fs/ext4/xattr.c

   1 /*
   2  * linux/fs/ext4/xattr.c
   3  *
   4  * Copyright (C) 2001-2003 Andreas Gruenbacher, <agruen@suse.de>
   5  *
   6  * Fix by Harrison Xing <harrison@mountainviewdata.com>.
   7  * Ext4 code with a lot of help from Eric Jarman <ejarman@acm.org>.
   8  * Extended attributes for symlinks and special files added per
   9  *  suggestion of Luka Renko <luka.renko@hermes.si>.
  10  * xattr consolidation Copyright (c) 2004 James Morris <jmorris@redhat.com>,
  11  *  Red Hat Inc.
  12  * ea-in-inode support by Alex Tomas <alex@clusterfs.com> aka bzzz
  13  *  and Andreas Gruenbacher <agruen@suse.de>.
  14  */
  15
  16 /*
  17  * Extended attributes are stored directly in inodes (on file systems with
  18  * inodes bigger than 128 bytes) and on additional disk blocks. The i_file_acl
  19  * field contains the block number if an inode uses an additional block. All
  20  * attributes must fit in the inode and one additional block. Blocks that
  21  * contain the identical set of attributes may be shared among several inodes.
  22  * Identical blocks are detected by keeping a cache of blocks that have
  23  * recently been accessed.
  24  *
  25  * The attributes in inodes and on blocks have a different header; the entries
  26  * are stored in the same format:
  27  *
  28  *   +------------------+
  29  *   | header           |
  30  *   | entry 1          | |
  31  *   | entry 2          | | growing downwards
  32  *   | entry 3          | v
  33  *   | four null bytes  |
  34  *   | . . .            |
  35  *   | value 1          | ^
  36  *   | value 3          | | growing upwards
  37  *   | value 2          | |
  38  *   +------------------+
  39  *
  40  * The header is followed by multiple entry descriptors. In disk blocks, the
  41  * entry descriptors are kept sorted. In inodes, they are unsorted. The
  42  * attribute values are aligned to the end of the block in no specific order.
  43  *
  44  * Locking strategy
  45  * ----------------
  46  * EXT4_I(inode)->i_file_acl is protected by EXT4_I(inode)->xattr_sem.
  47  * EA blocks are only changed if they are exclusive to an inode, so
  48  * holding xattr_sem also means that nothing but the EA block's reference
  49  * count can change. Multiple writers to the same block are synchronized
  50  * by the buffer lock.
  51  */
  52
  53 #include <linux/init.h>
  54 #include <linux/fs.h>
  55 #include <linux/slab.h>
  56 #include <linux/mbcache.h>
  57 #include <linux/quotaops.h>
  58 #include <linux/rwsem.h>
  59 #include "ext4_jbd2.h"
  60 #include "ext4.h"
  61 #include "xattr.h"
  62 #include "acl.h"
  63
  64 #define BHDR(bh) ((struct ext4_xattr_header *)((bh)->b_data))
  65 #define ENTRY(ptr) ((struct ext4_xattr_entry *)(ptr))
  66 #define BFIRST(bh) ENTRY(BHDR(bh)+1)
  67 #define IS_LAST_ENTRY(entry) (*(__u32 *)(entry) == 0)
  68
  69 #ifdef EXT4_XATTR_DEBUG
  70 # define ea_idebug(inode, f...) do { \
  71                 printk(KERN_DEBUG "inode %s:%lu: ", \
  72                         inode->i_sb->s_id, inode->i_ino); \
  73                 printk(f); \
  74                 printk("\n"); \
  75         } while (0)
  76 # define ea_bdebug(bh, f...) do { \
  77                 char b[BDEVNAME_SIZE]; \
  78                 printk(KERN_DEBUG "block %s:%lu: ", \
  79                         bdevname(bh->b_bdev, b), \
  80                         (unsigned long) bh->b_blocknr); \
  81                 printk(f); \
  82                 printk("\n"); \
  83         } while (0)
  84 #else
  85 # define ea_idebug(f...)
  86 # define ea_bdebug(f...)
  87 #endif
  88
  89 static void ext4_xattr_cache_insert(struct buffer_head *);
  90 static struct buffer_head *ext4_xattr_cache_find(struct inode *,
  91                                                  struct ext4_xattr_header *,
  92                                                  struct mb_cache_entry **);
  93 static void ext4_xattr_rehash(struct ext4_xattr_header *,
  94                               struct ext4_xattr_entry *);
  95 static int ext4_xattr_list(struct dentry *dentry, char *buffer,
  96                            size_t buffer_size);
  97
  98 static struct mb_cache *ext4_xattr_cache;
  99
 100 static struct xattr_handler *ext4_xattr_handler_map[] = {
 101         [EXT4_XATTR_INDEX_USER]              = &ext4_xattr_user_handler,
 102 #ifdef CONFIG_EXT4_FS_POSIX_ACL
 103         [EXT4_XATTR_INDEX_POSIX_ACL_ACCESS]  = &ext4_xattr_acl_access_handler,
 104         [EXT4_XATTR_INDEX_POSIX_ACL_DEFAULT] = &ext4_xattr_acl_default_handler,
 105 #endif
 106         [EXT4_XATTR_INDEX_TRUSTED]           = &ext4_xattr_trusted_handler,
 107 #ifdef CONFIG_EXT4_FS_SECURITY
 108         [EXT4_XATTR_INDEX_SECURITY]          = &ext4_xattr_security_handler,
 109 #endif
 110 };
 111
 112 struct xattr_handler *ext4_xattr_handlers[] = {
 113         &ext4_xattr_user_handler,
 114         &ext4_xattr_trusted_handler,
 115 #ifdef CONFIG_EXT4_FS_POSIX_ACL
 116         &ext4_xattr_acl_access_handler,
 117         &ext4_xattr_acl_default_handler,
 118 #endif
 119 #ifdef CONFIG_EXT4_FS_SECURITY
 120         &ext4_xattr_security_handler,
 121 #endif
 122         NULL
 123 };
 124
 125 static inline struct xattr_handler *
 126 ext4_xattr_handler(int name_index)
 127 {
 128         struct xattr_handler *handler = NULL;
 129
 130         if (name_index > 0 && name_index < ARRAY_SIZE(ext4_xattr_handler_map))
 131                 handler = ext4_xattr_handler_map[name_index];
 132         return handler;
 133 }
 134
 135 /*
 136  * Inode operation listxattr()
 137  *
 138  * dentry->d_inode->i_mutex: don't care
 139  */
 140 ssize_t
 141 ext4_listxattr(struct dentry *dentry, char *buffer, size_t size)
 142 {
 143         return ext4_xattr_list(dentry, buffer, size);
 144 }
 145
 146 static int
 147 ext4_xattr_check_names(struct ext4_xattr_entry *entry, void *end)
 148 {
 149         while (!IS_LAST_ENTRY(entry)) {
 150                 struct ext4_xattr_entry *next = EXT4_XATTR_NEXT(entry);
 151                 if ((void *)next >= end)
 152                         return -EIO;
 153                 entry = next;
 154         }
 155         return 0;
 156 }
 157
 158 static inline int
 159 ext4_xattr_check_block(struct buffer_head *bh)
 160 {
 161         int error;
 162
 163         if (BHDR(bh)->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC) ||
 164             BHDR(bh)->h_blocks != cpu_to_le32(1))
 165                 return -EIO;
 166         error = ext4_xattr_check_names(BFIRST(bh), bh->b_data + bh->b_size);
 167         return error;
 168 }
 169
 170 static inline int
 171 ext4_xattr_check_entry(struct ext4_xattr_entry *entry, size_t size)
 172 {
 173         size_t value_size = le32_to_cpu(entry->e_value_size);
 174
 175         if (entry->e_value_block != 0 || value_size > size ||
 176             le16_to_cpu(entry->e_value_offs) + value_size > size)
 177                 return -EIO;
 178         return 0;
 179 }
 180
 181 static int
 182 ext4_xattr_find_entry(struct ext4_xattr_entry **pentry, int name_index,
 183                       const char *name, size_t size, int sorted)
 184 {
 185         struct ext4_xattr_entry *entry;
 186         size_t name_len;
 187         int cmp = 1;
 188
 189         if (name == NULL)
 190                 return -EINVAL;
 191         name_len = strlen(name);
 192         entry = *pentry;
 193         for (; !IS_LAST_ENTRY(entry); entry = EXT4_XATTR_NEXT(entry)) {
 194                 cmp = name_index - entry->e_name_index;
 195                 if (!cmp)
 196                         cmp = name_len - entry->e_name_len;
 197                 if (!cmp)
 198                         cmp = memcmp(name, entry->e_name, name_len);
 199                 if (cmp <= 0 && (sorted || cmp == 0))
 200                         break;
 201         }
 202         *pentry = entry;
 203         if (!cmp && ext4_xattr_check_entry(entry, size))
 204                         return -EIO;
 205         return cmp ? -ENODATA : 0;
 206 }
 207
 208 static int
 209 ext4_xattr_block_get(struct inode *inode, int name_index, const char *name,
 210                      void *buffer, size_t buffer_size)
 211 {
 212         struct buffer_head *bh = NULL;
 213         struct ext4_xattr_entry *entry;
 214         size_t size;
 215         int error;
 216
 217         ea_idebug(inode, "name=%d.%s, buffer=%p, buffer_size=%ld",
 218                   name_index, name, buffer, (long)buffer_size);
 219
 220         error = -ENODATA;
 221         if (!EXT4_I(inode)->i_file_acl)
 222                 goto cleanup;
 223         ea_idebug(inode, "reading block %u", EXT4_I(inode)->i_file_acl);
 224         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
 225         if (!bh)
 226                 goto cleanup;
 227         ea_bdebug(bh, "b_count=%d, refcount=%d",
 228                 atomic_read(&(bh->b_count)), le32_to_cpu(BHDR(bh)->h_refcount));
 229         if (ext4_xattr_check_block(bh)) {
 230 bad_block:
 231                 ext4_error(inode->i_sb,
 232                            "inode %lu: bad block %llu", inode->i_ino,
 233                            EXT4_I(inode)->i_file_acl);
 234                 error = -EIO;
 235                 goto cleanup;
 236         }
 237         ext4_xattr_cache_insert(bh);
 238         entry = BFIRST(bh);
 239         error = ext4_xattr_find_entry(&entry, name_index, name, bh->b_size, 1);
 240         if (error == -EIO)
 241                 goto bad_block;
 242         if (error)
 243                 goto cleanup;
 244         size = le32_to_cpu(entry->e_value_size);
 245         if (buffer) {
 246                 error = -ERANGE;
 247                 if (size > buffer_size)
 248                         goto cleanup;
 249                 memcpy(buffer, bh->b_data + le16_to_cpu(entry->e_value_offs),
 250                        size);
 251         }
 252         error = size;
 253
 254 cleanup:
 255         brelse(bh);
 256         return error;
 257 }
 258
 259 static int
 260 ext4_xattr_ibody_get(struct inode *inode, int name_index, const char *name,
 261                      void *buffer, size_t buffer_size)
 262 {
 263         struct ext4_xattr_ibody_header *header;
 264         struct ext4_xattr_entry *entry;
 265         struct ext4_inode *raw_inode;
 266         struct ext4_iloc iloc;
 267         size_t size;
 268         void *end;
 269         int error;
 270
 271         if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR))
 272                 return -ENODATA;
 273         error = ext4_get_inode_loc(inode, &iloc);
 274         if (error)
 275                 return error;
 276         raw_inode = ext4_raw_inode(&iloc);
 277         header = IHDR(inode, raw_inode);
 278         entry = IFIRST(header);
 279         end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
 280         error = ext4_xattr_check_names(entry, end);
 281         if (error)
 282                 goto cleanup;
 283         error = ext4_xattr_find_entry(&entry, name_index, name,
 284                                       end - (void *)entry, 0);
 285         if (error)
 286                 goto cleanup;
 287         size = le32_to_cpu(entry->e_value_size);
 288         if (buffer) {
 289                 error = -ERANGE;
 290                 if (size > buffer_size)
 291                         goto cleanup;
 292                 memcpy(buffer, (void *)IFIRST(header) +
 293                        le16_to_cpu(entry->e_value_offs), size);
 294         }
 295         error = size;
 296
 297 cleanup:
 298         brelse(iloc.bh);
 299         return error;
 300 }
 301
 302 /*
 303  * ext4_xattr_get()
 304  *
 305  * Copy an extended attribute into the buffer
 306  * provided, or compute the buffer size required.
 307  * Buffer is NULL to compute the size of the buffer required.
 308  *
 309  * Returns a negative error number on failure, or the number of bytes
 310  * used / required on success.
 311  */
 312 int
 313 ext4_xattr_get(struct inode *inode, int name_index, const char *name,
 314                void *buffer, size_t buffer_size)
 315 {
 316         int error;
 317
 318         down_read(&EXT4_I(inode)->xattr_sem);
 319         error = ext4_xattr_ibody_get(inode, name_index, name, buffer,
 320                                      buffer_size);
 321         if (error == -ENODATA)
 322                 error = ext4_xattr_block_get(inode, name_index, name, buffer,
 323                                              buffer_size);
 324         up_read(&EXT4_I(inode)->xattr_sem);
 325         return error;
 326 }
 327
 328 static int
 329 ext4_xattr_list_entries(struct dentry *dentry, struct ext4_xattr_entry *entry,
 330                         char *buffer, size_t buffer_size)
 331 {
 332         size_t rest = buffer_size;
 333
 334         for (; !IS_LAST_ENTRY(entry); entry = EXT4_XATTR_NEXT(entry)) {
 335                 struct xattr_handler *handler =
 336                         ext4_xattr_handler(entry->e_name_index);
 337
 338                 if (handler) {
 339                         size_t size = handler->list(dentry, buffer, rest,
 340                                                     entry->e_name,
 341                                                     entry->e_name_len,
 342                                                     handler->flags);
 343                         if (buffer) {
 344                                 if (size > rest)
 345                                         return -ERANGE;
 346                                 buffer += size;
 347                         }
 348                         rest -= size;
 349                 }
 350         }
 351         return buffer_size - rest;
 352 }
 353
 354 static int
 355 ext4_xattr_block_list(struct dentry *dentry, char *buffer, size_t buffer_size)
 356 {
 357         struct inode *inode = dentry->d_inode;
 358         struct buffer_head *bh = NULL;
 359         int error;
 360
 361         ea_idebug(inode, "buffer=%p, buffer_size=%ld",
 362                   buffer, (long)buffer_size);
 363
 364         error = 0;
 365         if (!EXT4_I(inode)->i_file_acl)
 366                 goto cleanup;
 367         ea_idebug(inode, "reading block %u", EXT4_I(inode)->i_file_acl);
 368         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
 369         error = -EIO;
 370         if (!bh)
 371                 goto cleanup;
 372         ea_bdebug(bh, "b_count=%d, refcount=%d",
 373                 atomic_read(&(bh->b_count)), le32_to_cpu(BHDR(bh)->h_refcount));
 374         if (ext4_xattr_check_block(bh)) {
 375                 ext4_error(inode->i_sb,
 376                            "inode %lu: bad block %llu", inode->i_ino,
 377                            EXT4_I(inode)->i_file_acl);
 378                 error = -EIO;
 379                 goto cleanup;
 380         }
 381         ext4_xattr_cache_insert(bh);
 382         error = ext4_xattr_list_entries(dentry, BFIRST(bh), buffer, buffer_size);
 383
 384 cleanup:
 385         brelse(bh);
 386
 387         return error;
 388 }
 389
 390 static int
 391 ext4_xattr_ibody_list(struct dentry *dentry, char *buffer, size_t buffer_size)
 392 {
 393         struct inode *inode = dentry->d_inode;
 394         struct ext4_xattr_ibody_header *header;
 395         struct ext4_inode *raw_inode;
 396         struct ext4_iloc iloc;
 397         void *end;
 398         int error;
 399
 400         if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR))
 401                 return 0;
 402         error = ext4_get_inode_loc(inode, &iloc);
 403         if (error)
 404                 return error;
 405         raw_inode = ext4_raw_inode(&iloc);
 406         header = IHDR(inode, raw_inode);
 407         end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
 408         error = ext4_xattr_check_names(IFIRST(header), end);
 409         if (error)
 410                 goto cleanup;
 411         error = ext4_xattr_list_entries(dentry, IFIRST(header),
 412                                         buffer, buffer_size);
 413
 414 cleanup:
 415         brelse(iloc.bh);
 416         return error;
 417 }
 418
 419 /*
 420  * ext4_xattr_list()
 421  *
 422  * Copy a list of attribute names into the buffer
 423  * provided, or compute the buffer size required.
 424  * Buffer is NULL to compute the size of the buffer required.
 425  *
 426  * Returns a negative error number on failure, or the number of bytes
 427  * used / required on success.
 428  */
 429 static int
 430 ext4_xattr_list(struct dentry *dentry, char *buffer, size_t buffer_size)
 431 {
 432         int i_error, b_error;
 433
 434         down_read(&EXT4_I(dentry->d_inode)->xattr_sem);
 435         i_error = ext4_xattr_ibody_list(dentry, buffer, buffer_size);
 436         if (i_error < 0) {
 437                 b_error = 0;
 438         } else {
 439                 if (buffer) {
 440                         buffer += i_error;
 441                         buffer_size -= i_error;
 442                 }
 443                 b_error = ext4_xattr_block_list(dentry, buffer, buffer_size);
 444                 if (b_error < 0)
 445                         i_error = 0;
 446         }
 447         up_read(&EXT4_I(dentry->d_inode)->xattr_sem);
 448         return i_error + b_error;
 449 }
 450
 451 /*
 452  * If the EXT4_FEATURE_COMPAT_EXT_ATTR feature of this file system is
 453  * not set, set it.
 454  */
 455 static void ext4_xattr_update_super_block(handle_t *handle,
 456                                           struct super_block *sb)
 457 {
 458         if (EXT4_HAS_COMPAT_FEATURE(sb, EXT4_FEATURE_COMPAT_EXT_ATTR))
 459                 return;
 460
 461         if (ext4_journal_get_write_access(handle, EXT4_SB(sb)->s_sbh) == 0) {
 462                 EXT4_SET_COMPAT_FEATURE(sb, EXT4_FEATURE_COMPAT_EXT_ATTR);
 463                 sb->s_dirt = 1;
 464                 ext4_handle_dirty_metadata(handle, NULL, EXT4_SB(sb)->s_sbh);
 465         }
 466 }
 467
 468 /*
 469  * Release the xattr block BH: If the reference count is > 1, decrement
 470  * it; otherwise free the block.
 471  */
 472 static void
 473 ext4_xattr_release_block(handle_t *handle, struct inode *inode,
 474                          struct buffer_head *bh)
 475 {
 476         struct mb_cache_entry *ce = NULL;
 477         int error = 0;
 478
 479         ce = mb_cache_entry_get(ext4_xattr_cache, bh->b_bdev, bh->b_blocknr);
 480         error = ext4_journal_get_write_access(handle, bh);
 481         if (error)
 482                 goto out;
 483
 484         lock_buffer(bh);
 485         if (BHDR(bh)->h_refcount == cpu_to_le32(1)) {
 486                 ea_bdebug(bh, "refcount now=0; freeing");
 487                 if (ce)
 488                         mb_cache_entry_free(ce);
 489                 get_bh(bh);
 490                 ext4_free_blocks(handle, inode, bh, 0, 1,
 491                                  EXT4_FREE_BLOCKS_METADATA |
 492                                  EXT4_FREE_BLOCKS_FORGET);
 493         } else {
 494                 le32_add_cpu(&BHDR(bh)->h_refcount, -1);
 495                 error = ext4_handle_dirty_metadata(handle, inode, bh);
 496                 if (IS_SYNC(inode))
 497                         ext4_handle_sync(handle);
 498                 dquot_free_block(inode, 1);
 499                 ea_bdebug(bh, "refcount now=%d; releasing",
 500                           le32_to_cpu(BHDR(bh)->h_refcount));
 501                 if (ce)
 502                         mb_cache_entry_release(ce);
 503         }
 504         unlock_buffer(bh);
 505 out:
 506         ext4_std_error(inode->i_sb, error);
 507         return;
 508 }
 509
 510 /*
 511  * Find the available free space for EAs. This also returns the total number of
 512  * bytes used by EA entries.
 513  */
 514 static size_t ext4_xattr_free_space(struct ext4_xattr_entry *last,
 515                                     size_t *min_offs, void *base, int *total)
 516 {
 517         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 518                 *total += EXT4_XATTR_LEN(last->e_name_len);
 519                 if (!last->e_value_block && last->e_value_size) {
 520                         size_t offs = le16_to_cpu(last->e_value_offs);
 521                         if (offs < *min_offs)
 522                                 *min_offs = offs;
 523                 }
 524         }
 525         return (*min_offs - ((void *)last - base) - sizeof(__u32));
 526 }
 527
 528 struct ext4_xattr_info {
 529         int name_index;
 530         const char *name;
 531         const void *value;
 532         size_t value_len;
 533 };
 534
 535 struct ext4_xattr_search {
 536         struct ext4_xattr_entry *first;
 537         void *base;
 538         void *end;
 539         struct ext4_xattr_entry *here;
 540         int not_found;
 541 };
 542
 543 static int
 544 ext4_xattr_set_entry(struct ext4_xattr_info *i, struct ext4_xattr_search *s)
 545 {
 546         struct ext4_xattr_entry *last;
 547         size_t free, min_offs = s->end - s->base, name_len = strlen(i->name);
 548
 549         /* Compute min_offs and last. */
 550         last = s->first;
 551         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 552                 if (!last->e_value_block && last->e_value_size) {
 553                         size_t offs = le16_to_cpu(last->e_value_offs);
 554                         if (offs < min_offs)
 555                                 min_offs = offs;
 556                 }
 557         }
 558         free = min_offs - ((void *)last - s->base) - sizeof(__u32);
 559         if (!s->not_found) {
 560                 if (!s->here->e_value_block && s->here->e_value_size) {
 561                         size_t size = le32_to_cpu(s->here->e_value_size);
 562                         free += EXT4_XATTR_SIZE(size);
 563                 }
 564                 free += EXT4_XATTR_LEN(name_len);
 565         }
 566         if (i->value) {
 567                 if (free < EXT4_XATTR_SIZE(i->value_len) ||
 568                     free < EXT4_XATTR_LEN(name_len) +
 569                            EXT4_XATTR_SIZE(i->value_len))
 570                         return -ENOSPC;
 571         }
 572
 573         if (i->value && s->not_found) {
 574                 /* Insert the new name. */
 575                 size_t size = EXT4_XATTR_LEN(name_len);
 576                 size_t rest = (void *)last - (void *)s->here + sizeof(__u32);
 577                 memmove((void *)s->here + size, s->here, rest);
 578                 memset(s->here, 0, size);
 579                 s->here->e_name_index = i->name_index;
 580                 s->here->e_name_len = name_len;
 581                 memcpy(s->here->e_name, i->name, name_len);
 582         } else {
 583                 if (!s->here->e_value_block && s->here->e_value_size) {
 584                         void *first_val = s->base + min_offs;
 585                         size_t offs = le16_to_cpu(s->here->e_value_offs);
 586                         void *val = s->base + offs;
 587                         size_t size = EXT4_XATTR_SIZE(
 588                                 le32_to_cpu(s->here->e_value_size));
 589
 590                         if (i->value && size == EXT4_XATTR_SIZE(i->value_len)) {
 591                                 /* The old and the new value have the same
 592                                    size. Just replace. */
 593                                 s->here->e_value_size =
 594                                         cpu_to_le32(i->value_len);
 595                                 memset(val + size - EXT4_XATTR_PAD, 0,
 596                                        EXT4_XATTR_PAD); /* Clear pad bytes. */
 597                                 memcpy(val, i->value, i->value_len);
 598                                 return 0;
 599                         }
 600
 601                         /* Remove the old value. */
 602                         memmove(first_val + size, first_val, val - first_val);
 603                         memset(first_val, 0, size);
 604                         s->here->e_value_size = 0;
 605                         s->here->e_value_offs = 0;
 606                         min_offs += size;
 607
 608                         /* Adjust all value offsets. */
 609                         last = s->first;
 610                         while (!IS_LAST_ENTRY(last)) {
 611                                 size_t o = le16_to_cpu(last->e_value_offs);
 612                                 if (!last->e_value_block &&
 613                                     last->e_value_size && o < offs)
 614                                         last->e_value_offs =
 615                                                 cpu_to_le16(o + size);
 616                                 last = EXT4_XATTR_NEXT(last);
 617                         }
 618                 }
 619                 if (!i->value) {
 620                         /* Remove the old name. */
 621                         size_t size = EXT4_XATTR_LEN(name_len);
 622                         last = ENTRY((void *)last - size);
 623                         memmove(s->here, (void *)s->here + size,
 624                                 (void *)last - (void *)s->here + sizeof(__u32));
 625                         memset(last, 0, size);
 626                 }
 627         }
 628
 629         if (i->value) {
 630                 /* Insert the new value. */
 631                 s->here->e_value_size = cpu_to_le32(i->value_len);
 632                 if (i->value_len) {
 633                         size_t size = EXT4_XATTR_SIZE(i->value_len);
 634                         void *val = s->base + min_offs - size;
 635                         s->here->e_value_offs = cpu_to_le16(min_offs - size);
 636                         memset(val + size - EXT4_XATTR_PAD, 0,
 637                                EXT4_XATTR_PAD); /* Clear the pad bytes. */
 638                         memcpy(val, i->value, i->value_len);
 639                 }
 640         }
 641         return 0;
 642 }
 643
 644 struct ext4_xattr_block_find {
 645         struct ext4_xattr_search s;
 646         struct buffer_head *bh;
 647 };
 648
 649 static int
 650 ext4_xattr_block_find(struct inode *inode, struct ext4_xattr_info *i,
 651                       struct ext4_xattr_block_find *bs)
 652 {
 653         struct super_block *sb = inode->i_sb;
 654         int error;
 655
 656         ea_idebug(inode, "name=%d.%s, value=%p, value_len=%ld",
 657                   i->name_index, i->name, i->value, (long)i->value_len);
 658
 659         if (EXT4_I(inode)->i_file_acl) {
 660                 /* The inode already has an extended attribute block. */
 661                 bs->bh = sb_bread(sb, EXT4_I(inode)->i_file_acl);
 662                 error = -EIO;
 663                 if (!bs->bh)
 664                         goto cleanup;
 665                 ea_bdebug(bs->bh, "b_count=%d, refcount=%d",
 666                         atomic_read(&(bs->bh->b_count)),
 667                         le32_to_cpu(BHDR(bs->bh)->h_refcount));
 668                 if (ext4_xattr_check_block(bs->bh)) {
 669                         ext4_error(sb, "inode %lu: bad block %llu",
 670                                    inode->i_ino, EXT4_I(inode)->i_file_acl);
 671                         error = -EIO;
 672                         goto cleanup;
 673                 }
 674                 /* Find the named attribute. */
 675                 bs->s.base = BHDR(bs->bh);
 676                 bs->s.first = BFIRST(bs->bh);
 677                 bs->s.end = bs->bh->b_data + bs->bh->b_size;
 678                 bs->s.here = bs->s.first;
 679                 error = ext4_xattr_find_entry(&bs->s.here, i->name_index,
 680                                               i->name, bs->bh->b_size, 1);
 681                 if (error && error != -ENODATA)
 682                         goto cleanup;
 683                 bs->s.not_found = error;
 684         }
 685         error = 0;
 686
 687 cleanup:
 688         return error;
 689 }
 690
 691 static int
 692 ext4_xattr_block_set(handle_t *handle, struct inode *inode,
 693                      struct ext4_xattr_info *i,
 694                      struct ext4_xattr_block_find *bs)
 695 {
 696         struct super_block *sb = inode->i_sb;
 697         struct buffer_head *new_bh = NULL;
 698         struct ext4_xattr_search *s = &bs->s;
 699         struct mb_cache_entry *ce = NULL;
 700         int error = 0;
 701
 702 #define header(x) ((struct ext4_xattr_header *)(x))
 703
 704         if (i->value && i->value_len > sb->s_blocksize)
 705                 return -ENOSPC;
 706         if (s->base) {
 707                 ce = mb_cache_entry_get(ext4_xattr_cache, bs->bh->b_bdev,
 708                                         bs->bh->b_blocknr);
 709                 error = ext4_journal_get_write_access(handle, bs->bh);
 710                 if (error)
 711                         goto cleanup;
 712                 lock_buffer(bs->bh);
 713
 714                 if (header(s->base)->h_refcount == cpu_to_le32(1)) {
 715                         if (ce) {
 716                                 mb_cache_entry_free(ce);
 717                                 ce = NULL;
 718                         }
 719                         ea_bdebug(bs->bh, "modifying in-place");
 720                         error = ext4_xattr_set_entry(i, s);
 721                         if (!error) {
 722                                 if (!IS_LAST_ENTRY(s->first))
 723                                         ext4_xattr_rehash(header(s->base),
 724                                                           s->here);
 725                                 ext4_xattr_cache_insert(bs->bh);
 726                         }
 727                         unlock_buffer(bs->bh);
 728                         if (error == -EIO)
 729                                 goto bad_block;
 730                         if (!error)
 731                                 error = ext4_handle_dirty_metadata(handle,
 732                                                                    inode,
 733                                                                    bs->bh);
 734                         if (error)
 735                                 goto cleanup;
 736                         goto inserted;
 737                 } else {
 738                         int offset = (char *)s->here - bs->bh->b_data;
 739
 740                         unlock_buffer(bs->bh);
 741                         jbd2_journal_release_buffer(handle, bs->bh);
 742                         if (ce) {
 743                                 mb_cache_entry_release(ce);
 744                                 ce = NULL;
 745                         }
 746                         ea_bdebug(bs->bh, "cloning");
 747                         s->base = kmalloc(bs->bh->b_size, GFP_NOFS);
 748                         error = -ENOMEM;
 749                         if (s->base == NULL)
 750                                 goto cleanup;
 751                         memcpy(s->base, BHDR(bs->bh), bs->bh->b_size);
 752                         s->first = ENTRY(header(s->base)+1);
 753                         header(s->base)->h_refcount = cpu_to_le32(1);
 754                         s->here = ENTRY(s->base + offset);
 755                         s->end = s->base + bs->bh->b_size;
 756                 }
 757         } else {
 758                 /* Allocate a buffer where we construct the new block. */
 759                 s->base = kzalloc(sb->s_blocksize, GFP_NOFS);
 760                 /* assert(header == s->base) */
 761                 error = -ENOMEM;
 762                 if (s->base == NULL)
 763                         goto cleanup;
 764                 header(s->base)->h_magic = cpu_to_le32(EXT4_XATTR_MAGIC);
 765                 header(s->base)->h_blocks = cpu_to_le32(1);
 766                 header(s->base)->h_refcount = cpu_to_le32(1);
 767                 s->first = ENTRY(header(s->base)+1);
 768                 s->here = ENTRY(header(s->base)+1);
 769                 s->end = s->base + sb->s_blocksize;
 770         }
 771
 772         error = ext4_xattr_set_entry(i, s);
 773         if (error == -EIO)
 774                 goto bad_block;
 775         if (error)
 776                 goto cleanup;
 777         if (!IS_LAST_ENTRY(s->first))
 778                 ext4_xattr_rehash(header(s->base), s->here);
 779
 780 inserted:
 781         if (!IS_LAST_ENTRY(s->first)) {
 782                 new_bh = ext4_xattr_cache_find(inode, header(s->base), &ce);
 783                 if (new_bh) {
 784                         /* We found an identical block in the cache. */
 785                         if (new_bh == bs->bh)
 786                                 ea_bdebug(new_bh, "keeping");
 787                         else {
 788                                 /* The old block is released after updating
 789                                    the inode. */
 790                                 error = dquot_alloc_block(inode, 1);
 791                                 if (error)
 792                                         goto cleanup;
 793                                 error = ext4_journal_get_write_access(handle,
 794                                                                       new_bh);
 795                                 if (error)
 796                                         goto cleanup_dquot;
 797                                 lock_buffer(new_bh);
 798                                 le32_add_cpu(&BHDR(new_bh)->h_refcount, 1);
 799                                 ea_bdebug(new_bh, "reusing; refcount now=%d",
 800                                         le32_to_cpu(BHDR(new_bh)->h_refcount));
 801                                 unlock_buffer(new_bh);
 802                                 error = ext4_handle_dirty_metadata(handle,
 803                                                                    inode,
 804                                                                    new_bh);
 805                                 if (error)
 806                                         goto cleanup_dquot;
 807                         }
 808                         mb_cache_entry_release(ce);
 809                         ce = NULL;
 810                 } else if (bs->bh && s->base == bs->bh->b_data) {
 811                         /* We were modifying this block in-place. */
 812                         ea_bdebug(bs->bh, "keeping this block");
 813                         new_bh = bs->bh;
 814                         get_bh(new_bh);
 815                 } else {
 816                         /* We need to allocate a new block */
 817                         ext4_fsblk_t goal, block;
 818
 819                         goal = ext4_group_first_block_no(sb,
 820                                                 EXT4_I(inode)->i_block_group);
 821
 822                         /* non-extent files can't have physical blocks past 2^32 */
 823                         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL))
 824                                 goal = goal & EXT4_MAX_BLOCK_FILE_PHYS;
 825
 826                         block = ext4_new_meta_blocks(handle, inode,
 827                                                   goal, NULL, &error);
 828                         if (error)
 829                                 goto cleanup;
 830
 831                         if (!(EXT4_I(inode)->i_flags & EXT4_EXTENTS_FL))
 832                                 BUG_ON(block > EXT4_MAX_BLOCK_FILE_PHYS);
 833
 834                         ea_idebug(inode, "creating block %d", block);
 835
 836                         new_bh = sb_getblk(sb, block);
 837                         if (!new_bh) {
 838 getblk_failed:
 839                                 ext4_free_blocks(handle, inode, 0, block, 1,
 840                                                  EXT4_FREE_BLOCKS_METADATA);
 841                                 error = -EIO;
 842                                 goto cleanup;
 843                         }
 844                         lock_buffer(new_bh);
 845                         error = ext4_journal_get_create_access(handle, new_bh);
 846                         if (error) {
 847                                 unlock_buffer(new_bh);
 848                                 goto getblk_failed;
 849                         }
 850                         memcpy(new_bh->b_data, s->base, new_bh->b_size);
 851                         set_buffer_uptodate(new_bh);
 852                         unlock_buffer(new_bh);
 853                         ext4_xattr_cache_insert(new_bh);
 854                         error = ext4_handle_dirty_metadata(handle,
 855                                                            inode, new_bh);
 856                         if (error)
 857                                 goto cleanup;
 858                 }
 859         }
 860
 861         /* Update the inode. */
 862         EXT4_I(inode)->i_file_acl = new_bh ? new_bh->b_blocknr : 0;
 863
 864         /* Drop the previous xattr block. */
 865         if (bs->bh && bs->bh != new_bh)
 866                 ext4_xattr_release_block(handle, inode, bs->bh);
 867         error = 0;
 868
 869 cleanup:
 870         if (ce)
 871                 mb_cache_entry_release(ce);
 872         brelse(new_bh);
 873         if (!(bs->bh && s->base == bs->bh->b_data))
 874                 kfree(s->base);
 875
 876         return error;
 877
 878 cleanup_dquot:
 879         dquot_free_block(inode, 1);
 880         goto cleanup;
 881
 882 bad_block:
 883         ext4_error(inode->i_sb, "inode %lu: bad block %llu",
 884                    inode->i_ino, EXT4_I(inode)->i_file_acl);
 885         goto cleanup;
 886
 887 #undef header
 888 }
 889
 890 struct ext4_xattr_ibody_find {
 891         struct ext4_xattr_search s;
 892         struct ext4_iloc iloc;
 893 };
 894
 895 static int
 896 ext4_xattr_ibody_find(struct inode *inode, struct ext4_xattr_info *i,
 897                       struct ext4_xattr_ibody_find *is)
 898 {
 899         struct ext4_xattr_ibody_header *header;
 900         struct ext4_inode *raw_inode;
 901         int error;
 902
 903         if (EXT4_I(inode)->i_extra_isize == 0)
 904                 return 0;
 905         raw_inode = ext4_raw_inode(&is->iloc);
 906         header = IHDR(inode, raw_inode);
 907         is->s.base = is->s.first = IFIRST(header);
 908         is->s.here = is->s.first;
 909         is->s.end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
 910         if (ext4_test_inode_state(inode, EXT4_STATE_XATTR)) {
 911                 error = ext4_xattr_check_names(IFIRST(header), is->s.end);
 912                 if (error)
 913                         return error;
 914                 /* Find the named attribute. */
 915                 error = ext4_xattr_find_entry(&is->s.here, i->name_index,
 916                                               i->name, is->s.end -
 917                                               (void *)is->s.base, 0);
 918                 if (error && error != -ENODATA)
 919                         return error;
 920                 is->s.not_found = error;
 921         }
 922         return 0;
 923 }
 924
 925 static int
 926 ext4_xattr_ibody_set(handle_t *handle, struct inode *inode,
 927                      struct ext4_xattr_info *i,
 928                      struct ext4_xattr_ibody_find *is)
 929 {
 930         struct ext4_xattr_ibody_header *header;
 931         struct ext4_xattr_search *s = &is->s;
 932         int error;
 933
 934         if (EXT4_I(inode)->i_extra_isize == 0)
 935                 return -ENOSPC;
 936         error = ext4_xattr_set_entry(i, s);
 937         if (error)
 938                 return error;
 939         header = IHDR(inode, ext4_raw_inode(&is->iloc));
 940         if (!IS_LAST_ENTRY(s->first)) {
 941                 header->h_magic = cpu_to_le32(EXT4_XATTR_MAGIC);
 942                 ext4_set_inode_state(inode, EXT4_STATE_XATTR);
 943         } else {
 944                 header->h_magic = cpu_to_le32(0);
 945                 ext4_clear_inode_state(inode, EXT4_STATE_XATTR);
 946         }
 947         return 0;
 948 }
 949
 950 /*
 951  * ext4_xattr_set_handle()
 952  *
 953  * Create, replace or remove an extended attribute for this inode. Buffer
 954  * is NULL to remove an existing extended attribute, and non-NULL to
 955  * either replace an existing extended attribute, or create a new extended
 956  * attribute. The flags XATTR_REPLACE and XATTR_CREATE
 957  * specify that an extended attribute must exist and must not exist
 958  * previous to the call, respectively.
 959  *
 960  * Returns 0, or a negative error number on failure.
 961  */
 962 int
 963 ext4_xattr_set_handle(handle_t *handle, struct inode *inode, int name_index,
 964                       const char *name, const void *value, size_t value_len,
 965                       int flags)
 966 {
 967         struct ext4_xattr_info i = {
 968                 .name_index = name_index,
 969                 .name = name,
 970                 .value = value,
 971                 .value_len = value_len,
 972
 973         };
 974         struct ext4_xattr_ibody_find is = {
 975                 .s = { .not_found = -ENODATA, },
 976         };
 977         struct ext4_xattr_block_find bs = {
 978                 .s = { .not_found = -ENODATA, },
 979         };
 980         unsigned long no_expand;
 981         int error;
 982
 983         if (!name)
 984                 return -EINVAL;
 985         if (strlen(name) > 255)
 986                 return -ERANGE;
 987         down_write(&EXT4_I(inode)->xattr_sem);
 988         no_expand = ext4_test_inode_state(inode, EXT4_STATE_NO_EXPAND);
 989         ext4_set_inode_state(inode, EXT4_STATE_NO_EXPAND);
 990
 991         error = ext4_get_inode_loc(inode, &is.iloc);
 992         if (error)
 993                 goto cleanup;
 994
 995         error = ext4_journal_get_write_access(handle, is.iloc.bh);
 996         if (error)
 997                 goto cleanup;
 998
 999         if (ext4_test_inode_state(inode, EXT4_STATE_NEW)) {
1000                 struct ext4_inode *raw_inode = ext4_raw_inode(&is.iloc);
1001                 memset(raw_inode, 0, EXT4_SB(inode->i_sb)->s_inode_size);
1002                 ext4_clear_inode_state(inode, EXT4_STATE_NEW);
1003         }
1004
1005         error = ext4_xattr_ibody_find(inode, &i, &is);
1006         if (error)
1007                 goto cleanup;
1008         if (is.s.not_found)
1009                 error = ext4_xattr_block_find(inode, &i, &bs);
1010         if (error)
1011                 goto cleanup;
1012         if (is.s.not_found && bs.s.not_found) {
1013                 error = -ENODATA;
1014                 if (flags & XATTR_REPLACE)
1015                         goto cleanup;
1016                 error = 0;
1017                 if (!value)
1018                         goto cleanup;
1019         } else {
1020                 error = -EEXIST;
1021                 if (flags & XATTR_CREATE)
1022                         goto cleanup;
1023         }
1024         if (!value) {
1025                 if (!is.s.not_found)
1026                         error = ext4_xattr_ibody_set(handle, inode, &i, &is);
1027                 else if (!bs.s.not_found)
1028                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
1029         } else {
1030                 error = ext4_xattr_ibody_set(handle, inode, &i, &is);
1031                 if (!error && !bs.s.not_found) {
1032                         i.value = NULL;
1033                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
1034                 } else if (error == -ENOSPC) {
1035                         if (EXT4_I(inode)->i_file_acl && !bs.s.base) {
1036                                 error = ext4_xattr_block_find(inode, &i, &bs);
1037                                 if (error)
1038                                         goto cleanup;
1039                         }
1040                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
1041                         if (error)
1042                                 goto cleanup;
1043                         if (!is.s.not_found) {
1044                                 i.value = NULL;
1045                                 error = ext4_xattr_ibody_set(handle, inode, &i,
1046                                                              &is);
1047                         }
1048                 }
1049         }
1050         if (!error) {
1051                 ext4_xattr_update_super_block(handle, inode->i_sb);
1052                 inode->i_ctime = ext4_current_time(inode);
1053                 if (!value)
1054                         ext4_clear_inode_state(inode, EXT4_STATE_NO_EXPAND);
1055                 error = ext4_mark_iloc_dirty(handle, inode, &is.iloc);
1056                 /*
1057                  * The bh is consumed by ext4_mark_iloc_dirty, even with
1058                  * error != 0.
1059                  */
1060                 is.iloc.bh = NULL;
1061                 if (IS_SYNC(inode))
1062                         ext4_handle_sync(handle);
1063         }
1064
1065 cleanup:
1066         brelse(is.iloc.bh);
1067         brelse(bs.bh);
1068         if (no_expand == 0)
1069                 ext4_clear_inode_state(inode, EXT4_STATE_NO_EXPAND);
1070         up_write(&EXT4_I(inode)->xattr_sem);
1071         return error;
1072 }
1073
1074 /*
1075  * ext4_xattr_set()
1076  *
1077  * Like ext4_xattr_set_handle, but start from an inode. This extended
1078  * attribute modification is a filesystem transaction by itself.
1079  *
1080  * Returns 0, or a negative error number on failure.
1081  */
1082 int
1083 ext4_xattr_set(struct inode *inode, int name_index, const char *name,
1084                const void *value, size_t value_len, int flags)
1085 {
1086         handle_t *handle;
1087         int error, retries = 0;
1088
1089 retry:
1090         handle = ext4_journal_start(inode, EXT4_DATA_TRANS_BLOCKS(inode->i_sb));
1091         if (IS_ERR(handle)) {
1092                 error = PTR_ERR(handle);
1093         } else {
1094                 int error2;
1095
1096                 error = ext4_xattr_set_handle(handle, inode, name_index, name,
1097                                               value, value_len, flags);
1098                 error2 = ext4_journal_stop(handle);
1099                 if (error == -ENOSPC &&
1100                     ext4_should_retry_alloc(inode->i_sb, &retries))
1101                         goto retry;
1102                 if (error == 0)
1103                         error = error2;
1104         }
1105
1106         return error;
1107 }
1108
1109 /*
1110  * Shift the EA entries in the inode to create space for the increased
1111  * i_extra_isize.
1112  */
1113 static void ext4_xattr_shift_entries(struct ext4_xattr_entry *entry,
1114                                      int value_offs_shift, void *to,
1115                                      void *from, size_t n, int blocksize)
1116 {
1117         struct ext4_xattr_entry *last = entry;
1118         int new_offs;
1119
1120         /* Adjust the value offsets of the entries */
1121         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
1122                 if (!last->e_value_block && last->e_value_size) {
1123                         new_offs = le16_to_cpu(last->e_value_offs) +
1124                                                         value_offs_shift;
1125                         BUG_ON(new_offs + le32_to_cpu(last->e_value_size)
1126                                  > blocksize);
1127                         last->e_value_offs = cpu_to_le16(new_offs);
1128                 }
1129         }
1130         /* Shift the entries by n bytes */
1131         memmove(to, from, n);
1132 }
1133
1134 /*
1135  * Expand an inode by new_extra_isize bytes when EAs are present.
1136  * Returns 0 on success or negative error number on failure.
1137  */
1138 int ext4_expand_extra_isize_ea(struct inode *inode, int new_extra_isize,
1139                                struct ext4_inode *raw_inode, handle_t *handle)
1140 {
1141         struct ext4_xattr_ibody_header *header;
1142         struct ext4_xattr_entry *entry, *last, *first;
1143         struct buffer_head *bh = NULL;
1144         struct ext4_xattr_ibody_find *is = NULL;
1145         struct ext4_xattr_block_find *bs = NULL;
1146         char *buffer = NULL, *b_entry_name = NULL;
1147         size_t min_offs, free;
1148         int total_ino, total_blk;
1149         void *base, *start, *end;
1150         int extra_isize = 0, error = 0, tried_min_extra_isize = 0;
1151         int s_min_extra_isize = le16_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_min_extra_isize);
1152
1153         down_write(&EXT4_I(inode)->xattr_sem);
1154 retry:
1155         if (EXT4_I(inode)->i_extra_isize >= new_extra_isize) {
1156                 up_write(&EXT4_I(inode)->xattr_sem);
1157                 return 0;
1158         }
1159
1160         header = IHDR(inode, raw_inode);
1161         entry = IFIRST(header);
1162
1163         /*
1164          * Check if enough free space is available in the inode to shift the
1165          * entries ahead by new_extra_isize.
1166          */
1167
1168         base = start = entry;
1169         end = (void *)raw_inode + EXT4_SB(inode->i_sb)->s_inode_size;
1170         min_offs = end - base;
1171         last = entry;
1172         total_ino = sizeof(struct ext4_xattr_ibody_header);
1173
1174         free = ext4_xattr_free_space(last, &min_offs, base, &total_ino);
1175         if (free >= new_extra_isize) {
1176                 entry = IFIRST(header);
1177                 ext4_xattr_shift_entries(entry, EXT4_I(inode)->i_extra_isize
1178                                 - new_extra_isize, (void *)raw_inode +
1179                                 EXT4_GOOD_OLD_INODE_SIZE + new_extra_isize,
1180                                 (void *)header, total_ino,
1181                                 inode->i_sb->s_blocksize);
1182                 EXT4_I(inode)->i_extra_isize = new_extra_isize;
1183                 error = 0;
1184                 goto cleanup;
1185         }
1186
1187         /*
1188          * Enough free space isn't available in the inode, check if
1189          * EA block can hold new_extra_isize bytes.
1190          */
1191         if (EXT4_I(inode)->i_file_acl) {
1192                 bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
1193                 error = -EIO;
1194                 if (!bh)
1195                         goto cleanup;
1196                 if (ext4_xattr_check_block(bh)) {
1197                         ext4_error(inode->i_sb, "inode %lu: bad block %llu",
1198                                    inode->i_ino, EXT4_I(inode)->i_file_acl);
1199                         error = -EIO;
1200                         goto cleanup;
1201                 }
1202                 base = BHDR(bh);
1203                 first = BFIRST(bh);
1204                 end = bh->b_data + bh->b_size;
1205                 min_offs = end - base;
1206                 free = ext4_xattr_free_space(first, &min_offs, base,
1207                                              &total_blk);
1208                 if (free < new_extra_isize) {
1209                         if (!tried_min_extra_isize && s_min_extra_isize) {
1210                                 tried_min_extra_isize++;
1211                                 new_extra_isize = s_min_extra_isize;
1212                                 brelse(bh);
1213                                 goto retry;
1214                         }
1215                         error = -1;
1216                         goto cleanup;
1217                 }
1218         } else {
1219                 free = inode->i_sb->s_blocksize;
1220         }
1221
1222         while (new_extra_isize > 0) {
1223                 size_t offs, size, entry_size;
1224                 struct ext4_xattr_entry *small_entry = NULL;
1225                 struct ext4_xattr_info i = {
1226                         .value = NULL,
1227                         .value_len = 0,
1228                 };
1229                 unsigned int total_size;  /* EA entry size + value size */
1230                 unsigned int shift_bytes; /* No. of bytes to shift EAs by? */
1231                 unsigned int min_total_size = ~0U;
1232
1233                 is = kzalloc(sizeof(struct ext4_xattr_ibody_find), GFP_NOFS);
1234                 bs = kzalloc(sizeof(struct ext4_xattr_block_find), GFP_NOFS);
1235                 if (!is || !bs) {
1236                         error = -ENOMEM;
1237                         goto cleanup;
1238                 }
1239
1240                 is->s.not_found = -ENODATA;
1241                 bs->s.not_found = -ENODATA;
1242                 is->iloc.bh = NULL;
1243                 bs->bh = NULL;
1244
1245                 last = IFIRST(header);
1246                 /* Find the entry best suited to be pushed into EA block */
1247                 entry = NULL;
1248                 for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
1249                         total_size =
1250                         EXT4_XATTR_SIZE(le32_to_cpu(last->e_value_size)) +
1251                                         EXT4_XATTR_LEN(last->e_name_len);
1252                         if (total_size <= free && total_size < min_total_size) {
1253                                 if (total_size < new_extra_isize) {
1254                                         small_entry = last;
1255                                 } else {
1256                                         entry = last;
1257                                         min_total_size = total_size;
1258                                 }
1259                         }
1260                 }
1261
1262                 if (entry == NULL) {
1263                         if (small_entry) {
1264                                 entry = small_entry;
1265                         } else {
1266                                 if (!tried_min_extra_isize &&
1267                                     s_min_extra_isize) {
1268                                         tried_min_extra_isize++;
1269                                         new_extra_isize = s_min_extra_isize;
1270                                         goto retry;
1271                                 }
1272                                 error = -1;
1273                                 goto cleanup;
1274                         }
1275                 }
1276                 offs = le16_to_cpu(entry->e_value_offs);
1277                 size = le32_to_cpu(entry->e_value_size);
1278                 entry_size = EXT4_XATTR_LEN(entry->e_name_len);
1279                 i.name_index = entry->e_name_index,
1280                 buffer = kmalloc(EXT4_XATTR_SIZE(size), GFP_NOFS);
1281                 b_entry_name = kmalloc(entry->e_name_len + 1, GFP_NOFS);
1282                 if (!buffer || !b_entry_name) {
1283                         error = -ENOMEM;
1284                         goto cleanup;
1285                 }
1286                 /* Save the entry name and the entry value */
1287                 memcpy(buffer, (void *)IFIRST(header) + offs,
1288                        EXT4_XATTR_SIZE(size));
1289                 memcpy(b_entry_name, entry->e_name, entry->e_name_len);
1290                 b_entry_name[entry->e_name_len] = '\0';
1291                 i.name = b_entry_name;
1292
1293                 error = ext4_get_inode_loc(inode, &is->iloc);
1294                 if (error)
1295                         goto cleanup;
1296
1297                 error = ext4_xattr_ibody_find(inode, &i, is);
1298                 if (error)
1299                         goto cleanup;
1300
1301                 /* Remove the chosen entry from the inode */
1302                 error = ext4_xattr_ibody_set(handle, inode, &i, is);
1303                 if (error)
1304                         goto cleanup;
1305
1306                 entry = IFIRST(header);
1307                 if (entry_size + EXT4_XATTR_SIZE(size) >= new_extra_isize)
1308                         shift_bytes = new_extra_isize;
1309                 else
1310                         shift_bytes = entry_size + size;
1311                 /* Adjust the offsets and shift the remaining entries ahead */
1312                 ext4_xattr_shift_entries(entry, EXT4_I(inode)->i_extra_isize -
1313                         shift_bytes, (void *)raw_inode +
1314                         EXT4_GOOD_OLD_INODE_SIZE + extra_isize + shift_bytes,
1315                         (void *)header, total_ino - entry_size,
1316                         inode->i_sb->s_blocksize);
1317
1318                 extra_isize += shift_bytes;
1319                 new_extra_isize -= shift_bytes;
1320                 EXT4_I(inode)->i_extra_isize = extra_isize;
1321
1322                 i.name = b_entry_name;
1323                 i.value = buffer;
1324                 i.value_len = size;
1325                 error = ext4_xattr_block_find(inode, &i, bs);
1326                 if (error)
1327                         goto cleanup;
1328
1329                 /* Add entry which was removed from the inode into the block */
1330                 error = ext4_xattr_block_set(handle, inode, &i, bs);
1331                 if (error)
1332                         goto cleanup;
1333                 kfree(b_entry_name);
1334                 kfree(buffer);
1335                 b_entry_name = NULL;
1336                 buffer = NULL;
1337                 brelse(is->iloc.bh);
1338                 kfree(is);
1339                 kfree(bs);
1340         }
1341         brelse(bh);
1342         up_write(&EXT4_I(inode)->xattr_sem);
1343         return 0;
1344
1345 cleanup:
1346         kfree(b_entry_name);
1347         kfree(buffer);
1348         if (is)
1349                 brelse(is->iloc.bh);
1350         kfree(is);
1351         kfree(bs);
1352         brelse(bh);
1353         up_write(&EXT4_I(inode)->xattr_sem);
1354         return error;
1355 }
1356
1357
1358
1359 /*
1360  * ext4_xattr_delete_inode()
1361  *
1362  * Free extended attribute resources associated with this inode. This
1363  * is called immediately before an inode is freed. We have exclusive
1364  * access to the inode.
1365  */
1366 void
1367 ext4_xattr_delete_inode(handle_t *handle, struct inode *inode)
1368 {
1369         struct buffer_head *bh = NULL;
1370
1371         if (!EXT4_I(inode)->i_file_acl)
1372                 goto cleanup;
1373         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
1374         if (!bh) {
1375                 ext4_error(inode->i_sb, "inode %lu: block %llu read error",
1376                            inode->i_ino, EXT4_I(inode)->i_file_acl);
1377                 goto cleanup;
1378         }
1379         if (BHDR(bh)->h_magic != cpu_to_le32(EXT4_XATTR_MAGIC) ||
1380             BHDR(bh)->h_blocks != cpu_to_le32(1)) {
1381                 ext4_error(inode->i_sb, "inode %lu: bad block %llu",
1382                            inode->i_ino, EXT4_I(inode)->i_file_acl);
1383                 goto cleanup;
1384         }
1385         ext4_xattr_release_block(handle, inode, bh);
1386         EXT4_I(inode)->i_file_acl = 0;
1387
1388 cleanup:
1389         brelse(bh);
1390 }
1391
1392 /*
1393  * ext4_xattr_put_super()
1394  *
1395  * This is called when a file system is unmounted.
1396  */
1397 void
1398 ext4_xattr_put_super(struct super_block *sb)
1399 {
1400         mb_cache_shrink(sb->s_bdev);
1401 }
1402
1403 /*
1404  * ext4_xattr_cache_insert()
1405  *
1406  * Create a new entry in the extended attribute cache, and insert
1407  * it unless such an entry is already in the cache.
1408  *
1409  * Returns 0, or a negative error number on failure.
1410  */
1411 static void
1412 ext4_xattr_cache_insert(struct buffer_head *bh)
1413 {
1414         __u32 hash = le32_to_cpu(BHDR(bh)->h_hash);
1415         struct mb_cache_entry *ce;
1416         int error;
1417
1418         ce = mb_cache_entry_alloc(ext4_xattr_cache, GFP_NOFS);
1419         if (!ce) {
1420                 ea_bdebug(bh, "out of memory");
1421                 return;
1422         }
1423         error = mb_cache_entry_insert(ce, bh->b_bdev, bh->b_blocknr, &hash);
1424         if (error) {
1425                 mb_cache_entry_free(ce);
1426                 if (error == -EBUSY) {
1427                         ea_bdebug(bh, "already in cache");
1428                         error = 0;
1429                 }
1430         } else {
1431                 ea_bdebug(bh, "inserting [%x]", (int)hash);
1432                 mb_cache_entry_release(ce);
1433         }
1434 }
1435
1436 /*
1437  * ext4_xattr_cmp()
1438  *
1439  * Compare two extended attribute blocks for equality.
1440  *
1441  * Returns 0 if the blocks are equal, 1 if they differ, and
1442  * a negative error number on errors.
1443  */
1444 static int
1445 ext4_xattr_cmp(struct ext4_xattr_header *header1,
1446                struct ext4_xattr_header *header2)
1447 {
1448         struct ext4_xattr_entry *entry1, *entry2;
1449
1450         entry1 = ENTRY(header1+1);
1451         entry2 = ENTRY(header2+1);
1452         while (!IS_LAST_ENTRY(entry1)) {
1453                 if (IS_LAST_ENTRY(entry2))
1454                         return 1;
1455                 if (entry1->e_hash != entry2->e_hash ||
1456                     entry1->e_name_index != entry2->e_name_index ||
1457                     entry1->e_name_len != entry2->e_name_len ||
1458                     entry1->e_value_size != entry2->e_value_size ||
1459                     memcmp(entry1->e_name, entry2->e_name, entry1->e_name_len))
1460                         return 1;
1461                 if (entry1->e_value_block != 0 || entry2->e_value_block != 0)
1462                         return -EIO;
1463                 if (memcmp((char *)header1 + le16_to_cpu(entry1->e_value_offs),
1464                            (char *)header2 + le16_to_cpu(entry2->e_value_offs),
1465                            le32_to_cpu(entry1->e_value_size)))
1466                         return 1;
1467
1468                 entry1 = EXT4_XATTR_NEXT(entry1);
1469                 entry2 = EXT4_XATTR_NEXT(entry2);
1470         }
1471         if (!IS_LAST_ENTRY(entry2))
1472                 return 1;
1473         return 0;
1474 }
1475
1476 /*
1477  * ext4_xattr_cache_find()
1478  *
1479  * Find an identical extended attribute block.
1480  *
1481  * Returns a pointer to the block found, or NULL if such a block was
1482  * not found or an error occurred.
1483  */
1484 static struct buffer_head *
1485 ext4_xattr_cache_find(struct inode *inode, struct ext4_xattr_header *header,
1486                       struct mb_cache_entry **pce)
1487 {
1488         __u32 hash = le32_to_cpu(header->h_hash);
1489         struct mb_cache_entry *ce;
1490
1491         if (!header->h_hash)
1492                 return NULL;  /* never share */
1493         ea_idebug(inode, "looking for cached blocks [%x]", (int)hash);
1494 again:
1495         ce = mb_cache_entry_find_first(ext4_xattr_cache, 0,
1496                                        inode->i_sb->s_bdev, hash);
1497         while (ce) {
1498                 struct buffer_head *bh;
1499
1500                 if (IS_ERR(ce)) {
1501                         if (PTR_ERR(ce) == -EAGAIN)
1502                                 goto again;
1503                         break;
1504                 }
1505                 bh = sb_bread(inode->i_sb, ce->e_block);
1506                 if (!bh) {
1507                         ext4_error(inode->i_sb,
1508                                 "inode %lu: block %lu read error",
1509                                 inode->i_ino, (unsigned long) ce->e_block);
1510                 } else if (le32_to_cpu(BHDR(bh)->h_refcount) >=
1511                                 EXT4_XATTR_REFCOUNT_MAX) {
1512                         ea_idebug(inode, "block %lu refcount %d>=%d",
1513                                   (unsigned long) ce->e_block,
1514                                   le32_to_cpu(BHDR(bh)->h_refcount),
1515                                           EXT4_XATTR_REFCOUNT_MAX);
1516                 } else if (ext4_xattr_cmp(header, BHDR(bh)) == 0) {
1517                         *pce = ce;
1518                         return bh;
1519                 }
1520                 brelse(bh);
1521                 ce = mb_cache_entry_find_next(ce, 0, inode->i_sb->s_bdev, hash);
1522         }
1523         return NULL;
1524 }
1525
1526 #define NAME_HASH_SHIFT 5
1527 #define VALUE_HASH_SHIFT 16
1528
1529 /*
1530  * ext4_xattr_hash_entry()
1531  *
1532  * Compute the hash of an extended attribute.
1533  */
1534 static inline void ext4_xattr_hash_entry(struct ext4_xattr_header *header,
1535                                          struct ext4_xattr_entry *entry)
1536 {
1537         __u32 hash = 0;
1538         char *name = entry->e_name;
1539         int n;
1540
1541         for (n = 0; n < entry->e_name_len; n++) {
1542                 hash = (hash << NAME_HASH_SHIFT) ^
1543                        (hash >> (8*sizeof(hash) - NAME_HASH_SHIFT)) ^
1544                        *name++;
1545         }
1546
1547         if (entry->e_value_block == 0 && entry->e_value_size != 0) {
1548                 __le32 *value = (__le32 *)((char *)header +
1549                         le16_to_cpu(entry->e_value_offs));
1550                 for (n = (le32_to_cpu(entry->e_value_size) +
1551                      EXT4_XATTR_ROUND) >> EXT4_XATTR_PAD_BITS; n; n--) {
1552                         hash = (hash << VALUE_HASH_SHIFT) ^
1553                                (hash >> (8*sizeof(hash) - VALUE_HASH_SHIFT)) ^
1554                                le32_to_cpu(*value++);
1555                 }
1556         }
1557         entry->e_hash = cpu_to_le32(hash);
1558 }
1559
1560 #undef NAME_HASH_SHIFT
1561 #undef VALUE_HASH_SHIFT
1562
1563 #define BLOCK_HASH_SHIFT 16
1564
1565 /*
1566  * ext4_xattr_rehash()
1567  *
1568  * Re-compute the extended attribute hash value after an entry has changed.
1569  */
1570 static void ext4_xattr_rehash(struct ext4_xattr_header *header,
1571                               struct ext4_xattr_entry *entry)
1572 {
1573         struct ext4_xattr_entry *here;
1574         __u32 hash = 0;
1575
1576         ext4_xattr_hash_entry(header, entry);
1577         here = ENTRY(header+1);
1578         while (!IS_LAST_ENTRY(here)) {
1579                 if (!here->e_hash) {
1580                         /* Block is not shared if an entry's hash value == 0 */
1581                         hash = 0;
1582                         break;
1583                 }
1584                 hash = (hash << BLOCK_HASH_SHIFT) ^
1585                        (hash >> (8*sizeof(hash) - BLOCK_HASH_SHIFT)) ^
1586                        le32_to_cpu(here->e_hash);
1587                 here = EXT4_XATTR_NEXT(here);
1588         }
1589         header->h_hash = cpu_to_le32(hash);
1590 }
1591
1592 #undef BLOCK_HASH_SHIFT
1593
1594 int __init
1595 init_ext4_xattr(void)
1596 {
1597         ext4_xattr_cache = mb_cache_create("ext4_xattr", NULL,
1598                 sizeof(struct mb_cache_entry) +
1599                 sizeof(((struct mb_cache_entry *) 0)->e_indexes[0]), 1, 6);
1600         if (!ext4_xattr_cache)
1601                 return -ENOMEM;
1602         return 0;
1603 }
1604
1605 void
1606 exit_ext4_xattr(void)
1607 {
1608         if (ext4_xattr_cache)
1609                 mb_cache_destroy(ext4_xattr_cache);
1610         ext4_xattr_cache = NULL;
1611 }