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-3.10.0-123.13.2.el7.x86_64/fs/ext4/ext4.h
   8 ===================================================================
   9 --- linux-3.10.0-123.13.2.el7.x86_64.orig/fs/ext4/ext4.h
  10 +++ linux-3.10.0-123.13.2.el7.x86_64/fs/ext4/ext4.h
  11 @@ -1534,6 +1534,7 @@ 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 @@ -1934,6 +1935,12 @@ 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 @@ -1945,6 +1952,10 @@ 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 @@ -2157,6 +2168,7 @@ 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_discard_partial_page_buffers(handle_t *handle,
  51 Index: linux-3.10.0-123.13.2.el7.x86_64/fs/ext4/inode.c
  52 ===================================================================
  53 --- linux-3.10.0-123.13.2.el7.x86_64.orig/fs/ext4/inode.c
  54 +++ linux-3.10.0-123.13.2.el7.x86_64/fs/ext4/inode.c
  55 @@ -133,8 +183,6 @@ void ext4_evict_inode(struct inode *inod
  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 @@ -183,6 +183,8 @@ 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 @@ -235,8 +237,8 @@ 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 @@ -248,9 +250,36 @@ 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 @@ -305,8 +334,12 @@ 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 @@ -4841,7 +4874,7 @@ 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         int gdpblocks;
 145 Index: linux-3.10.0-123.13.2.el7.x86_64/fs/ext4/xattr.c
 146 ===================================================================
 147 --- linux-3.10.0-123.13.2.el7.x86_64.orig/fs/ext4/xattr.c
 148 +++ linux-3.10.0-123.13.2.el7.x86_64/fs/ext4/xattr.c
 149 @@ -220,19 +220,26 @@ ext4_xattr_check_block(struct inode *ino
 150  }
 151
 152  static inline int
 153 -ext4_xattr_check_entry(struct ext4_xattr_entry *entry, size_t size)
 154 +ext4_xattr_check_entry(struct ext4_xattr_entry *entry, size_t size,
 155 +                      struct inode *inode)
 156  {
 157         size_t value_size = le32_to_cpu(entry->e_value_size);
 158
 159 -       if (entry->e_value_block != 0 || value_size > size ||
 160 -           le16_to_cpu(entry->e_value_offs) + value_size > size)
 161 +       if ((entry->e_value_inum == 0) &&
 162 +          (le16_to_cpu(entry->e_value_offs) + value_size > size))
 163 +               return -EIO;
 164 +       if (entry->e_value_inum != 0 &&
 165 +           (le32_to_cpu(entry->e_value_inum) < EXT4_FIRST_INO(inode->i_sb) ||
 166 +            le32_to_cpu(entry->e_value_inum) >
 167 +            le32_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_inodes_count)))
 168                 return -EIO;
 169         return 0;
 170  }
 171
 172  static int
 173  ext4_xattr_find_entry(struct ext4_xattr_entry **pentry, int name_index,
 174 -                     const char *name, size_t size, int sorted)
 175 +                     const char *name, size_t size, int sorted,
 176 +                     struct inode *inode)
 177  {
 178         struct ext4_xattr_entry *entry;
 179         size_t name_len;
 180 @@ -252,11 +259,103 @@ ext4_xattr_find_entry(struct ext4_xattr_
 181                         break;
 182         }
 183         *pentry = entry;
 184 -       if (!cmp && ext4_xattr_check_entry(entry, size))
 185 +       if (!cmp && ext4_xattr_check_entry(entry, size, inode))
 186                         return -EIO;
 187         return cmp ? -ENODATA : 0;
 188  }
 189
 190 +/*
 191 + * Read the EA value from an inode.
 192 + */
 193 +static int
 194 +ext4_xattr_inode_read(struct inode *ea_inode, void *buf, size_t *size)
 195 +{
 196 +       unsigned long block = 0;
 197 +       struct buffer_head *bh = NULL;
 198 +       int err, blocksize;
 199 +       size_t csize, ret_size = 0;
 200 +
 201 +       if (*size == 0)
 202 +               return 0;
 203 +
 204 +       blocksize = ea_inode->i_sb->s_blocksize;
 205 +
 206 +       while (ret_size < *size) {
 207 +               csize = (*size - ret_size) > blocksize ? blocksize :
 208 +                                                       *size - ret_size;
 209 +               bh = ext4_bread(NULL, ea_inode, block, 0, &err);
 210 +               if (!bh) {
 211 +                       *size = ret_size;
 212 +                       return err;
 213 +               }
 214 +               memcpy(buf, bh->b_data, csize);
 215 +               brelse(bh);
 216 +
 217 +               buf += csize;
 218 +               block += 1;
 219 +               ret_size += csize;
 220 +       }
 221 +
 222 +       *size = ret_size;
 223 +
 224 +       return err;
 225 +}
 226 +
 227 +struct inode *ext4_xattr_inode_iget(struct inode *parent, int ea_ino, int *err)
 228 +{
 229 +       struct inode *ea_inode = NULL;
 230 +
 231 +       ea_inode = ext4_iget(parent->i_sb, ea_ino);
 232 +       if (IS_ERR(ea_inode) || is_bad_inode(ea_inode)) {
 233 +               ext4_error(parent->i_sb, "error while reading EA inode %d",
 234 +                          ea_ino);
 235 +               *err = -EIO;
 236 +               return NULL;
 237 +       }
 238 +
 239 +       if (ea_inode->i_xattr_inode_parent != parent->i_ino ||
 240 +           ea_inode->i_generation != parent->i_generation) {
 241 +               ext4_error(parent->i_sb, "Backpointer from EA inode %d "
 242 +                          "to parent invalid.", ea_ino);
 243 +               *err = -EINVAL;
 244 +               goto error;
 245 +       }
 246 +
 247 +       if (!(EXT4_I(ea_inode)->i_flags & EXT4_EA_INODE_FL)) {
 248 +               ext4_error(parent->i_sb, "EA inode %d does not have "
 249 +                          "EXT4_EA_INODE_FL flag set.\n", ea_ino);
 250 +               *err = -EINVAL;
 251 +               goto error;
 252 +       }
 253 +
 254 +       *err = 0;
 255 +       return ea_inode;
 256 +
 257 +error:
 258 +       iput(ea_inode);
 259 +       return NULL;
 260 +}
 261 +
 262 +/*
 263 + * Read the value from the EA inode.
 264 + */
 265 +static int
 266 +ext4_xattr_inode_get(struct inode *inode, int ea_ino, void *buffer,
 267 +                    size_t *size)
 268 +{
 269 +       struct inode *ea_inode = NULL;
 270 +       int err;
 271 +
 272 +       ea_inode = ext4_xattr_inode_iget(inode, ea_ino, &err);
 273 +       if (err)
 274 +               return err;
 275 +
 276 +       err = ext4_xattr_inode_read(ea_inode, buffer, size);
 277 +       iput(ea_inode);
 278 +
 279 +       return err;
 280 +}
 281 +
 282  static int
 283  ext4_xattr_block_get(struct inode *inode, int name_index, const char *name,
 284                      void *buffer, size_t buffer_size)
 285 @@ -288,7 +387,8 @@ bad_block:
 286         }
 287         ext4_xattr_cache_insert(bh);
 288         entry = BFIRST(bh);
 289 -       error = ext4_xattr_find_entry(&entry, name_index, name, bh->b_size, 1);
 290 +       error = ext4_xattr_find_entry(&entry, name_index, name, bh->b_size, 1,
 291 +                                     inode);
 292         if (error == -EIO)
 293                 goto bad_block;
 294         if (error)
 295 @@ -298,8 +398,16 @@ bad_block:
 296                 error = -ERANGE;
 297                 if (size > buffer_size)
 298                         goto cleanup;
 299 -               memcpy(buffer, bh->b_data + le16_to_cpu(entry->e_value_offs),
 300 -                      size);
 301 +               if (entry->e_value_inum != 0) {
 302 +                       error = ext4_xattr_inode_get(inode,
 303 +                                            le32_to_cpu(entry->e_value_inum),
 304 +                                            buffer, &size);
 305 +                       if (error)
 306 +                               goto cleanup;
 307 +               } else {
 308 +                       memcpy(buffer, bh->b_data +
 309 +                              le16_to_cpu(entry->e_value_offs), size);
 310 +               }
 311         }
 312         error = size;
 313
 314 @@ -333,7 +441,7 @@ ext4_xattr_ibody_get(struct inode *inode
 315         if (error)
 316                 goto cleanup;
 317         error = ext4_xattr_find_entry(&entry, name_index, name,
 318 -                                     end - (void *)entry, 0);
 319 +                                     end - (void *)entry, 0, inode);
 320         if (error)
 321                 goto cleanup;
 322         size = le32_to_cpu(entry->e_value_size);
 323 @@ -341,8 +449,16 @@ ext4_xattr_ibody_get(struct inode *inode
 324                 error = -ERANGE;
 325                 if (size > buffer_size)
 326                         goto cleanup;
 327 -               memcpy(buffer, (void *)IFIRST(header) +
 328 -                      le16_to_cpu(entry->e_value_offs), size);
 329 +               if (entry->e_value_inum != 0) {
 330 +                       error = ext4_xattr_inode_get(inode,
 331 +                                            le32_to_cpu(entry->e_value_inum),
 332 +                                            buffer, &size);
 333 +                       if (error)
 334 +                               goto cleanup;
 335 +               } else {
 336 +                       memcpy(buffer, (void *)IFIRST(header) +
 337 +                              le16_to_cpu(entry->e_value_offs), size);
 338 +               }
 339         }
 340         error = size;
 341
 342 @@ -568,7 +684,7 @@ static size_t ext4_xattr_free_space(stru
 343                 *total += EXT4_XATTR_LEN(last->e_name_len);
 344  {
 345         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 346 -               if (!last->e_value_block && last->e_value_size) {
 347 +               if (last->e_value_inum == 0 && last->e_value_size > 0) {
 348                         size_t offs = le16_to_cpu(last->e_value_offs);
 349                         if (offs < *min_offs)
 350                                 *min_offs = offs;
 351 @@ -577,16 +693,172 @@ static size_t ext4_xattr_free_space(stru
 352         return (*min_offs - ((void *)last - base) - sizeof(__u32));
 353  }
 354
 355 +/*
 356 + * Write the value of the EA in an inode.
 357 + */
 358 +static int
 359 +ext4_xattr_inode_write(handle_t *handle, struct inode *ea_inode,
 360 +                      const void *buf, int bufsize)
 361 +{
 362 +       struct buffer_head *bh = NULL;
 363 +       unsigned long block = 0;
 364 +       unsigned blocksize = ea_inode->i_sb->s_blocksize;
 365 +       unsigned max_blocks = (bufsize + blocksize - 1) >> ea_inode->i_blkbits;
 366 +       int csize, wsize = 0;
 367 +       int ret = 0;
 368 +       int retries = 0;
 369 +
 370 +retry:
 371 +       while (ret >= 0 && ret < max_blocks) {
 372 +               struct ext4_map_blocks map;
 373 +               map.m_lblk = block += ret;
 374 +               map.m_len = max_blocks -= ret;
 375 +
 376 +               ret = ext4_map_blocks(handle, ea_inode, &map,
 377 +                                     EXT4_GET_BLOCKS_CREATE);
 378 +               if (ret <= 0) {
 379 +                       ext4_mark_inode_dirty(handle, ea_inode);
 380 +                       if (ret == -ENOSPC &&
 381 +                           ext4_should_retry_alloc(ea_inode->i_sb, &retries)) {
 382 +                               ret = 0;
 383 +                               goto retry;
 384 +                       }
 385 +                       break;
 386 +               }
 387 +       }
 388 +
 389 +       if (ret < 0)
 390 +               return ret;
 391 +
 392 +       block = 0;
 393 +       while (wsize < bufsize) {
 394 +               if (bh != NULL)
 395 +                       brelse(bh);
 396 +               csize = (bufsize - wsize) > blocksize ? blocksize :
 397 +                                                               bufsize - wsize;
 398 +               bh = ext4_getblk(handle, ea_inode, block, 0, &ret);
 399 +               if (!bh)
 400 +                       goto out;
 401 +               ret = ext4_journal_get_write_access(handle, bh);
 402 +               if (ret)
 403 +                       goto out;
 404 +
 405 +               memcpy(bh->b_data, buf, csize);
 406 +               set_buffer_uptodate(bh);
 407 +               ext4_handle_dirty_metadata(handle, ea_inode, bh);
 408 +
 409 +               buf += csize;
 410 +               wsize += csize;
 411 +               block += 1;
 412 +       }
 413 +
 414 +       i_size_write(ea_inode, wsize);
 415 +       ext4_update_i_disksize(ea_inode, wsize);
 416 +
 417 +       ext4_mark_inode_dirty(handle, ea_inode);
 418 +
 419 +out:
 420 +       brelse(bh);
 421 +
 422 +       return ret;
 423 +}
 424 +
 425 +/*
 426 + * Create an inode to store the value of a large EA.
 427 + */
 428 +static struct inode *
 429 +ext4_xattr_inode_create(handle_t *handle, struct inode *inode)
 430 +{
 431 +       struct inode *ea_inode = NULL;
 432 +
 433 +       /*
 434 +        * Let the next inode be the goal, so we try and allocate the EA inode
 435 +        * in the same group, or nearby one.
 436 +        */
 437 +       ea_inode = ext4_new_inode(handle, inode->i_sb->s_root->d_inode,
 438 +                                 S_IFREG|0600, NULL, inode->i_ino + 1, NULL);
 439 +
 440 +       if (!IS_ERR(ea_inode)) {
 441 +               ea_inode->i_op = &ext4_file_inode_operations;
 442 +               ea_inode->i_fop = &ext4_file_operations;
 443 +               ext4_set_aops(ea_inode);
 444 +               ea_inode->i_generation = inode->i_generation;
 445 +               EXT4_I(ea_inode)->i_flags |= EXT4_EA_INODE_FL;
 446 +
 447 +               /*
 448 +                * A back-pointer from EA inode to parent inode will be useful
 449 +                * for e2fsck.
 450 +                */
 451 +               ea_inode->i_xattr_inode_parent = inode->i_ino;
 452 +               unlock_new_inode(ea_inode);
 453 +       }
 454 +
 455 +       return ea_inode;
 456 +}
 457 +
 458 +/*
 459 + * Unlink the inode storing the value of the EA.
 460 + */
 461 +int
 462 +ext4_xattr_inode_unlink(struct inode *inode, int ea_ino)
 463 +{
 464 +       struct inode *ea_inode = NULL;
 465 +       int err;
 466 +
 467 +       ea_inode = ext4_xattr_inode_iget(inode, ea_ino, &err);
 468 +       if (err)
 469 +               return err;
 470 +
 471 +       clear_nlink(ea_inode);
 472 +       iput(ea_inode);
 473 +
 474 +       return 0;
 475 +}
 476 +
 477 +/*
 478 + * Add value of the EA in an inode.
 479 + */
 480  static int
 481 -ext4_xattr_set_entry(struct ext4_xattr_info *i, struct ext4_xattr_search *s)
 482 +ext4_xattr_inode_set(handle_t *handle, struct inode *inode, int *ea_ino,
 483 +                    const void *value, size_t value_len)
 484 +{
 485 +       struct inode *ea_inode = NULL;
 486 +       int err;
 487 +
 488 +       /* Create an inode for the EA value */
 489 +       ea_inode = ext4_xattr_inode_create(handle, inode);
 490 +       if (IS_ERR(ea_inode))
 491 +               return -1;
 492 +
 493 +       err = ext4_xattr_inode_write(handle, ea_inode, value, value_len);
 494 +       if (err)
 495 +               clear_nlink(ea_inode);
 496 +       else
 497 +               *ea_ino = ea_inode->i_ino;
 498 +
 499 +       iput(ea_inode);
 500 +
 501 +       return err;
 502 +}
 503 +
 504 +static int
 505 +ext4_xattr_set_entry(struct ext4_xattr_info *i, struct ext4_xattr_search *s,
 506 +                    handle_t *handle, struct inode *inode)
 507  {
 508         struct ext4_xattr_entry *last;
 509         size_t free, min_offs = s->end - s->base, name_len = strlen(i->name);
 510 +       int in_inode = i->in_inode;
 511 +
 512 +       if (EXT4_HAS_INCOMPAT_FEATURE(inode->i_sb,
 513 +                EXT4_FEATURE_INCOMPAT_EA_INODE) &&
 514 +           (EXT4_XATTR_SIZE(i->value_len) >
 515 +            EXT4_XATTR_MIN_LARGE_EA_SIZE(inode->i_sb->s_blocksize)))
 516 +               in_inode = 1;
 517
 518         /* Compute min_offs and last. */
 519         last = s->first;
 520         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 521 -               if (!last->e_value_block && last->e_value_size) {
 522 +               if (last->e_value_inum == 0 && last->e_value_size > 0) {
 523                         size_t offs = le16_to_cpu(last->e_value_offs);
 524                         if (offs < min_offs)
 525                                 min_offs = offs;
 526 @@ -594,16 +865,21 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 527         }
 528         free = min_offs - ((void *)last - s->base) - sizeof(__u32);
 529         if (!s->not_found) {
 530 -               if (!s->here->e_value_block && s->here->e_value_size) {
 531 +               if (!in_inode && s->here->e_value_inum == 0 &&
 532 +                   s->here->e_value_size > 0) {
 533                         size_t size = le32_to_cpu(s->here->e_value_size);
 534                         free += EXT4_XATTR_SIZE(size);
 535                 }
 536                 free += EXT4_XATTR_LEN(name_len);
 537         }
 538         if (i->value) {
 539 -               if (free < EXT4_XATTR_SIZE(i->value_len) ||
 540 -                   free < EXT4_XATTR_LEN(name_len) +
 541 -                          EXT4_XATTR_SIZE(i->value_len))
 542 +               size_t value_len = EXT4_XATTR_SIZE(i->value_len);
 543 +
 544 +               if (in_inode)
 545 +                       value_len = 0;
 546 +
 547 +               if (free < value_len ||
 548 +                   free < EXT4_XATTR_LEN(name_len) + value_len)
 549                         return -ENOSPC;
 550         }
 551
 552 @@ -617,7 +893,8 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 553                 s->here->e_name_len = name_len;
 554                 memcpy(s->here->e_name, i->name, name_len);
 555         } else {
 556 -               if (!s->here->e_value_block && s->here->e_value_size) {
 557 +               if (s->here->e_value_offs > 0 && s->here->e_value_inum == 0 &&
 558 +                   s->here->e_value_size > 0) {
 559                         void *first_val = s->base + min_offs;
 560                         size_t offs = le16_to_cpu(s->here->e_value_offs);
 561                         void *val = s->base + offs;
 562 @@ -651,13 +928,17 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 563                         last = s->first;
 564                         while (!IS_LAST_ENTRY(last)) {
 565                                 size_t o = le16_to_cpu(last->e_value_offs);
 566 -                               if (!last->e_value_block &&
 567 -                                   last->e_value_size && o < offs)
 568 +                               if (last->e_value_size > 0 && o < offs)
 569                                         last->e_value_offs =
 570                                                 cpu_to_le16(o + size);
 571                                 last = EXT4_XATTR_NEXT(last);
 572                         }
 573                 }
 574 +               if (s->here->e_value_inum != 0) {
 575 +                       ext4_xattr_inode_unlink(inode,
 576 +                                       le32_to_cpu(s->here->e_value_inum));
 577 +                       s->here->e_value_inum = 0;
 578 +               }
 579                 if (!i->value) {
 580                         /* Remove the old name. */
 581                         size_t size = EXT4_XATTR_LEN(name_len);
 582 @@ -671,10 +952,17 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 583         if (i->value) {
 584                 /* Insert the new value. */
 585                 s->here->e_value_size = cpu_to_le32(i->value_len);
 586 -               if (i->value_len) {
 587 +               if (in_inode) {
 588 +                       int ea_ino = le32_to_cpu(s->here->e_value_inum);
 589 +                       ext4_xattr_inode_set(handle, inode, &ea_ino, i->value,
 590 +                                            i->value_len);
 591 +                       s->here->e_value_inum = cpu_to_le32(ea_ino);
 592 +                       s->here->e_value_offs = 0;
 593 +               } else if (i->value_len) {
 594                         size_t size = EXT4_XATTR_SIZE(i->value_len);
 595                         void *val = s->base + min_offs - size;
 596                         s->here->e_value_offs = cpu_to_le16(min_offs - size);
 597 +                       s->here->e_value_inum = 0;
 598                         if (i->value == EXT4_ZERO_XATTR_VALUE) {
 599                                 memset(val, 0, size);
 600                         } else {
 601 @@ -724,7 +1012,7 @@ ext4_xattr_block_find(struct inode *inod
 602                 bs->s.end = bs->bh->b_data + bs->bh->b_size;
 603                 bs->s.here = bs->s.first;
 604                 error = ext4_xattr_find_entry(&bs->s.here, i->name_index,
 605 -                                             i->name, bs->bh->b_size, 1);
 606 +                                            i->name, bs->bh->b_size, 1, inode);
 607                 if (error && error != -ENODATA)
 608                         goto cleanup;
 609                 bs->s.not_found = error;
 610 @@ -748,8 +1036,6 @@ ext4_xattr_block_set(handle_t *handle, s
 611
 612  #define header(x) ((struct ext4_xattr_header *)(x))
 613
 614 -       if (i->value && i->value_len > sb->s_blocksize)
 615 -               return -ENOSPC;
 616         if (s->base) {
 617                 ce = mb_cache_entry_get(ext4_xattr_cache, bs->bh->b_bdev,
 618                                         bs->bh->b_blocknr);
 619 @@ -764,7 +1050,7 @@ ext4_xattr_block_set(handle_t *handle, s
 620                                 ce = NULL;
 621                         }
 622                         ea_bdebug(bs->bh, "modifying in-place");
 623 -                       error = ext4_xattr_set_entry(i, s);
 624 +                       error = ext4_xattr_set_entry(i, s, handle, inode);
 625                         if (!error) {
 626                                 if (!IS_LAST_ENTRY(s->first))
 627                                         ext4_xattr_rehash(header(s->base),
 628 @@ -815,7 +1101,7 @@ ext4_xattr_block_set(handle_t *handle, s
 629                 s->end = s->base + sb->s_blocksize;
 630         }
 631
 632 -       error = ext4_xattr_set_entry(i, s);
 633 +       error = ext4_xattr_set_entry(i, s, handle, inode);
 634         if (error == -EIO)
 635                 goto bad_block;
 636         if (error)
 637 @@ -963,7 +1249,7 @@ int ext4_xattr_ibody_find(struct inode *
 638                 /* Find the named attribute. */
 639                 error = ext4_xattr_find_entry(&is->s.here, i->name_index,
 640                                               i->name, is->s.end -
 641 -                                             (void *)is->s.base, 0);
 642 +                                             (void *)is->s.base, 0, inode);
 643                 if (error && error != -ENODATA)
 644                         return error;
 645                 is->s.not_found = error;
 646 @@ -981,7 +1267,7 @@ int ext4_xattr_ibody_inline_set(handle_t
 647
 648         if (EXT4_I(inode)->i_extra_isize == 0)
 649                 return -ENOSPC;
 650 -       error = ext4_xattr_set_entry(i, s);
 651 +       error = ext4_xattr_set_entry(i, s, handle, inode);
 652         if (error) {
 653                 if (error == -ENOSPC &&
 654                     ext4_has_inline_data(inode)) {
 655 @@ -993,7 +1279,7 @@ int ext4_xattr_ibody_inline_set(handle_t
 656                         error = ext4_xattr_ibody_find(inode, i, is);
 657                         if (error)
 658                                 return error;
 659 -                       error = ext4_xattr_set_entry(i, s);
 660 +                       error = ext4_xattr_set_entry(i, s, handle, inode);
 661                 }
 662                 if (error)
 663                         return error;
 664 @@ -1019,7 +1305,7 @@ static int ext4_xattr_ibody_set(handle_t
 665
 666         if (EXT4_I(inode)->i_extra_isize == 0)
 667                 return -ENOSPC;
 668 -       error = ext4_xattr_set_entry(i, s);
 669 +       error = ext4_xattr_set_entry(i, s, handle, inode);
 670         if (error)
 671                 return error;
 672         header = IHDR(inode, ext4_raw_inode(&is->iloc));
 673 @@ -1055,7 +1341,7 @@ ext4_xattr_set_handle(handle_t *handle,
 674                 .name = name,
 675                 .value = value,
 676                 .value_len = value_len,
 677 -
 678 +               .in_inode = 0,
 679         };
 680         struct ext4_xattr_ibody_find is = {
 681                 .s = { .not_found = -ENODATA, },
 682 @@ -1120,6 +1406,15 @@ ext4_xattr_set_handle(handle_t *handle,
 683                                         goto cleanup;
 684                         }
 685                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
 686 +                       if (EXT4_HAS_INCOMPAT_FEATURE(inode->i_sb,
 687 +                                       EXT4_FEATURE_INCOMPAT_EA_INODE) &&
 688 +                           error == -ENOSPC) {
 689 +                               /* xattr not fit to block, store at external
 690 +                                * inode */
 691 +                               i.in_inode = 1;
 692 +                               error = ext4_xattr_ibody_set(handle, inode,
 693 +                                                            &i, &is);
 694 +                       }
 695                         if (error)
 696                                 goto cleanup;
 697                         if (!is.s.not_found) {
 698 @@ -1166,9 +1461,22 @@ ext4_xattr_set(struct inode *inode, int
 699                const void *value, size_t value_len, int flags)
 700  {
 701         handle_t *handle;
 702 +       struct super_block *sb = inode->i_sb;
 703         int error, retries = 0;
 704         int credits = ext4_jbd2_credits_xattr(inode);
 705
 706 +       if ((value_len >= EXT4_XATTR_MIN_LARGE_EA_SIZE(sb->s_blocksize)) &&
 707 +           EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EA_INODE)) {
 708 +               int nrblocks = (value_len + sb->s_blocksize - 1) >>
 709 +                                       sb->s_blocksize_bits;
 710 +
 711 +               /* For new inode */
 712 +               credits += EXT4_SINGLEDATA_TRANS_BLOCKS(sb) + 3;
 713 +
 714 +               /* For data blocks of EA inode */
 715 +               credits += ext4_meta_trans_blocks(inode, nrblocks, 0);
 716 +       }
 717 +
 718  retry:
 719         handle = ext4_journal_start(inode, EXT4_HT_XATTR, credits);
 720         if (IS_ERR(handle)) {
 721 @@ -1180,7 +1488,7 @@ retry:
 722                                               value, value_len, flags);
 723                 error2 = ext4_journal_stop(handle);
 724                 if (error == -ENOSPC &&
 725 -                   ext4_should_retry_alloc(inode->i_sb, &retries))
 726 +                   ext4_should_retry_alloc(sb, &retries))
 727                         goto retry;
 728                 if (error == 0)
 729                         error = error2;
 730 @@ -1202,7 +1510,7 @@ static void ext4_xattr_shift_entries(str
 731
 732         /* Adjust the value offsets of the entries */
 733         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 734 -               if (!last->e_value_block && last->e_value_size) {
 735 +               if (last->e_value_inum == 0 && last->e_value_size > 0) {
 736                         new_offs = le16_to_cpu(last->e_value_offs) +
 737                                                         value_offs_shift;
 738                         BUG_ON(new_offs + le32_to_cpu(last->e_value_size)
 739 @@ -1441,19 +1749,91 @@ cleanup:
 740  }
 741
 742
 743 -
 744 +#define EIA_INCR 16 /* must be 2^n */
 745 +#define EIA_MASK (EIA_INCR - 1)
 746 +/* Add the large xattr @ino into @lea_ino_array for later deletion.
 747 + * If @lea_ino_array is new or full it will be grown and the old
 748 + * contents copied over.
 749 + */
 750 +static int
 751 +ext4_expand_ino_array(struct ext4_xattr_ino_array **lea_ino_array, __u32 ino)
 752 +{
 753 +       if (*lea_ino_array == NULL) {
 754 +               /*
 755 +                * Start with 15 inodes, so it fits into a power-of-two size.
 756 +                * If *lea_ino_array is NULL, this is essentially offsetof()
 757 +                */
 758 +               (*lea_ino_array) =
 759 +                       kmalloc(offsetof(struct ext4_xattr_ino_array,
 760 +                                        xia_inodes[EIA_MASK]),
 761 +                               GFP_NOFS);
 762 +               if (*lea_ino_array == NULL)
 763 +                       return -ENOMEM;
 764 +               (*lea_ino_array)->xia_count = 0;
 765 +       } else if (((*lea_ino_array)->xia_count & EIA_MASK) == EIA_MASK) {
 766 +               /* expand the array once all 15 + n * 16 slots are full */
 767 +               struct ext4_xattr_ino_array *new_array = NULL;
 768 +               int count = (*lea_ino_array)->xia_count;
 769 +
 770 +               /* if new_array is NULL, this is essentially offsetof() */
 771 +               new_array = kmalloc(
 772 +                               offsetof(struct ext4_xattr_ino_array,
 773 +                                        xia_inodes[count + EIA_INCR]),
 774 +                               GFP_NOFS);
 775 +               if (new_array == NULL)
 776 +                       return -ENOMEM;
 777 +               memcpy(new_array, *lea_ino_array,
 778 +                      offsetof(struct ext4_xattr_ino_array,
 779 +                               xia_inodes[count]));
 780 +               kfree(*lea_ino_array);
 781 +               *lea_ino_array = new_array;
 782 +       }
 783 +       (*lea_ino_array)->xia_inodes[(*lea_ino_array)->xia_count++] = ino;
 784 +       return 0;
 785 +}
 786  /*
 787   * ext4_xattr_delete_inode()
 788   *
 789 - * Free extended attribute resources associated with this inode. This
 790 + * Free extended attribute resources associated with this inode. Traverse
 791 + * all entries and unlink any xattr inodes associated with this inode. This
 792   * is called immediately before an inode is freed. We have exclusive
 793 - * access to the inode.
 794 + * access to the inode. If an orphan inode is deleted it will also delete any
 795 + * xattr block and all xattr inodes. They are checked by ext4_xattr_inode_iget()
 796 + * to ensure they belong to the parent inode and were not deleted already.
 797   */
 798 -void
 799 -ext4_xattr_delete_inode(handle_t *handle, struct inode *inode)
 800 +int
 801 +ext4_xattr_delete_inode(handle_t *handle, struct inode *inode,
 802 +                       struct ext4_xattr_ino_array **lea_ino_array)
 803  {
 804         struct buffer_head *bh = NULL;
 805 +       struct ext4_xattr_ibody_header *header;
 806 +       struct ext4_inode *raw_inode;
 807 +       struct ext4_iloc iloc;
 808 +       struct ext4_xattr_entry *entry;
 809 +       int credits = 3, error = 0;
 810 +
 811 +       if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR))
 812 +               goto delete_external_ea;
 813 +
 814 +       error = ext4_get_inode_loc(inode, &iloc);
 815 +       if (error)
 816 +               goto cleanup;
 817 +       raw_inode = ext4_raw_inode(&iloc);
 818 +       header = IHDR(inode, raw_inode);
 819 +       entry = IFIRST(header);
 820 +       for (; !IS_LAST_ENTRY(entry); entry = EXT4_XATTR_NEXT(entry)) {
 821 +               if (entry->e_value_inum == 0)
 822 +                       continue;
 823 +               if (ext4_expand_ino_array(lea_ino_array,
 824 +                                         entry->e_value_inum) != 0) {
 825 +                       brelse(iloc.bh);
 826 +                       goto cleanup;
 827 +               }
 828 +               entry->e_value_inum = 0;
 829 +       }
 830 +       brelse(iloc.bh);
 831
 832 +delete_external_ea:
 833         if (!EXT4_I(inode)->i_file_acl)
 834                 goto cleanup;
 835         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
 836 @@ -1468,11 +1848,88 @@ ext4_xattr_delete_inode(handle_t *handle
 837                                  EXT4_I(inode)->i_file_acl);
 838                 goto cleanup;
 839         }
 840 +
 841 +       entry = BFIRST(bh);
 842 +       for (; !IS_LAST_ENTRY(entry); entry = EXT4_XATTR_NEXT(entry)) {
 843 +               if (entry->e_value_inum == 0)
 844 +                       continue;
 845 +               if (ext4_expand_ino_array(lea_ino_array,
 846 +                                         entry->e_value_inum) != 0)
 847 +                       goto cleanup;
 848 +               entry->e_value_inum = 0;
 849 +       }
 850 +
 851 +       /* add xattr inode to orphan list */
 852 +       if (*lea_ino_array != NULL) {
 853 +               struct inode *ea_inode = NULL;
 854 +               int idx = 0;
 855 +
 856 +               for (; idx < (*lea_ino_array)->xia_count; ++idx) {
 857 +                       if (!ext4_handle_has_enough_credits(handle, 3)) {
 858 +                               error = ext4_journal_extend(handle, 3);
 859 +                               if (error > 0)
 860 +                                       error = ext4_journal_restart(handle, 3);
 861 +                               if (error != 0) {
 862 +                                       ext4_warning(inode->i_sb,
 863 +                                               "couldn't extend journal "
 864 +                                               "(err %d)", error);
 865 +                                       goto cleanup;
 866 +                               }
 867 +                       }
 868 +                       ea_inode = ext4_xattr_inode_iget(inode,
 869 +                               (*lea_ino_array)->xia_inodes[idx], &error);
 870 +                       if (error)
 871 +                               continue;
 872 +                       ext4_orphan_add(handle, ea_inode);
 873 +                       /* the inode's i_count will be released by caller */
 874 +               }
 875 +       }
 876 +
 877 +       if (!IS_NOQUOTA(inode))
 878 +               credits += 2 * EXT4_QUOTA_DEL_BLOCKS(inode->i_sb);
 879 +
 880 +       if (!ext4_handle_has_enough_credits(handle, credits)) {
 881 +               error = ext4_journal_extend(handle, credits);
 882 +               if (error > 0)
 883 +                       error = ext4_journal_restart(handle, credits);
 884 +               if (error != 0) {
 885 +                       ext4_warning(inode->i_sb,
 886 +                               "couldn't extend journal (err %d)", error);
 887 +                       goto cleanup;
 888 +               }
 889 +       }
 890 +
 891         ext4_xattr_release_block(handle, inode, bh);
 892         EXT4_I(inode)->i_file_acl = 0;
 893
 894  cleanup:
 895         brelse(bh);
 896 +
 897 +       return error;
 898 +}
 899 +
 900 +void
 901 +ext4_xattr_inode_array_free(struct inode *inode,
 902 +                           struct ext4_xattr_ino_array *lea_ino_array)
 903 +{
 904 +       struct inode    *ea_inode = NULL;
 905 +       int             idx = 0;
 906 +       int             err;
 907 +
 908 +       if (lea_ino_array == NULL)
 909 +               return;
 910 +
 911 +       for (; idx < lea_ino_array->xia_count; ++idx) {
 912 +               ea_inode = ext4_xattr_inode_iget(inode,
 913 +                               lea_ino_array->xia_inodes[idx], &err);
 914 +               if (err)
 915 +                       continue;
 916 +               clear_nlink(ea_inode);
 917 +               iput(ea_inode);
 918 +               /* for inode's i_count get from ext4_xattr_delete_inode */
 919 +               iput(ea_inode);
 920 +       }
 921 +       kfree(lea_ino_array);
 922  }
 923
 924  /*
 925 @@ -1542,10 +1999,9 @@ ext4_xattr_cmp(struct ext4_xattr_header
 926                     entry1->e_name_index != entry2->e_name_index ||
 927                     entry1->e_name_len != entry2->e_name_len ||
 928                     entry1->e_value_size != entry2->e_value_size ||
 929 +                   entry1->e_value_inum != entry2->e_value_inum ||
 930                     memcmp(entry1->e_name, entry2->e_name, entry1->e_name_len))
 931                         return 1;
 932 -               if (entry1->e_value_block != 0 || entry2->e_value_block != 0)
 933 -                       return -EIO;
 934                 if (memcmp((char *)header1 + le16_to_cpu(entry1->e_value_offs),
 935                            (char *)header2 + le16_to_cpu(entry2->e_value_offs),
 936                            le32_to_cpu(entry1->e_value_size)))
 937 @@ -1629,7 +2085,7 @@ static inline void ext4_xattr_hash_entry
 938                        *name++;
 939         }
 940
 941 -       if (entry->e_value_block == 0 && entry->e_value_size != 0) {
 942 +       if (entry->e_value_inum == 0 && entry->e_value_size != 0) {
 943                 __le32 *value = (__le32 *)((char *)header +
 944                         le16_to_cpu(entry->e_value_offs));
 945                 for (n = (le32_to_cpu(entry->e_value_size) +
 946 Index: linux-3.10.0-123.13.2.el7.x86_64/fs/ext4/xattr.h
 947 ===================================================================
 948 --- linux-3.10.0-123.13.2.el7.x86_64.orig/fs/ext4/xattr.h
 949 +++ linux-3.10.0-123.13.2.el7.x86_64/fs/ext4/xattr.h
 950 @@ -42,7 +42,7 @@ struct ext4_xattr_entry {
 951         __u8    e_name_len;     /* length of name */
 952         __u8    e_name_index;   /* attribute name index */
 953         __le16  e_value_offs;   /* offset in disk block of value */
 954 -       __le32  e_value_block;  /* disk block attribute is stored on (n/i) */
 955 +       __le32  e_value_inum;   /* inode in which the value is stored */
 956         __le32  e_value_size;   /* size of attribute value */
 957         __le32  e_hash;         /* hash value of name and value */
 958         char    e_name[0];      /* attribute name */
 959 @@ -67,6 +67,15 @@ struct ext4_xattr_entry {
 960                 EXT4_I(inode)->i_extra_isize))
 961  #define IFIRST(hdr) ((struct ext4_xattr_entry *)((hdr)+1))
 962
 963 +#define i_xattr_inode_parent i_mtime.tv_sec
 964 +
 965 +/*
 966 + * The minimum size of EA value when you start storing it in an external inode
 967 + * size of block - size of header - size of 1 entry - 4 null bytes
 968 +*/
 969 +#define EXT4_XATTR_MIN_LARGE_EA_SIZE(b)                                        \
 970 +       ((b) - EXT4_XATTR_LEN(3) - sizeof(struct ext4_xattr_header) - 4)
 971 +
 972  #define BHDR(bh) ((struct ext4_xattr_header *)((bh)->b_data))
 973  #define ENTRY(ptr) ((struct ext4_xattr_entry *)(ptr))
 974  #define BFIRST(bh) ENTRY(BHDR(bh)+1)
 975 @@ -75,10 +84,11 @@ struct ext4_xattr_entry {
 976  #define EXT4_ZERO_XATTR_VALUE ((void *)-1)
 977
 978  struct ext4_xattr_info {
 979 -       int name_index;
 980         const char *name;
 981         const void *value;
 982         size_t value_len;
 983 +       int name_index;
 984 +       int in_inode;
 985  };
 986
 987  struct ext4_xattr_search {
 988 @@ -106,7 +116,13 @@ extern int ext4_xattr_get(struct inode *
 989  extern int ext4_xattr_set(struct inode *, int, const char *, const void *, size_t, int);
 990  extern int ext4_xattr_set_handle(handle_t *, struct inode *, int, const char *, const void *, size_t, int);
 991
 992 -extern void ext4_xattr_delete_inode(handle_t *, struct inode *);
 993 +extern struct inode *ext4_xattr_inode_iget(struct inode *parent, int ea_ino,
 994 +                                          int *err);
 995 +extern int ext4_xattr_inode_unlink(struct inode *inode, int ea_ino);
 996 +extern int ext4_xattr_delete_inode(handle_t *handle, struct inode *inode,
 997 +                                  struct ext4_xattr_ino_array **array);
 998 +extern void ext4_xattr_inode_array_free(struct inode *inode,
 999 +                                       struct ext4_xattr_ino_array *array);
1000  extern void ext4_xattr_put_super(struct super_block *);
1001
1002  extern int ext4_expand_extra_isize_ea(struct inode *inode, int new_extra_isize,
1003 Index: linux-3.10.0-123.13.2.el7.x86_64/fs/ext4/ialloc.c
1004 ===================================================================
1005 --- linux-3.10.0-123.13.2.el7.x86_64.orig/fs/ext4/ialloc.c
1006 +++ linux-3.10.0-123.13.2.el7.x86_64/fs/ext4/ialloc.c
1007 @@ -250,7 +250,6 @@ void ext4_free_inode(handle_t *handle, s
1008          * as writing the quota to disk may need the lock as well.
1009          */
1010         dquot_initialize(inode);
1011 -       ext4_xattr_delete_inode(handle, inode);
1012         dquot_free_inode(inode);
1013         dquot_drop(inode);
1014
1015 Index: linux-3.10.0-123.13.2.el7.x86_64/fs/ext4/inline.c
1016 ===================================================================
1017 --- linux-3.10.0-123.13.2.el7.x86_64.orig/fs/ext4/inline.c
1018 +++ linux-3.10.0-123.13.2.el7.x86_64/fs/ext4/inline.c
1019 @@ -59,7 +59,7 @@ static int get_max_inline_xattr_value_si
1020
1021         /* Compute min_offs. */
1022         for (; !IS_LAST_ENTRY(entry); entry = EXT4_XATTR_NEXT(entry)) {
1023 -               if (!entry->e_value_block && entry->e_value_size) {
1024 +               if (!entry->e_value_inum && entry->e_value_size) {
1025                         size_t offs = le16_to_cpu(entry->e_value_offs);
1026                         if (offs < min_offs)
1027                                 min_offs = offs;