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

   1 Index: linux-stage/fs/ext4/ext4.h
   2 ===================================================================
   3 --- linux-stage.orig/fs/ext4/ext4.h
   4 +++ linux-stage/fs/ext4/ext4.h
   5 @@ -1329,6 +1329,7 @@ EXT4_INODE_BIT_FNS(state, state_flags)
   6  #define EXT4_FEATURE_INCOMPAT_64BIT            0x0080
   7  #define EXT4_FEATURE_INCOMPAT_MMP               0x0100
   8  #define EXT4_FEATURE_INCOMPAT_FLEX_BG          0x0200
   9 +#define EXT4_FEATURE_INCOMPAT_EA_INODE         0x0400
  10  #define EXT4_FEATURE_INCOMPAT_DIRDATA          0x1000
  11
  12  #define EXT4_FEATURE_COMPAT_SUPP       EXT2_FEATURE_COMPAT_EXT_ATTR
  13 @@ -1338,6 +1339,7 @@ EXT4_INODE_BIT_FNS(state, state_flags)
  14                                          EXT4_FEATURE_INCOMPAT_EXTENTS| \
  15                                          EXT4_FEATURE_INCOMPAT_64BIT| \
  16                                          EXT4_FEATURE_INCOMPAT_FLEX_BG| \
  17 +                                        EXT4_FEATURE_INCOMPAT_EA_INODE| \
  18                                          EXT4_FEATURE_INCOMPAT_MMP| \
  19                                          EXT4_FEATURE_INCOMPAT_DIRDATA)
  20
  21 @@ -1706,6 +1714,10 @@ struct mmpd_data {
  22  # define ATTRIB_NORET  __attribute__((noreturn))
  23  # define NORET_AND     noreturn,
  24
  25 +struct ext4_xattr_ino_array {
  26 +       unsigned int xia_count;         /* # of used item in the array */
  27 +       unsigned int xia_inodes[0];
  28 +};
  29  /* bitmap.c */
  30  extern unsigned int ext4_count_free(struct buffer_head *, unsigned);
  31
  32 Index: linux-stage/fs/ext4/xattr.c
  33 ===================================================================
  34 --- linux-stage.orig/fs/ext4/xattr.c
  35 +++ linux-stage/fs/ext4/xattr.c
  36 @@ -168,19 +168,26 @@ ext4_xattr_check_block(struct buffer_hea
  37  }
  38
  39  static inline int
  40 -ext4_xattr_check_entry(struct ext4_xattr_entry *entry, size_t size)
  41 +ext4_xattr_check_entry(struct ext4_xattr_entry *entry, size_t size,
  42 +                      struct inode *inode)
  43  {
  44         size_t value_size = le32_to_cpu(entry->e_value_size);
  45
  46 -       if (entry->e_value_block != 0 || value_size > size ||
  47 +       if (!entry->e_value_inum &&
  48             le16_to_cpu(entry->e_value_offs) + value_size > size)
  49 +               return -EIO;
  50 +       if (entry->e_value_inum &&
  51 +           (le32_to_cpu(entry->e_value_inum) < EXT4_FIRST_INO(inode->i_sb) ||
  52 +            le32_to_cpu(entry->e_value_inum) >
  53 +            le32_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_inodes_count)))
  54                 return -EIO;
  55         return 0;
  56  }
  57
  58  static int
  59  ext4_xattr_find_entry(struct ext4_xattr_entry **pentry, int name_index,
  60 -                     const char *name, size_t size, int sorted)
  61 +                     const char *name, size_t size, int sorted,
  62 +                     struct inode *inode)
  63  {
  64         struct ext4_xattr_entry *entry;
  65         size_t name_len;
  66 @@ -200,11 +207,104 @@ ext4_xattr_find_entry(struct ext4_xattr_
  67                         break;
  68         }
  69         *pentry = entry;
  70 -       if (!cmp && ext4_xattr_check_entry(entry, size))
  71 +       if (!cmp && ext4_xattr_check_entry(entry, size, inode))
  72                         return -EIO;
  73         return cmp ? -ENODATA : 0;
  74  }
  75
  76 +/*
  77 + * Read the EA value from an inode.
  78 + */
  79 +static int
  80 +ext4_xattr_inode_read(struct inode *ea_inode, void *buf, size_t *size)
  81 +{
  82 +       unsigned long block = 0;
  83 +       struct buffer_head *bh = NULL;
  84 +       int err, blocksize;
  85 +       size_t csize, ret_size = 0;
  86 +
  87 +       if (*size == 0)
  88 +               return 0;
  89 +
  90 +       blocksize = ea_inode->i_sb->s_blocksize;
  91 +
  92 +       while (ret_size < *size) {
  93 +               csize = (*size - ret_size) > blocksize ? blocksize :
  94 +                                                       *size - ret_size;
  95 +               bh = ext4_bread(NULL, ea_inode, block, 0, &err);
  96 +               if (!bh) {
  97 +                       *size = ret_size;
  98 +                       return err;
  99 +               }
 100 +               memcpy(buf, bh->b_data, csize);
 101 +               brelse(bh);
 102 +
 103 +               buf += csize;
 104 +               block += 1;
 105 +               ret_size += csize;
 106 +       }
 107 +
 108 +       *size = ret_size;
 109 +
 110 +       return err;
 111 +}
 112 +
 113 +struct inode *ext4_xattr_inode_iget(struct inode *parent, unsigned long ea_ino, int *err)
 114 +{
 115 +       struct inode *ea_inode = NULL;
 116 +
 117 +       ea_inode = ext4_iget(parent->i_sb, ea_ino);
 118 +       if (IS_ERR(ea_inode) || is_bad_inode(ea_inode)) {
 119 +               int rc = IS_ERR(ea_inode) ? PTR_ERR(ea_inode) : 0;
 120 +               ext4_error(parent->i_sb, "error while reading EA inode %lu "
 121 +                          "/ %d %d", ea_ino, rc, is_bad_inode(ea_inode));
 122 +               *err = rc != 0 ? rc : -EIO;
 123 +               return NULL;
 124 +       }
 125 +
 126 +       if (EXT4_XATTR_INODE_GET_PARENT(ea_inode) != parent->i_ino ||
 127 +           ea_inode->i_generation != parent->i_generation) {
 128 +               ext4_error(parent->i_sb, "Backpointer from EA inode %lu "
 129 +                          "to parent invalid.", ea_ino);
 130 +               *err = -EINVAL;
 131 +               goto error;
 132 +       }
 133 +
 134 +       if (!(EXT4_I(ea_inode)->i_flags & EXT4_EA_INODE_FL)) {
 135 +               ext4_error(parent->i_sb, "EA inode %lu does not have "
 136 +                          "EXT4_EA_INODE_FL flag set.\n", ea_ino);
 137 +               *err = -EINVAL;
 138 +               goto error;
 139 +       }
 140 +
 141 +       *err = 0;
 142 +       return ea_inode;
 143 +
 144 +error:
 145 +       iput(ea_inode);
 146 +       return NULL;
 147 +}
 148 +
 149 +/*
 150 + * Read the value from the EA inode.
 151 + */
 152 +static int
 153 +ext4_xattr_inode_get(struct inode *inode, unsigned long ea_ino, void *buffer,
 154 +                    size_t *size)
 155 +{
 156 +       struct inode *ea_inode = NULL;
 157 +       int err;
 158 +
 159 +       ea_inode = ext4_xattr_inode_iget(inode, ea_ino, &err);
 160 +       if (err)
 161 +               return err;
 162 +
 163 +       err = ext4_xattr_inode_read(ea_inode, buffer, size);
 164 +       iput(ea_inode);
 165 +
 166 +       return err;
 167 +}
 168 +
 169  static int
 170  ext4_xattr_block_get(struct inode *inode, int name_index, const char *name,
 171                      void *buffer, size_t buffer_size)
 172 @@ -236,7 +335,8 @@ bad_block:
 173         }
 174         ext4_xattr_cache_insert(bh);
 175         entry = BFIRST(bh);
 176 -       error = ext4_xattr_find_entry(&entry, name_index, name, bh->b_size, 1);
 177 +       error = ext4_xattr_find_entry(&entry, name_index, name, bh->b_size, 1,
 178 +                                     inode);
 179         if (error == -EIO)
 180                 goto bad_block;
 181         if (error)
 182 @@ -246,8 +346,16 @@ bad_block:
 183                 error = -ERANGE;
 184                 if (size > buffer_size)
 185                         goto cleanup;
 186 -               memcpy(buffer, bh->b_data + le16_to_cpu(entry->e_value_offs),
 187 -                      size);
 188 +               if (entry->e_value_inum) {
 189 +                       error = ext4_xattr_inode_get(inode,
 190 +                                            le32_to_cpu(entry->e_value_inum),
 191 +                                            buffer, &size);
 192 +                       if (error)
 193 +                               goto cleanup;
 194 +               } else {
 195 +                       memcpy(buffer, bh->b_data +
 196 +                              le16_to_cpu(entry->e_value_offs), size);
 197 +               }
 198         }
 199         error = size;
 200
 201 @@ -281,7 +389,7 @@ ext4_xattr_ibody_get(struct inode *inode
 202         if (error)
 203                 goto cleanup;
 204         error = ext4_xattr_find_entry(&entry, name_index, name,
 205 -                                     end - (void *)entry, 0);
 206 +                                     end - (void *)entry, 0, inode);
 207         if (error)
 208                 goto cleanup;
 209         size = le32_to_cpu(entry->e_value_size);
 210 @@ -289,8 +397,16 @@ ext4_xattr_ibody_get(struct inode *inode
 211                 error = -ERANGE;
 212                 if (size > buffer_size)
 213                         goto cleanup;
 214 -               memcpy(buffer, (void *)IFIRST(header) +
 215 -                      le16_to_cpu(entry->e_value_offs), size);
 216 +               if (entry->e_value_inum) {
 217 +                       error = ext4_xattr_inode_get(inode,
 218 +                                            le32_to_cpu(entry->e_value_inum),
 219 +                                            buffer, &size);
 220 +                       if (error)
 221 +                               goto cleanup;
 222 +               } else {
 223 +                       memcpy(buffer, (void *)IFIRST(header) +
 224 +                              le16_to_cpu(entry->e_value_offs), size);
 225 +               }
 226         }
 227         error = size;
 228
 229 @@ -513,7 +629,7 @@ static size_t ext4_xattr_free_space(stru
 230  {
 231         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 232                 *total += EXT4_XATTR_LEN(last->e_name_len);
 233 -               if (!last->e_value_block && last->e_value_size) {
 234 +               if (!last->e_value_inum && last->e_value_size) {
 235                         size_t offs = le16_to_cpu(last->e_value_offs);
 236                         if (offs < *min_offs)
 237                                 *min_offs = offs;
 238 @@ -522,11 +638,159 @@ static size_t ext4_xattr_free_space(stru
 239         return (*min_offs - ((void *)last - base) - sizeof(__u32));
 240  }
 241
 242 +/*
 243 + * Write the value of the EA in an inode.
 244 + */
 245 +static int
 246 +ext4_xattr_inode_write(handle_t *handle, struct inode *ea_inode,
 247 +                      const void *buf, int bufsize)
 248 +{
 249 +       struct buffer_head *bh = NULL, dummy;
 250 +       unsigned long block = 0;
 251 +       unsigned blocksize = ea_inode->i_sb->s_blocksize;
 252 +       unsigned max_blocks = (bufsize + blocksize - 1) >> ea_inode->i_blkbits;
 253 +       int csize, wsize = 0;
 254 +       int ret = 0;
 255 +       int retries = 0;
 256 +
 257 +retry:
 258 +       while (ret >= 0 && ret < max_blocks) {
 259 +               block += ret;
 260 +               max_blocks -= ret;
 261 +
 262 +               ret = ext4_get_blocks(handle, ea_inode, block, max_blocks,
 263 +                                     &dummy, EXT4_GET_BLOCKS_CREATE);
 264 +               if (ret <= 0) {
 265 +                       ext4_mark_inode_dirty(handle, ea_inode);
 266 +                       if (ret == -ENOSPC &&
 267 +                           ext4_should_retry_alloc(ea_inode->i_sb, &retries)) {
 268 +                               ret = 0;
 269 +                               goto retry;
 270 +                       }
 271 +                       break;
 272 +               }
 273 +       }
 274 +
 275 +       if (ret < 0)
 276 +               return ret;
 277 +
 278 +       block = 0;
 279 +       while (wsize < bufsize) {
 280 +               if (bh != NULL)
 281 +                       brelse(bh);
 282 +               csize = (bufsize - wsize) > blocksize ? blocksize :
 283 +                                                               bufsize - wsize;
 284 +               bh = ext4_getblk(handle, ea_inode, block, 0, &ret);
 285 +               if (!bh)
 286 +                       goto out;
 287 +               ret = ext4_journal_get_write_access(handle, bh);
 288 +               if (ret)
 289 +                       goto out;
 290 +
 291 +               memcpy(bh->b_data, buf, csize);
 292 +               set_buffer_uptodate(bh);
 293 +               ext4_handle_dirty_metadata(handle, ea_inode, bh);
 294 +
 295 +               buf += csize;
 296 +               wsize += csize;
 297 +               block += 1;
 298 +       }
 299 +
 300 +       i_size_write(ea_inode, wsize);
 301 +       ext4_update_i_disksize(ea_inode, wsize);
 302 +
 303 +       ext4_mark_inode_dirty(handle, ea_inode);
 304 +
 305 +out:
 306 +       brelse(bh);
 307 +
 308 +       return ret;
 309 +}
 310 +
 311 +/*
 312 + * Create an inode to store the value of a large EA.
 313 + */
 314 +static struct inode *
 315 +ext4_xattr_inode_create(handle_t *handle, struct inode *inode)
 316 +{
 317 +       struct inode *ea_inode = NULL;
 318 +
 319 +       /*
 320 +        * Let the next inode be the goal, so we try and allocate the EA inode
 321 +        * in the same group, or nearby one.
 322 +        */
 323 +       ea_inode = ext4_new_inode(handle, inode->i_sb->s_root->d_inode,
 324 +                                 S_IFREG|0600, NULL, inode->i_ino + 1);
 325 +
 326 +       if (!IS_ERR(ea_inode)) {
 327 +               ea_inode->i_op = &ext4_file_inode_operations;
 328 +               ea_inode->i_fop = &ext4_file_operations;
 329 +               ext4_set_aops(ea_inode);
 330 +               ea_inode->i_generation = inode->i_generation;
 331 +               EXT4_I(ea_inode)->i_flags |= EXT4_EA_INODE_FL;
 332 +
 333 +               /*
 334 +                * A back-pointer from EA inode to parent inode will be useful
 335 +                * for e2fsck.
 336 +                */
 337 +               EXT4_XATTR_INODE_SET_PARENT(ea_inode, inode->i_ino);
 338 +               unlock_new_inode(ea_inode);
 339 +       }
 340 +
 341 +       return ea_inode;
 342 +}
 343 +
 344 +/*
 345 + * Unlink the inode storing the value of the EA.
 346 + */
 347 +int
 348 +ext4_xattr_inode_unlink(struct inode *inode, unsigned long ea_ino)
 349 +{
 350 +       struct inode *ea_inode = NULL;
 351 +       int err;
 352 +
 353 +       ea_inode = ext4_xattr_inode_iget(inode, ea_ino, &err);
 354 +       if (err)
 355 +               return err;
 356 +
 357 +       ea_inode->i_nlink = 0;
 358 +       iput(ea_inode);
 359 +
 360 +       return 0;
 361 +}
 362 +
 363 +/*
 364 + * Add value of the EA in an inode.
 365 + */
 366 +static int
 367 +ext4_xattr_inode_set(handle_t *handle, struct inode *inode, unsigned long *ea_ino,
 368 +                    const void *value, size_t value_len)
 369 +{
 370 +       struct inode *ea_inode = NULL;
 371 +       int err;
 372 +
 373 +       /* Create an inode for the EA value */
 374 +       ea_inode = ext4_xattr_inode_create(handle, inode);
 375 +       if (IS_ERR(ea_inode))
 376 +               return -1;
 377 +
 378 +       err = ext4_xattr_inode_write(handle, ea_inode, value, value_len);
 379 +       if (err)
 380 +               ea_inode->i_nlink = 0;
 381 +       else
 382 +               *ea_ino = ea_inode->i_ino;
 383 +
 384 +       iput(ea_inode);
 385 +
 386 +       return err;
 387 +}
 388 +
 389  struct ext4_xattr_info {
 390 -       int name_index;
 391         const char *name;
 392         const void *value;
 393         size_t value_len;
 394 +       int name_index;
 395 +       int in_inode;
 396  };
 397
 398  struct ext4_xattr_search {
 399 @@ -538,15 +802,23 @@ struct ext4_xattr_search {
 400  };
 401
 402  static int
 403 -ext4_xattr_set_entry(struct ext4_xattr_info *i, struct ext4_xattr_search *s)
 404 +ext4_xattr_set_entry(struct ext4_xattr_info *i, struct ext4_xattr_search *s,
 405 +                    handle_t *handle, struct inode *inode)
 406  {
 407         struct ext4_xattr_entry *last;
 408         size_t free, min_offs = s->end - s->base, name_len = strlen(i->name);
 409 +       int in_inode = i->in_inode;
 410 +
 411 +       if (EXT4_HAS_INCOMPAT_FEATURE(inode->i_sb,
 412 +                EXT4_FEATURE_INCOMPAT_EA_INODE) &&
 413 +           (EXT4_XATTR_SIZE(i->value_len) >
 414 +            EXT4_XATTR_MIN_LARGE_EA_SIZE(inode->i_sb->s_blocksize)))
 415 +               in_inode = 1;
 416
 417         /* Compute min_offs and last. */
 418         last = s->first;
 419         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 420 -               if (!last->e_value_block && last->e_value_size) {
 421 +               if (!last->e_value_inum && last->e_value_size) {
 422                         size_t offs = le16_to_cpu(last->e_value_offs);
 423                         if (offs < min_offs)
 424                                 min_offs = offs;
 425 @@ -554,16 +826,21 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 426         }
 427         free = min_offs - ((void *)last - s->base) - sizeof(__u32);
 428         if (!s->not_found) {
 429 -               if (!s->here->e_value_block && s->here->e_value_size) {
 430 +               if (!in_inode &&
 431 +                   !s->here->e_value_inum && s->here->e_value_size) {
 432                         size_t size = le32_to_cpu(s->here->e_value_size);
 433                         free += EXT4_XATTR_SIZE(size);
 434                 }
 435                 free += EXT4_XATTR_LEN(name_len);
 436         }
 437         if (i->value) {
 438 -               if (free < EXT4_XATTR_SIZE(i->value_len) ||
 439 -                   free < EXT4_XATTR_LEN(name_len) +
 440 -                          EXT4_XATTR_SIZE(i->value_len))
 441 +               size_t value_len = EXT4_XATTR_SIZE(i->value_len);
 442 +
 443 +               if (in_inode)
 444 +                       value_len = 0;
 445 +
 446 +               if (free < value_len ||
 447 +                   free < EXT4_XATTR_LEN(name_len) + value_len)
 448                         return -ENOSPC;
 449         }
 450
 451 @@ -577,7 +854,8 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 452                 s->here->e_name_len = name_len;
 453                 memcpy(s->here->e_name, i->name, name_len);
 454         } else {
 455 -               if (!s->here->e_value_block && s->here->e_value_size) {
 456 +               if (!s->here->e_value_inum && s->here->e_value_size &&
 457 +                   s->here->e_value_offs > 0) {
 458                         void *first_val = s->base + min_offs;
 459                         size_t offs = le16_to_cpu(s->here->e_value_offs);
 460                         void *val = s->base + offs;
 461 @@ -606,13 +884,18 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 462                         last = s->first;
 463                         while (!IS_LAST_ENTRY(last)) {
 464                                 size_t o = le16_to_cpu(last->e_value_offs);
 465 -                               if (!last->e_value_block &&
 466 +                               if (!last->e_value_inum &&
 467                                     last->e_value_size && o < offs)
 468                                         last->e_value_offs =
 469                                                 cpu_to_le16(o + size);
 470                                 last = EXT4_XATTR_NEXT(last);
 471                         }
 472                 }
 473 +               if (s->here->e_value_inum) {
 474 +                       ext4_xattr_inode_unlink(inode,
 475 +                                       le32_to_cpu(s->here->e_value_inum));
 476 +                       s->here->e_value_inum = 0;
 477 +               }
 478                 if (!i->value) {
 479                         /* Remove the old name. */
 480                         size_t size = EXT4_XATTR_LEN(name_len);
 481 @@ -626,10 +908,17 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 482         if (i->value) {
 483                 /* Insert the new value. */
 484                 s->here->e_value_size = cpu_to_le32(i->value_len);
 485 -               if (i->value_len) {
 486 +               if (in_inode) {
 487 +                       unsigned long ea_ino = le32_to_cpu(s->here->e_value_inum);
 488 +                       ext4_xattr_inode_set(handle, inode, &ea_ino, i->value,
 489 +                                            i->value_len);
 490 +                       s->here->e_value_inum = cpu_to_le32(ea_ino);
 491 +                       s->here->e_value_offs = 0;
 492 +               } else if (i->value_len) {
 493                         size_t size = EXT4_XATTR_SIZE(i->value_len);
 494                         void *val = s->base + min_offs - size;
 495                         s->here->e_value_offs = cpu_to_le16(min_offs - size);
 496 +                       s->here->e_value_inum = 0;
 497                         memset(val + size - EXT4_XATTR_PAD, 0,
 498                                EXT4_XATTR_PAD); /* Clear the pad bytes. */
 499                         memcpy(val, i->value, i->value_len);
 500 @@ -674,7 +963,7 @@ ext4_xattr_block_find(struct inode *inod
 501                 bs->s.end = bs->bh->b_data + bs->bh->b_size;
 502                 bs->s.here = bs->s.first;
 503                 error = ext4_xattr_find_entry(&bs->s.here, i->name_index,
 504 -                                             i->name, bs->bh->b_size, 1);
 505 +                                            i->name, bs->bh->b_size, 1, inode);
 506                 if (error && error != -ENODATA)
 507                         goto cleanup;
 508                 bs->s.not_found = error;
 509 @@ -698,8 +987,6 @@ ext4_xattr_block_set(handle_t *handle, s
 510
 511  #define header(x) ((struct ext4_xattr_header *)(x))
 512
 513 -       if (i->value && i->value_len > sb->s_blocksize)
 514 -               return -ENOSPC;
 515         if (s->base) {
 516                 ce = mb_cache_entry_get(ext4_xattr_cache, bs->bh->b_bdev,
 517                                         bs->bh->b_blocknr);
 518 @@ -714,7 +1001,7 @@ ext4_xattr_block_set(handle_t *handle, s
 519                                 ce = NULL;
 520                         }
 521                         ea_bdebug(bs->bh, "modifying in-place");
 522 -                       error = ext4_xattr_set_entry(i, s);
 523 +                       error = ext4_xattr_set_entry(i, s, handle, inode);
 524                         if (!error) {
 525                                 if (!IS_LAST_ENTRY(s->first))
 526                                         ext4_xattr_rehash(header(s->base),
 527 @@ -766,7 +1053,7 @@ ext4_xattr_block_set(handle_t *handle, s
 528                 s->end = s->base + sb->s_blocksize;
 529         }
 530
 531 -       error = ext4_xattr_set_entry(i, s);
 532 +       error = ext4_xattr_set_entry(i, s, handle, inode);
 533         if (error == -EIO)
 534                 goto bad_block;
 535         if (error)
 536 @@ -917,7 +1204,7 @@ ext4_xattr_ibody_find(struct inode *inod
 537                 /* Find the named attribute. */
 538                 error = ext4_xattr_find_entry(&is->s.here, i->name_index,
 539                                               i->name, is->s.end -
 540 -                                             (void *)is->s.base, 0);
 541 +                                             (void *)is->s.base, 0, inode);
 542                 if (error && error != -ENODATA)
 543                         return error;
 544                 is->s.not_found = error;
 545 @@ -936,7 +1223,7 @@ ext4_xattr_ibody_set(handle_t *handle, s
 546
 547         if (EXT4_I(inode)->i_extra_isize == 0)
 548                 return -ENOSPC;
 549 -       error = ext4_xattr_set_entry(i, s);
 550 +       error = ext4_xattr_set_entry(i, s, handle, inode);
 551         if (error)
 552                 return error;
 553         header = IHDR(inode, ext4_raw_inode(&is->iloc));
 554 @@ -972,7 +1259,7 @@ ext4_xattr_set_handle(handle_t *handle,
 555                 .name = name,
 556                 .value = value,
 557                 .value_len = value_len,
 558 -
 559 +               .in_inode = 0,
 560         };
 561         struct ext4_xattr_ibody_find is = {
 562                 .s = { .not_found = -ENODATA, },
 563 @@ -1041,6 +1328,15 @@ ext4_xattr_set_handle(handle_t *handle,
 564                                         goto cleanup;
 565                         }
 566                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
 567 +                       if (EXT4_HAS_INCOMPAT_FEATURE(inode->i_sb,
 568 +                                       EXT4_FEATURE_INCOMPAT_EA_INODE) &&
 569 +                           error == -ENOSPC) {
 570 +                               /* xattr not fit to block, store at external
 571 +                                * inode */
 572 +                               i.in_inode = 1;
 573 +                               error = ext4_xattr_ibody_set(handle, inode,
 574 +                                                            &i, &is);
 575 +                       }
 576                         if (error)
 577                                 goto cleanup;
 578                         if (!is.s.not_found) {
 579 @@ -1087,10 +1383,25 @@ ext4_xattr_set(struct inode *inode, int
 580                const void *value, size_t value_len, int flags)
 581  {
 582         handle_t *handle;
 583 +       struct super_block *sb = inode->i_sb;
 584 +       int buffer_credits;
 585         int error, retries = 0;
 586
 587 +       buffer_credits = EXT4_DATA_TRANS_BLOCKS(sb);
 588 +       if ((value_len >= EXT4_XATTR_MIN_LARGE_EA_SIZE(sb->s_blocksize)) &&
 589 +           EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EA_INODE)) {
 590 +               int nrblocks = (value_len + sb->s_blocksize - 1) >>
 591 +                                       sb->s_blocksize_bits;
 592 +
 593 +               /* For new inode */
 594 +               buffer_credits += EXT4_SINGLEDATA_TRANS_BLOCKS(sb) + 3;
 595 +
 596 +               /* For data blocks of EA inode */
 597 +               buffer_credits += ext4_meta_trans_blocks(inode, nrblocks, 0);
 598 +       }
 599 +
 600  retry:
 601 -       handle = ext4_journal_start(inode, EXT4_DATA_TRANS_BLOCKS(inode->i_sb));
 602 +       handle = ext4_journal_start(inode, buffer_credits);
 603         if (IS_ERR(handle)) {
 604                 error = PTR_ERR(handle);
 605         } else {
 606 @@ -1100,7 +1411,7 @@ retry:
 607                                               value, value_len, flags);
 608                 error2 = ext4_journal_stop(handle);
 609                 if (error == -ENOSPC &&
 610 -                   ext4_should_retry_alloc(inode->i_sb, &retries))
 611 +                   ext4_should_retry_alloc(sb, &retries))
 612                         goto retry;
 613                 if (error == 0)
 614                         error = error2;
 615 @@ -1122,7 +1433,7 @@ static void ext4_xattr_shift_entries(str
 616
 617         /* Adjust the value offsets of the entries */
 618         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 619 -               if (!last->e_value_block && last->e_value_size) {
 620 +               if (!last->e_value_inum && last->e_value_size) {
 621                         new_offs = le16_to_cpu(last->e_value_offs) +
 622                                                         value_offs_shift;
 623                         BUG_ON(new_offs + le32_to_cpu(last->e_value_size)
 624 @@ -1355,22 +1666,135 @@ cleanup:
 625         return error;
 626  }
 627
 628 +#define EIA_INCR 16 /* must be 2^n */
 629 +#define EIA_MASK (EIA_INCR - 1)
 630 +/* Add the large xattr @ino into @lea_ino_array for later deletion.
 631 + * If @lea_ino_array is new or full it will be grown and the old
 632 + * contents copied over.
 633 + */
 634 +static int
 635 +ext4_expand_ino_array(struct ext4_xattr_ino_array **lea_ino_array, __u32 ino)
 636 +{
 637 +       if (*lea_ino_array == NULL) {
 638 +               /*
 639 +                * Start with 15 inodes, so it fits into a power-of-two size.
 640 +                * If *lea_ino_array is NULL, this is essentially offsetof()
 641 +                */
 642 +               (*lea_ino_array) =
 643 +                       kmalloc(offsetof(struct ext4_xattr_ino_array,
 644 +                                        xia_inodes[EIA_MASK]),
 645 +                               GFP_NOFS);
 646 +               if (*lea_ino_array == NULL)
 647 +                       return -ENOMEM;
 648 +               (*lea_ino_array)->xia_count = 0;
 649 +       } else if (((*lea_ino_array)->xia_count & EIA_MASK) == EIA_MASK) {
 650 +               /* expand the array once all 15 + n * 16 slots are full */
 651 +               struct ext4_xattr_ino_array *new_array = NULL;
 652 +               int count = (*lea_ino_array)->xia_count;
 653 +
 654 +               /* if new_array is NULL, this is essentially offsetof() */
 655 +               new_array = kmalloc(
 656 +                               offsetof(struct ext4_xattr_ino_array,
 657 +                                        xia_inodes[count + EIA_INCR]),
 658 +                               GFP_NOFS);
 659 +               if (new_array == NULL)
 660 +                       return -ENOMEM;
 661 +               memcpy(new_array, *lea_ino_array,
 662 +                      offsetof(struct ext4_xattr_ino_array,
 663 +                               xia_inodes[count]));
 664 +               kfree(*lea_ino_array);
 665 +               *lea_ino_array = new_array;
 666 +       }
 667 +       (*lea_ino_array)->xia_inodes[(*lea_ino_array)->xia_count++] = ino;
 668 +       return 0;
 669 +}
 670
 671 +/**
 672 + * Add xattr inode to orphan list
 673 + */
 674 +static int
 675 +ext4_xattr_inode_orphan_add(handle_t *handle, struct inode *inode,
 676 +                       int credits, struct ext4_xattr_ino_array *lea_ino_array)
 677 +{
 678 +       struct inode *ea_inode = NULL;
 679 +       int idx = 0, error = 0;
 680 +
 681 +       if (lea_ino_array == NULL)
 682 +               return 0;
 683 +
 684 +       for (; idx < lea_ino_array->xia_count; ++idx) {
 685 +               if (!ext4_handle_has_enough_credits(handle, credits)) {
 686 +                       error = ext4_journal_extend(handle, credits);
 687 +                       if (error > 0)
 688 +                               error = ext4_journal_restart(handle, credits);
 689 +
 690 +                       if (error != 0) {
 691 +                               ext4_warning(inode->i_sb,
 692 +                                       "couldn't extend journal "
 693 +                                       "(err %d)", error);
 694 +                               return error;
 695 +                       }
 696 +               }
 697 +               ea_inode = ext4_xattr_inode_iget(inode,
 698 +                               lea_ino_array->xia_inodes[idx], &error);
 699 +               if (error)
 700 +                       continue;
 701 +               ext4_orphan_add(handle, ea_inode);
 702 +               /* the inode's i_count will be released by caller */
 703 +       }
 704 +
 705 +       return 0;
 706 +}
 707
 708  /*
 709   * ext4_xattr_delete_inode()
 710   *
 711 - * Free extended attribute resources associated with this inode. This
 712 + * Free extended attribute resources associated with this inode. Traverse
 713 + * all entries and unlink any xattr inodes associated with this inode. This
 714   * is called immediately before an inode is freed. We have exclusive
 715 - * access to the inode.
 716 + * access to the inode. If an orphan inode is deleted it will also delete any
 717 + * xattr block and all xattr inodes. They are checked by ext4_xattr_inode_iget()
 718 + * to ensure they belong to the parent inode and were not deleted already.
 719   */
 720 -void
 721 -ext4_xattr_delete_inode(handle_t *handle, struct inode *inode)
 722 +int
 723 +ext4_xattr_delete_inode(handle_t *handle, struct inode *inode,
 724 +                       struct ext4_xattr_ino_array **lea_ino_array)
 725  {
 726         struct buffer_head *bh = NULL;
 727 +       struct ext4_xattr_ibody_header *header;
 728 +       struct ext4_inode *raw_inode;
 729 +       struct ext4_iloc iloc;
 730 +       struct ext4_xattr_entry *entry;
 731 +       int credits = 3, error = 0;
 732
 733 -       if (!EXT4_I(inode)->i_file_acl)
 734 +       if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR))
 735 +               goto delete_external_ea;
 736 +
 737 +       error = ext4_get_inode_loc(inode, &iloc);
 738 +       if (error)
 739 +               goto cleanup;
 740 +       raw_inode = ext4_raw_inode(&iloc);
 741 +       header = IHDR(inode, raw_inode);
 742 +       for (entry = IFIRST(header); !IS_LAST_ENTRY(entry);
 743 +            entry = EXT4_XATTR_NEXT(entry)) {
 744 +               if (!entry->e_value_inum)
 745 +                       continue;
 746 +               if (ext4_expand_ino_array(lea_ino_array,
 747 +                                         entry->e_value_inum) != 0) {
 748 +                       brelse(iloc.bh);
 749 +                       goto cleanup;
 750 +               }
 751 +               entry->e_value_inum = 0;
 752 +       }
 753 +       brelse(iloc.bh);
 754 +
 755 +delete_external_ea:
 756 +       if (!EXT4_I(inode)->i_file_acl) {
 757 +               /* add xattr inode to orphan list */
 758 +               ext4_xattr_inode_orphan_add(handle, inode, credits,
 759 +                                               *lea_ino_array);
 760                 goto cleanup;
 761 +       }
 762         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
 763         if (!bh) {
 764                 ext4_error(inode->i_sb, "inode %lu: block %llu read error",
 765 @@ -1383,11 +1807,71 @@ ext4_xattr_delete_inode(handle_t *handle
 766                            inode->i_ino, EXT4_I(inode)->i_file_acl);
 767                 goto cleanup;
 768         }
 769 +
 770 +       for (entry = BFIRST(bh); !IS_LAST_ENTRY(entry);
 771 +            entry = EXT4_XATTR_NEXT(entry)) {
 772 +               if (!entry->e_value_inum)
 773 +                       continue;
 774 +               if (ext4_expand_ino_array(lea_ino_array,
 775 +                                         entry->e_value_inum) != 0)
 776 +                       goto cleanup;
 777 +               entry->e_value_inum = 0;
 778 +       }
 779 +
 780 +       /* add xattr inode to orphan list */
 781 +       error = ext4_xattr_inode_orphan_add(handle, inode, credits,
 782 +                                       *lea_ino_array);
 783 +       if (error != 0)
 784 +               goto cleanup;
 785 +
 786 +       if (!IS_NOQUOTA(inode))
 787 +               credits += 2 * EXT4_QUOTA_DEL_BLOCKS(inode->i_sb);
 788 +
 789 +       if (!ext4_handle_has_enough_credits(handle, credits)) {
 790 +               error = ext4_journal_extend(handle, credits);
 791 +               if (error > 0)
 792 +                       error = ext4_journal_restart(handle, credits);
 793 +               if (error != 0) {
 794 +                       ext4_warning(inode->i_sb,
 795 +                               "couldn't extend journal (err %d)", error);
 796 +                       goto cleanup;
 797 +               }
 798 +       }
 799 +
 800         ext4_xattr_release_block(handle, inode, bh);
 801         EXT4_I(inode)->i_file_acl = 0;
 802
 803  cleanup:
 804         brelse(bh);
 805 +
 806 +       return error;
 807 +}
 808 +
 809 +void
 810 +ext4_xattr_inode_array_free(struct inode *inode,
 811 +                           struct ext4_xattr_ino_array *lea_ino_array)
 812 +{
 813 +       struct inode    *ea_inode = NULL;
 814 +       int             idx = 0;
 815 +       int             err;
 816 +
 817 +       if (lea_ino_array == NULL)
 818 +               return;
 819 +
 820 +       for (; idx < lea_ino_array->xia_count; ++idx) {
 821 +               ea_inode = ext4_xattr_inode_iget(inode,
 822 +                               lea_ino_array->xia_inodes[idx], &err);
 823 +               if (err)
 824 +                       continue;
 825 +
 826 +               /* for inode's i_count get from ext4_xattr_delete_inode */
 827 +               if (!list_empty(&EXT4_I(ea_inode)->i_orphan))
 828 +                       iput(ea_inode);
 829 +
 830 +               ea_inode->i_nlink = 0;
 831 +               iput(ea_inode);
 832 +       }
 833 +       kfree(lea_ino_array);
 834  }
 835
 836  /*
 837 @@ -1457,10 +1941,9 @@ ext4_xattr_cmp(struct ext4_xattr_header
 838                     entry1->e_name_index != entry2->e_name_index ||
 839                     entry1->e_name_len != entry2->e_name_len ||
 840                     entry1->e_value_size != entry2->e_value_size ||
 841 +                   entry1->e_value_inum != entry2->e_value_inum ||
 842                     memcmp(entry1->e_name, entry2->e_name, entry1->e_name_len))
 843                         return 1;
 844 -               if (entry1->e_value_block != 0 || entry2->e_value_block != 0)
 845 -                       return -EIO;
 846                 if (memcmp((char *)header1 + le16_to_cpu(entry1->e_value_offs),
 847                            (char *)header2 + le16_to_cpu(entry2->e_value_offs),
 848                            le32_to_cpu(entry1->e_value_size)))
 849 @@ -1545,7 +2028,7 @@ static inline void ext4_xattr_hash_entry
 850                        *name++;
 851         }
 852
 853 -       if (entry->e_value_block == 0 && entry->e_value_size != 0) {
 854 +       if (!entry->e_value_inum && entry->e_value_size) {
 855                 __le32 *value = (__le32 *)((char *)header +
 856                         le16_to_cpu(entry->e_value_offs));
 857                 for (n = (le32_to_cpu(entry->e_value_size) +
 858 Index: linux-stage/fs/ext4/xattr.h
 859 ===================================================================
 860 --- linux-stage.orig/fs/ext4/xattr.h
 861 +++ linux-stage/fs/ext4/xattr.h
 862 @@ -38,7 +38,7 @@ struct ext4_xattr_entry {
 863         __u8    e_name_len;     /* length of name */
 864         __u8    e_name_index;   /* attribute name index */
 865         __le16  e_value_offs;   /* offset in disk block of value */
 866 -       __le32  e_value_block;  /* disk block attribute is stored on (n/i) */
 867 +       __le32  e_value_inum;   /* inode in which the value is stored */
 868         __le32  e_value_size;   /* size of attribute value */
 869         __le32  e_hash;         /* hash value of name and value */
 870         char    e_name[0];      /* attribute name */
 871 @@ -63,6 +63,26 @@ struct ext4_xattr_entry {
 872                 EXT4_I(inode)->i_extra_isize))
 873  #define IFIRST(hdr) ((struct ext4_xattr_entry *)((hdr)+1))
 874
 875 +/*
 876 + * Link EA inode back to parent one using i_mtime field.
 877 + * Extra integer type conversion added to ignore higher
 878 + * bits in i_mtime.tv_sec which might be set by ext4_get()
 879 + */
 880 +#define EXT4_XATTR_INODE_SET_PARENT(inode, inum)      \
 881 +do {                                                  \
 882 +      (inode)->i_mtime.tv_sec = inum;                 \
 883 +} while(0)
 884 +
 885 +#define EXT4_XATTR_INODE_GET_PARENT(inode)            \
 886 +((__u32)(inode)->i_mtime.tv_sec)
 887 +
 888 +/*
 889 + * The minimum size of EA value when you start storing it in an external inode
 890 + * size of block - size of header - size of 1 entry - 4 null bytes
 891 +*/
 892 +#define EXT4_XATTR_MIN_LARGE_EA_SIZE(b)                                        \
 893 +       ((b) - EXT4_XATTR_LEN(3) - sizeof(struct ext4_xattr_header) - 4)
 894 +
 895  # ifdef CONFIG_EXT4_FS_XATTR
 896
 897  extern struct xattr_handler ext4_xattr_user_handler;
 898 @@ -77,7 +86,13 @@ extern int ext4_xattr_get(struct inode *
 899  extern int ext4_xattr_set(struct inode *, int, const char *, const void *, size_t, int);
 900  extern int ext4_xattr_set_handle(handle_t *, struct inode *, int, const char *, const void *, size_t, int);
 901
 902 -extern void ext4_xattr_delete_inode(handle_t *, struct inode *);
 903 +extern struct inode *ext4_xattr_inode_iget(struct inode *parent, unsigned long ea_ino,
 904 +                                          int *err);
 905 +extern int ext4_xattr_inode_unlink(struct inode *inode, unsigned long ea_ino);
 906 +extern int ext4_xattr_delete_inode(handle_t *handle, struct inode *inode,
 907 +                                  struct ext4_xattr_ino_array **array);
 908 +extern void ext4_xattr_inode_array_free(struct inode *inode,
 909 +                                       struct ext4_xattr_ino_array *array);
 910  extern void ext4_xattr_put_super(struct super_block *);
 911
 912  extern int ext4_expand_extra_isize_ea(struct inode *inode, int new_extra_isize,
 913 @@ -111,9 +126,11 @@ ext4_xattr_set_handle(handle_t *handle,
 914         return -EOPNOTSUPP;
 915  }
 916
 917 -static inline void
 918 -ext4_xattr_delete_inode(handle_t *handle, struct inode *inode)
 919 +inline int
 920 +ext4_xattr_delete_inode(handle_t *handle, struct inode *inode,
 921 +                       struct ext4_xattr_ino_array **array)
 922  {
 923 +       return -EOPNOTSUPP;
 924  }
 925
 926  static inline void
 927 Index: linux-stage/fs/ext4/inode.c
 928 ===================================================================
 929 --- linux-stage.orig/fs/ext4/inode.c
 930 +++ linux-stage/fs/ext4/inode.c
 931 @@ -222,6 +222,8 @@ void ext4_delete_inode(struct inode *ino
 932  {
 933         handle_t *handle;
 934         int err;
 935 +       int extra_credits = 3;
 936 +       struct ext4_xattr_ino_array *lea_ino_array = NULL;
 937
 938         if (ext4_should_order_data(inode))
 939                 ext4_begin_ordered_truncate(inode, 0);
 940 @@ -235,7 +237,8 @@ void ext4_delete_inode(struct inode *ino
 941          * protection against it
 942          */
 943         sb_start_intwrite(inode->i_sb);
 944 -       handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
 945 +
 946 +       handle = ext4_journal_start(inode, extra_credits);
 947         if (IS_ERR(handle)) {
 948                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
 949                 /*
 950 @@ -247,9 +250,36 @@ void ext4_delete_inode(struct inode *ino
 951                 sb_end_intwrite(inode->i_sb);
 952                 goto no_delete;
 953         }
 954 -
 955         if (IS_SYNC(inode))
 956                 ext4_handle_sync(handle);
 957 +
 958 +       /*
 959 +        * Delete xattr inode before deleting the main inode.
 960 +        */
 961 +       err = ext4_xattr_delete_inode(handle, inode, &lea_ino_array);
 962 +       if (err) {
 963 +               ext4_warning(inode->i_sb,
 964 +                            "couldn't delete inode's xattr (err %d)", err);
 965 +               goto stop_handle;
 966 +       }
 967 +
 968 +       if (!IS_NOQUOTA(inode))
 969 +               extra_credits += 2 * EXT4_QUOTA_DEL_BLOCKS(inode->i_sb);
 970 +
 971 +       if (!ext4_handle_has_enough_credits(handle,
 972 +                               blocks_for_truncate(inode) + extra_credits)) {
 973 +               err = ext4_journal_extend(handle,
 974 +                               blocks_for_truncate(inode) + extra_credits);
 975 +               if (err > 0)
 976 +                       err = ext4_journal_restart(handle,
 977 +                               blocks_for_truncate(inode) + extra_credits);
 978 +               if (err != 0) {
 979 +                       ext4_warning(inode->i_sb,
 980 +                                    "couldn't extend journal (err %d)", err);
 981 +                       goto stop_handle;
 982 +               }
 983 +       }
 984 +
 985         inode->i_size = 0;
 986         err = ext4_mark_inode_dirty(handle, inode);
 987         if (err) {
 988 @@ -266,10 +296,10 @@ void ext4_delete_inode(struct inode *ino
 989          * enough credits left in the handle to remove the inode from
 990          * the orphan list and set the dtime field.
 991          */
 992 -       if (!ext4_handle_has_enough_credits(handle, 3)) {
 993 -               err = ext4_journal_extend(handle, 3);
 994 +       if (!ext4_handle_has_enough_credits(handle, extra_credits)) {
 995 +               err = ext4_journal_extend(handle, extra_credits);
 996                 if (err > 0)
 997 -                       err = ext4_journal_restart(handle, 3);
 998 +                       err = ext4_journal_restart(handle, extra_credits);
 999                 if (err != 0) {
1000                         ext4_warning(inode->i_sb,
1001                                      "couldn't extend journal (err %d)", err);
1002 @@ -303,8 +333,12 @@ void ext4_delete_inode(struct inode *ino
1003                 clear_inode(inode);
1004         else
1005                 ext4_free_inode(handle, inode);
1006 +
1007         ext4_journal_stop(handle);
1008         sb_end_intwrite(inode->i_sb);
1009 +
1010 +       if (lea_ino_array != NULL)
1011 +               ext4_xattr_inode_array_free(inode, lea_ino_array);
1012         return;
1013  no_delete:
1014         clear_inode(inode);     /* We must guarantee clearing of inode... */
1015 Index: linux-stage/fs/ext4/ialloc.c
1016 ===================================================================
1017 --- linux-stage.orig/fs/ext4/ialloc.c
1018 +++ linux-stage/fs/ext4/ialloc.c
1019 @@ -219,7 +219,6 @@ void ext4_free_inode(handle_t *handle, s
1020          * as writing the quota to disk may need the lock as well.
1021          */
1022         vfs_dq_init(inode);
1023 -       ext4_xattr_delete_inode(handle, inode);
1024         vfs_dq_free_inode(inode);
1025         vfs_dq_drop(inode);
1026