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 @@ -1695,6 +1697,12 @@ struct mmpd_data {
  22  #define EXT4_MMP_MAX_CHECK_INTERVAL    300UL
  23
  24  /*
  25 + * Maximum size of xattr attributes for FEATURE_INCOMPAT_EA_INODE 1Mb
  26 + * This limit is arbitrary, but is reasonable for the xattr API.
  27 + */
  28 +#define EXT4_XATTR_MAX_LARGE_EA_SIZE    (1024 * 1024)
  29 +
  30 +/*
  31   * Function prototypes
  32   */
  33
  34 @@ -1706,6 +1714,10 @@ struct mmpd_data {
  35  # define ATTRIB_NORET  __attribute__((noreturn))
  36  # define NORET_AND     noreturn,
  37
  38 +struct ext4_xattr_ino_array {
  39 +       unsigned int xia_count;         /* # of used item in the array */
  40 +       unsigned int xia_inodes[0];
  41 +};
  42  /* bitmap.c */
  43  extern unsigned int ext4_count_free(struct buffer_head *, unsigned);
  44
  45 Index: linux-stage/fs/ext4/xattr.c
  46 ===================================================================
  47 --- linux-stage.orig/fs/ext4/xattr.c
  48 +++ linux-stage/fs/ext4/xattr.c
  49 @@ -168,19 +168,26 @@ ext4_xattr_check_block(struct buffer_hea
  50  }
  51
  52  static inline int
  53 -ext4_xattr_check_entry(struct ext4_xattr_entry *entry, size_t size)
  54 +ext4_xattr_check_entry(struct ext4_xattr_entry *entry, size_t size,
  55 +                      struct inode *inode)
  56  {
  57         size_t value_size = le32_to_cpu(entry->e_value_size);
  58
  59 -       if (entry->e_value_block != 0 || value_size > size ||
  60 -           le16_to_cpu(entry->e_value_offs) + value_size > size)
  61 +       if ((entry->e_value_inum == 0) &&
  62 +          (le16_to_cpu(entry->e_value_offs) + value_size > size))
  63 +               return -EIO;
  64 +       if (entry->e_value_inum != 0 &&
  65 +           (le32_to_cpu(entry->e_value_inum) < EXT4_FIRST_INO(inode->i_sb) ||
  66 +            le32_to_cpu(entry->e_value_inum) >
  67 +            le32_to_cpu(EXT4_SB(inode->i_sb)->s_es->s_inodes_count)))
  68                 return -EIO;
  69         return 0;
  70  }
  71
  72  static int
  73  ext4_xattr_find_entry(struct ext4_xattr_entry **pentry, int name_index,
  74 -                     const char *name, size_t size, int sorted)
  75 +                     const char *name, size_t size, int sorted,
  76 +                     struct inode *inode)
  77  {
  78         struct ext4_xattr_entry *entry;
  79         size_t name_len;
  80 @@ -200,11 +207,104 @@ ext4_xattr_find_entry(struct ext4_xattr_
  81                         break;
  82         }
  83         *pentry = entry;
  84 -       if (!cmp && ext4_xattr_check_entry(entry, size))
  85 +       if (!cmp && ext4_xattr_check_entry(entry, size, inode))
  86                         return -EIO;
  87         return cmp ? -ENODATA : 0;
  88  }
  89
  90 +/*
  91 + * Read the EA value from an inode.
  92 + */
  93 +static int
  94 +ext4_xattr_inode_read(struct inode *ea_inode, void *buf, size_t *size)
  95 +{
  96 +       unsigned long block = 0;
  97 +       struct buffer_head *bh = NULL;
  98 +       int err, blocksize;
  99 +       size_t csize, ret_size = 0;
 100 +
 101 +       if (*size == 0)
 102 +               return 0;
 103 +
 104 +       blocksize = ea_inode->i_sb->s_blocksize;
 105 +
 106 +       while (ret_size < *size) {
 107 +               csize = (*size - ret_size) > blocksize ? blocksize :
 108 +                                                       *size - ret_size;
 109 +               bh = ext4_bread(NULL, ea_inode, block, 0, &err);
 110 +               if (!bh) {
 111 +                       *size = ret_size;
 112 +                       return err;
 113 +               }
 114 +               memcpy(buf, bh->b_data, csize);
 115 +               brelse(bh);
 116 +
 117 +               buf += csize;
 118 +               block += 1;
 119 +               ret_size += csize;
 120 +       }
 121 +
 122 +       *size = ret_size;
 123 +
 124 +       return err;
 125 +}
 126 +
 127 +struct inode *ext4_xattr_inode_iget(struct inode *parent, unsigned long ea_ino, int *err)
 128 +{
 129 +       struct inode *ea_inode = NULL;
 130 +
 131 +       ea_inode = ext4_iget(parent->i_sb, ea_ino);
 132 +       if (IS_ERR(ea_inode) || is_bad_inode(ea_inode)) {
 133 +               int rc = IS_ERR(ea_inode) ? PTR_ERR(ea_inode) : 0;
 134 +               ext4_error(parent->i_sb, "error while reading EA inode %lu "
 135 +                          "/ %d %d", ea_ino, rc, is_bad_inode(ea_inode));
 136 +               *err = rc != 0 ? rc : -EIO;
 137 +               return NULL;
 138 +       }
 139 +
 140 +       if (ea_inode->i_xattr_inode_parent != parent->i_ino ||
 141 +           ea_inode->i_generation != parent->i_generation) {
 142 +               ext4_error(parent->i_sb, "Backpointer from EA inode %lu "
 143 +                          "to parent invalid.", ea_ino);
 144 +               *err = -EINVAL;
 145 +               goto error;
 146 +       }
 147 +
 148 +       if (!(EXT4_I(ea_inode)->i_flags & EXT4_EA_INODE_FL)) {
 149 +               ext4_error(parent->i_sb, "EA inode %lu does not have "
 150 +                          "EXT4_EA_INODE_FL flag set.\n", ea_ino);
 151 +               *err = -EINVAL;
 152 +               goto error;
 153 +       }
 154 +
 155 +       *err = 0;
 156 +       return ea_inode;
 157 +
 158 +error:
 159 +       iput(ea_inode);
 160 +       return NULL;
 161 +}
 162 +
 163 +/*
 164 + * Read the value from the EA inode.
 165 + */
 166 +static int
 167 +ext4_xattr_inode_get(struct inode *inode, unsigned long ea_ino, void *buffer,
 168 +                    size_t *size)
 169 +{
 170 +       struct inode *ea_inode = NULL;
 171 +       int err;
 172 +
 173 +       ea_inode = ext4_xattr_inode_iget(inode, ea_ino, &err);
 174 +       if (err)
 175 +               return err;
 176 +
 177 +       err = ext4_xattr_inode_read(ea_inode, buffer, size);
 178 +       iput(ea_inode);
 179 +
 180 +       return err;
 181 +}
 182 +
 183  static int
 184  ext4_xattr_block_get(struct inode *inode, int name_index, const char *name,
 185                      void *buffer, size_t buffer_size)
 186 @@ -236,7 +335,8 @@ bad_block:
 187         }
 188         ext4_xattr_cache_insert(bh);
 189         entry = BFIRST(bh);
 190 -       error = ext4_xattr_find_entry(&entry, name_index, name, bh->b_size, 1);
 191 +       error = ext4_xattr_find_entry(&entry, name_index, name, bh->b_size, 1,
 192 +                                     inode);
 193         if (error == -EIO)
 194                 goto bad_block;
 195         if (error)
 196 @@ -246,8 +346,16 @@ bad_block:
 197                 error = -ERANGE;
 198                 if (size > buffer_size)
 199                         goto cleanup;
 200 -               memcpy(buffer, bh->b_data + le16_to_cpu(entry->e_value_offs),
 201 -                      size);
 202 +               if (entry->e_value_inum != 0) {
 203 +                       error = ext4_xattr_inode_get(inode,
 204 +                                            le32_to_cpu(entry->e_value_inum),
 205 +                                            buffer, &size);
 206 +                       if (error)
 207 +                               goto cleanup;
 208 +               } else {
 209 +                       memcpy(buffer, bh->b_data +
 210 +                              le16_to_cpu(entry->e_value_offs), size);
 211 +               }
 212         }
 213         error = size;
 214
 215 @@ -281,7 +389,7 @@ ext4_xattr_ibody_get(struct inode *inode
 216         if (error)
 217                 goto cleanup;
 218         error = ext4_xattr_find_entry(&entry, name_index, name,
 219 -                                     end - (void *)entry, 0);
 220 +                                     end - (void *)entry, 0, inode);
 221         if (error)
 222                 goto cleanup;
 223         size = le32_to_cpu(entry->e_value_size);
 224 @@ -289,8 +397,16 @@ ext4_xattr_ibody_get(struct inode *inode
 225                 error = -ERANGE;
 226                 if (size > buffer_size)
 227                         goto cleanup;
 228 -               memcpy(buffer, (void *)IFIRST(header) +
 229 -                      le16_to_cpu(entry->e_value_offs), size);
 230 +               if (entry->e_value_inum != 0) {
 231 +                       error = ext4_xattr_inode_get(inode,
 232 +                                            le32_to_cpu(entry->e_value_inum),
 233 +                                            buffer, &size);
 234 +                       if (error)
 235 +                               goto cleanup;
 236 +               } else {
 237 +                       memcpy(buffer, (void *)IFIRST(header) +
 238 +                              le16_to_cpu(entry->e_value_offs), size);
 239 +               }
 240         }
 241         error = size;
 242
 243 @@ -513,7 +629,7 @@ static size_t ext4_xattr_free_space(stru
 244  {
 245         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 246                 *total += EXT4_XATTR_LEN(last->e_name_len);
 247 -               if (!last->e_value_block && last->e_value_size) {
 248 +               if (last->e_value_inum == 0 && last->e_value_size > 0) {
 249                         size_t offs = le16_to_cpu(last->e_value_offs);
 250                         if (offs < *min_offs)
 251                                 *min_offs = offs;
 252 @@ -522,11 +638,159 @@ static size_t ext4_xattr_free_space(stru
 253         return (*min_offs - ((void *)last - base) - sizeof(__u32));
 254  }
 255
 256 +/*
 257 + * Write the value of the EA in an inode.
 258 + */
 259 +static int
 260 +ext4_xattr_inode_write(handle_t *handle, struct inode *ea_inode,
 261 +                      const void *buf, int bufsize)
 262 +{
 263 +       struct buffer_head *bh = NULL, dummy;
 264 +       unsigned long block = 0;
 265 +       unsigned blocksize = ea_inode->i_sb->s_blocksize;
 266 +       unsigned max_blocks = (bufsize + blocksize - 1) >> ea_inode->i_blkbits;
 267 +       int csize, wsize = 0;
 268 +       int ret = 0;
 269 +       int retries = 0;
 270 +
 271 +retry:
 272 +       while (ret >= 0 && ret < max_blocks) {
 273 +               block += ret;
 274 +               max_blocks -= ret;
 275 +
 276 +               ret = ext4_get_blocks(handle, ea_inode, block, max_blocks,
 277 +                                     &dummy, EXT4_GET_BLOCKS_CREATE);
 278 +               if (ret <= 0) {
 279 +                       ext4_mark_inode_dirty(handle, ea_inode);
 280 +                       if (ret == -ENOSPC &&
 281 +                           ext4_should_retry_alloc(ea_inode->i_sb, &retries)) {
 282 +                               ret = 0;
 283 +                               goto retry;
 284 +                       }
 285 +                       break;
 286 +               }
 287 +       }
 288 +
 289 +       if (ret < 0)
 290 +               return ret;
 291 +
 292 +       block = 0;
 293 +       while (wsize < bufsize) {
 294 +               if (bh != NULL)
 295 +                       brelse(bh);
 296 +               csize = (bufsize - wsize) > blocksize ? blocksize :
 297 +                                                               bufsize - wsize;
 298 +               bh = ext4_getblk(handle, ea_inode, block, 0, &ret);
 299 +               if (!bh)
 300 +                       goto out;
 301 +               ret = ext4_journal_get_write_access(handle, bh);
 302 +               if (ret)
 303 +                       goto out;
 304 +
 305 +               memcpy(bh->b_data, buf, csize);
 306 +               set_buffer_uptodate(bh);
 307 +               ext4_handle_dirty_metadata(handle, ea_inode, bh);
 308 +
 309 +               buf += csize;
 310 +               wsize += csize;
 311 +               block += 1;
 312 +       }
 313 +
 314 +       i_size_write(ea_inode, wsize);
 315 +       ext4_update_i_disksize(ea_inode, wsize);
 316 +
 317 +       ext4_mark_inode_dirty(handle, ea_inode);
 318 +
 319 +out:
 320 +       brelse(bh);
 321 +
 322 +       return ret;
 323 +}
 324 +
 325 +/*
 326 + * Create an inode to store the value of a large EA.
 327 + */
 328 +static struct inode *
 329 +ext4_xattr_inode_create(handle_t *handle, struct inode *inode)
 330 +{
 331 +       struct inode *ea_inode = NULL;
 332 +
 333 +       /*
 334 +        * Let the next inode be the goal, so we try and allocate the EA inode
 335 +        * in the same group, or nearby one.
 336 +        */
 337 +       ea_inode = ext4_new_inode(handle, inode->i_sb->s_root->d_inode,
 338 +                                 S_IFREG|0600, NULL, inode->i_ino + 1);
 339 +
 340 +       if (!IS_ERR(ea_inode)) {
 341 +               ea_inode->i_op = &ext4_file_inode_operations;
 342 +               ea_inode->i_fop = &ext4_file_operations;
 343 +               ext4_set_aops(ea_inode);
 344 +               ea_inode->i_generation = inode->i_generation;
 345 +               EXT4_I(ea_inode)->i_flags |= EXT4_EA_INODE_FL;
 346 +
 347 +               /*
 348 +                * A back-pointer from EA inode to parent inode will be useful
 349 +                * for e2fsck.
 350 +                */
 351 +               ea_inode->i_xattr_inode_parent = inode->i_ino;
 352 +               unlock_new_inode(ea_inode);
 353 +       }
 354 +
 355 +       return ea_inode;
 356 +}
 357 +
 358 +/*
 359 + * Unlink the inode storing the value of the EA.
 360 + */
 361 +int
 362 +ext4_xattr_inode_unlink(struct inode *inode, unsigned long ea_ino)
 363 +{
 364 +       struct inode *ea_inode = NULL;
 365 +       int err;
 366 +
 367 +       ea_inode = ext4_xattr_inode_iget(inode, ea_ino, &err);
 368 +       if (err)
 369 +               return err;
 370 +
 371 +       ea_inode->i_nlink = 0;
 372 +       iput(ea_inode);
 373 +
 374 +       return 0;
 375 +}
 376 +
 377 +/*
 378 + * Add value of the EA in an inode.
 379 + */
 380 +static int
 381 +ext4_xattr_inode_set(handle_t *handle, struct inode *inode, unsigned long *ea_ino,
 382 +                    const void *value, size_t value_len)
 383 +{
 384 +       struct inode *ea_inode = NULL;
 385 +       int err;
 386 +
 387 +       /* Create an inode for the EA value */
 388 +       ea_inode = ext4_xattr_inode_create(handle, inode);
 389 +       if (IS_ERR(ea_inode))
 390 +               return -1;
 391 +
 392 +       err = ext4_xattr_inode_write(handle, ea_inode, value, value_len);
 393 +       if (err)
 394 +               ea_inode->i_nlink = 0;
 395 +       else
 396 +               *ea_ino = ea_inode->i_ino;
 397 +
 398 +       iput(ea_inode);
 399 +
 400 +       return err;
 401 +}
 402 +
 403  struct ext4_xattr_info {
 404 -       int name_index;
 405         const char *name;
 406         const void *value;
 407         size_t value_len;
 408 +       int name_index;
 409 +       int in_inode;
 410  };
 411
 412  struct ext4_xattr_search {
 413 @@ -538,15 +802,23 @@ struct ext4_xattr_search {
 414  };
 415
 416  static int
 417 -ext4_xattr_set_entry(struct ext4_xattr_info *i, struct ext4_xattr_search *s)
 418 +ext4_xattr_set_entry(struct ext4_xattr_info *i, struct ext4_xattr_search *s,
 419 +                    handle_t *handle, struct inode *inode)
 420  {
 421         struct ext4_xattr_entry *last;
 422         size_t free, min_offs = s->end - s->base, name_len = strlen(i->name);
 423 +       int in_inode = i->in_inode;
 424 +
 425 +       if (EXT4_HAS_INCOMPAT_FEATURE(inode->i_sb,
 426 +                EXT4_FEATURE_INCOMPAT_EA_INODE) &&
 427 +           (EXT4_XATTR_SIZE(i->value_len) >
 428 +            EXT4_XATTR_MIN_LARGE_EA_SIZE(inode->i_sb->s_blocksize)))
 429 +               in_inode = 1;
 430
 431         /* Compute min_offs and last. */
 432         last = s->first;
 433         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 434 -               if (!last->e_value_block && last->e_value_size) {
 435 +               if (last->e_value_inum == 0 && last->e_value_size > 0) {
 436                         size_t offs = le16_to_cpu(last->e_value_offs);
 437                         if (offs < min_offs)
 438                                 min_offs = offs;
 439 @@ -554,16 +826,21 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 440         }
 441         free = min_offs - ((void *)last - s->base) - sizeof(__u32);
 442         if (!s->not_found) {
 443 -               if (!s->here->e_value_block && s->here->e_value_size) {
 444 +               if (!in_inode && s->here->e_value_inum == 0 &&
 445 +                   s->here->e_value_size > 0) {
 446                         size_t size = le32_to_cpu(s->here->e_value_size);
 447                         free += EXT4_XATTR_SIZE(size);
 448                 }
 449                 free += EXT4_XATTR_LEN(name_len);
 450         }
 451         if (i->value) {
 452 -               if (free < EXT4_XATTR_SIZE(i->value_len) ||
 453 -                   free < EXT4_XATTR_LEN(name_len) +
 454 -                          EXT4_XATTR_SIZE(i->value_len))
 455 +               size_t value_len = EXT4_XATTR_SIZE(i->value_len);
 456 +
 457 +               if (in_inode)
 458 +                       value_len = 0;
 459 +
 460 +               if (free < value_len ||
 461 +                   free < EXT4_XATTR_LEN(name_len) + value_len)
 462                         return -ENOSPC;
 463         }
 464
 465 @@ -577,7 +854,8 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 466                 s->here->e_name_len = name_len;
 467                 memcpy(s->here->e_name, i->name, name_len);
 468         } else {
 469 -               if (!s->here->e_value_block && s->here->e_value_size) {
 470 +               if (s->here->e_value_offs > 0 && s->here->e_value_inum == 0 &&
 471 +                   s->here->e_value_size > 0) {
 472                         void *first_val = s->base + min_offs;
 473                         size_t offs = le16_to_cpu(s->here->e_value_offs);
 474                         void *val = s->base + offs;
 475 @@ -606,13 +884,18 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 476                         last = s->first;
 477                         while (!IS_LAST_ENTRY(last)) {
 478                                 size_t o = le16_to_cpu(last->e_value_offs);
 479 -                               if (!last->e_value_block &&
 480 +                               if (!last->e_value_inum &&
 481                                     last->e_value_size && o < offs)
 482                                         last->e_value_offs =
 483                                                 cpu_to_le16(o + size);
 484                                 last = EXT4_XATTR_NEXT(last);
 485                         }
 486                 }
 487 +               if (s->here->e_value_inum != 0) {
 488 +                       ext4_xattr_inode_unlink(inode,
 489 +                                       le32_to_cpu(s->here->e_value_inum));
 490 +                       s->here->e_value_inum = 0;
 491 +               }
 492                 if (!i->value) {
 493                         /* Remove the old name. */
 494                         size_t size = EXT4_XATTR_LEN(name_len);
 495 @@ -626,10 +908,17 @@ ext4_xattr_set_entry(struct ext4_xattr_i
 496         if (i->value) {
 497                 /* Insert the new value. */
 498                 s->here->e_value_size = cpu_to_le32(i->value_len);
 499 -               if (i->value_len) {
 500 +               if (in_inode) {
 501 +                       unsigned long ea_ino = le32_to_cpu(s->here->e_value_inum);
 502 +                       ext4_xattr_inode_set(handle, inode, &ea_ino, i->value,
 503 +                                            i->value_len);
 504 +                       s->here->e_value_inum = cpu_to_le32(ea_ino);
 505 +                       s->here->e_value_offs = 0;
 506 +               } else if (i->value_len) {
 507                         size_t size = EXT4_XATTR_SIZE(i->value_len);
 508                         void *val = s->base + min_offs - size;
 509                         s->here->e_value_offs = cpu_to_le16(min_offs - size);
 510 +                       s->here->e_value_inum = 0;
 511                         memset(val + size - EXT4_XATTR_PAD, 0,
 512                                EXT4_XATTR_PAD); /* Clear the pad bytes. */
 513                         memcpy(val, i->value, i->value_len);
 514 @@ -674,7 +963,7 @@ ext4_xattr_block_find(struct inode *inod
 515                 bs->s.end = bs->bh->b_data + bs->bh->b_size;
 516                 bs->s.here = bs->s.first;
 517                 error = ext4_xattr_find_entry(&bs->s.here, i->name_index,
 518 -                                             i->name, bs->bh->b_size, 1);
 519 +                                            i->name, bs->bh->b_size, 1, inode);
 520                 if (error && error != -ENODATA)
 521                         goto cleanup;
 522                 bs->s.not_found = error;
 523 @@ -698,8 +987,6 @@ ext4_xattr_block_set(handle_t *handle, s
 524
 525  #define header(x) ((struct ext4_xattr_header *)(x))
 526
 527 -       if (i->value && i->value_len > sb->s_blocksize)
 528 -               return -ENOSPC;
 529         if (s->base) {
 530                 ce = mb_cache_entry_get(ext4_xattr_cache, bs->bh->b_bdev,
 531                                         bs->bh->b_blocknr);
 532 @@ -714,7 +1001,7 @@ ext4_xattr_block_set(handle_t *handle, s
 533                                 ce = NULL;
 534                         }
 535                         ea_bdebug(bs->bh, "modifying in-place");
 536 -                       error = ext4_xattr_set_entry(i, s);
 537 +                       error = ext4_xattr_set_entry(i, s, handle, inode);
 538                         if (!error) {
 539                                 if (!IS_LAST_ENTRY(s->first))
 540                                         ext4_xattr_rehash(header(s->base),
 541 @@ -766,7 +1053,7 @@ ext4_xattr_block_set(handle_t *handle, s
 542                 s->end = s->base + sb->s_blocksize;
 543         }
 544
 545 -       error = ext4_xattr_set_entry(i, s);
 546 +       error = ext4_xattr_set_entry(i, s, handle, inode);
 547         if (error == -EIO)
 548                 goto bad_block;
 549         if (error)
 550 @@ -917,7 +1204,7 @@ ext4_xattr_ibody_find(struct inode *inod
 551                 /* Find the named attribute. */
 552                 error = ext4_xattr_find_entry(&is->s.here, i->name_index,
 553                                               i->name, is->s.end -
 554 -                                             (void *)is->s.base, 0);
 555 +                                             (void *)is->s.base, 0, inode);
 556                 if (error && error != -ENODATA)
 557                         return error;
 558                 is->s.not_found = error;
 559 @@ -936,7 +1223,7 @@ ext4_xattr_ibody_set(handle_t *handle, s
 560
 561         if (EXT4_I(inode)->i_extra_isize == 0)
 562                 return -ENOSPC;
 563 -       error = ext4_xattr_set_entry(i, s);
 564 +       error = ext4_xattr_set_entry(i, s, handle, inode);
 565         if (error)
 566                 return error;
 567         header = IHDR(inode, ext4_raw_inode(&is->iloc));
 568 @@ -972,7 +1259,7 @@ ext4_xattr_set_handle(handle_t *handle,
 569                 .name = name,
 570                 .value = value,
 571                 .value_len = value_len,
 572 -
 573 +               .in_inode = 0,
 574         };
 575         struct ext4_xattr_ibody_find is = {
 576                 .s = { .not_found = -ENODATA, },
 577 @@ -1041,6 +1328,15 @@ ext4_xattr_set_handle(handle_t *handle,
 578                                         goto cleanup;
 579                         }
 580                         error = ext4_xattr_block_set(handle, inode, &i, &bs);
 581 +                       if (EXT4_HAS_INCOMPAT_FEATURE(inode->i_sb,
 582 +                                       EXT4_FEATURE_INCOMPAT_EA_INODE) &&
 583 +                           error == -ENOSPC) {
 584 +                               /* xattr not fit to block, store at external
 585 +                                * inode */
 586 +                               i.in_inode = 1;
 587 +                               error = ext4_xattr_ibody_set(handle, inode,
 588 +                                                            &i, &is);
 589 +                       }
 590                         if (error)
 591                                 goto cleanup;
 592                         if (!is.s.not_found) {
 593 @@ -1087,10 +1383,25 @@ ext4_xattr_set(struct inode *inode, int
 594                const void *value, size_t value_len, int flags)
 595  {
 596         handle_t *handle;
 597 +       struct super_block *sb = inode->i_sb;
 598 +       int buffer_credits;
 599         int error, retries = 0;
 600
 601 +       buffer_credits = EXT4_DATA_TRANS_BLOCKS(sb);
 602 +       if ((value_len >= EXT4_XATTR_MIN_LARGE_EA_SIZE(sb->s_blocksize)) &&
 603 +           EXT4_HAS_INCOMPAT_FEATURE(sb, EXT4_FEATURE_INCOMPAT_EA_INODE)) {
 604 +               int nrblocks = (value_len + sb->s_blocksize - 1) >>
 605 +                                       sb->s_blocksize_bits;
 606 +
 607 +               /* For new inode */
 608 +               buffer_credits += EXT4_SINGLEDATA_TRANS_BLOCKS(sb) + 3;
 609 +
 610 +               /* For data blocks of EA inode */
 611 +               buffer_credits += ext4_meta_trans_blocks(inode, nrblocks, 0);
 612 +       }
 613 +
 614  retry:
 615 -       handle = ext4_journal_start(inode, EXT4_DATA_TRANS_BLOCKS(inode->i_sb));
 616 +       handle = ext4_journal_start(inode, buffer_credits);
 617         if (IS_ERR(handle)) {
 618                 error = PTR_ERR(handle);
 619         } else {
 620 @@ -1100,7 +1411,7 @@ retry:
 621                                               value, value_len, flags);
 622                 error2 = ext4_journal_stop(handle);
 623                 if (error == -ENOSPC &&
 624 -                   ext4_should_retry_alloc(inode->i_sb, &retries))
 625 +                   ext4_should_retry_alloc(sb, &retries))
 626                         goto retry;
 627                 if (error == 0)
 628                         error = error2;
 629 @@ -1122,7 +1433,7 @@ static void ext4_xattr_shift_entries(str
 630
 631         /* Adjust the value offsets of the entries */
 632         for (; !IS_LAST_ENTRY(last); last = EXT4_XATTR_NEXT(last)) {
 633 -               if (!last->e_value_block && last->e_value_size) {
 634 +               if (last->e_value_inum == 0 && last->e_value_size > 0) {
 635                         new_offs = le16_to_cpu(last->e_value_offs) +
 636                                                         value_offs_shift;
 637                         BUG_ON(new_offs + le32_to_cpu(last->e_value_size)
 638 @@ -1355,22 +1666,135 @@ cleanup:
 639         return error;
 640  }
 641
 642 +#define EIA_INCR 16 /* must be 2^n */
 643 +#define EIA_MASK (EIA_INCR - 1)
 644 +/* Add the large xattr @ino into @lea_ino_array for later deletion.
 645 + * If @lea_ino_array is new or full it will be grown and the old
 646 + * contents copied over.
 647 + */
 648 +static int
 649 +ext4_expand_ino_array(struct ext4_xattr_ino_array **lea_ino_array, __u32 ino)
 650 +{
 651 +       if (*lea_ino_array == NULL) {
 652 +               /*
 653 +                * Start with 15 inodes, so it fits into a power-of-two size.
 654 +                * If *lea_ino_array is NULL, this is essentially offsetof()
 655 +                */
 656 +               (*lea_ino_array) =
 657 +                       kmalloc(offsetof(struct ext4_xattr_ino_array,
 658 +                                        xia_inodes[EIA_MASK]),
 659 +                               GFP_NOFS);
 660 +               if (*lea_ino_array == NULL)
 661 +                       return -ENOMEM;
 662 +               (*lea_ino_array)->xia_count = 0;
 663 +       } else if (((*lea_ino_array)->xia_count & EIA_MASK) == EIA_MASK) {
 664 +               /* expand the array once all 15 + n * 16 slots are full */
 665 +               struct ext4_xattr_ino_array *new_array = NULL;
 666 +               int count = (*lea_ino_array)->xia_count;
 667 +
 668 +               /* if new_array is NULL, this is essentially offsetof() */
 669 +               new_array = kmalloc(
 670 +                               offsetof(struct ext4_xattr_ino_array,
 671 +                                        xia_inodes[count + EIA_INCR]),
 672 +                               GFP_NOFS);
 673 +               if (new_array == NULL)
 674 +                       return -ENOMEM;
 675 +               memcpy(new_array, *lea_ino_array,
 676 +                      offsetof(struct ext4_xattr_ino_array,
 677 +                               xia_inodes[count]));
 678 +               kfree(*lea_ino_array);
 679 +               *lea_ino_array = new_array;
 680 +       }
 681 +       (*lea_ino_array)->xia_inodes[(*lea_ino_array)->xia_count++] = ino;
 682 +       return 0;
 683 +}
 684
 685 +/**
 686 + * Add xattr inode to orphan list
 687 + */
 688 +static int
 689 +ext4_xattr_inode_orphan_add(handle_t *handle, struct inode *inode,
 690 +                       int credits, struct ext4_xattr_ino_array *lea_ino_array)
 691 +{
 692 +       struct inode *ea_inode = NULL;
 693 +       int idx = 0, error = 0;
 694 +
 695 +       if (lea_ino_array == NULL)
 696 +               return 0;
 697 +
 698 +       for (; idx < lea_ino_array->xia_count; ++idx) {
 699 +               if (!ext4_handle_has_enough_credits(handle, credits)) {
 700 +                       error = ext4_journal_extend(handle, credits);
 701 +                       if (error > 0)
 702 +                               error = ext4_journal_restart(handle, credits);
 703 +
 704 +                       if (error != 0) {
 705 +                               ext4_warning(inode->i_sb,
 706 +                                       "couldn't extend journal "
 707 +                                       "(err %d)", error);
 708 +                               return error;
 709 +                       }
 710 +               }
 711 +               ea_inode = ext4_xattr_inode_iget(inode,
 712 +                               lea_ino_array->xia_inodes[idx], &error);
 713 +               if (error)
 714 +                       continue;
 715 +               ext4_orphan_add(handle, ea_inode);
 716 +               /* the inode's i_count will be released by caller */
 717 +       }
 718 +
 719 +       return 0;
 720 +}
 721
 722  /*
 723   * ext4_xattr_delete_inode()
 724   *
 725 - * Free extended attribute resources associated with this inode. This
 726 + * Free extended attribute resources associated with this inode. Traverse
 727 + * all entries and unlink any xattr inodes associated with this inode. This
 728   * is called immediately before an inode is freed. We have exclusive
 729 - * access to the inode.
 730 + * access to the inode. If an orphan inode is deleted it will also delete any
 731 + * xattr block and all xattr inodes. They are checked by ext4_xattr_inode_iget()
 732 + * to ensure they belong to the parent inode and were not deleted already.
 733   */
 734 -void
 735 -ext4_xattr_delete_inode(handle_t *handle, struct inode *inode)
 736 +int
 737 +ext4_xattr_delete_inode(handle_t *handle, struct inode *inode,
 738 +                       struct ext4_xattr_ino_array **lea_ino_array)
 739  {
 740         struct buffer_head *bh = NULL;
 741 +       struct ext4_xattr_ibody_header *header;
 742 +       struct ext4_inode *raw_inode;
 743 +       struct ext4_iloc iloc;
 744 +       struct ext4_xattr_entry *entry;
 745 +       int credits = 3, error = 0;
 746
 747 -       if (!EXT4_I(inode)->i_file_acl)
 748 +       if (!ext4_test_inode_state(inode, EXT4_STATE_XATTR))
 749 +               goto delete_external_ea;
 750 +
 751 +       error = ext4_get_inode_loc(inode, &iloc);
 752 +       if (error)
 753 +               goto cleanup;
 754 +       raw_inode = ext4_raw_inode(&iloc);
 755 +       header = IHDR(inode, raw_inode);
 756 +       entry = IFIRST(header);
 757 +       for (; !IS_LAST_ENTRY(entry); entry = EXT4_XATTR_NEXT(entry)) {
 758 +               if (entry->e_value_inum == 0)
 759 +                       continue;
 760 +               if (ext4_expand_ino_array(lea_ino_array,
 761 +                                         entry->e_value_inum) != 0) {
 762 +                       brelse(iloc.bh);
 763 +                       goto cleanup;
 764 +               }
 765 +               entry->e_value_inum = 0;
 766 +       }
 767 +       brelse(iloc.bh);
 768 +
 769 +delete_external_ea:
 770 +       if (!EXT4_I(inode)->i_file_acl) {
 771 +               /* add xattr inode to orphan list */
 772 +               ext4_xattr_inode_orphan_add(handle, inode, credits,
 773 +                                               *lea_ino_array);
 774                 goto cleanup;
 775 +       }
 776         bh = sb_bread(inode->i_sb, EXT4_I(inode)->i_file_acl);
 777         if (!bh) {
 778                 ext4_error(inode->i_sb, "inode %lu: block %llu read error",
 779 @@ -1383,11 +1807,71 @@ ext4_xattr_delete_inode(handle_t *handle
 780                            inode->i_ino, EXT4_I(inode)->i_file_acl);
 781                 goto cleanup;
 782         }
 783 +
 784 +       entry = BFIRST(bh);
 785 +       for (; !IS_LAST_ENTRY(entry); entry = EXT4_XATTR_NEXT(entry)) {
 786 +               if (entry->e_value_inum == 0)
 787 +                       continue;
 788 +               if (ext4_expand_ino_array(lea_ino_array,
 789 +                                         entry->e_value_inum) != 0)
 790 +                       goto cleanup;
 791 +               entry->e_value_inum = 0;
 792 +       }
 793 +
 794 +       /* add xattr inode to orphan list */
 795 +       error = ext4_xattr_inode_orphan_add(handle, inode, credits,
 796 +                                       *lea_ino_array);
 797 +       if (error != 0)
 798 +               goto cleanup;
 799 +
 800 +       if (!IS_NOQUOTA(inode))
 801 +               credits += 2 * EXT4_QUOTA_DEL_BLOCKS(inode->i_sb);
 802 +
 803 +       if (!ext4_handle_has_enough_credits(handle, credits)) {
 804 +               error = ext4_journal_extend(handle, credits);
 805 +               if (error > 0)
 806 +                       error = ext4_journal_restart(handle, credits);
 807 +               if (error != 0) {
 808 +                       ext4_warning(inode->i_sb,
 809 +                               "couldn't extend journal (err %d)", error);
 810 +                       goto cleanup;
 811 +               }
 812 +       }
 813 +
 814         ext4_xattr_release_block(handle, inode, bh);
 815         EXT4_I(inode)->i_file_acl = 0;
 816
 817  cleanup:
 818         brelse(bh);
 819 +
 820 +       return error;
 821 +}
 822 +
 823 +void
 824 +ext4_xattr_inode_array_free(struct inode *inode,
 825 +                           struct ext4_xattr_ino_array *lea_ino_array)
 826 +{
 827 +       struct inode    *ea_inode = NULL;
 828 +       int             idx = 0;
 829 +       int             err;
 830 +
 831 +       if (lea_ino_array == NULL)
 832 +               return;
 833 +
 834 +       for (; idx < lea_ino_array->xia_count; ++idx) {
 835 +               ea_inode = ext4_xattr_inode_iget(inode,
 836 +                               lea_ino_array->xia_inodes[idx], &err);
 837 +               if (err)
 838 +                       continue;
 839 +
 840 +               /* for inode's i_count get from ext4_xattr_delete_inode */
 841 +               if (!list_empty(&EXT4_I(ea_inode)->i_orphan))
 842 +                       iput(ea_inode);
 843 +
 844 +               ea_inode->i_nlink = 0;
 845 +               iput(ea_inode);
 846 +       }
 847 +       kfree(lea_ino_array);
 848  }
 849
 850  /*
 851 @@ -1457,10 +1941,9 @@ ext4_xattr_cmp(struct ext4_xattr_header
 852                     entry1->e_name_index != entry2->e_name_index ||
 853                     entry1->e_name_len != entry2->e_name_len ||
 854                     entry1->e_value_size != entry2->e_value_size ||
 855 +                   entry1->e_value_inum != entry2->e_value_inum ||
 856                     memcmp(entry1->e_name, entry2->e_name, entry1->e_name_len))
 857                         return 1;
 858 -               if (entry1->e_value_block != 0 || entry2->e_value_block != 0)
 859 -                       return -EIO;
 860                 if (memcmp((char *)header1 + le16_to_cpu(entry1->e_value_offs),
 861                            (char *)header2 + le16_to_cpu(entry2->e_value_offs),
 862                            le32_to_cpu(entry1->e_value_size)))
 863 @@ -1545,7 +2028,7 @@ static inline void ext4_xattr_hash_entry
 864                        *name++;
 865         }
 866
 867 -       if (entry->e_value_block == 0 && entry->e_value_size != 0) {
 868 +       if (entry->e_value_inum == 0 && entry->e_value_size != 0) {
 869                 __le32 *value = (__le32 *)((char *)header +
 870                         le16_to_cpu(entry->e_value_offs));
 871                 for (n = (le32_to_cpu(entry->e_value_size) +
 872 Index: linux-stage/fs/ext4/xattr.h
 873 ===================================================================
 874 --- linux-stage.orig/fs/ext4/xattr.h
 875 +++ linux-stage/fs/ext4/xattr.h
 876 @@ -38,7 +38,7 @@ struct ext4_xattr_entry {
 877         __u8    e_name_len;     /* length of name */
 878         __u8    e_name_index;   /* attribute name index */
 879         __le16  e_value_offs;   /* offset in disk block of value */
 880 -       __le32  e_value_block;  /* disk block attribute is stored on (n/i) */
 881 +       __le32  e_value_inum;   /* inode in which the value is stored */
 882         __le32  e_value_size;   /* size of attribute value */
 883         __le32  e_hash;         /* hash value of name and value */
 884         char    e_name[0];      /* attribute name */
 885 @@ -63,6 +63,15 @@ struct ext4_xattr_entry {
 886                 EXT4_I(inode)->i_extra_isize))
 887  #define IFIRST(hdr) ((struct ext4_xattr_entry *)((hdr)+1))
 888
 889 +#define i_xattr_inode_parent i_mtime.tv_sec
 890 +
 891 +/*
 892 + * The minimum size of EA value when you start storing it in an external inode
 893 + * size of block - size of header - size of 1 entry - 4 null bytes
 894 +*/
 895 +#define EXT4_XATTR_MIN_LARGE_EA_SIZE(b)                                        \
 896 +       ((b) - EXT4_XATTR_LEN(3) - sizeof(struct ext4_xattr_header) - 4)
 897 +
 898  # ifdef CONFIG_EXT4_FS_XATTR
 899
 900  extern struct xattr_handler ext4_xattr_user_handler;
 901 @@ -77,7 +86,13 @@ extern int ext4_xattr_get(struct inode *
 902  extern int ext4_xattr_set(struct inode *, int, const char *, const void *, size_t, int);
 903  extern int ext4_xattr_set_handle(handle_t *, struct inode *, int, const char *, const void *, size_t, int);
 904
 905 -extern void ext4_xattr_delete_inode(handle_t *, struct inode *);
 906 +extern struct inode *ext4_xattr_inode_iget(struct inode *parent, unsigned long ea_ino,
 907 +                                          int *err);
 908 +extern int ext4_xattr_inode_unlink(struct inode *inode, unsigned long ea_ino);
 909 +extern int ext4_xattr_delete_inode(handle_t *handle, struct inode *inode,
 910 +                                  struct ext4_xattr_ino_array **array);
 911 +extern void ext4_xattr_inode_array_free(struct inode *inode,
 912 +                                       struct ext4_xattr_ino_array *array);
 913  extern void ext4_xattr_put_super(struct super_block *);
 914
 915  extern int ext4_expand_extra_isize_ea(struct inode *inode, int new_extra_isize,
 916 @@ -111,9 +126,11 @@ ext4_xattr_set_handle(handle_t *handle,
 917         return -EOPNOTSUPP;
 918  }
 919
 920 -static inline void
 921 -ext4_xattr_delete_inode(handle_t *handle, struct inode *inode)
 922 +inline int
 923 +ext4_xattr_delete_inode(handle_t *handle, struct inode *inode,
 924 +                       struct ext4_xattr_ino_array **array)
 925  {
 926 +       return -EOPNOTSUPP;
 927  }
 928
 929  static inline void
 930 Index: linux-stage/fs/ext4/inode.c
 931 ===================================================================
 932 --- linux-stage.orig/fs/ext4/inode.c
 933 +++ linux-stage/fs/ext4/inode.c
 934 @@ -222,6 +222,8 @@ void ext4_delete_inode(struct inode *ino
 935  {
 936         handle_t *handle;
 937         int err;
 938 +       int extra_credits = 3;
 939 +       struct ext4_xattr_ino_array *lea_ino_array = NULL;
 940
 941         if (ext4_should_order_data(inode))
 942                 ext4_begin_ordered_truncate(inode, 0);
 943 @@ -235,7 +237,8 @@ void ext4_delete_inode(struct inode *ino
 944          * protection against it
 945          */
 946         sb_start_intwrite(inode->i_sb);
 947 -       handle = ext4_journal_start(inode, blocks_for_truncate(inode)+3);
 948 +
 949 +       handle = ext4_journal_start(inode, extra_credits);
 950         if (IS_ERR(handle)) {
 951                 ext4_std_error(inode->i_sb, PTR_ERR(handle));
 952                 /*
 953 @@ -247,9 +250,36 @@ void ext4_delete_inode(struct inode *ino
 954                 sb_end_intwrite(inode->i_sb);
 955                 goto no_delete;
 956         }
 957 -
 958         if (IS_SYNC(inode))
 959                 ext4_handle_sync(handle);
 960 +
 961 +       /*
 962 +        * Delete xattr inode before deleting the main inode.
 963 +        */
 964 +       err = ext4_xattr_delete_inode(handle, inode, &lea_ino_array);
 965 +       if (err) {
 966 +               ext4_warning(inode->i_sb,
 967 +                            "couldn't delete inode's xattr (err %d)", err);
 968 +               goto stop_handle;
 969 +       }
 970 +
 971 +       if (!IS_NOQUOTA(inode))
 972 +               extra_credits += 2 * EXT4_QUOTA_DEL_BLOCKS(inode->i_sb);
 973 +
 974 +       if (!ext4_handle_has_enough_credits(handle,
 975 +                               blocks_for_truncate(inode) + extra_credits)) {
 976 +               err = ext4_journal_extend(handle,
 977 +                               blocks_for_truncate(inode) + extra_credits);
 978 +               if (err > 0)
 979 +                       err = ext4_journal_restart(handle,
 980 +                               blocks_for_truncate(inode) + extra_credits);
 981 +               if (err != 0) {
 982 +                       ext4_warning(inode->i_sb,
 983 +                                    "couldn't extend journal (err %d)", err);
 984 +                       goto stop_handle;
 985 +               }
 986 +       }
 987 +
 988         inode->i_size = 0;
 989         err = ext4_mark_inode_dirty(handle, inode);
 990         if (err) {
 991 @@ -303,8 +333,12 @@ void ext4_delete_inode(struct inode *ino
 992                 clear_inode(inode);
 993         else
 994                 ext4_free_inode(handle, inode);
 995 +
 996         ext4_journal_stop(handle);
 997         sb_end_intwrite(inode->i_sb);
 998 +
 999 +       if (lea_ino_array != NULL)
1000 +               ext4_xattr_inode_array_free(inode, lea_ino_array);
1001         return;
1002  no_delete:
1003         clear_inode(inode);     /* We must guarantee clearing of inode... */
1004 Index: linux-stage/fs/ext4/ialloc.c
1005 ===================================================================
1006 --- linux-stage.orig/fs/ext4/ialloc.c
1007 +++ linux-stage/fs/ext4/ialloc.c
1008 @@ -219,7 +219,6 @@ void ext4_free_inode(handle_t *handle, s
1009          * as writing the quota to disk may need the lock as well.
1010          */
1011         vfs_dq_init(inode);
1012 -       ext4_xattr_delete_inode(handle, inode);
1013         vfs_dq_free_inode(inode);
1014         vfs_dq_drop(inode);
1015