ldiskfs/kernel_patches/patches/sles11sp2/ext4-large-eas.patch

   1 This patch implements the large EA support in ext4. If the size of
   2 an EA value is larger than the blocksize, then the EA value would
   3 not be saved in the external EA block, instead it would be saved
   4 in an external EA inode. So, the patch also helps support a larger
   5 number of EAs.
   6
   7 Index: linux-stage/fs/ext4/ext4.h
   8 ===================================================================
   9 --- linux-stage.orig/fs/ext4/ext4.h
  10 +++ linux-stage/fs/ext4/ext4.h
  11 @@ -1408,6 +1408,7 @@ static inline void ext4_clear_state_flag
  12                                          EXT4_FEATURE_INCOMPAT_EXTENTS| \
  13                                          EXT4_FEATURE_INCOMPAT_64BIT| \
  14                                          EXT4_FEATURE_INCOMPAT_FLEX_BG| \
  15 +                                        EXT4_FEATURE_INCOMPAT_EA_INODE| \
  16                                          EXT4_FEATURE_INCOMPAT_MMP| \
  17                                          EXT4_FEATURE_INCOMPAT_DIRDATA)
  18
  19 @@ -1775,6 +1782,10 @@ struct mmpd_data {
  20  # define ATTRIB_NORET  __attribute__((noreturn))
  21  # define NORET_AND     noreturn,
  22
  23 +struct ext4_xattr_ino_array {
  24 +       unsigned int xia_count;         /* # of used item in the array */
  25 +       unsigned int xia_inodes[0];
  26 +};
  27  /* bitmap.c */
  28  extern unsigned int ext4_count_free(char *bitmap, unsigned numchars);
  29
  30 @@ -1893,6 +1904,7 @@ extern void ext4_set_inode_flags(struct
  31  extern void ext4_get_inode_flags(struct ext4_inode_info *);
  32  extern int ext4_alloc_da_blocks(struct inode *inode);
  33  extern void ext4_set_aops(struct inode *inode);
  34 +extern int ext4_meta_trans_blocks(struct inode *, int nrblocks, int chunk);
  35  extern int ext4_writepage_trans_blocks(struct inode *);
  36  extern int ext4_chunk_trans_blocks(struct inode *, int nrblocks);
  37  extern int ext4_block_truncate_page(handle_t *handle,
  38 Index: linux-stage/fs/ext4/inode.c
  39 ===================================================================
  40 --- linux-stage.orig/fs/ext4/inode.c
  41 +++ linux-stage/fs/ext4/inode.c
  42 @@ -187,6 +187,8 @@ void ext4_evict_inode(struct inode *inod
  43  {
  44         handle_t *handle;
  45         int err;
  46 +       int extra_credits = 3;
  47 +       struct ext4_xattr_ino_array *lea_ino_array = NULL;
  48
  49         trace_ext4_evict_inode(inode);
  50
  51 @@ -207,7 +209,10 @@ void ext4_evict_inode(struct inode *inod
  52         if (is_bad_inode(inode))
  53                 goto no_delete;
  54
  55 -       handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
  56 +       /*
  57 +        * Delete xattr inode before deleting the main inode.
  58 +        */
  59 +       handle = ext4_journal_start(inode, extra_credits);
  60         if (IS_ERR(handle)) {
  61                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
  62                 /*
  63 @@ -218,9 +223,33 @@ void ext4_evict_inode(struct inode *inod
  64                 ext4_orphan_del(NULL, inode);
  65                 goto no_delete;
  66         }
  67 -
  68         if (IS_SYNC(inode))
  69                 ext4_handle_sync(handle);
  70 +
  71 +       err = ext4_xattr_delete_inode(handle, inode, &lea_ino_array);
  72 +       if (err) {
  73 +               ext4_warning(inode->i_sb,
  74 +                            "couldn't delete inode's xattr (err %d)", err);
  75 +               goto stop_handle;
  76 +       }
  77 +
  78 +       if (!IS_NOQUOTA(inode))
  79 +               extra_credits += 2 * EXT4_QUOTA_DEL_BLOCKS(inode->i_sb);
  80 +
  81 +       if (!ext4_handle_has_enough_credits(handle,
  82 +                               blocks_for_truncate(inode) + extra_credits)) {
  83 +               err = ext4_journal_extend(handle,
  84 +                               blocks_for_truncate(inode) + extra_credits);
  85 +               if (err > 0)
  86 +                       err = ext4_journal_restart(handle,
  87 +                               blocks_for_truncate(inode) + extra_credits);
  88 +               if (err != 0) {
  89 +                       ext4_warning(inode->i_sb,
  90 +                                    "couldn't extend journal (err %d)", err);
  91 +                       goto stop_handle;
  92 +               }
  93 +       }
  94 +
  95         inode->i_size = 0;
  96         err = ext4_mark_inode_dirty(handle, inode);
  97         if (err) {
  98 @@ -237,10 +266,10 @@ void ext4_evict_inode(struct inode *inod
  99          * enough credits left in the handle to remove the inode from
 100          * the orphan list and set the dtime field.
 101          */
 102 -       if (!ext4_handle_has_enough_credits(handle, 3)) {
 103 -               err = ext4_journal_extend(handle, 3);
 104 +       if (!ext4_handle_has_enough_credits(handle, extra_credits)) {
 105 +               err = ext4_journal_extend(handle, extra_credits);
 106                 if (err > 0)
 107 -                       err = ext4_journal_restart(handle, 3);
 108 +                       err = ext4_journal_restart(handle, extra_credits);
 109                 if (err != 0) {
 110                         ext4_warning(inode->i_sb,
 111                                      "couldn't extend journal (err %d)", err);
 112 @@ -274,7 +303,11 @@ void ext4_evict_inode(struct inode *inod
 113                 ext4_clear_inode(inode);
 114         else
 115                 ext4_free_inode(handle, inode);
 116 +
 117         ext4_journal_stop(handle);
 118 +
 119 +       if (lea_ino_array != NULL)
 120 +               ext4_xattr_inode_array_free(inode, lea_ino_array);
 121         return;
 122  no_delete:
 123         ext4_clear_inode(inode);        /* We must guarantee clearing of inode... */
 124 @@ -5552,7 +5585,7 @@ static int ext4_index_trans_blocks(struc
 125   *
 126   * Also account for superblock, inode, quota and xattr blocks
 127   */
 128 -static int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
 129 +int ext4_meta_trans_blocks(struct inode *inode, int nrblocks, int chunk)
 130  {
 131         ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
 132         int gdpblocks;
 133 Index: linux-stage/fs/ext4/xattr.c
 134 ===================================================================
 135 --- linux-stage.orig/fs/ext4/xattr.c
 136 +++ linux-stage/fs/ext4/xattr.c
 137 @@ -168,19 +168,26 @@ ext4_xattr_check_block(struct buffer_hea
 138  }
 139
 140  static inline int
 141 -ext4_xattr_check_entry(struct ext4_xattr_entry *entry, size_t size)
 142 +ext4_xattr_check_entry(struct ext4_xattr_entry *entry, size_t size,
 143 +                      struct inode *inode)
 144  {
 145         size_t value_size = le32_to_cpu(entry->e_value_size);
 146
 147 -       if (entry->e_value_block != 0 || value_size > size ||
 148 +       if (!entry->e_value_inum &&
 149             le16_to_cpu(entry->e_value_offs) + value_size > size)
 150 +               return -EIO;
 151 +       if (entry->e_value_inum &&
 152 +           (le32_to_cpu(entry->e_value_inum) < EXT4_FIRST_INO(inode->i_sb) ||
 153 +            le32_to_cpu(entry->e_value_inum) >
 154 +            le32_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_inodes_count)))
 155                 return -EIO;
 156         return 0;
 157  }
 158
 159  static int
 160  ext4_xattr_find_entry(struct ext4_xattr_entry **pentry, int name_index,
 161 -                     const char *name, size_t size, int sorted)
 162 +                     const char *name, size_t size, int sorted,
 163 +                     struct inode *inode)
 164  {
 165         struct ext4_xattr_entry *entry;
 166         size_t name_len;
 167 @@ -200,11 +207,104 @@ ext4_xattr_find_entry(struct ext4_xattr_
 168                         break;
 169         }
 170         *pentry = entry;
 171 -       if (!cmp && ext4_xattr_check_entry(entry, size))
 172 +       if (!cmp && ext4_xattr_check_entry(entry, size, inode))
 173                         return -EIO;
 174         return cmp ? -ENODATA : 0;
 175  }
 176
 177 +/*
 178 + * Read the EA value from an inode.
 179 + */
 180 +static int
 181 +ext4_xattr_inode_read(struct inode *ea_inode, void *buf, size_t *size)
 182 +{
 183 +       unsigned long block = 0;
 184 +       struct buffer_head *bh = NULL;
 185 +       int err, blocksize;
 186 +       size_t csize, ret_size = 0;
 187 +
 188 +       if (*size == 0)
 189 +               return 0;
 190 +
 191 +       blocksize = ea_inode->i_sb->s_blocksize;
 192 +
 193 +       while (ret_size < *size) {
 194 +               csize = (*size - ret_size) > blocksize ? blocksize :
 195 +                                                       *size - ret_size;
 196 +               bh = ext4_bread(NULL, ea_inode, block, 0, &err);
 197 +               if (!bh) {
 198 +                       *size = ret_size;
 199 +                       return err;
 200 +               }
 201 +               memcpy(buf, bh->b_data, csize);
 202 +               brelse(bh);
 203 +
 204 +               buf += csize;
 205 +               block += 1;
 206 +               ret_size += csize;
 207 +       }
 208 +
 209 +       *size = ret_size;
 210 +
 211 +       return err;
 212 +}
 213 +
 214 +struct inode *ext4_xattr_inode_iget(struct inode *parent, unsigned long ea_ino, int *err)
 215 +{
 216 +       struct inode *ea_inode = NULL;
 217 +
 218 +       ea_inode = ext4_iget(parent->i_sb, ea_ino);
 219 +       if (IS_ERR(ea_inode) || is_bad_inode(ea_inode)) {
 220 +               int rc = IS_ERR(ea_inode) ? PTR_ERR(ea_inode) : 0;
 221 +               ext4_error(parent->i_sb, "error while reading EA inode %lu "
 222 +                          "/ %d %d", ea_ino, rc, is_bad_inode(ea_inode));
 223 +               *err = rc != 0 ? rc : -EIO;
 224 +               return NULL;
 225 +       }
 226 +
 227 +       if (EXT4_XATTR_INODE_GET_PARENT(ea_inode) != parent->i_ino ||
 228 +           ea_inode->i_generation != parent->i_generation) {
 229 +               ext4_error(parent->i_sb, "Backpointer from EA inode %lu "
 230 +                          "to parent invalid.", ea_ino);
 231 +               *err = -EINVAL;
 232 +               goto error;
 233 +       }
 234 +
 235 +       if (!(EXT4_I(ea_inode)->i_flags & EXT4_EA_INODE_FL)) {
 236 +               ext4_error(parent->i_sb, "EA inode %lu does not have "
 237 +                          "EXT4_EA_INODE_FL flag set.\n", ea_ino);
 238 +               *err = -EINVAL;
 239 +               goto error;
 240 +       }
 241 +
 242 +       *err = 0;
 243 +       return ea_inode;
 244 +
 245 +error:
 246 +       iput(ea_inode);
 247 +       return NULL;
 248 +}
 249 +
 250 +/*
 251 + * Read the value from the EA inode.
 252 + */
 253 +static int
 254 +ext4_xattr_inode_get(struct inode *inode, unsigned long ea_ino, void *buffer,
 255 +                    size_t *size)
 256 +{
 257 +       struct inode *ea_inode = NULL;
 258 +       int err;
 259 +
 260 +       ea_inode = ext4_xattr_inode_iget(inode, ea_ino, &err);
 261 +       if (err)
 262 +               return err;
 263 +
 264 +       err = ext4_xattr_inode_read(ea_inode, buffer, size);
 265 +       iput(ea_inode);
 266 +
 267 +       return err;
 268 +}
 269 +
 270  static int
 271  ext4_xattr_block_get(struct inode *inode, int name_index, const char *name,
 272                      void *buffer, size_t buffer_size)
 273 @@ -235,7 +334,8 @@ bad_block:
 274         }
 275         ext4_xattr_cache_insert(bh);
 276         entry = BFIRST(bh);
 277 -       error = ext4_xattr_find_entry(&entry, name_index, name, bh->b_size, 1);
 278 +       error = ext4_xattr_find_entry(&entry, name_index, name, bh->b_size, 1,
 279 +                                     inode);
 280         if (error == -EIO)
 281                 goto bad_block;
 282         if (error)
 283 @@ -245,8 +345,16 @@ bad_block:
 284                 error = -ERANGE;
 285                 if (size > buffer_size)
 286                         goto cleanup;
 287 -               memcpy(buffer, bh->b_data + le16_to_cpu(entry->e_value_offs),
 288 -                      size);
 289 +               if (entry->e_value_inum) {
 290 +                       error = ext4_xattr_inode_get(inode,
 291 +                                            le32_to_cpu(entry->e_value_inum),
 292 +                                            buffer, &size);
 293 +                       if (error)
 294 +                               goto cleanup;
 295 +               } else {
 296 +                       memcpy(buffer, bh->b_data +
 297 +                              le16_to_cpu(entry->e_value_offs), size);
 298 +               }
 299         }
 300         error = size;
 301
 302 @@ -280,7 +388,7 @@ ext4_xattr_ibody_get(struct inode *inode
 303         if (error)
 304                 goto cleanup;
 305         error = ext4_xattr_find_entry(&entry, name_index, name,
 306 -                                     end - (void *)entry, 0);
 307 +                                     end - (void *)entry, 0, inode);
 308         if (error)
 309                 goto cleanup;
 310         size = le32_to_cpu(entry->e_value_size);
 311 @@ -288,8 +396,16 @@ ext4_xattr_ibody_get(struct inode *inode
 312                 error = -ERANGE;
 313                 if (size > buffer_size)
 314                         goto cleanup;
 315 -               memcpy(buffer, (void *)IFIRST(header) +
 316 -                      le16_to_cpu(entry->e_value_offs), size);
 317 +               if (entry->e_value_inum) {
 318 +                       error = ext4_xattr_inode_get(inode,
 319 +                                            le32_to_cpu(entry->e_value_inum),
 320 +                                            buffer, &size);
 321 +                       if (error)
 322 +                               goto cleanup;
 323 +               } else {
 324 +                       memcpy(buffer, (void *)IFIRST(header) +
 325 +                              le16_to_cpu(entry->e_value_offs), size);
 326 +               }
 327         }
 328         error = size;
 329
 330 @@ -514,7 +630,7 @@ static size_t ext4_xattr_free_space(stru
 331  {
 332         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 333                 *total += EXT4_XATTR_LEN(last->e_name_len);
 334 -               if (!last->e_value_block && last->e_value_size) {
 335 +               if (!last->e_value_inum && last->e_value_size) {
 336                         size_t offs = le16_to_cpu(last->e_value_offs);
 337                         if (offs < *min_offs)
 338                                 *min_offs = offs;
 339 @@ -523,11 +639,162 @@ static size_t ext4_xattr_free_space(stru
 340         return (*min_offs - ((void *)last - base) - sizeof(__u32));
 341  }
 342
 343 +/*
 344 + * Write the value of the EA in an inode.
 345 + */
 346 +static int
 347 +ext4_xattr_inode_write(handle_t *handle, struct inode *ea_inode,
 348 +                      const void *buf, int bufsize)
 349 +{
 350 +       struct buffer_head *bh = NULL;
 351 +       struct ext4_map_blocks map;
 352 +       unsigned long block = 0;
 353 +       unsigned blocksize = ea_inode->i_sb->s_blocksize;
 354 +       unsigned max_blocks = (bufsize + blocksize - 1) >> ea_inode->i_blkbits;
 355 +       int csize, wsize = 0;
 356 +       int ret = 0;
 357 +       int retries = 0;
 358 +
 359 +retry:
 360 +       while (ret >= 0 && ret < max_blocks) {
 361 +               block += ret;
 362 +               max_blocks -= ret;
 363 +
 364 +               map.m_lblk = block;
 365 +               map.m_len = max_blocks;
 366 +               ret = ext4_map_blocks(handle, ea_inode, &map,
 367 +                                     EXT4_GET_BLOCKS_CREATE);
 368 +               if (ret <= 0) {
 369 +                       ext4_mark_inode_dirty(handle, ea_inode);
 370 +                       if (ret == -ENOSPC &&
 371 +                           ext4_should_retry_alloc(ea_inode->i_sb, &retries)) {
 372 +                               ret = 0;
 373 +                               goto retry;
 374 +                       }
 375 +                       break;
 376 +               }
 377 +       }
 378 +
 379 +       if (ret < 0)
 380 +               return ret;
 381 +
 382 +       block = 0;
 383 +       while (wsize < bufsize) {
 384 +               if (bh != NULL)
 385 +                       brelse(bh);
 386 +               csize = (bufsize - wsize) > blocksize ? blocksize :
 387 +                                                               bufsize - wsize;
 388 +               bh = ext4_getblk(handle, ea_inode, block, 0, &ret);
 389 +               if (!bh)
 390 +                       goto out;
 391 +               ret = ext4_journal_get_write_access(handle, bh);
 392 +               if (ret)
 393 +                       goto out;
 394 +
 395 +               memcpy(bh->b_data, buf, csize);
 396 +               set_buffer_uptodate(bh);
 397 +               ext4_handle_dirty_metadata(handle, ea_inode, bh);
 398 +
 399 +               buf += csize;
 400 +               wsize += csize;
 401 +               block += 1;
 402 +       }
 403 +
 404 +       i_size_write(ea_inode, wsize);
 405 +       ext4_update_i_disksize(ea_inode, wsize);
 406 +
 407 +       ext4_mark_inode_dirty(handle, ea_inode);
 408 +
 409 +out:
 410 +       brelse(bh);
 411 +
 412 +       return ret;
 413 +}
 414 +
 415 +/*
 416 + * Create an inode to store the value of a large EA.
 417 + */
 418 +static struct inode *
 419 +ext4_xattr_inode_create(handle_t *handle, struct inode *inode)
 420 +{
 421 +       struct inode *ea_inode = NULL;
 422 +
 423 +       /*
 424 +        * Let the next inode be the goal, so we try and allocate the EA inode
 425 +        * in the same group, or nearby one.
 426 +        */
 427 +       ea_inode = ext4_new_inode(handle, inode->i_sb->s_root->d_inode,
 428 +                                 S_IFREG|0600, NULL, inode->i_ino + 1);
 429 +
 430 +       if (!IS_ERR(ea_inode)) {
 431 +               ea_inode->i_op = &ext4_file_inode_operations;
 432 +               ea_inode->i_fop = &ext4_file_operations;
 433 +               ext4_set_aops(ea_inode);
 434 +               ea_inode->i_generation = inode->i_generation;
 435 +               EXT4_I(ea_inode)->i_flags |= EXT4_EA_INODE_FL;
 436 +
 437 +               /*
 438 +                * A back-pointer from EA inode to parent inode will be useful
 439 +                * for e2fsck.
 440 +                */
 441 +               EXT4_XATTR_INODE_SET_PARENT(ea_inode, inode->i_ino);
 442 +               unlock_new_inode(ea_inode);
 443 +       }
 444 +
 445 +       return ea_inode;
 446 +}
 447 +
 448 +/*
 449 + * Unlink the inode storing the value of the EA.
 450 + */
 451 +int
 452 +ext4_xattr_inode_unlink(struct inode *inode, unsigned long ea_ino)
 453 +{
 454 +       struct inode *ea_inode = NULL;
 455 +       int err;
 456 +
 457 +       ea_inode = ext4_xattr_inode_iget(inode, ea_ino, &err);
 458 +       if (err)
 459 +               return err;
 460 +
 461 +       ea_inode->i_nlink = 0;
 462 +       iput(ea_inode);
 463 +
 464 +       return 0;
 465 +}
 466 +
 467 +/*
 468 + * Add value of the EA in an inode.
 469 + */
 470 +static int
 471 +ext4_xattr_inode_set(handle_t *handle, struct inode *inode, unsigned long *ea_ino,
 472 +                    const void *value, size_t value_len)
 473 +{
 474 +       struct inode *ea_inode = NULL;
 475 +       int err;
 476 +
 477 +       /* Create an inode for the EA value */
 478 +       ea_inode = ext4_xattr_inode_create(handle, inode);
 479 +       if (IS_ERR(ea_inode))
 480 +               return -1;
 481 +
 482 +       err = ext4_xattr_inode_write(handle, ea_inode, value, value_len);
 483 +       if (err)
 484 +               ea_inode->i_nlink = 0;
 485 +       else
 486 +               *ea_ino = ea_inode->i_ino;
 487 +
 488 +       iput(ea_inode);
 489 +
 490 +       return err;
 491 +}
 492 +
 493  struct ext4_xattr_info {
 494 -       int name_index;
 495         const char *name;
 496         const void *value;
 497         size_t value_len;
 498 +       int name_index;
 499 +       int in_inode;
 500  };
 501
 502  struct ext4_xattr_search {
 503 @@ -539,15 +806,23 @@ struct ext4_xattr_search {
 504  };
 505
 506  static int
 507 -ext4_xattr_set_entry(struct ext4_xattr_info *i, struct ext4_xattr_search *s)
 508 +ext4_xattr_set_entry(struct ext4_xattr_info *i, struct ext4_xattr_search *s,
 509 +                    handle_t *handle, struct inode *inode)
 510  {
 511         struct ext4_xattr_entry *last;
 512         size_t free, min_offs = s->end - s->base, name_len = strlen(i->name);
 513 +       int in_inode = i->in_inode;
 514 +
 515 +       if (EXT4_HAS_INCOMPAT_FEATURE(inode->i_sb,
 516 +                EXT4_FEATURE_INCOMPAT_EA_INODE) &&
 517 +           (EXT4_XATTR_SIZE(i->value_len) >
 518 +            EXT4_XATTR_MIN_LARGE_EA_SIZE(inode->i_sb->s_blocksize)))
 519 +               in_inode = 1;
 520
 521         /* Compute min_offs and last. */
 522         last = s->first;
 523         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 524 -               if (!last->e_value_block && last->e_value_size) {
 525 +               if (!last->e_value_inum && last->e_value_size) {
 526                         size_t offs = le16_to_cpu(last->e_value_offs);
 527                         if (offs < min_offs)
 528                                 min_offs = offs;
 529 @@ -555,16 +830,21 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 530         }
 531         free = min_offs - ((void *)last - s->base) - sizeof(__u32);
 532         if (!s->not_found) {
 533 -               if (!s->here->e_value_block && s->here->e_value_size) {
 534 +               if (!in_inode &&
 535 +                   !s->here->e_value_inum && s->here->e_value_size) {
 536                         size_t size = le32_to_cpu(s->here->e_value_size);
 537                         free += EXT4_XATTR_SIZE(size);
 538                 }
 539                 free += EXT4_XATTR_LEN(name_len);
 540         }
 541         if (i->value) {
 542 -               if (free < EXT4_XATTR_SIZE(i->value_len) ||
 543 -                   free < EXT4_XATTR_LEN(name_len) +
 544 -                          EXT4_XATTR_SIZE(i->value_len))
 545 +               size_t value_len = EXT4_XATTR_SIZE(i->value_len);
 546 +
 547 +               if (in_inode)
 548 +                       value_len = 0;
 549 +
 550 +               if (free < value_len ||
 551 +                   free < EXT4_XATTR_LEN(name_len) + value_len)
 552                         return -ENOSPC;
 553         }
 554
 555 @@ -578,7 +858,8 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 556                 s->here->e_name_len = name_len;
 557                 memcpy(s->here->e_name, i->name, name_len);
 558         } else {
 559 -               if (!s->here->e_value_block && s->here->e_value_size) {
 560 +               if (!s->here->e_value_inum && s->here->e_value_size &&
 561 +                   s->here->e_value_offs > 0) {
 562                         void *first_val = s->base + min_offs;
 563                         size_t offs = le16_to_cpu(s->here->e_value_offs);
 564                         void *val = s->base + offs;
 565 @@ -607,13 +888,18 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 566                         last = s->first;
 567                         while (!IS_LAST_ENTRY(last)) {
 568                                 size_t o = le16_to_cpu(last->e_value_offs);
 569 -                               if (!last->e_value_block &&
 570 +                               if (!last->e_value_inum &&
 571                                     last->e_value_size && o < offs)
 572                                         last->e_value_offs =
 573                                                 cpu_to_le16(o + size);
 574                                 last = EXT4_XATTR_NEXT(last);
 575                         }
 576                 }
 577 +               if (s->here->e_value_inum) {
 578 +                       ext4_xattr_inode_unlink(inode,
 579 +                                       le32_to_cpu(s->here->e_value_inum));
 580 +                       s->here->e_value_inum = 0;
 581 +               }
 582                 if (!i->value) {
 583                         /* Remove the old name. */
 584                         size_t size = EXT4_XATTR_LEN(name_len);
 585 @@ -627,10 +912,17 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 586         if (i->value) {
 587                 /* Insert the new value. */
 588                 s->here->e_value_size = cpu_to_le32(i->value_len);
 589 -               if (i->value_len) {
 590 +               if (in_inode) {
 591 +                       unsigned long ea_ino = le32_to_cpu(s->here->e_value_inum);
 592 +                       ext4_xattr_inode_set(handle, inode, &ea_ino, i->value,
 593 +                                            i->value_len);
 594 +                       s->here->e_value_inum = cpu_to_le32(ea_ino);
 595 +                       s->here->e_value_offs = 0;
 596 +               } else if (i->value_len) {
 597                         size_t size = EXT4_XATTR_SIZE(i->value_len);
 598                         void *val = s->base + min_offs - size;
 599                         s->here->e_value_offs = cpu_to_le16(min_offs - size);
 600 +                       s->here->e_value_inum = 0;
 601                         memset(val + size - EXT4_XATTR_PAD, 0,
 602                                EXT4_XATTR_PAD); /* Clear the pad bytes. */
 603                         memcpy(val, i->value, i->value_len);
 604 @@ -675,7 +967,7 @@ ext4_xattr_block_find(struct inode *inod
 605                 bs->s.end = bs->bh->b_data + bs->bh->b_size;
 606                 bs->s.here = bs->s.first;
 607                 error = ext4_xattr_find_entry(&bs->s.here, i->name_index,
 608 -                                             i->name, bs->bh->b_size, 1);
 609 +                                            i->name, bs->bh->b_size, 1, inode);
 610                 if (error && error != -ENODATA)
 611                         goto cleanup;
 612                 bs->s.not_found = error;
 613 @@ -699,8 +991,6 @@ ext4_xattr_block_set(handle_t *handle, s
 614
 615  #define header(x) ((struct ext4_xattr_header *)(x))
 616
 617 -       if (i->value && i->value_len > sb->s_blocksize)
 618 -               return -ENOSPC;
 619         if (s->base) {
 620                 ce = mb_cache_entry_get(ext4_xattr_cache, bs->bh->b_bdev,
 621                                         bs->bh->b_blocknr);
 622 @@ -715,7 +1005,7 @@ ext4_xattr_block_set(handle_t *handle, s
 623                                 ce = NULL;
 624                         }
 625                         ea_bdebug(bs->bh, "modifying in-place");
 626 -                       error = ext4_xattr_set_entry(i, s);
 627 +                       error = ext4_xattr_set_entry(i, s, handle, inode);
 628                         if (!error) {
 629                                 if (!IS_LAST_ENTRY(s->first))
 630                                         ext4_xattr_rehash(header(s->base),
 631 @@ -767,7 +1057,7 @@ ext4_xattr_block_set(handle_t *handle, s
 632                 s->end = s->base + sb->s_blocksize;
 633         }
 634
 635 -       error = ext4_xattr_set_entry(i, s);
 636 +       error = ext4_xattr_set_entry(i, s, handle, inode);
 637         if (error == -EIO)
 638                 goto bad_block;
 639         if (error)
 640 @@ -918,7 +1208,7 @@ ext4_xattr_ibody_find(struct inode *inod
 641                 /* Find the named attribute. */
 642                 error = ext4_xattr_find_entry(&is->s.here, i->name_index,
 643                                               i->name, is->s.end -
 644 -                                             (void *)is->s.base, 0);
 645 +                                             (void *)is->s.base, 0, inode);
 646                 if (error && error != -ENODATA)
 647                         return error;
 648                 is->s.not_found = error;
 649 @@ -937,7 +1227,7 @@ ext4_xattr_ibody_set(handle_t *handle, s
 650
 651         if (EXT4_I(inode)->i_extra_isize == 0)
 652                 return -ENOSPC;
 653 -       error = ext4_xattr_set_entry(i, s);
 654 +       error = ext4_xattr_set_entry(i, s, handle, inode);
 655         if (error)
 656                 return error;
 657         header = IHDR(inode, ext4_raw_inode(&is->iloc));
 658 @@ -973,7 +1263,7 @@ ext4_xattr_set_handle(handle_t *handle,
 659                 .name = name,
 660                 .value = value,
 661                 .value_len = value_len,
 662 -
 663 +               .in_inode = 0,
 664         };
 665         struct ext4_xattr_ibody_find is = {
 666                 .s = { .not_found = -ENODATA, },
 667 @@ -1042,6 +1332,15 @@ ext4_xattr_set_handle(handle_t *handle,
 668                                         goto cleanup;
 669                         }
 670                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
 671 +                       if (EXT4_HAS_INCOMPAT_FEATURE(inode->i_sb,
 672 +                                       EXT4_FEATURE_INCOMPAT_EA_INODE) &&
 673 +                           error == -ENOSPC) {
 674 +                               /* xattr not fit to block, store at external
 675 +                                * inode */
 676 +                               i.in_inode = 1;
 677 +                               error = ext4_xattr_ibody_set(handle, inode,
 678 +                                                            &i, &is);
 679 +                       }
 680                         if (error)
 681                                 goto cleanup;
 682                         if (!is.s.not_found) {
 683 @@ -1088,10 +1387,25 @@ ext4_xattr_set(struct inode *inode, int
 684                const void *value, size_t value_len, int flags)
 685  {
 686         handle_t *handle;
 687 +       struct super_block *sb = inode->i_sb;
 688 +       int buffer_credits;
 689         int error, retries = 0;
 690
 691 +       buffer_credits = EXT4_DATA_TRANS_BLOCKS(sb);
 692 +       if ((value_len >= EXT4_XATTR_MIN_LARGE_EA_SIZE(sb->s_blocksize)) &&
 693 +           EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EA_INODE)) {
 694 +               int nrblocks = (value_len + sb->s_blocksize - 1) >>
 695 +                                       sb->s_blocksize_bits;
 696 +
 697 +               /* For new inode */
 698 +               buffer_credits += EXT4_SINGLEDATA_TRANS_BLOCKS(sb) + 3;
 699 +
 700 +               /* For data blocks of EA inode */
 701 +               buffer_credits += ext4_meta_trans_blocks(inode, nrblocks, 0);
 702 +       }
 703 +
 704  retry:
 705 -       handle = ext4_journal_start(inode, EXT4_DATA_TRANS_BLOCKS(inode->i_sb));
 706 +       handle = ext4_journal_start(inode, buffer_credits);
 707         if (IS_ERR(handle)) {
 708                 error = PTR_ERR(handle);
 709         } else {
 710 @@ -1101,7 +1415,7 @@ retry:
 711                                               value, value_len, flags);
 712                 error2 = ext4_journal_stop(handle);
 713                 if (error == -ENOSPC &&
 714 -                   ext4_should_retry_alloc(inode->i_sb, &retries))
 715 +                   ext4_should_retry_alloc(sb, &retries))
 716                         goto retry;
 717                 if (error == 0)
 718                         error = error2;
 719 @@ -1123,7 +1437,7 @@ static void ext4_xattr_shift_entries(str
 720
 721         /* Adjust the value offsets of the entries */
 722         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 723 -               if (!last->e_value_block && last->e_value_size) {
 724 +               if (!last->e_value_inum && last->e_value_size) {
 725                         new_offs = le16_to_cpu(last->e_value_offs) +
 726                                                         value_offs_shift;
 727                         BUG_ON(new_offs + le32_to_cpu(last->e_value_size)
 728 @@ -1358,22 +1672,135 @@ cleanup:
 729         return error;
 730  }
 731
 732 +#define EIA_INCR 16 /* must be 2^n */
 733 +#define EIA_MASK (EIA_INCR - 1)
 734 +/* Add the large xattr @ino into @lea_ino_array for later deletion.
 735 + * If @lea_ino_array is new or full it will be grown and the old
 736 + * contents copied over.
 737 + */
 738 +static int
 739 +ext4_expand_ino_array(struct ext4_xattr_ino_array **lea_ino_array, __u32 ino)
 740 +{
 741 +       if (*lea_ino_array == NULL) {
 742 +               /*
 743 +                * Start with 15 inodes, so it fits into a powr-of-two size.
 744 +                * If *lea_ino_array is NULL, this is essentially offsetof()
 745 +                */
 746 +               (*lea_ino_array) =
 747 +                       kmalloc(offsetof(struct ext4_xattr_ino_array,
 748 +                                        xia_inodes[EIA_MASK]),
 749 +                               GFP_NOFS);
 750 +               if (*lea_ino_array == NULL)
 751 +                       return -ENOMEM;
 752 +               (*lea_ino_array)->xia_count = 0;
 753 +       } else if (((*lea_ino_array)->xia_count & EIA_MASK) == EIA_MASK) {
 754 +               /* expand the array once all 15 + n * 16 slots are full */
 755 +               struct ext4_xattr_ino_array *new_array = NULL;
 756 +               int count = (*lea_ino_array)->xia_count;
 757 +
 758 +               /* if new_array is NULL, this is essentially offsetof() */
 759 +               new_array = kmalloc(
 760 +                               offsetof(struct ext4_xattr_ino_array,
 761 +                                        xia_inodes[count + EIA_INCR]),
 762 +                               GFP_NOFS);
 763 +               if (new_array == NULL)
 764 +                       return -ENOMEM;
 765 +               memcpy(new_array, *lea_ino_array,
 766 +                      offsetof(struct ext4_xattr_ino_array,
 767 +                               xia_inodes[count]));
 768 +               kfree(*lea_ino_array);
 769 +               *lea_ino_array = new_array;
 770 +       }
 771 +       (*lea_ino_array)->xia_inodes[(*lea_ino_array)->xia_count++] = ino;
 772 +       return 0;
 773 +}
 774 +
 775 +/**
 776 + * Add xattr inode to orphan list
 777 + */
 778 +static int
 779 +ext4_xattr_inode_orphan_add(handle_t *handle, struct inode *inode,
 780 +                       int credits, struct ext4_xattr_ino_array *lea_ino_array)
 781 +{
 782 +       struct inode *ea_inode = NULL;
 783 +       int idx = 0, error = 0;
 784 +
 785 +       if (lea_ino_array == NULL)
 786 +               return 0;
 787
 788 +       for (; idx < lea_ino_array->xia_count; ++idx) {
 789 +               if (!ext4_handle_has_enough_credits(handle, credits)) {
 790 +                       error = ext4_journal_extend(handle, credits);
 791 +                       if (error > 0)
 792 +                               error = ext4_journal_restart(handle, credits);
 793 +
 794 +                       if (error != 0) {
 795 +                               ext4_warning(inode->i_sb,
 796 +                                       "couldn't extend journal "
 797 +                                       "(err %d)", error);
 798 +                               return error;
 799 +                       }
 800 +               }
 801 +               ea_inode = ext4_xattr_inode_iget(inode,
 802 +                               lea_ino_array->xia_inodes[idx], &error);
 803 +               if (error)
 804 +                       continue;
 805 +               ext4_orphan_add(handle, ea_inode);
 806 +               /* the inode's i_count will be released by caller */
 807 +       }
 808 +
 809 +       return 0;
 810 +}
 811
 812  /*
 813   * ext4_xattr_delete_inode()
 814   *
 815 - * Free extended attribute resources associated with this inode. This
 816 + * Free extended attribute resources associated with this inode. Traverse
 817 + * all entries and unlink any xattr inodes associated with this inode. This
 818   * is called immediately before an inode is freed. We have exclusive
 819 - * access to the inode.
 820 + * access to the inode. If an orphan inode is deleted it will also delete any
 821 + * xattr block and all xattr inodes. They are checked by ext4_xattr_inode_iget()
 822 + * to ensure they belong to the parent inode and were not deleted already.
 823   */
 824 -void
 825 -ext4_xattr_delete_inode(handle_t *handle, struct inode *inode)
 826 +int
 827 +ext4_xattr_delete_inode(handle_t *handle, struct inode *inode,
 828 +                       struct ext4_xattr_ino_array **lea_ino_array)
 829  {
 830         struct buffer_head *bh = NULL;
 831 +       struct ext4_xattr_ibody_header *header;
 832 +       struct ext4_inode *raw_inode;
 833 +       struct ext4_iloc iloc;
 834 +       struct ext4_xattr_entry *entry;
 835 +       int credits = 3, error = 0;
 836
 837 -       if (!EXT4_I(inode)->i_file_acl)
 838 +       if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR))
 839 +               goto delete_external_ea;
 840 +
 841 +       error = ext4_get_inode_loc(inode, &iloc);
 842 +       if (error)
 843                 goto cleanup;
 844 +       raw_inode = ext4_raw_inode(&iloc);
 845 +       header = IHDR(inode, raw_inode);
 846 +       for (entry = IFIRST(header); !IS_LAST_ENTRY(entry);
 847 +            entry = EXT4_XATTR_NEXT(entry)) {
 848 +               if (!entry->e_value_inum)
 849 +                       continue;
 850 +               if (ext4_expand_ino_array(lea_ino_array,
 851 +                                         entry->e_value_inum) != 0) {
 852 +                       brelse(iloc.bh);
 853 +                       goto cleanup;
 854 +               }
 855 +               entry->e_value_inum = 0;
 856 +       }
 857 +       brelse(iloc.bh);
 858 +
 859 +delete_external_ea:
 860 +       if (!EXT4_I(inode)->i_file_acl) {
 861 +               /* add xattr inode to orphan list */
 862 +               ext4_xattr_inode_orphan_add(handle, inode, credits,
 863 +                                               *lea_ino_array);
 864 +               goto cleanup;
 865 +       }
 866         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
 867         if (!bh) {
 868                 EXT4_ERROR_INODE(inode, "block %llu read error",
 869 @@ -1386,11 +1813,69 @@ ext4_xattr_delete_inode(handle_t *handle
 870                                  EXT4_I(inode)->i_file_acl);
 871                 goto cleanup;
 872         }
 873 +
 874 +       for (entry = BFIRST(bh); !IS_LAST_ENTRY(entry);
 875 +            entry = EXT4_XATTR_NEXT(entry)) {
 876 +               if (!entry->e_value_inum)
 877 +                       continue;
 878 +               if (ext4_expand_ino_array(lea_ino_array,
 879 +                                         entry->e_value_inum) != 0)
 880 +                       goto cleanup;
 881 +               entry->e_value_inum = 0;
 882 +       }
 883 +
 884 +       /* adding xattr inode to orphan list */
 885 +       error = ext4_xattr_inode_orphan_add(handle, inode, credits,
 886 +                                       *lea_ino_array);
 887 +       if (error != 0)
 888 +               goto cleanup;
 889 +
 890 +       if (!IS_NOQUOTA(inode))
 891 +               credits += 2 * EXT4_QUOTA_DEL_BLOCKS(inode->i_sb);
 892 +
 893 +       if (!ext4_handle_has_enough_credits(handle, credits)) {
 894 +               error = ext4_journal_extend(handle, credits);
 895 +               if (error > 0)
 896 +                       error = ext4_journal_restart(handle, credits);
 897 +               if (error != 0) {
 898 +                       ext4_warning(inode->i_sb,
 899 +                               "couldn't extend journal (err %d)", error);
 900 +                       goto cleanup;
 901 +               }
 902 +       }
 903 +
 904         ext4_xattr_release_block(handle, inode, bh);
 905         EXT4_I(inode)->i_file_acl = 0;
 906
 907  cleanup:
 908         brelse(bh);
 909 +
 910 +       return error;
 911 +}
 912 +
 913 +void
 914 +ext4_xattr_inode_array_free(struct inode *inode,
 915 +                           struct ext4_xattr_ino_array *lea_ino_array)
 916 +{
 917 +       struct inode    *ea_inode = NULL;
 918 +       int             idx = 0;
 919 +       int             err;
 920 +
 921 +       if (lea_ino_array == NULL)
 922 +               return;
 923 +
 924 +       for (; idx < lea_ino_array->xia_count; ++idx) {
 925 +               ea_inode = ext4_xattr_inode_iget(inode,
 926 +                               lea_ino_array->xia_inodes[idx], &err);
 927 +               if (err)
 928 +                       continue;
 929 +               /* for inode's i_count get from ext4_xattr_delete_inode */
 930 +               if (!list_empty(&EXT4_I(ea_inode)->i_orphan))
 931 +                       iput(ea_inode);
 932 +               ea_inode->i_nlink = 0;
 933 +               iput(ea_inode);
 934 +       }
 935 +       kfree(lea_ino_array);
 936  }
 937
 938  /*
 939 @@ -1460,10 +1945,9 @@ ext4_xattr_cmp(struct ext4_xattr_header
 940                     entry1->e_name_index != entry2->e_name_index ||
 941                     entry1->e_name_len != entry2->e_name_len ||
 942                     entry1->e_value_size != entry2->e_value_size ||
 943 +                   entry1->e_value_inum != entry2->e_value_inum ||
 944                     memcmp(entry1->e_name, entry2->e_name, entry1->e_name_len))
 945                         return 1;
 946 -               if (entry1->e_value_block != 0 || entry2->e_value_block != 0)
 947 -                       return -EIO;
 948                 if (memcmp((char *)header1 + le16_to_cpu(entry1->e_value_offs),
 949                            (char *)header2 + le16_to_cpu(entry2->e_value_offs),
 950                            le32_to_cpu(entry1->e_value_size)))
 951 @@ -1547,7 +2031,7 @@ static inline void ext4_xattr_hash_entry
 952                        *name++;
 953         }
 954
 955 -       if (entry->e_value_block == 0 && entry->e_value_size != 0) {
 956 +       if (!entry->e_value_inum && entry->e_value_size) {
 957                 __le32 *value = (__le32 *)((char *)header +
 958                         le16_to_cpu(entry->e_value_offs));
 959                 for (n = (le32_to_cpu(entry->e_value_size) +
 960 Index: linux-stage/fs/ext4/xattr.h
 961 ===================================================================
 962 --- linux-stage.orig/fs/ext4/xattr.h
 963 +++ linux-stage/fs/ext4/xattr.h
 964 @@ -38,7 +38,7 @@ struct ext4_xattr_entry {
 965         __u8    e_name_len;     /* length of name */
 966         __u8    e_name_index;   /* attribute name index */
 967         __le16  e_value_offs;   /* offset in disk block of value */
 968 -       __le32  e_value_block;  /* disk block attribute is stored on (n/i) */
 969 +       __le32  e_value_inum;   /* inode in which the value is stored */
 970         __le32  e_value_size;   /* size of attribute value */
 971         __le32  e_hash;         /* hash value of name and value */
 972         char    e_name[0];      /* attribute name */
 973 @@ -63,6 +63,26 @@ struct ext4_xattr_entry {
 974                 EXT4_I(inode)->i_extra_isize))
 975  #define IFIRST(hdr) ((struct ext4_xattr_entry *)((hdr)+1))
 976
 977 +/*
 978 + * Link EA inode back to parent one using i_mtime field.
 979 + * Extra integer type conversion added to ignore higher
 980 + * bits in i_mtime.tv_sec which might be set by ext4_get()
 981 + */
 982 +#define EXT4_XATTR_INODE_SET_PARENT(inode, inum)      \
 983 +do {                                                  \
 984 +      (inode)->i_mtime.tv_sec = inum;                 \
 985 +} while(0)
 986 +
 987 +#define EXT4_XATTR_INODE_GET_PARENT(inode)            \
 988 +((__u32)(inode)->i_mtime.tv_sec)
 989 +
 990 +/*
 991 + * The minimum size of EA value when you start storing it in an external inode
 992 + * size of block - size of header - size of 1 entry - 4 null bytes
 993 +*/
 994 +#define EXT4_XATTR_MIN_LARGE_EA_SIZE(b)                                        \
 995 +       ((b) - EXT4_XATTR_LEN(3) - sizeof(struct ext4_xattr_header) - 4)
 996 +
 997  # ifdef CONFIG_EXT4_FS_XATTR
 998
 999  extern const struct xattr_handler ext4_xattr_user_handler;
1000 @@ -77,7 +86,13 @@ extern int ext4_xattr_get(struct inode *
1001  extern int ext4_xattr_set(struct inode *, int, const char *, const void *, size_t, int);
1002  extern int ext4_xattr_set_handle(handle_t *, struct inode *, int, const char *, const void *, size_t, int);
1003
1004 -extern void ext4_xattr_delete_inode(handle_t *, struct inode *);
1005 +extern struct inode *ext4_xattr_inode_iget(struct inode *parent, unsigned long ea_ino,
1006 +                                          int *err);
1007 +extern int ext4_xattr_inode_unlink(struct inode *inode, unsigned long ea_ino);
1008 +extern int ext4_xattr_delete_inode(handle_t *handle, struct inode *inode,
1009 +                                  struct ext4_xattr_ino_array **array);
1010 +extern void ext4_xattr_inode_array_free(struct inode *inode,
1011 +                                       struct ext4_xattr_ino_array *array);
1012  extern void ext4_xattr_put_super(struct super_block *);
1013
1014  extern int ext4_expand_extra_isize_ea(struct inode *inode, int new_extra_isize,
1015 @@ -111,9 +126,11 @@ ext4_xattr_set_handle(handle_t *handle,
1016         return -EOPNOTSUPP;
1017  }
1018
1019 -static inline void
1020 -ext4_xattr_delete_inode(handle_t *handle, struct inode *inode)
1021 +inline int
1022 +ext4_xattr_delete_inode(handle_t *handle, struct inode *inode,
1023 +                       struct ext4_xattr_ino_array *array)
1024  {
1025 +       return -EOPNOTSUPP;
1026  }
1027
1028  static inline void
1029 Index: linux-stage/fs/ext4/ialloc.c
1030 ===================================================================
1031 --- linux-stage.orig/fs/ext4/ialloc.c
1032 +++ linux-stage/fs/ext4/ialloc.c
1033 @@ -221,7 +221,6 @@ void ext4_free_inode(handle_t *handle, s
1034          * as writing the quota to disk may need the lock as well.
1035          */
1036         dquot_initialize(inode);
1037 -       ext4_xattr_delete_inode(handle, inode);
1038         dquot_free_inode(inode);
1039         dquot_drop(inode);
1040