_name_index ||
1424                     entry1->e_name_len != entry2->e_name_len ||
1425                     entry1->e_value_size != entry2->e_value_size ||
1426                     memcmp(entry1->e_name, entry2->e_name, entry1->e_name_len))
1427                         return 1;
1428                 if (entry1->e_value_block != 0 || entry2->e_value_block != 0)
1429                         return -EIO;
1430                 if (memcmp((char *)header1 + le16_to_cpu(entry1->e_value_offs),
1431                            (char *)header2 + le16_to_cpu(entry2->e_value_offs),
1432                            le32_to_cpu(entry1->e_value_size)))
1433                         return 1;
1434
1435                 entry1 = EXT4_XATTR_NEXT(entry1);
1436                 entry2 = EXT4_XATTR_NEXT(entry2);
1437         }
1438         if (!IS_LAST_ENTRY(entry2))
1439                 return 1;
1440         return 0;
1441 }
1442
1443 /*
1444  * ext4_xattr_cache_find()
1445  *
1446  * Find an identical extended attribute block.
1447  *
1448  * Returns a pointer to the block found, or NULL if such a block was
1449  * not found or an error occurred.
1450  */
1451 static struct buffer_head *
1452 ext4_xattr_cache_find(struct inode *inode, struct ext4_xattr_header *header,
1453                       struct mb_cache_entry **pce)
1454 {
1455         __u32 hash = le32_to_cpu(header->h_hash);
1456         struct mb_cache_entry *ce;
1457
1458         if (!header->h_hash)
1459                 return NULL;  /* never share */
1460         ea_idebug(inode, "looking for cached blocks [%x]", (int)hash);
1461 again:
1462         ce = mb_cache_entry_find_first(ext4_xattr_cache, 0,
1463                                        inode->i_sb->s_bdev, hash);
1464         while (ce) {
1465                 struct buffer_head *bh;
1466
1467                 if (IS_ERR(ce)) {
1468                         if (PTR_ERR(ce) == -EAGAIN)
1469                                 goto again;
1470                         break;
1471                 }
1472                 bh = sb_bread(inode->i_sb, ce->e_block);
1473                 if (!bh) {
1474                         ext4_error(inode->i_sb, __FUNCTION__,
1475                                 "inode %lu: block %lu read error",
1476                                 inode->i_ino, (unsigned long) ce->e_block);
1477                 } else if (le32_to_cpu(BHDR(bh)->h_refcount) >=
1478                                 EXT4_XATTR_REFCOUNT_MAX) {
1479                         ea_idebug(inode, "block %lu refcount %d>=%d",
1480                                   (unsigned long) ce->e_block,
1481                                   le32_to_cpu(BHDR(bh)->h_refcount),
1482                                           EXT4_XATTR_REFCOUNT_MAX);
1483                 } else if (ext4_xattr_cmp(header, BHDR(bh)) == 0) {
1484                         *pce = ce;
1485                         return bh;
1486                 }
1487                 brelse(bh);
1488                 ce = mb_cache_entry_find_next(ce, 0, inode->i_sb->s_bdev, hash);
1489         }
1490         return NULL;
1491 }
1492
1493 #define NAME_HASH_SHIFT 5
1494 #define VALUE_HASH_SHIFT 16
1495
1496 /*
1497  * ext4_xattr_hash_entry()
1498  *
1499  * Compute the hash of an extended attribute.
1500  */
1501 static inline void ext4_xattr_hash_entry(struct ext4_xattr_header *header,
1502                                          struct ext4_xattr_entry *entry)
1503 {
1504         __u32 hash = 0;
1505         char *name = entry->e_name;
1506         int n;
1507
1508         for (n=0; n < entry->e_name_len; n++) {
1509                 hash = (hash << NAME_HASH_SHIFT) ^
1510                        (hash >> (8*sizeof(hash) - NAME_HASH_SHIFT)) ^
1511                        *name++;
1512         }
1513
1514         if (entry->e_value_block == 0 && entry->e_value_size != 0) {
1515                 __le32 *value = (__le32 *)((char *)header +
1516                         le16_to_cpu(entry->e_value_offs));
1517                 for (n = (le32_to_cpu(entry->e_value_size) +
1518                      EXT4_XATTR_ROUND) >> EXT4_XATTR_PAD_BITS; n; n--) {
1519                         hash = (hash << VALUE_HASH_SHIFT) ^
1520                                (hash >> (8*sizeof(hash) - VALUE_HASH_SHIFT)) ^
1521                                le32_to_cpu(*value++);
1522                 }
1523         }
1524         entry->e_hash = cpu_to_le32(hash);
1525 }
1526
1527 #undef NAME_HASH_SHIFT
1528 #undef VALUE_HASH_SHIFT
1529
1530 #define BLOCK_HASH_SHIFT 16
1531
1532 /*
1533  * ext4_xattr_rehash()
1534  *
1535  * Re-compute the extended attribute hash value after an entry has changed.
1536  */
1537 static void ext4_xattr_rehash(struct ext4_xattr_header *header,
1538                               struct ext4_xattr_entry *entry)
1539 {
1540         struct ext4_xattr_entry *here;
1541         __u32 hash = 0;
1542
1543         ext4_xattr_hash_entry(header, entry);
1544         here = ENTRY(header+1);
1545         while (!IS_LAST_ENTRY(here)) {
1546                 if (!here->e_hash) {
1547                         /* Block is not shared if an entry's hash value == 0 */
1548                         hash = 0;
1549                         break;
1550                 }
1551                 hash = (hash << BLOCK_HASH_SHIFT) ^
1552                        (hash >> (8*sizeof(hash) - BLOCK_HASH_SHIFT)) ^
1553                        le32_to_cpu(here->e_hash);
1554                 here = EXT4_XATTR_NEXT(here);
1555         }
1556         header->h_hash = cpu_to_le32(hash);
1557 }
1558
1559 #undef BLOCK_HASH_SHIFT
1560
1561 int __init
1562 init_ext4_xattr(void)
1563 {
1564         ext4_xattr_cache = mb_cache_create("ext4_xattr", NULL,
1565                 sizeof(struct mb_cache_entry) +
1566                 sizeof(((struct mb_cache_entry *) 0)->e_indexes[0]), 1, 6);
1567         if (!ext4_xattr_cache)
1568                 return -ENOMEM;
1569         return 0;
1570 }
1571
1572 void
1573 exit_ext4_xattr(void)
1574 {
1575         if (ext4_xattr_cache)
1576                 mb_cache_destroy(ext4_xattr_cache);
1577         ext4_xattr_cache = NULL;
1578 }