ldiskfs/kernel_patches/patches/rhel6.3/ext4-large-eas.patch

   1 Index: linux-stage/fs/ext4/ext4.h
   2 ===================================================================
   3 --- linux-stage.orig/fs/ext4/ext4.h
   4 +++ linux-stage/fs/ext4/ext4.h
   5 @@ -1333,6 +1333,7 @@ EXT4_INODE_BIT_FNS(state, state_flags)
   6  #define EXT4_FEATURE_INCOMPAT_64BIT            0x0080
   7  #define EXT4_FEATURE_INCOMPAT_MMP               0x0100
   8  #define EXT4_FEATURE_INCOMPAT_FLEX_BG          0x0200
   9 +#define EXT4_FEATURE_INCOMPAT_EA_INODE         0x0400
  10  #define EXT4_FEATURE_INCOMPAT_DIRDATA          0x1000
  11
  12  #define EXT4_FEATURE_COMPAT_SUPP       EXT2_FEATURE_COMPAT_EXT_ATTR
  13 @@ -1342,6 +1343,7 @@ EXT4_INODE_BIT_FNS(state, state_flags)
  14                                          EXT4_FEATURE_INCOMPAT_EXTENTS| \
  15                                          EXT4_FEATURE_INCOMPAT_64BIT| \
  16                                          EXT4_FEATURE_INCOMPAT_FLEX_BG| \
  17 +                                        EXT4_FEATURE_INCOMPAT_EA_INODE| \
  18                                          EXT4_FEATURE_INCOMPAT_MMP| \
  19                                          EXT4_FEATURE_INCOMPAT_DIRDATA)
  20
  21 @@ -1706,6 +1708,12 @@ struct mmpd_data {
  22  #endif
  23
  24  /*
  25 + * Maximum size of xattr attributes for FEATURE_INCOMPAT_EA_INODE 1Mb
  26 + * This limit is arbitrary, but is reasonable for the xattr API.
  27 + */
  28 +#define EXT4_XATTR_MAX_LARGE_EA_SIZE    (1024 * 1024)
  29 +
  30 +/*
  31   * Function prototypes
  32   */
  33
  34 @@ -1717,6 +1725,10 @@ struct mmpd_data {
  35  # define ATTRIB_NORET  __attribute__((noreturn))
  36  # define NORET_AND     noreturn,
  37
  38 +struct ext4_xattr_ino_array {
  39 +       unsigned int xia_count;         /* # of used item in the array */
  40 +       unsigned int xia_inodes[0];
  41 +};
  42  /* bitmap.c */
  43  extern unsigned int ext4_count_free(struct buffer_head *, unsigned);
  44
  45 Index: linux-stage/fs/ext4/xattr.c
  46 ===================================================================
  47 --- linux-stage.orig/fs/ext4/xattr.c
  48 +++ linux-stage/fs/ext4/xattr.c
  49 @@ -168,19 +168,26 @@ ext4_xattr_check_block(struct buffer_hea
  50  }
  51
  52  static inline int
  53 -ext4_xattr_check_entry(struct ext4_xattr_entry *entry, size_t size)
  54 +ext4_xattr_check_entry(struct ext4_xattr_entry *entry, size_t size,
  55 +                      struct inode *inode)
  56  {
  57         size_t value_size = le32_to_cpu(entry->e_value_size);
  58
  59 -       if (entry->e_value_block != 0 || value_size > size ||
  60 -           le16_to_cpu(entry->e_value_offs) + value_size > size)
  61 +       if ((entry->e_value_inum == 0) &&
  62 +          (le16_to_cpu(entry->e_value_offs) + value_size > size))
  63 +               return -EIO;
  64 +       if (entry->e_value_inum != 0 &&
  65 +           (le32_to_cpu(entry->e_value_inum) < EXT4_FIRST_INO(inode->i_sb) ||
  66 +            le32_to_cpu(entry->e_value_inum) >
  67 +            le32_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_inodes_count)))
  68                 return -EIO;
  69         return 0;
  70  }
  71
  72  static int
  73  ext4_xattr_find_entry(struct ext4_xattr_entry **pentry, int name_index,
  74 -                     const char *name, size_t size, int sorted)
  75 +                     const char *name, size_t size, int sorted,
  76 +                     struct inode *inode)
  77  {
  78         struct ext4_xattr_entry *entry;
  79         size_t name_len;
  80 @@ -200,11 +207,103 @@ ext4_xattr_find_entry(struct ext4_xattr_
  81                         break;
  82         }
  83         *pentry = entry;
  84 -       if (!cmp && ext4_xattr_check_entry(entry, size))
  85 +       if (!cmp && ext4_xattr_check_entry(entry, size, inode))
  86                         return -EIO;
  87         return cmp ? -ENODATA : 0;
  88  }
  89
  90 +/*
  91 + * Read the EA value from an inode.
  92 + */
  93 +static int
  94 +ext4_xattr_inode_read(struct inode *ea_inode, void *buf, size_t *size)
  95 +{
  96 +       unsigned long block = 0;
  97 +       struct buffer_head *bh = NULL;
  98 +       int err, blocksize;
  99 +       size_t csize, ret_size = 0;
 100 +
 101 +       if (*size == 0)
 102 +               return 0;
 103 +
 104 +       blocksize = ea_inode->i_sb->s_blocksize;
 105 +
 106 +       while (ret_size < *size) {
 107 +               csize = (*size - ret_size) > blocksize ? blocksize :
 108 +                                                       *size - ret_size;
 109 +               bh = ext4_bread(NULL, ea_inode, block, 0, &err);
 110 +               if (!bh) {
 111 +                       *size = ret_size;
 112 +                       return err;
 113 +               }
 114 +               memcpy(buf, bh->b_data, csize);
 115 +               brelse(bh);
 116 +
 117 +               buf += csize;
 118 +               block += 1;
 119 +               ret_size += csize;
 120 +       }
 121 +
 122 +       *size = ret_size;
 123 +
 124 +       return err;
 125 +}
 126 +
 127 +struct inode *ext4_xattr_inode_iget(struct inode *parent, int ea_ino, int *err)
 128 +{
 129 +       struct inode *ea_inode = NULL;
 130 +
 131 +       ea_inode = ext4_iget(parent->i_sb, ea_ino);
 132 +       if (IS_ERR(ea_inode) || is_bad_inode(ea_inode)) {
 133 +               ext4_error(parent->i_sb, "error while reading EA inode %d",
 134 +                          ea_ino);
 135 +               *err = -EIO;
 136 +               return NULL;
 137 +       }
 138 +
 139 +       if (ea_inode->i_xattr_inode_parent != parent->i_ino ||
 140 +           ea_inode->i_generation != parent->i_generation) {
 141 +               ext4_error(parent->i_sb, "Backpointer from EA inode %d "
 142 +                          "to parent invalid.", ea_ino);
 143 +               *err = -EINVAL;
 144 +               goto error;
 145 +       }
 146 +
 147 +       if (!(EXT4_I(ea_inode)->i_flags & EXT4_EA_INODE_FL)) {
 148 +               ext4_error(parent->i_sb, "EA inode %d does not have "
 149 +                          "EXT4_EA_INODE_FL flag set.\n", ea_ino);
 150 +               *err = -EINVAL;
 151 +               goto error;
 152 +       }
 153 +
 154 +       *err = 0;
 155 +       return ea_inode;
 156 +
 157 +error:
 158 +       iput(ea_inode);
 159 +       return NULL;
 160 +}
 161 +
 162 +/*
 163 + * Read the value from the EA inode.
 164 + */
 165 +static int
 166 +ext4_xattr_inode_get(struct inode *inode, int ea_ino, void *buffer,
 167 +                    size_t *size)
 168 +{
 169 +       struct inode *ea_inode = NULL;
 170 +       int err;
 171 +
 172 +       ea_inode = ext4_xattr_inode_iget(inode, ea_ino, &err);
 173 +       if (err)
 174 +               return err;
 175 +
 176 +       err = ext4_xattr_inode_read(ea_inode, buffer, size);
 177 +       iput(ea_inode);
 178 +
 179 +       return err;
 180 +}
 181 +
 182  static int
 183  ext4_xattr_block_get(struct inode *inode, int name_index, const char *name,
 184                      void *buffer, size_t buffer_size)
 185 @@ -236,7 +335,8 @@ bad_block:
 186         }
 187         ext4_xattr_cache_insert(bh);
 188         entry = BFIRST(bh);
 189 -       error = ext4_xattr_find_entry(&entry, name_index, name, bh->b_size, 1);
 190 +       error = ext4_xattr_find_entry(&entry, name_index, name, bh->b_size, 1,
 191 +                                     inode);
 192         if (error == -EIO)
 193                 goto bad_block;
 194         if (error)
 195 @@ -246,8 +346,16 @@ bad_block:
 196                 error = -ERANGE;
 197                 if (size > buffer_size)
 198                         goto cleanup;
 199 -               memcpy(buffer, bh->b_data + le16_to_cpu(entry->e_value_offs),
 200 -                      size);
 201 +               if (entry->e_value_inum != 0) {
 202 +                       error = ext4_xattr_inode_get(inode,
 203 +                                            le32_to_cpu(entry->e_value_inum),
 204 +                                            buffer, &size);
 205 +                       if (error)
 206 +                               goto cleanup;
 207 +               } else {
 208 +                       memcpy(buffer, bh->b_data +
 209 +                              le16_to_cpu(entry->e_value_offs), size);
 210 +               }
 211         }
 212         error = size;
 213
 214 @@ -281,7 +389,7 @@ ext4_xattr_ibody_get(struct inode *inode
 215         if (error)
 216                 goto cleanup;
 217         error = ext4_xattr_find_entry(&entry, name_index, name,
 218 -                                     end - (void *)entry, 0);
 219 +                                     end - (void *)entry, 0, inode);
 220         if (error)
 221                 goto cleanup;
 222         size = le32_to_cpu(entry->e_value_size);
 223 @@ -289,8 +397,16 @@ ext4_xattr_ibody_get(struct inode *inode
 224                 error = -ERANGE;
 225                 if (size > buffer_size)
 226                         goto cleanup;
 227 -               memcpy(buffer, (void *)IFIRST(header) +
 228 -                      le16_to_cpu(entry->e_value_offs), size);
 229 +               if (entry->e_value_inum != 0) {
 230 +                       error = ext4_xattr_inode_get(inode,
 231 +                                            le32_to_cpu(entry->e_value_inum),
 232 +                                            buffer, &size);
 233 +                       if (error)
 234 +                               goto cleanup;
 235 +               } else {
 236 +                       memcpy(buffer, (void *)IFIRST(header) +
 237 +                              le16_to_cpu(entry->e_value_offs), size);
 238 +               }
 239         }
 240         error = size;
 241
 242 @@ -513,7 +629,7 @@ static size_t ext4_xattr_free_space(stru
 243  {
 244         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 245                 *total += EXT4_XATTR_LEN(last->e_name_len);
 246 -               if (!last->e_value_block && last->e_value_size) {
 247 +               if (last->e_value_inum == 0 && last->e_value_size > 0) {
 248                         size_t offs = le16_to_cpu(last->e_value_offs);
 249                         if (offs < *min_offs)
 250                                 *min_offs = offs;
 251 @@ -522,11 +638,159 @@ static size_t ext4_xattr_free_space(stru
 252         return (*min_offs - ((void *)last - base) - sizeof(__u32));
 253  }
 254
 255 +/*
 256 + * Write the value of the EA in an inode.
 257 + */
 258 +static int
 259 +ext4_xattr_inode_write(handle_t *handle, struct inode *ea_inode,
 260 +                      const void *buf, int bufsize)
 261 +{
 262 +       struct buffer_head *bh = NULL, dummy;
 263 +       unsigned long block = 0;
 264 +       unsigned blocksize = ea_inode->i_sb->s_blocksize;
 265 +       unsigned max_blocks = (bufsize + blocksize - 1) >> ea_inode->i_blkbits;
 266 +       int csize, wsize = 0;
 267 +       int ret = 0;
 268 +       int retries = 0;
 269 +
 270 +retry:
 271 +       while (ret >= 0 && ret < max_blocks) {
 272 +               block += ret;
 273 +               max_blocks -= ret;
 274 +
 275 +               ret = ext4_get_blocks(handle, ea_inode, block, max_blocks,
 276 +                                     &dummy, EXT4_GET_BLOCKS_CREATE);
 277 +               if (ret <= 0) {
 278 +                       ext4_mark_inode_dirty(handle, ea_inode);
 279 +                       if (ret == -ENOSPC &&
 280 +                           ext4_should_retry_alloc(ea_inode->i_sb, &retries)) {
 281 +                               ret = 0;
 282 +                               goto retry;
 283 +                       }
 284 +                       break;
 285 +               }
 286 +       }
 287 +
 288 +       if (ret < 0)
 289 +               return ret;
 290 +
 291 +       block = 0;
 292 +       while (wsize < bufsize) {
 293 +               if (bh != NULL)
 294 +                       brelse(bh);
 295 +               csize = (bufsize - wsize) > blocksize ? blocksize :
 296 +                                                               bufsize - wsize;
 297 +               bh = ext4_getblk(handle, ea_inode, block, 0, &ret);
 298 +               if (!bh)
 299 +                       goto out;
 300 +               ret = ext4_journal_get_write_access(handle, bh);
 301 +               if (ret)
 302 +                       goto out;
 303 +
 304 +               memcpy(bh->b_data, buf, csize);
 305 +               set_buffer_uptodate(bh);
 306 +               ext4_handle_dirty_metadata(handle, ea_inode, bh);
 307 +
 308 +               buf += csize;
 309 +               wsize += csize;
 310 +               block += 1;
 311 +       }
 312 +
 313 +       i_size_write(ea_inode, wsize);
 314 +       ext4_update_i_disksize(ea_inode, wsize);
 315 +
 316 +       ext4_mark_inode_dirty(handle, ea_inode);
 317 +
 318 +out:
 319 +       brelse(bh);
 320 +
 321 +       return ret;
 322 +}
 323 +
 324 +/*
 325 + * Create an inode to store the value of a large EA.
 326 + */
 327 +static struct inode *
 328 +ext4_xattr_inode_create(handle_t *handle, struct inode *inode)
 329 +{
 330 +       struct inode *ea_inode = NULL;
 331 +
 332 +       /*
 333 +        * Let the next inode be the goal, so we try and allocate the EA inode
 334 +        * in the same group, or nearby one.
 335 +        */
 336 +       ea_inode = ext4_new_inode(handle, inode->i_sb->s_root->d_inode,
 337 +                                 S_IFREG|0600, NULL, inode->i_ino + 1);
 338 +
 339 +       if (!IS_ERR(ea_inode)) {
 340 +               ea_inode->i_op = &ext4_file_inode_operations;
 341 +               ea_inode->i_fop = &ext4_file_operations;
 342 +               ext4_set_aops(ea_inode);
 343 +               ea_inode->i_generation = inode->i_generation;
 344 +               EXT4_I(ea_inode)->i_flags |= EXT4_EA_INODE_FL;
 345 +
 346 +               /*
 347 +                * A back-pointer from EA inode to parent inode will be useful
 348 +                * for e2fsck.
 349 +                */
 350 +               ea_inode->i_xattr_inode_parent = inode->i_ino;
 351 +               unlock_new_inode(ea_inode);
 352 +       }
 353 +
 354 +       return ea_inode;
 355 +}
 356 +
 357 +/*
 358 + * Unlink the inode storing the value of the EA.
 359 + */
 360 +int
 361 +ext4_xattr_inode_unlink(struct inode *inode, int ea_ino)
 362 +{
 363 +       struct inode *ea_inode = NULL;
 364 +       int err;
 365 +
 366 +       ea_inode = ext4_xattr_inode_iget(inode, ea_ino, &err);
 367 +       if (err)
 368 +               return err;
 369 +
 370 +       ea_inode->i_nlink = 0;
 371 +       iput(ea_inode);
 372 +
 373 +       return 0;
 374 +}
 375 +
 376 +/*
 377 + * Add value of the EA in an inode.
 378 + */
 379 +static int
 380 +ext4_xattr_inode_set(handle_t *handle, struct inode *inode, int *ea_ino,
 381 +                    const void *value, size_t value_len)
 382 +{
 383 +       struct inode *ea_inode = NULL;
 384 +       int err;
 385 +
 386 +       /* Create an inode for the EA value */
 387 +       ea_inode = ext4_xattr_inode_create(handle, inode);
 388 +       if (IS_ERR(ea_inode))
 389 +               return -1;
 390 +
 391 +       err = ext4_xattr_inode_write(handle, ea_inode, value, value_len);
 392 +       if (err)
 393 +               ea_inode->i_nlink = 0;
 394 +       else
 395 +               *ea_ino = ea_inode->i_ino;
 396 +
 397 +       iput(ea_inode);
 398 +
 399 +       return err;
 400 +}
 401 +
 402  struct ext4_xattr_info {
 403 -       int name_index;
 404         const char *name;
 405         const void *value;
 406         size_t value_len;
 407 +       int name_index;
 408 +       int in_inode;
 409  };
 410
 411  struct ext4_xattr_search {
 412 @@ -538,15 +802,23 @@ struct ext4_xattr_search {
 413  };
 414
 415  static int
 416 -ext4_xattr_set_entry(struct ext4_xattr_info *i, struct ext4_xattr_search *s)
 417 +ext4_xattr_set_entry(struct ext4_xattr_info *i, struct ext4_xattr_search *s,
 418 +                    handle_t *handle, struct inode *inode)
 419  {
 420         struct ext4_xattr_entry *last;
 421         size_t free, min_offs = s->end - s->base, name_len = strlen(i->name);
 422 +       int in_inode = i->in_inode;
 423 +
 424 +       if (EXT4_HAS_INCOMPAT_FEATURE(inode->i_sb,
 425 +                EXT4_FEATURE_INCOMPAT_EA_INODE) &&
 426 +           (EXT4_XATTR_SIZE(i->value_len) >
 427 +            EXT4_XATTR_MIN_LARGE_EA_SIZE(inode->i_sb->s_blocksize)))
 428 +               in_inode = 1;
 429
 430         /* Compute min_offs and last. */
 431         last = s->first;
 432         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 433 -               if (!last->e_value_block && last->e_value_size) {
 434 +               if (last->e_value_inum == 0 && last->e_value_size > 0) {
 435                         size_t offs = le16_to_cpu(last->e_value_offs);
 436                         if (offs < min_offs)
 437                                 min_offs = offs;
 438 @@ -554,16 +826,21 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 439         }
 440         free = min_offs - ((void *)last - s->base) - sizeof(__u32);
 441         if (!s->not_found) {
 442 -               if (!s->here->e_value_block && s->here->e_value_size) {
 443 +               if (!in_inode && s->here->e_value_inum == 0 &&
 444 +                   s->here->e_value_size > 0) {
 445                         size_t size = le32_to_cpu(s->here->e_value_size);
 446                         free += EXT4_XATTR_SIZE(size);
 447                 }
 448                 free += EXT4_XATTR_LEN(name_len);
 449         }
 450         if (i->value) {
 451 -               if (free < EXT4_XATTR_SIZE(i->value_len) ||
 452 -                   free < EXT4_XATTR_LEN(name_len) +
 453 -                          EXT4_XATTR_SIZE(i->value_len))
 454 +               size_t value_len = EXT4_XATTR_SIZE(i->value_len);
 455 +
 456 +               if (in_inode)
 457 +                       value_len = 0;
 458 +
 459 +               if (free < value_len ||
 460 +                   free < EXT4_XATTR_LEN(name_len) + value_len)
 461                         return -ENOSPC;
 462         }
 463
 464 @@ -577,7 +854,8 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 465                 s->here->e_name_len = name_len;
 466                 memcpy(s->here->e_name, i->name, name_len);
 467         } else {
 468 -               if (!s->here->e_value_block && s->here->e_value_size) {
 469 +               if (s->here->e_value_offs > 0 && s->here->e_value_inum == 0 &&
 470 +                   s->here->e_value_size > 0) {
 471                         void *first_val = s->base + min_offs;
 472                         size_t offs = le16_to_cpu(s->here->e_value_offs);
 473                         void *val = s->base + offs;
 474 @@ -606,13 +884,17 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 475                         last = s->first;
 476                         while (!IS_LAST_ENTRY(last)) {
 477                                 size_t o = le16_to_cpu(last->e_value_offs);
 478 -                               if (!last->e_value_block &&
 479 -                                   last->e_value_size && o < offs)
 480 +                               if (last->e_value_size > 0 && o < offs)
 481                                         last->e_value_offs =
 482                                                 cpu_to_le16(o + size);
 483                                 last = EXT4_XATTR_NEXT(last);
 484                         }
 485                 }
 486 +               if (s->here->e_value_inum != 0) {
 487 +                       ext4_xattr_inode_unlink(inode,
 488 +                                       le32_to_cpu(s->here->e_value_inum));
 489 +                       s->here->e_value_inum = 0;
 490 +               }
 491                 if (!i->value) {
 492                         /* Remove the old name. */
 493                         size_t size = EXT4_XATTR_LEN(name_len);
 494 @@ -626,10 +908,17 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 495         if (i->value) {
 496                 /* Insert the new value. */
 497                 s->here->e_value_size = cpu_to_le32(i->value_len);
 498 -               if (i->value_len) {
 499 +               if (in_inode) {
 500 +                       int ea_ino = le32_to_cpu(s->here->e_value_inum);
 501 +                       ext4_xattr_inode_set(handle, inode, &ea_ino, i->value,
 502 +                                            i->value_len);
 503 +                       s->here->e_value_inum = cpu_to_le32(ea_ino);
 504 +                       s->here->e_value_offs = 0;
 505 +               } else if (i->value_len) {
 506                         size_t size = EXT4_XATTR_SIZE(i->value_len);
 507                         void *val = s->base + min_offs - size;
 508                         s->here->e_value_offs = cpu_to_le16(min_offs - size);
 509 +                       s->here->e_value_inum = 0;
 510                         memset(val + size - EXT4_XATTR_PAD, 0,
 511                                EXT4_XATTR_PAD); /* Clear the pad bytes. */
 512                         memcpy(val, i->value, i->value_len);
 513 @@ -674,7 +963,7 @@ ext4_xattr_block_find(struct inode *inod
 514                 bs->s.end = bs->bh->b_data + bs->bh->b_size;
 515                 bs->s.here = bs->s.first;
 516                 error = ext4_xattr_find_entry(&bs->s.here, i->name_index,
 517 -                                             i->name, bs->bh->b_size, 1);
 518 +                                            i->name, bs->bh->b_size, 1, inode);
 519                 if (error && error != -ENODATA)
 520                         goto cleanup;
 521                 bs->s.not_found = error;
 522 @@ -698,8 +987,6 @@ ext4_xattr_block_set(handle_t *handle, s
 523
 524  #define header(x) ((struct ext4_xattr_header *)(x))
 525
 526 -       if (i->value && i->value_len > sb->s_blocksize)
 527 -               return -ENOSPC;
 528         if (s->base) {
 529                 ce = mb_cache_entry_get(ext4_xattr_cache, bs->bh->b_bdev,
 530                                         bs->bh->b_blocknr);
 531 @@ -714,7 +1001,7 @@ ext4_xattr_block_set(handle_t *handle, s
 532                                 ce = NULL;
 533                         }
 534                         ea_bdebug(bs->bh, "modifying in-place");
 535 -                       error = ext4_xattr_set_entry(i, s);
 536 +                       error = ext4_xattr_set_entry(i, s, handle, inode);
 537                         if (!error) {
 538                                 if (!IS_LAST_ENTRY(s->first))
 539                                         ext4_xattr_rehash(header(s->base),
 540 @@ -766,7 +1053,7 @@ ext4_xattr_block_set(handle_t *handle, s
 541                 s->end = s->base + sb->s_blocksize;
 542         }
 543
 544 -       error = ext4_xattr_set_entry(i, s);
 545 +       error = ext4_xattr_set_entry(i, s, handle, inode);
 546         if (error == -EIO)
 547                 goto bad_block;
 548         if (error)
 549 @@ -917,7 +1204,7 @@ ext4_xattr_ibody_find(struct inode *inod
 550                 /* Find the named attribute. */
 551                 error = ext4_xattr_find_entry(&is->s.here, i->name_index,
 552                                               i->name, is->s.end -
 553 -                                             (void *)is->s.base, 0);
 554 +                                             (void *)is->s.base, 0, inode);
 555                 if (error && error != -ENODATA)
 556                         return error;
 557                 is->s.not_found = error;
 558 @@ -936,7 +1223,7 @@ ext4_xattr_ibody_set(handle_t *handle, s
 559
 560         if (EXT4_I(inode)->i_extra_isize == 0)
 561                 return -ENOSPC;
 562 -       error = ext4_xattr_set_entry(i, s);
 563 +       error = ext4_xattr_set_entry(i, s, handle, inode);
 564         if (error)
 565                 return error;
 566         header = IHDR(inode, ext4_raw_inode(&is->iloc));
 567 @@ -972,7 +1259,7 @@ ext4_xattr_set_handle(handle_t *handle,
 568                 .name = name,
 569                 .value = value,
 570                 .value_len = value_len,
 571 -
 572 +               .in_inode = 0,
 573         };
 574         struct ext4_xattr_ibody_find is = {
 575                 .s = { .not_found = -ENODATA, },
 576 @@ -1041,6 +1328,15 @@ ext4_xattr_set_handle(handle_t *handle,
 577                                         goto cleanup;
 578                         }
 579                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
 580 +                       if (EXT4_HAS_INCOMPAT_FEATURE(inode->i_sb,
 581 +                                       EXT4_FEATURE_INCOMPAT_EA_INODE) &&
 582 +                           error == -ENOSPC) {
 583 +                               /* xattr not fit to block, store at external
 584 +                                * inode */
 585 +                               i.in_inode = 1;
 586 +                               error = ext4_xattr_ibody_set(handle, inode,
 587 +                                                            &i, &is);
 588 +                       }
 589                         if (error)
 590                                 goto cleanup;
 591                         if (!is.s.not_found) {
 592 @@ -1088,10 +1384,25 @@ ext4_xattr_set(struct inode *inode, int
 593                const void *value, size_t value_len, int flags)
 594  {
 595         handle_t *handle;
 596 +       struct super_block *sb = inode->i_sb;
 597 +       int buffer_credits;
 598         int error, retries = 0;
 599
 600 +       buffer_credits = EXT4_DATA_TRANS_BLOCKS(sb);
 601 +       if ((value_len >= EXT4_XATTR_MIN_LARGE_EA_SIZE(sb->s_blocksize)) &&
 602 +           EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EA_INODE)) {
 603 +               int nrblocks = (value_len + sb->s_blocksize - 1) >>
 604 +                                       sb->s_blocksize_bits;
 605 +
 606 +               /* For new inode */
 607 +               buffer_credits += EXT4_SINGLEDATA_TRANS_BLOCKS(sb) + 3;
 608 +
 609 +               /* For data blocks of EA inode */
 610 +               buffer_credits += ext4_meta_trans_blocks(inode, nrblocks, 0);
 611 +       }
 612 +
 613  retry:
 614 -       handle = ext4_journal_start(inode, EXT4_DATA_TRANS_BLOCKS(inode->i_sb));
 615 +       handle = ext4_journal_start(inode, buffer_credits);
 616         if (IS_ERR(handle)) {
 617                 error = PTR_ERR(handle);
 618         } else {
 619 @@ -1101,7 +1412,7 @@ retry:
 620                                               value, value_len, flags);
 621                 error2 = ext4_journal_stop(handle);
 622                 if (error == -ENOSPC &&
 623 -                   ext4_should_retry_alloc(inode->i_sb, &retries))
 624 +                   ext4_should_retry_alloc(sb, &retries))
 625                         goto retry;
 626                 if (error == 0)
 627                         error = error2;
 628 @@ -1123,7 +1434,7 @@ static void ext4_xattr_shift_entries(str
 629
 630         /* Adjust the value offsets of the entries */
 631         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 632 -               if (!last->e_value_block && last->e_value_size) {
 633 +               if (last->e_value_inum == 0 && last->e_value_size > 0) {
 634                         new_offs = le16_to_cpu(last->e_value_offs) +
 635                                                         value_offs_shift;
 636                         BUG_ON(new_offs + le32_to_cpu(last->e_value_size)
 637 @@ -1356,20 +1667,92 @@ cleanup:
 638         return error;
 639  }
 640
 641 -
 642 +#define EIA_INCR 16 /* must be 2^n */
 643 +#define EIA_MASK (EIA_INCR - 1)
 644 +/* Add the large xattr @ino into @lea_ino_array for later deletion.
 645 + * If @lea_ino_array is new or full it will be grown and the old
 646 + * contents copied over.
 647 + */
 648 +static int
 649 +ext4_expand_ino_array(struct ext4_xattr_ino_array **lea_ino_array, __u32 ino)
 650 +{
 651 +       if (*lea_ino_array == NULL) {
 652 +               /*
 653 +                * Start with 15 inodes, so it fits into a power-of-two size.
 654 +                * If *lea_ino_array is NULL, this is essentially offsetof()
 655 +                */
 656 +               (*lea_ino_array) =
 657 +                       kmalloc(offsetof(struct ext4_xattr_ino_array,
 658 +                                        xia_inodes[EIA_MASK]),
 659 +                               GFP_NOFS);
 660 +               if (*lea_ino_array == NULL)
 661 +                       return -ENOMEM;
 662 +               (*lea_ino_array)->xia_count = 0;
 663 +       } else if (((*lea_ino_array)->xia_count & EIA_MASK) == EIA_MASK) {
 664 +               /* expand the array once all 15 + n * 16 slots are full */
 665 +               struct ext4_xattr_ino_array *new_array = NULL;
 666 +               int count = (*lea_ino_array)->xia_count;
 667 +
 668 +               /* if new_array is NULL, this is essentially offsetof() */
 669 +               new_array = kmalloc(
 670 +                               offsetof(struct ext4_xattr_ino_array,
 671 +                                        xia_inodes[count + EIA_INCR]),
 672 +                               GFP_NOFS);
 673 +               if (new_array == NULL)
 674 +                       return -ENOMEM;
 675 +               memcpy(new_array, *lea_ino_array,
 676 +                      offsetof(struct ext4_xattr_ino_array,
 677 +                               xia_inodes[count]));
 678 +               kfree(*lea_ino_array);
 679 +               *lea_ino_array = new_array;
 680 +       }
 681 +       (*lea_ino_array)->xia_inodes[(*lea_ino_array)->xia_count++] = ino;
 682 +       return 0;
 683 +}
 684
 685  /*
 686   * ext4_xattr_delete_inode()
 687   *
 688 - * Free extended attribute resources associated with this inode. This
 689 + * Free extended attribute resources associated with this inode. Traverse
 690 + * all entries and unlink any xattr inodes associated with this inode. This
 691   * is called immediately before an inode is freed. We have exclusive
 692 - * access to the inode.
 693 + * access to the inode. If an orphan inode is deleted it will also delete any
 694 + * xattr block and all xattr inodes. They are checked by ext4_xattr_inode_iget()
 695 + * to ensure they belong to the parent inode and were not deleted already.
 696   */
 697 -void
 698 -ext4_xattr_delete_inode(handle_t *handle, struct inode *inode)
 699 +int
 700 +ext4_xattr_delete_inode(handle_t *handle, struct inode *inode,
 701 +                       struct ext4_xattr_ino_array **lea_ino_array)
 702  {
 703         struct buffer_head *bh = NULL;
 704 +       struct ext4_xattr_ibody_header *header;
 705 +       struct ext4_inode *raw_inode;
 706 +       struct ext4_iloc iloc;
 707 +       struct ext4_xattr_entry *entry;
 708 +       int error = 0;
 709 +
 710 +       if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR))
 711 +               goto delete_external_ea;
 712 +
 713 +       error = ext4_get_inode_loc(inode, &iloc);
 714 +       if (error)
 715 +               goto cleanup;
 716 +       raw_inode = ext4_raw_inode(&iloc);
 717 +       header = IHDR(inode, raw_inode);
 718 +       entry = IFIRST(header);
 719 +       for (; !IS_LAST_ENTRY(entry); entry = EXT4_XATTR_NEXT(entry)) {
 720 +               if (entry->e_value_inum == 0)
 721 +                       continue;
 722 +               if (ext4_expand_ino_array(lea_ino_array,
 723 +                                         entry->e_value_inum) != 0) {
 724 +                       brelse(iloc.bh);
 725 +                       goto cleanup;
 726 +               }
 727 +               entry->e_value_inum = 0;
 728 +       }
 729 +       brelse(iloc.bh);
 730
 731 +delete_external_ea:
 732         if (!EXT4_I(inode)->i_file_acl)
 733                 goto cleanup;
 734         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
 735 @@ -1384,11 +1764,74 @@ ext4_xattr_delete_inode(handle_t *handle
 736                            inode->i_ino, EXT4_I(inode)->i_file_acl);
 737                 goto cleanup;
 738         }
 739 +
 740 +       entry = BFIRST(bh);
 741 +       for (; !IS_LAST_ENTRY(entry); entry = EXT4_XATTR_NEXT(entry)) {
 742 +               if (entry->e_value_inum == 0)
 743 +                       continue;
 744 +               if (ext4_expand_ino_array(lea_ino_array,
 745 +                                         entry->e_value_inum) != 0)
 746 +                       goto cleanup;
 747 +               entry->e_value_inum = 0;
 748 +       }
 749 +
 750 +       /* add xattr inode to orphan list */
 751 +       if (*lea_ino_array != NULL) {
 752 +               struct inode *ea_inode = NULL;
 753 +               int idx = 0;
 754 +
 755 +               for (; idx < (*lea_ino_array)->xia_count; ++idx) {
 756 +                       if (!ext4_handle_has_enough_credits(handle, 3)) {
 757 +                               error = ext4_journal_extend(handle, 3);
 758 +                               if (error > 0)
 759 +                                       error = ext4_journal_restart(handle, 3);
 760 +                               if (error != 0) {
 761 +                                       ext4_warning(inode->i_sb,
 762 +                                               "couldn't extend journal "
 763 +                                               "(err %d)", error);
 764 +                                       goto cleanup;
 765 +                               }
 766 +                       }
 767 +                       ea_inode = ext4_xattr_inode_iget(inode,
 768 +                               (*lea_ino_array)->xia_inodes[idx], &error);
 769 +                       if (error)
 770 +                               continue;
 771 +                       ext4_orphan_add(handle, ea_inode);
 772 +                       /* the inode's i_count will be released by caller */
 773 +               }
 774 +       }
 775 +
 776         ext4_xattr_release_block(handle, inode, bh);
 777         EXT4_I(inode)->i_file_acl = 0;
 778
 779  cleanup:
 780         brelse(bh);
 781 +
 782 +       return error;
 783 +}
 784 +
 785 +void
 786 +ext4_xattr_inode_array_free(struct inode *inode,
 787 +                           struct ext4_xattr_ino_array *lea_ino_array)
 788 +{
 789 +       struct inode    *ea_inode = NULL;
 790 +       int             idx = 0;
 791 +       int             err;
 792 +
 793 +       if (lea_ino_array == NULL)
 794 +               return;
 795 +
 796 +       for (; idx < lea_ino_array->xia_count; ++idx) {
 797 +               ea_inode = ext4_xattr_inode_iget(inode,
 798 +                               lea_ino_array->xia_inodes[idx], &err);
 799 +               if (err)
 800 +                       continue;
 801 +               ea_inode->i_nlink = 0;
 802 +               iput(ea_inode);
 803 +               /* for inode's i_count get from ext4_xattr_delete_inode */
 804 +               iput(ea_inode);
 805 +       }
 806 +       kfree(lea_ino_array);
 807  }
 808
 809  /*
 810 @@ -1458,10 +1901,9 @@ ext4_xattr_cmp(struct ext4_xattr_header
 811                     entry1->e_name_index != entry2->e_name_index ||
 812                     entry1->e_name_len != entry2->e_name_len ||
 813                     entry1->e_value_size != entry2->e_value_size ||
 814 +                   entry1->e_value_inum != entry2->e_value_inum ||
 815                     memcmp(entry1->e_name, entry2->e_name, entry1->e_name_len))
 816                         return 1;
 817 -               if (entry1->e_value_block != 0 || entry2->e_value_block != 0)
 818 -                       return -EIO;
 819                 if (memcmp((char *)header1 + le16_to_cpu(entry1->e_value_offs),
 820                            (char *)header2 + le16_to_cpu(entry2->e_value_offs),
 821                            le32_to_cpu(entry1->e_value_size)))
 822 @@ -1546,7 +1988,7 @@ static inline void ext4_xattr_hash_entry
 823                        *name++;
 824         }
 825
 826 -       if (entry->e_value_block == 0 && entry->e_value_size != 0) {
 827 +       if (entry->e_value_inum == 0 && entry->e_value_size != 0) {
 828                 __le32 *value = (__le32 *)((char *)header +
 829                         le16_to_cpu(entry->e_value_offs));
 830                 for (n = (le32_to_cpu(entry->e_value_size) +
 831 Index: linux-stage/fs/ext4/xattr.h
 832 ===================================================================
 833 --- linux-stage.orig/fs/ext4/xattr.h
 834 +++ linux-stage/fs/ext4/xattr.h
 835 @@ -38,7 +38,7 @@ struct ext4_xattr_entry {
 836         __u8    e_name_len;     /* length of name */
 837         __u8    e_name_index;   /* attribute name index */
 838         __le16  e_value_offs;   /* offset in disk block of value */
 839 -       __le32  e_value_block;  /* disk block attribute is stored on (n/i) */
 840 +       __le32  e_value_inum;   /* inode in which the value is stored */
 841         __le32  e_value_size;   /* size of attribute value */
 842         __le32  e_hash;         /* hash value of name and value */
 843         char    e_name[0];      /* attribute name */
 844 @@ -63,6 +63,15 @@ struct ext4_xattr_entry {
 845                 EXT4_I(inode)->i_extra_isize))
 846  #define IFIRST(hdr) ((struct ext4_xattr_entry *)((hdr)+1))
 847
 848 +#define i_xattr_inode_parent i_mtime.tv_sec
 849 +
 850 +/*
 851 + * The minimum size of EA value when you start storing it in an external inode
 852 + * size of block - size of header - size of 1 entry - 4 null bytes
 853 +*/
 854 +#define EXT4_XATTR_MIN_LARGE_EA_SIZE(b)                                        \
 855 +       ((b) - EXT4_XATTR_LEN(3) - sizeof(struct ext4_xattr_header) - 4)
 856 +
 857  # ifdef CONFIG_EXT4_FS_XATTR
 858
 859  extern struct xattr_handler ext4_xattr_user_handler;
 860 @@ -77,7 +86,13 @@ extern int ext4_xattr_get(struct inode *
 861  extern int ext4_xattr_set(struct inode *, int, const char *, const void *, size_t, int);
 862  extern int ext4_xattr_set_handle(handle_t *, struct inode *, int, const char *, const void *, size_t, int);
 863
 864 -extern void ext4_xattr_delete_inode(handle_t *, struct inode *);
 865 +extern struct inode *ext4_xattr_inode_iget(struct inode *parent, int ea_ino,
 866 +                                          int *err);
 867 +extern int ext4_xattr_inode_unlink(struct inode *inode, int ea_ino);
 868 +extern int ext4_xattr_delete_inode(handle_t *handle, struct inode *inode,
 869 +                                  struct ext4_xattr_ino_array **array);
 870 +extern void ext4_xattr_inode_array_free(struct inode *inode,
 871 +                                       struct ext4_xattr_ino_array *array);
 872  extern void ext4_xattr_put_super(struct super_block *);
 873
 874  extern int ext4_expand_extra_isize_ea(struct inode *inode, int new_extra_isize,
 875 @@ -111,9 +126,11 @@ ext4_xattr_set_handle(handle_t *handle,
 876         return -EOPNOTSUPP;
 877  }
 878
 879 -static inline void
 880 -ext4_xattr_delete_inode(handle_t *handle, struct inode *inode)
 881 +inline int
 882 +ext4_xattr_delete_inode(handle_t *handle, struct inode *inode,
 883 +                       struct ext4_xattr_ino_array **array)
 884  {
 885 +       return -EOPNOTSUPP;
 886  }
 887
 888  static inline void
 889 Index: linux-stage/fs/ext4/inode.c
 890 ===================================================================
 891 --- linux-stage.orig/fs/ext4/inode.c
 892 +++ linux-stage/fs/ext4/inode.c
 893 @@ -223,6 +223,7 @@ void ext4_delete_inode(struct inode *ino
 894         handle_t *handle;
 895         int err;
 896         int extra_credits = 3;
 897 +       struct ext4_xattr_ino_array *lea_ino_array = NULL;
 898
 899         if (ext4_should_order_data(inode))
 900                 ext4_begin_ordered_truncate(inode, 0);
 901 @@ -238,8 +239,8 @@ void ext4_delete_inode(struct inode *ino
 902          * protection against it
 903          */
 904         sb_start_intwrite(inode->i_sb);
 905 -       handle = ext4_journal_start(inode,
 906 -                       blocks_for_truncate(inode) + extra_credits);
 907 +
 908 +       handle = ext4_journal_start(inode, extra_credits);
 909         if (IS_ERR(handle)) {
 910                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
 911                 /*
 912 @@ -251,9 +252,33 @@ void ext4_delete_inode(struct inode *ino
 913                 sb_end_intwrite(inode->i_sb);
 914                 goto no_delete;
 915         }
 916 -
 917         if (IS_SYNC(inode))
 918                 ext4_handle_sync(handle);
 919 +
 920 +       /*
 921 +        * Delete xattr inode before deleting the main inode.
 922 +        */
 923 +       err = ext4_xattr_delete_inode(handle, inode, &lea_ino_array);
 924 +       if (err) {
 925 +               ext4_warning(inode->i_sb,
 926 +                            "couldn't delete inode's xattr (err %d)", err);
 927 +               goto stop_handle;
 928 +       }
 929 +
 930 +       if (!ext4_handle_has_enough_credits(handle,
 931 +                               blocks_for_truncate(inode) + extra_credits)) {
 932 +               err = ext4_journal_extend(handle,
 933 +                               blocks_for_truncate(inode) + extra_credits);
 934 +               if (err > 0)
 935 +                       err = ext4_journal_restart(handle,
 936 +                               blocks_for_truncate(inode) + extra_credits);
 937 +               if (err != 0) {
 938 +                       ext4_warning(inode->i_sb,
 939 +                                    "couldn't extend journal (err %d)", err);
 940 +                       goto stop_handle;
 941 +               }
 942 +       }
 943 +
 944         inode->i_size = 0;
 945         err = ext4_mark_inode_dirty(handle, inode);
 946         if (err) {
 947 @@ -307,8 +332,12 @@ void ext4_delete_inode(struct inode *ino
 948                 clear_inode(inode);
 949         else
 950                 ext4_free_inode(handle, inode);
 951 +
 952         ext4_journal_stop(handle);
 953         sb_end_intwrite(inode->i_sb);
 954 +
 955 +       if (lea_ino_array != NULL)
 956 +               ext4_xattr_inode_array_free(inode, lea_ino_array);
 957         return;
 958  no_delete:
 959         clear_inode(inode);     /* We must guarantee clearing of inode... */
 960 Index: linux-stage/fs/ext4/ialloc.c
 961 ===================================================================
 962 --- linux-stage.orig/fs/ext4/ialloc.c
 963 +++ linux-stage/fs/ext4/ialloc.c
 964 @@ -219,7 +219,6 @@ void ext4_free_inode(handle_t *handle, s
 965          * as writing the quota to disk may need the lock as well.
 966          */
 967         vfs_dq_init(inode);
 968 -       ext4_xattr_delete_inode(handle, inode);
 969         vfs_dq_free_inode(inode);
 970         vfs_dq_drop(inode);
 971