ldiskfs/kernel_patches/patches/rhel7/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 @@ -1579,6 +1579,7 @@ 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                                          EXT4_FEATURE_INCOMPAT_INLINE_DATA)
  19 @@ -1979,6 +1980,12 @@ struct mmpd_data {
  20  #define EXT4_MMP_MAX_CHECK_INTERVAL    300UL
  21
  22  /*
  23 + * Maximum size of xattr attributes for FEATURE_INCOMPAT_EA_INODE 1Mb
  24 + * This limit is arbitrary, but is reasonable for the xattr API.
  25 + */
  26 +#define EXT4_XATTR_MAX_LARGE_EA_SIZE    (1024 * 1024)
  27 +
  28 +/*
  29   * Function prototypes
  30   */
  31
  32 @@ -1990,6 +1997,10 @@ struct mmpd_data {
  33  # define ATTRIB_NORET  __attribute__((noreturn))
  34  # define NORET_AND     noreturn,
  35
  36 +struct ext4_xattr_ino_array {
  37 +       unsigned int xia_count;         /* # of used item in the array */
  38 +       unsigned int xia_inodes[0];
  39 +};
  40  /* bitmap.c */
  41  extern unsigned int ext4_count_free(char *bitmap, unsigned numchars);
  42  void ext4_inode_bitmap_csum_set(struct super_block *sb, ext4_group_t group,
  43 @@ -2194,6 +2205,7 @@ extern void ext4_set_inode_flags(struct
  44  extern void ext4_get_inode_flags(struct ext4_inode_info *);
  45  extern int ext4_alloc_da_blocks(struct inode *inode);
  46  extern void ext4_set_aops(struct inode *inode);
  47 +extern int ext4_meta_trans_blocks(struct inode *, int nrblocks, int chunk);
  48  extern int ext4_writepage_trans_blocks(struct inode *);
  49  extern int ext4_chunk_trans_blocks(struct inode *, int nrblocks);
  50  extern int ext4_zero_partial_blocks(handle_t *handle, struct inode *inode,
  51 Index: linux-stage/fs/ext4/inode.c
  52 ===================================================================
  53 --- linux-stage.orig/fs/ext4/inode.c
  54 +++ linux-stage/fs/ext4/inode.c
  55 @@ -134,8 +134,6 @@ static void ext4_invalidatepage(struct p
  56                                 unsigned int length);
  57  static int __ext4_journalled_writepage(struct page *page, unsigned int len);
  58  static int ext4_bh_delay_or_unwritten(handle_t *handle, struct buffer_head *bh);
  59 -static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
  60 -                                 int pextents);
  61
  62  /*
  63   * Test whether an inode is a fast symlink.
  64 @@ -184,6 +182,8 @@ void ext4_evict_inode(struct inode *inod
  65  {
  66         handle_t *handle;
  67         int err;
  68 +       int extra_credits = 3;
  69 +       struct ext4_xattr_ino_array *lea_ino_array = NULL;
  70
  71         trace_ext4_evict_inode(inode);
  72
  73 @@ -236,8 +236,8 @@ void ext4_evict_inode(struct inode *inod
  74          * protection against it
  75          */
  76         sb_start_intwrite(inode->i_sb);
  77 -       handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE,
  78 -                                   ext4_blocks_for_truncate(inode)+3);
  79 +
  80 +       handle = ext4_journal_start(inode, EXT4_HT_TRUNCATE, extra_credits);
  81         if (IS_ERR(handle)) {
  82                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
  83                 /*
  84 @@ -249,9 +249,36 @@ void ext4_evict_inode(struct inode *inod
  85                 sb_end_intwrite(inode->i_sb);
  86                 goto no_delete;
  87         }
  88 -
  89         if (IS_SYNC(inode))
  90                 ext4_handle_sync(handle);
  91 +
  92 +       /*
  93 +        * Delete xattr inode before deleting the main inode.
  94 +        */
  95 +       err = ext4_xattr_delete_inode(handle, inode, &lea_ino_array);
  96 +       if (err) {
  97 +               ext4_warning(inode->i_sb,
  98 +                            "couldn't delete inode's xattr (err %d)", err);
  99 +               goto stop_handle;
 100 +       }
 101 +
 102 +       if (!IS_NOQUOTA(inode))
 103 +               extra_credits += 2 * EXT4_QUOTA_DEL_BLOCKS(inode->i_sb);
 104 +
 105 +       if (!ext4_handle_has_enough_credits(handle,
 106 +                       ext4_blocks_for_truncate(inode) + extra_credits)) {
 107 +               err = ext4_journal_extend(handle,
 108 +                       ext4_blocks_for_truncate(inode) + extra_credits);
 109 +               if (err > 0)
 110 +                       err = ext4_journal_restart(handle,
 111 +                       ext4_blocks_for_truncate(inode) + extra_credits);
 112 +               if (err != 0) {
 113 +                       ext4_warning(inode->i_sb,
 114 +                                    "couldn't extend journal (err %d)", err);
 115 +                       goto stop_handle;
 116 +               }
 117 +       }
 118 +
 119         inode->i_size = 0;
 120         err = ext4_mark_inode_dirty(handle, inode);
 121         if (err) {
 122 @@ -306,8 +333,12 @@ void ext4_evict_inode(struct inode *inod
 123                 ext4_clear_inode(inode);
 124         else
 125                 ext4_free_inode(handle, inode);
 126 +
 127         ext4_journal_stop(handle);
 128         sb_end_intwrite(inode->i_sb);
 129 +
 130 +       if (lea_ino_array != NULL)
 131 +               ext4_xattr_inode_array_free(inode, lea_ino_array);
 132         return;
 133  no_delete:
 134         ext4_clear_inode(inode);        /* We must guarantee clearing of inode... */
 135 @@ -4681,7 +4712,7 @@ static int ext4_index_trans_blocks(struc
 136   *
 137   * Also account for superblock, inode, quota and xattr blocks
 138   */
 139 -static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
 140 +int ext4_meta_trans_blocks(struct inode *inode, int lblocks,
 141                                   int pextents)
 142  {
 143         ext4_group_t groups, ngroups = ext4_get_groups_count(inode->i_sb);
 144 Index: linux-stage/fs/ext4/xattr.c
 145 ===================================================================
 146 --- linux-stage.orig/fs/ext4/xattr.c
 147 +++ linux-stage/fs/ext4/xattr.c
 148 @@ -201,6 +201,7 @@ ext4_xattr_check_names(struct ext4_xattr
 149
 150         while (!IS_LAST_ENTRY(entry)) {
 151                 if (entry->e_value_size != 0 &&
 152 +                   entry->e_value_inum == 0 &&
 153                     (value_start + le16_to_cpu(entry->e_value_offs) <
 154                      (void *)e + sizeof(__u32) ||
 155                      value_start + le16_to_cpu(entry->e_value_offs) +
 156 @@ -233,19 +234,26 @@ ext4_xattr_check_block(struct inode *ino
 157  }
 158
 159  static inline int
 160 -ext4_xattr_check_entry(struct ext4_xattr_entry *entry, size_t size)
 161 +ext4_xattr_check_entry(struct ext4_xattr_entry *entry, size_t size,
 162 +                      struct inode *inode)
 163  {
 164         size_t value_size = le32_to_cpu(entry->e_value_size);
 165
 166 -       if (entry->e_value_block != 0 || value_size > size ||
 167 +       if (!entry->e_value_inum &&
 168             le16_to_cpu(entry->e_value_offs) + value_size > size)
 169 +               return -EIO;
 170 +       if (entry->e_value_inum &&
 171 +           (le32_to_cpu(entry->e_value_inum) < EXT4_FIRST_INO(inode->i_sb) ||
 172 +            le32_to_cpu(entry->e_value_inum) >
 173 +            le32_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_inodes_count)))
 174                 return -EIO;
 175         return 0;
 176  }
 177
 178  static int
 179  ext4_xattr_find_entry(struct ext4_xattr_entry **pentry, int name_index,
 180 -                     const char *name, size_t size, int sorted)
 181 +                     const char *name, size_t size, int sorted,
 182 +                     struct inode *inode)
 183  {
 184         struct ext4_xattr_entry *entry;
 185         size_t name_len;
 186 @@ -265,11 +273,104 @@ ext4_xattr_find_entry(struct ext4_xattr_
 187                         break;
 188         }
 189         *pentry = entry;
 190 -       if (!cmp && ext4_xattr_check_entry(entry, size))
 191 +       if (!cmp && ext4_xattr_check_entry(entry, size, inode))
 192                         return -EIO;
 193         return cmp ? -ENODATA : 0;
 194  }
 195
 196 +/*
 197 + * Read the EA value from an inode.
 198 + */
 199 +static int
 200 +ext4_xattr_inode_read(struct inode *ea_inode, void *buf, size_t *size)
 201 +{
 202 +       unsigned long block = 0;
 203 +       struct buffer_head *bh = NULL;
 204 +       int err, blocksize;
 205 +       size_t csize, ret_size = 0;
 206 +
 207 +       if (*size == 0)
 208 +               return 0;
 209 +
 210 +       blocksize = ea_inode->i_sb->s_blocksize;
 211 +
 212 +       while (ret_size < *size) {
 213 +               csize = (*size - ret_size) > blocksize ? blocksize :
 214 +                                                       *size - ret_size;
 215 +               bh = ext4_bread(NULL, ea_inode, block, 0, &err);
 216 +               if (!bh) {
 217 +                       *size = ret_size;
 218 +                       return err;
 219 +               }
 220 +               memcpy(buf, bh->b_data, csize);
 221 +               brelse(bh);
 222 +
 223 +               buf += csize;
 224 +               block += 1;
 225 +               ret_size += csize;
 226 +       }
 227 +
 228 +       *size = ret_size;
 229 +
 230 +       return err;
 231 +}
 232 +
 233 +struct inode *ext4_xattr_inode_iget(struct inode *parent, unsigned long ea_ino, int *err)
 234 +{
 235 +       struct inode *ea_inode = NULL;
 236 +
 237 +       ea_inode = ext4_iget(parent->i_sb, ea_ino);
 238 +       if (IS_ERR(ea_inode) || is_bad_inode(ea_inode)) {
 239 +               int rc = IS_ERR(ea_inode) ? PTR_ERR(ea_inode) : 0;
 240 +               ext4_error(parent->i_sb, "error while reading EA inode %lu "
 241 +                          "/ %d %d", ea_ino, rc, is_bad_inode(ea_inode));
 242 +               *err = rc != 0 ? rc : -EIO;
 243 +               return NULL;
 244 +       }
 245 +
 246 +       if (ea_inode->i_xattr_inode_parent != parent->i_ino ||
 247 +           ea_inode->i_generation != parent->i_generation) {
 248 +               ext4_error(parent->i_sb, "Backpointer from EA inode %lu "
 249 +                          "to parent invalid.", ea_ino);
 250 +               *err = -EINVAL;
 251 +               goto error;
 252 +       }
 253 +
 254 +       if (!(EXT4_I(ea_inode)->i_flags & EXT4_EA_INODE_FL)) {
 255 +               ext4_error(parent->i_sb, "EA inode %lu does not have "
 256 +                          "EXT4_EA_INODE_FL flag set.\n", ea_ino);
 257 +               *err = -EINVAL;
 258 +               goto error;
 259 +       }
 260 +
 261 +       *err = 0;
 262 +       return ea_inode;
 263 +
 264 +error:
 265 +       iput(ea_inode);
 266 +       return NULL;
 267 +}
 268 +
 269 +/*
 270 + * Read the value from the EA inode.
 271 + */
 272 +static int
 273 +ext4_xattr_inode_get(struct inode *inode, unsigned long ea_ino, void *buffer,
 274 +                    size_t *size)
 275 +{
 276 +       struct inode *ea_inode = NULL;
 277 +       int err;
 278 +
 279 +       ea_inode = ext4_xattr_inode_iget(inode, ea_ino, &err);
 280 +       if (err)
 281 +               return err;
 282 +
 283 +       err = ext4_xattr_inode_read(ea_inode, buffer, size);
 284 +       iput(ea_inode);
 285 +
 286 +       return err;
 287 +}
 288 +
 289  static int
 290  ext4_xattr_block_get(struct inode *inode, int name_index, const char *name,
 291                      void *buffer, size_t buffer_size)
 292 @@ -301,7 +401,8 @@ bad_block:
 293         }
 294         ext4_xattr_cache_insert(bh);
 295         entry = BFIRST(bh);
 296 -       error = ext4_xattr_find_entry(&entry, name_index, name, bh->b_size, 1);
 297 +       error = ext4_xattr_find_entry(&entry, name_index, name, bh->b_size, 1,
 298 +                                     inode);
 299         if (error == -EIO)
 300                 goto bad_block;
 301         if (error)
 302 @@ -311,8 +412,16 @@ bad_block:
 303                 error = -ERANGE;
 304                 if (size > buffer_size)
 305                         goto cleanup;
 306 -               memcpy(buffer, bh->b_data + le16_to_cpu(entry->e_value_offs),
 307 -                      size);
 308 +               if (entry->e_value_inum) {
 309 +                       error = ext4_xattr_inode_get(inode,
 310 +                                            le32_to_cpu(entry->e_value_inum),
 311 +                                            buffer, &size);
 312 +                       if (error)
 313 +                               goto cleanup;
 314 +               } else {
 315 +                       memcpy(buffer, bh->b_data +
 316 +                              le16_to_cpu(entry->e_value_offs), size);
 317 +               }
 318         }
 319         error = size;
 320
 321 @@ -346,7 +455,7 @@ ext4_xattr_ibody_get(struct inode *inode
 322         if (error)
 323                 goto cleanup;
 324         error = ext4_xattr_find_entry(&entry, name_index, name,
 325 -                                     end - (void *)entry, 0);
 326 +                                     end - (void *)entry, 0, inode);
 327         if (error)
 328                 goto cleanup;
 329         size = le32_to_cpu(entry->e_value_size);
 330 @@ -354,8 +463,16 @@ ext4_xattr_ibody_get(struct inode *inode
 331                 error = -ERANGE;
 332                 if (size > buffer_size)
 333                         goto cleanup;
 334 -               memcpy(buffer, (void *)IFIRST(header) +
 335 -                      le16_to_cpu(entry->e_value_offs), size);
 336 +               if (entry->e_value_inum) {
 337 +                       error = ext4_xattr_inode_get(inode,
 338 +                                            le32_to_cpu(entry->e_value_inum),
 339 +                                            buffer, &size);
 340 +                       if (error)
 341 +                               goto cleanup;
 342 +               } else {
 343 +                       memcpy(buffer, (void *)IFIRST(header) +
 344 +                              le16_to_cpu(entry->e_value_offs), size);
 345 +               }
 346         }
 347         error = size;
 348
 349 @@ -600,7 +717,7 @@ static size_t ext4_xattr_free_space(stru
 350                                     size_t *min_offs, void *base, int *total)
 351  {
 352         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 353 -               if (!last->e_value_block && last->e_value_size) {
 354 +               if (!last->e_value_inum && last->e_value_size) {
 355                         size_t offs = le16_to_cpu(last->e_value_offs);
 356                         if (offs < *min_offs)
 357                                 *min_offs = offs;
 358 @@ -611,16 +728,174 @@ static size_t ext4_xattr_free_space(stru
 359         return (*min_offs - ((void *)last - base) - sizeof(__u32));
 360  }
 361
 362 +/*
 363 + * Write the value of the EA in an inode.
 364 + */
 365  static int
 366 -ext4_xattr_set_entry(struct ext4_xattr_info *i, struct ext4_xattr_search *s)
 367 +ext4_xattr_inode_write(handle_t *handle, struct inode *ea_inode,
 368 +                      const void *buf, int bufsize)
 369 +{
 370 +       struct buffer_head *bh = NULL;
 371 +       unsigned long block = 0;
 372 +       unsigned blocksize = ea_inode->i_sb->s_blocksize;
 373 +       unsigned max_blocks = (bufsize + blocksize - 1) >> ea_inode->i_blkbits;
 374 +       int csize, wsize = 0;
 375 +       int ret = 0;
 376 +       int retries = 0;
 377 +
 378 +retry:
 379 +       while (ret >= 0 && ret < max_blocks) {
 380 +               struct ext4_map_blocks map;
 381 +               map.m_lblk = block += ret;
 382 +               map.m_len = max_blocks -= ret;
 383 +
 384 +               ret = ext4_map_blocks(handle, ea_inode, &map,
 385 +                                     EXT4_GET_BLOCKS_CREATE);
 386 +               if (ret <= 0) {
 387 +                       ext4_mark_inode_dirty(handle, ea_inode);
 388 +                       if (ret == -ENOSPC &&
 389 +                           ext4_should_retry_alloc(ea_inode->i_sb, &retries)) {
 390 +                               ret = 0;
 391 +                               goto retry;
 392 +                       }
 393 +                       break;
 394 +               }
 395 +       }
 396 +
 397 +       if (ret < 0)
 398 +               return ret;
 399 +
 400 +       block = 0;
 401 +       while (wsize < bufsize) {
 402 +               if (bh != NULL)
 403 +                       brelse(bh);
 404 +               csize = (bufsize - wsize) > blocksize ? blocksize :
 405 +                                                               bufsize - wsize;
 406 +               bh = ext4_getblk(handle, ea_inode, block, 0, &ret);
 407 +               if (!bh)
 408 +                       goto out;
 409 +               ret = ext4_journal_get_write_access(handle, bh);
 410 +               if (ret)
 411 +                       goto out;
 412 +
 413 +               memcpy(bh->b_data, buf, csize);
 414 +               set_buffer_uptodate(bh);
 415 +               ext4_handle_dirty_metadata(handle, ea_inode, bh);
 416 +
 417 +               buf += csize;
 418 +               wsize += csize;
 419 +               block += 1;
 420 +       }
 421 +
 422 +       mutex_lock(&ea_inode->i_mutex);
 423 +       i_size_write(ea_inode, wsize);
 424 +       ext4_update_i_disksize(ea_inode, wsize);
 425 +       mutex_unlock(&ea_inode->i_mutex);
 426 +
 427 +       ext4_mark_inode_dirty(handle, ea_inode);
 428 +
 429 +out:
 430 +       brelse(bh);
 431 +
 432 +       return ret;
 433 +}
 434 +
 435 +/*
 436 + * Create an inode to store the value of a large EA.
 437 + */
 438 +static struct inode *
 439 +ext4_xattr_inode_create(handle_t *handle, struct inode *inode)
 440 +{
 441 +       struct inode *ea_inode = NULL;
 442 +
 443 +       /*
 444 +        * Let the next inode be the goal, so we try and allocate the EA inode
 445 +        * in the same group, or nearby one.
 446 +        */
 447 +       ea_inode = ext4_new_inode(handle, inode->i_sb->s_root->d_inode,
 448 +                                 S_IFREG|0600, NULL, inode->i_ino + 1, NULL);
 449 +
 450 +       if (!IS_ERR(ea_inode)) {
 451 +               ea_inode->i_op = &ext4_file_inode_operations;
 452 +               ea_inode->i_fop = &ext4_file_operations;
 453 +               ext4_set_aops(ea_inode);
 454 +               ea_inode->i_generation = inode->i_generation;
 455 +               EXT4_I(ea_inode)->i_flags |= EXT4_EA_INODE_FL;
 456 +
 457 +               /*
 458 +                * A back-pointer from EA inode to parent inode will be useful
 459 +                * for e2fsck.
 460 +                */
 461 +               ea_inode->i_xattr_inode_parent = inode->i_ino;
 462 +               unlock_new_inode(ea_inode);
 463 +       }
 464 +
 465 +       return ea_inode;
 466 +}
 467 +
 468 +/*
 469 + * Unlink the inode storing the value of the EA.
 470 + */
 471 +int
 472 +ext4_xattr_inode_unlink(struct inode *inode, unsigned long ea_ino)
 473 +{
 474 +       struct inode *ea_inode = NULL;
 475 +       int err;
 476 +
 477 +       ea_inode = ext4_xattr_inode_iget(inode, ea_ino, &err);
 478 +       if (err)
 479 +               return err;
 480 +
 481 +       clear_nlink(ea_inode);
 482 +       iput(ea_inode);
 483 +
 484 +       return 0;
 485 +}
 486 +
 487 +/*
 488 + * Add value of the EA in an inode.
 489 + */
 490 +static int
 491 +ext4_xattr_inode_set(handle_t *handle, struct inode *inode, unsigned long *ea_ino,
 492 +                    const void *value, size_t value_len)
 493 +{
 494 +       struct inode *ea_inode = NULL;
 495 +       int err;
 496 +
 497 +       /* Create an inode for the EA value */
 498 +       ea_inode = ext4_xattr_inode_create(handle, inode);
 499 +       if (IS_ERR(ea_inode))
 500 +               return -1;
 501 +
 502 +       err = ext4_xattr_inode_write(handle, ea_inode, value, value_len);
 503 +       if (err)
 504 +               clear_nlink(ea_inode);
 505 +       else
 506 +               *ea_ino = ea_inode->i_ino;
 507 +
 508 +       iput(ea_inode);
 509 +
 510 +       return err;
 511 +}
 512 +
 513 +static int
 514 +ext4_xattr_set_entry(struct ext4_xattr_info *i, struct ext4_xattr_search *s,
 515 +                    handle_t *handle, struct inode *inode)
 516  {
 517         struct ext4_xattr_entry *last;
 518         size_t free, min_offs = s->end - s->base, name_len = strlen(i->name);
 519 +       int in_inode = i->in_inode;
 520 +
 521 +       if (EXT4_HAS_INCOMPAT_FEATURE(inode->i_sb,
 522 +                EXT4_FEATURE_INCOMPAT_EA_INODE) &&
 523 +           (EXT4_XATTR_SIZE(i->value_len) >
 524 +            EXT4_XATTR_MIN_LARGE_EA_SIZE(inode->i_sb->s_blocksize)))
 525 +               in_inode = 1;
 526
 527         /* Compute min_offs and last. */
 528         last = s->first;
 529         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 530 -               if (!last->e_value_block && last->e_value_size) {
 531 +               if (!last->e_value_inum && last->e_value_size) {
 532                         size_t offs = le16_to_cpu(last->e_value_offs);
 533                         if (offs < min_offs)
 534                                 min_offs = offs;
 535 @@ -628,16 +903,21 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 536         }
 537         free = min_offs - ((void *)last - s->base) - sizeof(__u32);
 538         if (!s->not_found) {
 539 -               if (!s->here->e_value_block && s->here->e_value_size) {
 540 +               if (!in_inode &&
 541 +                   !s->here->e_value_inum && s->here->e_value_size) {
 542                         size_t size = le32_to_cpu(s->here->e_value_size);
 543                         free += EXT4_XATTR_SIZE(size);
 544                 }
 545                 free += EXT4_XATTR_LEN(name_len);
 546         }
 547         if (i->value) {
 548 -               if (free < EXT4_XATTR_SIZE(i->value_len) ||
 549 -                   free < EXT4_XATTR_LEN(name_len) +
 550 -                          EXT4_XATTR_SIZE(i->value_len))
 551 +               size_t value_len = EXT4_XATTR_SIZE(i->value_len);
 552 +
 553 +               if (in_inode)
 554 +                       value_len = 0;
 555 +
 556 +               if (free < value_len ||
 557 +                   free < EXT4_XATTR_LEN(name_len) + value_len)
 558                         return -ENOSPC;
 559         }
 560
 561 @@ -651,7 +931,8 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 562                 s->here->e_name_len = name_len;
 563                 memcpy(s->here->e_name, i->name, name_len);
 564         } else {
 565 -               if (!s->here->e_value_block && s->here->e_value_size) {
 566 +               if (!s->here->e_value_inum && s->here->e_value_size &&
 567 +                   s->here->e_value_offs > 0) {
 568                         void *first_val = s->base + min_offs;
 569                         size_t offs = le16_to_cpu(s->here->e_value_offs);
 570                         void *val = s->base + offs;
 571 @@ -685,13 +966,18 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 572                         last = s->first;
 573                         while (!IS_LAST_ENTRY(last)) {
 574                                 size_t o = le16_to_cpu(last->e_value_offs);
 575 -                               if (!last->e_value_block &&
 576 +                               if (!last->e_value_inum &&
 577                                     last->e_value_size && o < offs)
 578                                         last->e_value_offs =
 579                                                 cpu_to_le16(o + size);
 580                                 last = EXT4_XATTR_NEXT(last);
 581                         }
 582                 }
 583 +               if (s->here->e_value_inum) {
 584 +                       ext4_xattr_inode_unlink(inode,
 585 +                                       le32_to_cpu(s->here->e_value_inum));
 586 +                       s->here->e_value_inum = 0;
 587 +               }
 588                 if (!i->value) {
 589                         /* Remove the old name. */
 590                         size_t size = EXT4_XATTR_LEN(name_len);
 591 @@ -705,10 +990,17 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 592         if (i->value) {
 593                 /* Insert the new value. */
 594                 s->here->e_value_size = cpu_to_le32(i->value_len);
 595 -               if (i->value_len) {
 596 +               if (in_inode) {
 597 +                       unsigned long ea_ino = le32_to_cpu(s->here->e_value_inum);
 598 +                       ext4_xattr_inode_set(handle, inode, &ea_ino, i->value,
 599 +                                            i->value_len);
 600 +                       s->here->e_value_inum = cpu_to_le32(ea_ino);
 601 +                       s->here->e_value_offs = 0;
 602 +               } else if (i->value_len) {
 603                         size_t size = EXT4_XATTR_SIZE(i->value_len);
 604                         void *val = s->base + min_offs - size;
 605                         s->here->e_value_offs = cpu_to_le16(min_offs - size);
 606 +                       s->here->e_value_inum = 0;
 607                         if (i->value == EXT4_ZERO_XATTR_VALUE) {
 608                                 memset(val, 0, size);
 609                         } else {
 610 @@ -758,7 +1050,7 @@ ext4_xattr_block_find(struct inode *inod
 611                 bs->s.end = bs->bh->b_data + bs->bh->b_size;
 612                 bs->s.here = bs->s.first;
 613                 error = ext4_xattr_find_entry(&bs->s.here, i->name_index,
 614 -                                             i->name, bs->bh->b_size, 1);
 615 +                                            i->name, bs->bh->b_size, 1, inode);
 616                 if (error && error != -ENODATA)
 617                         goto cleanup;
 618                 bs->s.not_found = error;
 619 @@ -782,8 +1074,6 @@ ext4_xattr_block_set(handle_t *handle, s
 620
 621  #define header(x) ((struct ext4_xattr_header *)(x))
 622
 623 -       if (i->value && i->value_len > sb->s_blocksize)
 624 -               return -ENOSPC;
 625         if (s->base) {
 626                 ce = mb_cache_entry_get(ext4_xattr_cache, bs->bh->b_bdev,
 627                                         bs->bh->b_blocknr);
 628 @@ -799,7 +1089,7 @@ ext4_xattr_block_set(handle_t *handle, s
 629                                 ce = NULL;
 630                         }
 631                         ea_bdebug(bs->bh, "modifying in-place");
 632 -                       error = ext4_xattr_set_entry(i, s);
 633 +                       error = ext4_xattr_set_entry(i, s, handle, inode);
 634                         if (!error) {
 635                                 if (!IS_LAST_ENTRY(s->first))
 636                                         ext4_xattr_rehash(header(s->base),
 637 @@ -850,7 +1140,7 @@ ext4_xattr_block_set(handle_t *handle, s
 638                 s->end = s->base + sb->s_blocksize;
 639         }
 640
 641 -       error = ext4_xattr_set_entry(i, s);
 642 +       error = ext4_xattr_set_entry(i, s, handle, inode);
 643         if (error == -EIO)
 644                 goto bad_block;
 645         if (error)
 646 @@ -1000,7 +1290,7 @@ int ext4_xattr_ibody_find(struct inode *
 647                 /* Find the named attribute. */
 648                 error = ext4_xattr_find_entry(&is->s.here, i->name_index,
 649                                               i->name, is->s.end -
 650 -                                             (void *)is->s.base, 0);
 651 +                                             (void *)is->s.base, 0, inode);
 652                 if (error && error != -ENODATA)
 653                         return error;
 654                 is->s.not_found = error;
 655 @@ -1018,7 +1308,7 @@ int ext4_xattr_ibody_inline_set(handle_t
 656
 657         if (EXT4_I(inode)->i_extra_isize == 0)
 658                 return -ENOSPC;
 659 -       error = ext4_xattr_set_entry(i, s);
 660 +       error = ext4_xattr_set_entry(i, s, handle, inode);
 661         if (error) {
 662                 if (error == -ENOSPC &&
 663                     ext4_has_inline_data(inode)) {
 664 @@ -1030,7 +1320,7 @@ int ext4_xattr_ibody_inline_set(handle_t
 665                         error = ext4_xattr_ibody_find(inode, i, is);
 666                         if (error)
 667                                 return error;
 668 -                       error = ext4_xattr_set_entry(i, s);
 669 +                       error = ext4_xattr_set_entry(i, s, handle, inode);
 670                 }
 671                 if (error)
 672                         return error;
 673 @@ -1056,7 +1346,7 @@ static int ext4_xattr_ibody_set(handle_t
 674
 675         if (EXT4_I(inode)->i_extra_isize == 0)
 676                 return -ENOSPC;
 677 -       error = ext4_xattr_set_entry(i, s);
 678 +       error = ext4_xattr_set_entry(i, s, handle, inode);
 679         if (error)
 680                 return error;
 681         header = IHDR(inode, ext4_raw_inode(&is->iloc));
 682 @@ -1092,7 +1382,7 @@ ext4_xattr_set_handle(handle_t *handle,
 683                 .name = name,
 684                 .value = value,
 685                 .value_len = value_len,
 686 -
 687 +               .in_inode = 0,
 688         };
 689         struct ext4_xattr_ibody_find is = {
 690                 .s = { .not_found = -ENODATA, },
 691 @@ -1157,6 +1447,15 @@ ext4_xattr_set_handle(handle_t *handle,
 692                                         goto cleanup;
 693                         }
 694                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
 695 +                       if (EXT4_HAS_INCOMPAT_FEATURE(inode->i_sb,
 696 +                                       EXT4_FEATURE_INCOMPAT_EA_INODE) &&
 697 +                           error == -ENOSPC) {
 698 +                               /* xattr not fit to block, store at external
 699 +                                * inode */
 700 +                               i.in_inode = 1;
 701 +                               error = ext4_xattr_ibody_set(handle, inode,
 702 +                                                            &i, &is);
 703 +                       }
 704                         if (error)
 705                                 goto cleanup;
 706                         if (!is.s.not_found) {
 707 @@ -1203,9 +1502,22 @@ ext4_xattr_set(struct inode *inode, int
 708                const void *value, size_t value_len, int flags)
 709  {
 710         handle_t *handle;
 711 +       struct super_block *sb = inode->i_sb;
 712         int error, retries = 0;
 713         int credits = ext4_jbd2_credits_xattr(inode);
 714
 715 +       if ((value_len >= EXT4_XATTR_MIN_LARGE_EA_SIZE(sb->s_blocksize)) &&
 716 +           EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EA_INODE)) {
 717 +               int nrblocks = (value_len + sb->s_blocksize - 1) >>
 718 +                                       sb->s_blocksize_bits;
 719 +
 720 +               /* For new inode */
 721 +               credits += EXT4_SINGLEDATA_TRANS_BLOCKS(sb) + 3;
 722 +
 723 +               /* For data blocks of EA inode */
 724 +               credits += ext4_meta_trans_blocks(inode, nrblocks, 0);
 725 +       }
 726 +
 727  retry:
 728         handle = ext4_journal_start(inode, EXT4_HT_XATTR, credits);
 729         if (IS_ERR(handle)) {
 730 @@ -1217,7 +1529,7 @@ retry:
 731                                               value, value_len, flags);
 732                 error2 = ext4_journal_stop(handle);
 733                 if (error == -ENOSPC &&
 734 -                   ext4_should_retry_alloc(inode->i_sb, &retries))
 735 +                   ext4_should_retry_alloc(sb, &retries))
 736                         goto retry;
 737                 if (error == 0)
 738                         error = error2;
 739 @@ -1239,7 +1551,7 @@ static void ext4_xattr_shift_entries(str
 740
 741         /* Adjust the value offsets of the entries */
 742         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 743 -               if (!last->e_value_block && last->e_value_size) {
 744 +               if (!last->e_value_inum && last->e_value_size) {
 745                         new_offs = le16_to_cpu(last->e_value_offs) +
 746                                                         value_offs_shift;
 747                         BUG_ON(new_offs + le32_to_cpu(last->e_value_size)
 748 @@ -1477,21 +1789,135 @@ cleanup:
 749  }
 750
 751
 752 +#define EIA_INCR 16 /* must be 2^n */
 753 +#define EIA_MASK (EIA_INCR - 1)
 754 +/* Add the large xattr @ino into @lea_ino_array for later deletion.
 755 + * If @lea_ino_array is new or full it will be grown and the old
 756 + * contents copied over.
 757 + */
 758 +static int
 759 +ext4_expand_ino_array(struct ext4_xattr_ino_array **lea_ino_array, __u32 ino)
 760 +{
 761 +       if (*lea_ino_array == NULL) {
 762 +               /*
 763 +                * Start with 15 inodes, so it fits into a power-of-two size.
 764 +                * If *lea_ino_array is NULL, this is essentially offsetof()
 765 +                */
 766 +               (*lea_ino_array) =
 767 +                       kmalloc(offsetof(struct ext4_xattr_ino_array,
 768 +                                        xia_inodes[EIA_MASK]),
 769 +                               GFP_NOFS);
 770 +               if (*lea_ino_array == NULL)
 771 +                       return -ENOMEM;
 772 +               (*lea_ino_array)->xia_count = 0;
 773 +       } else if (((*lea_ino_array)->xia_count & EIA_MASK) == EIA_MASK) {
 774 +               /* expand the array once all 15 + n * 16 slots are full */
 775 +               struct ext4_xattr_ino_array *new_array = NULL;
 776 +               int count = (*lea_ino_array)->xia_count;
 777 +
 778 +               /* if new_array is NULL, this is essentially offsetof() */
 779 +               new_array = kmalloc(
 780 +                               offsetof(struct ext4_xattr_ino_array,
 781 +                                        xia_inodes[count + EIA_INCR]),
 782 +                               GFP_NOFS);
 783 +               if (new_array == NULL)
 784 +                       return -ENOMEM;
 785 +               memcpy(new_array, *lea_ino_array,
 786 +                      offsetof(struct ext4_xattr_ino_array,
 787 +                               xia_inodes[count]));
 788 +               kfree(*lea_ino_array);
 789 +               *lea_ino_array = new_array;
 790 +       }
 791 +       (*lea_ino_array)->xia_inodes[(*lea_ino_array)->xia_count++] = ino;
 792 +       return 0;
 793 +}
 794 +
 795 +/**
 796 + * Add xattr inode to orphan list
 797 + */
 798 +static int
 799 +ext4_xattr_inode_orphan_add(handle_t *handle, struct inode *inode,
 800 +                       int credits, struct ext4_xattr_ino_array *lea_ino_array)
 801 +{
 802 +       struct inode *ea_inode = NULL;
 803 +       int idx = 0, error = 0;
 804 +
 805 +       if (lea_ino_array == NULL)
 806 +               return 0;
 807 +
 808 +       for (; idx < lea_ino_array->xia_count; ++idx) {
 809 +               if (!ext4_handle_has_enough_credits(handle, credits)) {
 810 +                       error = ext4_journal_extend(handle, credits);
 811 +                       if (error > 0)
 812 +                               error = ext4_journal_restart(handle, credits);
 813 +
 814 +                       if (error != 0) {
 815 +                               ext4_warning(inode->i_sb,
 816 +                                       "couldn't extend journal "
 817 +                                       "(err %d)", error);
 818 +                               return error;
 819 +                       }
 820 +               }
 821 +               ea_inode = ext4_xattr_inode_iget(inode,
 822 +                               lea_ino_array->xia_inodes[idx], &error);
 823 +               if (error)
 824 +                       continue;
 825 +               ext4_orphan_add(handle, ea_inode);
 826 +               /* the inode's i_count will be released by caller */
 827 +       }
 828 +
 829 +       return 0;
 830 +}
 831
 832  /*
 833   * ext4_xattr_delete_inode()
 834   *
 835 - * Free extended attribute resources associated with this inode. This
 836 + * Free extended attribute resources associated with this inode. Traverse
 837 + * all entries and unlink any xattr inodes associated with this inode. This
 838   * is called immediately before an inode is freed. We have exclusive
 839 - * access to the inode.
 840 + * access to the inode. If an orphan inode is deleted it will also delete any
 841 + * xattr block and all xattr inodes. They are checked by ext4_xattr_inode_iget()
 842 + * to ensure they belong to the parent inode and were not deleted already.
 843   */
 844 -void
 845 -ext4_xattr_delete_inode(handle_t *handle, struct inode *inode)
 846 +int
 847 +ext4_xattr_delete_inode(handle_t *handle, struct inode *inode,
 848 +                       struct ext4_xattr_ino_array **lea_ino_array)
 849  {
 850         struct buffer_head *bh = NULL;
 851 +       struct ext4_xattr_ibody_header *header;
 852 +       struct ext4_inode *raw_inode;
 853 +       struct ext4_iloc iloc;
 854 +       struct ext4_xattr_entry *entry;
 855 +       int credits = 3, error = 0;
 856
 857 -       if (!EXT4_I(inode)->i_file_acl)
 858 +       if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR))
 859 +               goto delete_external_ea;
 860 +
 861 +       error = ext4_get_inode_loc(inode, &iloc);
 862 +       if (error)
 863                 goto cleanup;
 864 +       raw_inode = ext4_raw_inode(&iloc);
 865 +       header = IHDR(inode, raw_inode);
 866 +       for (entry = IFIRST(header); !IS_LAST_ENTRY(entry);
 867 +            entry = EXT4_XATTR_NEXT(entry)) {
 868 +               if (!entry->e_value_inum)
 869 +                       continue;
 870 +               if (ext4_expand_ino_array(lea_ino_array,
 871 +                                         entry->e_value_inum) != 0) {
 872 +                       brelse(iloc.bh);
 873 +                       goto cleanup;
 874 +               }
 875 +               entry->e_value_inum = 0;
 876 +       }
 877 +       brelse(iloc.bh);
 878 +
 879 +delete_external_ea:
 880 +       if (!EXT4_I(inode)->i_file_acl) {
 881 +               /* add xattr inode to orphan list */
 882 +               ext4_xattr_inode_orphan_add(handle, inode, credits,
 883 +                                               *lea_ino_array);
 884 +               goto cleanup;
 885 +       }
 886         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
 887         if (!bh) {
 888                 EXT4_ERROR_INODE(inode, "block %llu read error",
 889 @@ -1504,11 +1930,69 @@ ext4_xattr_delete_inode(handle_t *handle
 890                                  EXT4_I(inode)->i_file_acl);
 891                 goto cleanup;
 892         }
 893 +
 894 +       for (entry = BFIRST(bh); !IS_LAST_ENTRY(entry);
 895 +            entry = EXT4_XATTR_NEXT(entry)) {
 896 +               if (!entry->e_value_inum)
 897 +                       continue;
 898 +               if (ext4_expand_ino_array(lea_ino_array,
 899 +                                         entry->e_value_inum) != 0)
 900 +                       goto cleanup;
 901 +               entry->e_value_inum = 0;
 902 +       }
 903 +
 904 +       /* add xattr inode to orphan list */
 905 +       error = ext4_xattr_inode_orphan_add(handle, inode, credits,
 906 +                                       *lea_ino_array);
 907 +       if (error != 0)
 908 +               goto cleanup;
 909 +
 910 +       if (!IS_NOQUOTA(inode))
 911 +               credits += 2 * EXT4_QUOTA_DEL_BLOCKS(inode->i_sb);
 912 +
 913 +       if (!ext4_handle_has_enough_credits(handle, credits)) {
 914 +               error = ext4_journal_extend(handle, credits);
 915 +               if (error > 0)
 916 +                       error = ext4_journal_restart(handle, credits);
 917 +               if (error != 0) {
 918 +                       ext4_warning(inode->i_sb,
 919 +                               "couldn't extend journal (err %d)", error);
 920 +                       goto cleanup;
 921 +               }
 922 +       }
 923 +
 924         ext4_xattr_release_block(handle, inode, bh);
 925         EXT4_I(inode)->i_file_acl = 0;
 926
 927  cleanup:
 928         brelse(bh);
 929 +
 930 +       return error;
 931 +}
 932 +
 933 +void
 934 +ext4_xattr_inode_array_free(struct inode *inode,
 935 +                           struct ext4_xattr_ino_array *lea_ino_array)
 936 +{
 937 +       struct inode    *ea_inode = NULL;
 938 +       int             idx = 0;
 939 +       int             err;
 940 +
 941 +       if (lea_ino_array == NULL)
 942 +               return;
 943 +
 944 +       for (; idx < lea_ino_array->xia_count; ++idx) {
 945 +               ea_inode = ext4_xattr_inode_iget(inode,
 946 +                               lea_ino_array->xia_inodes[idx], &err);
 947 +               if (err)
 948 +                       continue;
 949 +               /* for inode's i_count get from ext4_xattr_delete_inode */
 950 +               if (!list_empty(&EXT4_I(ea_inode)->i_orphan))
 951 +                       iput(ea_inode);
 952 +               clear_nlink(ea_inode);
 953 +               iput(ea_inode);
 954 +       }
 955 +       kfree(lea_ino_array);
 956  }
 957
 958  /*
 959 @@ -1578,10 +2062,9 @@ ext4_xattr_cmp(struct ext4_xattr_header
 960                     entry1->e_name_index != entry2->e_name_index ||
 961                     entry1->e_name_len != entry2->e_name_len ||
 962                     entry1->e_value_size != entry2->e_value_size ||
 963 +                   entry1->e_value_inum != entry2->e_value_inum ||
 964                     memcmp(entry1->e_name, entry2->e_name, entry1->e_name_len))
 965                         return 1;
 966 -               if (entry1->e_value_block != 0 || entry2->e_value_block != 0)
 967 -                       return -EIO;
 968                 if (memcmp((char *)header1 + le16_to_cpu(entry1->e_value_offs),
 969                            (char *)header2 + le16_to_cpu(entry2->e_value_offs),
 970                            le32_to_cpu(entry1->e_value_size)))
 971 @@ -1665,7 +2148,7 @@ static inline void ext4_xattr_hash_entry
 972                        *name++;
 973         }
 974
 975 -       if (entry->e_value_block == 0 && entry->e_value_size != 0) {
 976 +       if (!entry->e_value_inum && entry->e_value_size) {
 977                 __le32 *value = (__le32 *)((char *)header +
 978                         le16_to_cpu(entry->e_value_offs));
 979                 for (n = (le32_to_cpu(entry->e_value_size) +
 980 Index: linux-stage/fs/ext4/xattr.h
 981 ===================================================================
 982 --- linux-stage.orig/fs/ext4/xattr.h
 983 +++ linux-stage/fs/ext4/xattr.h
 984 @@ -42,7 +42,7 @@ struct ext4_xattr_entry {
 985         __u8    e_name_len;     /* length of name */
 986         __u8    e_name_index;   /* attribute name index */
 987         __le16  e_value_offs;   /* offset in disk block of value */
 988 -       __le32  e_value_block;  /* disk block attribute is stored on (n/i) */
 989 +       __le32  e_value_inum;   /* inode in which the value is stored */
 990         __le32  e_value_size;   /* size of attribute value */
 991         __le32  e_hash;         /* hash value of name and value */
 992         char    e_name[0];      /* attribute name */
 993 @@ -67,6 +67,15 @@ struct ext4_xattr_entry {
 994                 EXT4_I(inode)->i_extra_isize))
 995  #define IFIRST(hdr) ((struct ext4_xattr_entry *)((hdr)+1))
 996
 997 +#define i_xattr_inode_parent i_mtime.tv_sec
 998 +
 999 +/*
1000 + * The minimum size of EA value when you start storing it in an external inode
1001 + * size of block - size of header - size of 1 entry - 4 null bytes
1002 +*/
1003 +#define EXT4_XATTR_MIN_LARGE_EA_SIZE(b)                                        \
1004 +       ((b) - EXT4_XATTR_LEN(3) - sizeof(struct ext4_xattr_header) - 4)
1005 +
1006  #define BHDR(bh) ((struct ext4_xattr_header *)((bh)->b_data))
1007  #define ENTRY(ptr) ((struct ext4_xattr_entry *)(ptr))
1008  #define BFIRST(bh) ENTRY(BHDR(bh)+1)
1009 @@ -75,10 +84,11 @@ struct ext4_xattr_entry {
1010  #define EXT4_ZERO_XATTR_VALUE ((void *)-1)
1011
1012  struct ext4_xattr_info {
1013 -       int name_index;
1014         const char *name;
1015         const void *value;
1016         size_t value_len;
1017 +       int name_index;
1018 +       int in_inode;
1019  };
1020
1021  struct ext4_xattr_search {
1022 @@ -106,7 +116,13 @@ extern int ext4_xattr_get(struct inode *
1023  extern int ext4_xattr_set(struct inode *, int, const char *, const void *, size_t, int);
1024  extern int ext4_xattr_set_handle(handle_t *, struct inode *, int, const char *, const void *, size_t, int);
1025
1026 -extern void ext4_xattr_delete_inode(handle_t *, struct inode *);
1027 +extern struct inode *ext4_xattr_inode_iget(struct inode *parent, unsigned long ea_ino,
1028 +                                          int *err);
1029 +extern int ext4_xattr_inode_unlink(struct inode *inode, unsigned long ea_ino);
1030 +extern int ext4_xattr_delete_inode(handle_t *handle, struct inode *inode,
1031 +                                  struct ext4_xattr_ino_array **array);
1032 +extern void ext4_xattr_inode_array_free(struct inode *inode,
1033 +                                       struct ext4_xattr_ino_array *array);
1034  extern void ext4_xattr_put_super(struct super_block *);
1035
1036  extern int ext4_expand_extra_isize_ea(struct inode *inode, int new_extra_isize,
1037 Index: linux-stage/fs/ext4/ialloc.c
1038 ===================================================================
1039 --- linux-stage.orig/fs/ext4/ialloc.c
1040 +++ linux-stage/fs/ext4/ialloc.c
1041 @@ -269,7 +269,6 @@ void ext4_free_inode(handle_t *handle, s
1042          * as writing the quota to disk may need the lock as well.
1043          */
1044         dquot_initialize(inode);
1045 -       ext4_xattr_delete_inode(handle, inode);
1046         dquot_free_inode(inode);
1047         dquot_drop(inode);
1048
1049 Index: linux-stage/fs/ext4/inline.c
1050 ===================================================================
1051 --- linux-stage.orig/fs/ext4/inline.c
1052 +++ linux-stage/fs/ext4/inline.c
1053 @@ -59,7 +59,7 @@ static int get_max_inline_xattr_value_si
1054
1055         /* Compute min_offs. */
1056         for (; !IS_LAST_ENTRY(entry); entry = EXT4_XATTR_NEXT(entry)) {
1057 -               if (!entry->e_value_block && entry->e_value_size) {
1058 +               if (!entry->e_value_inum && entry->e_value_size) {
1059                         size_t offs = le16_to_cpu(entry->e_value_offs);
1060                         if (offs < min_offs)
1061                                 min_offs = offs;