fs/reiserfs/do_balan.c

   1 /*
   2  * Copyright 2000 by Hans Reiser, licensing governed by reiserfs/README
   3  */
   4
   5 /*
   6  * Now we have all buffers that must be used in balancing of the tree
   7  * Further calculations can not cause schedule(), and thus the buffer
   8  * tree will be stable until the balancing will be finished
   9  * balance the tree according to the analysis made before,
  10  * and using buffers obtained after all above.
  11  */
  12
  13 #include <asm/uaccess.h>
  14 #include <linux/time.h>
  15 #include "reiserfs.h"
  16 #include <linux/buffer_head.h>
  17 #include <linux/kernel.h>
  18
  19 static inline void buffer_info_init_left(struct tree_balance *tb,
  20                                          struct buffer_info *bi)
  21 {
  22         bi->tb          = tb;
  23         bi->bi_bh       = tb->L[0];
  24         bi->bi_parent   = tb->FL[0];
  25         bi->bi_position = get_left_neighbor_position(tb, 0);
  26 }
  27
  28 static inline void buffer_info_init_right(struct tree_balance *tb,
  29                                           struct buffer_info *bi)
  30 {
  31         bi->tb          = tb;
  32         bi->bi_bh       = tb->R[0];
  33         bi->bi_parent   = tb->FR[0];
  34         bi->bi_position = get_right_neighbor_position(tb, 0);
  35 }
  36
  37 static inline void buffer_info_init_tbS0(struct tree_balance *tb,
  38                                          struct buffer_info *bi)
  39 {
  40         bi->tb          = tb;
  41         bi->bi_bh        = PATH_PLAST_BUFFER(tb->tb_path);
  42         bi->bi_parent   = PATH_H_PPARENT(tb->tb_path, 0);
  43         bi->bi_position = PATH_H_POSITION(tb->tb_path, 1);
  44 }
  45
  46 static inline void buffer_info_init_bh(struct tree_balance *tb,
  47                                        struct buffer_info *bi,
  48                                        struct buffer_head *bh)
  49 {
  50         bi->tb          = tb;
  51         bi->bi_bh       = bh;
  52         bi->bi_parent   = NULL;
  53         bi->bi_position = 0;
  54 }
  55
  56 inline void do_balance_mark_leaf_dirty(struct tree_balance *tb,
  57                                        struct buffer_head *bh, int flag)
  58 {
  59         journal_mark_dirty(tb->transaction_handle, bh);
  60 }
  61
  62 #define do_balance_mark_internal_dirty do_balance_mark_leaf_dirty
  63 #define do_balance_mark_sb_dirty do_balance_mark_leaf_dirty
  64
  65 /*
  66  * summary:
  67  *  if deleting something ( tb->insert_size[0] < 0 )
  68  *    return(balance_leaf_when_delete()); (flag d handled here)
  69  *  else
  70  *    if lnum is larger than 0 we put items into the left node
  71  *    if rnum is larger than 0 we put items into the right node
  72  *    if snum1 is larger than 0 we put items into the new node s1
  73  *    if snum2 is larger than 0 we put items into the new node s2
  74  * Note that all *num* count new items being created.
  75  *
  76  * It would be easier to read balance_leaf() if each of these summary
  77  * lines was a separate procedure rather than being inlined.  I think
  78  * that there are many passages here and in balance_leaf_when_delete() in
  79  * which two calls to one procedure can replace two passages, and it
  80  * might save cache space and improve software maintenance costs to do so.
  81  *
  82  * Vladimir made the perceptive comment that we should offload most of
  83  * the decision making in this function into fix_nodes/check_balance, and
  84  * then create some sort of structure in tb that says what actions should
  85  * be performed by do_balance.
  86  *
  87  * -Hans
  88  */
  89
  90 /*
  91  * Balance leaf node in case of delete or cut: insert_size[0] < 0
  92  *
  93  * lnum, rnum can have values >= -1
  94  *      -1 means that the neighbor must be joined with S
  95  *       0 means that nothing should be done with the neighbor
  96  *      >0 means to shift entirely or partly the specified number of items
  97  *         to the neighbor
  98  */
  99 static int balance_leaf_when_delete(struct tree_balance *tb, int flag)
 100 {
 101         struct buffer_head *tbS0 = PATH_PLAST_BUFFER(tb->tb_path);
 102         int item_pos = PATH_LAST_POSITION(tb->tb_path);
 103         int pos_in_item = tb->tb_path->pos_in_item;
 104         struct buffer_info bi;
 105         int n;
 106         struct item_head *ih;
 107
 108         RFALSE(tb->FR[0] && B_LEVEL(tb->FR[0]) != DISK_LEAF_NODE_LEVEL + 1,
 109                "vs- 12000: level: wrong FR %z", tb->FR[0]);
 110         RFALSE(tb->blknum[0] > 1,
 111                "PAP-12005: tb->blknum == %d, can not be > 1", tb->blknum[0]);
 112         RFALSE(!tb->blknum[0] && !PATH_H_PPARENT(tb->tb_path, 0),
 113                "PAP-12010: tree can not be empty");
 114
 115         ih = item_head(tbS0, item_pos);
 116         buffer_info_init_tbS0(tb, &bi);
 117
 118         /* Delete or truncate the item */
 119
 120         switch (flag) {
 121         case M_DELETE:          /* delete item in S[0] */
 122
 123                 RFALSE(ih_item_len(ih) + IH_SIZE != -tb->insert_size[0],
 124                        "vs-12013: mode Delete, insert size %d, ih to be deleted %h",
 125                        -tb->insert_size[0], ih);
 126
 127                 leaf_delete_items(&bi, 0, item_pos, 1, -1);
 128
 129                 if (!item_pos && tb->CFL[0]) {
 130                         if (B_NR_ITEMS(tbS0)) {
 131                                 replace_key(tb, tb->CFL[0], tb->lkey[0], tbS0,
 132                                             0);
 133                         } else {
 134                                 if (!PATH_H_POSITION(tb->tb_path, 1))
 135                                         replace_key(tb, tb->CFL[0], tb->lkey[0],
 136                                                     PATH_H_PPARENT(tb->tb_path,
 137                                                                    0), 0);
 138                         }
 139                 }
 140
 141                 RFALSE(!item_pos && !tb->CFL[0],
 142                        "PAP-12020: tb->CFL[0]==%p, tb->L[0]==%p", tb->CFL[0],
 143                        tb->L[0]);
 144
 145                 break;
 146
 147         case M_CUT:{            /* cut item in S[0] */
 148                         if (is_direntry_le_ih(ih)) {
 149
 150                                 /*
 151                                  * UFS unlink semantics are such that you
 152                                  * can only delete one directory entry at
 153                                  * a time.
 154                                  */
 155
 156                                 /*
 157                                  * when we cut a directory tb->insert_size[0]
 158                                  * means number of entries to be cut (always 1)
 159                                  */
 160                                 tb->insert_size[0] = -1;
 161                                 leaf_cut_from_buffer(&bi, item_pos, pos_in_item,
 162                                                      -tb->insert_size[0]);
 163
 164                                 RFALSE(!item_pos && !pos_in_item && !tb->CFL[0],
 165                                        "PAP-12030: can not change delimiting key. CFL[0]=%p",
 166                                        tb->CFL[0]);
 167
 168                                 if (!item_pos && !pos_in_item && tb->CFL[0]) {
 169                                         replace_key(tb, tb->CFL[0], tb->lkey[0],
 170                                                     tbS0, 0);
 171                                 }
 172                         } else {
 173                                 leaf_cut_from_buffer(&bi, item_pos, pos_in_item,
 174                                                      -tb->insert_size[0]);
 175
 176                                 RFALSE(!ih_item_len(ih),
 177                                        "PAP-12035: cut must leave non-zero dynamic length of item");
 178                         }
 179                         break;
 180                 }
 181
 182         default:
 183                 print_cur_tb("12040");
 184                 reiserfs_panic(tb->tb_sb, "PAP-12040",
 185                                "unexpected mode: %s(%d)",
 186                                (flag ==
 187                                 M_PASTE) ? "PASTE" : ((flag ==
 188                                                        M_INSERT) ? "INSERT" :
 189                                                       "UNKNOWN"), flag);
 190         }
 191
 192         /*
 193          * the rule is that no shifting occurs unless by shifting
 194          * a node can be freed
 195          */
 196         n = B_NR_ITEMS(tbS0);
 197         /* L[0] takes part in balancing */
 198         if (tb->lnum[0]) {
 199                 /* L[0] must be joined with S[0] */
 200                 if (tb->lnum[0] == -1) {
 201                         /* R[0] must be also joined with S[0] */
 202                         if (tb->rnum[0] == -1) {
 203                                 if (tb->FR[0] == PATH_H_PPARENT(tb->tb_path, 0)) {
 204                                         /*
 205                                          * all contents of all the 3 buffers
 206                                          * will be in L[0]
 207                                          */
 208                                         if (PATH_H_POSITION(tb->tb_path, 1) == 0
 209                                             && 1 < B_NR_ITEMS(tb->FR[0]))
 210                                                 replace_key(tb, tb->CFL[0],
 211                                                             tb->lkey[0],
 212                                                             tb->FR[0], 1);
 213
 214                                         leaf_move_items(LEAF_FROM_S_TO_L, tb, n,
 215                                                         -1, NULL);
 216                                         leaf_move_items(LEAF_FROM_R_TO_L, tb,
 217                                                         B_NR_ITEMS(tb->R[0]),
 218                                                         -1, NULL);
 219
 220                                         reiserfs_invalidate_buffer(tb, tbS0);
 221                                         reiserfs_invalidate_buffer(tb,
 222                                                                    tb->R[0]);
 223
 224                                         return 0;
 225                                 }
 226                                 /*
 227                                  * all contents of all the 3 buffers will
 228                                  * be in R[0]
 229                                  */
 230                                 leaf_move_items(LEAF_FROM_S_TO_R, tb, n, -1,
 231                                                 NULL);
 232                                 leaf_move_items(LEAF_FROM_L_TO_R, tb,
 233                                                 B_NR_ITEMS(tb->L[0]), -1, NULL);
 234
 235                                 /* right_delimiting_key is correct in R[0] */
 236                                 replace_key(tb, tb->CFR[0], tb->rkey[0],
 237                                             tb->R[0], 0);
 238
 239                                 reiserfs_invalidate_buffer(tb, tbS0);
 240                                 reiserfs_invalidate_buffer(tb, tb->L[0]);
 241
 242                                 return -1;
 243                         }
 244
 245                         RFALSE(tb->rnum[0] != 0,
 246                                "PAP-12045: rnum must be 0 (%d)", tb->rnum[0]);
 247                         /* all contents of L[0] and S[0] will be in L[0] */
 248                         leaf_shift_left(tb, n, -1);
 249
 250                         reiserfs_invalidate_buffer(tb, tbS0);
 251
 252                         return 0;
 253                 }
 254
 255                 /*
 256                  * a part of contents of S[0] will be in L[0] and the
 257                  * rest part of S[0] will be in R[0]
 258                  */
 259
 260                 RFALSE((tb->lnum[0] + tb->rnum[0] < n) ||
 261                        (tb->lnum[0] + tb->rnum[0] > n + 1),
 262                        "PAP-12050: rnum(%d) and lnum(%d) and item number(%d) in S[0] are not consistent",
 263                        tb->rnum[0], tb->lnum[0], n);
 264                 RFALSE((tb->lnum[0] + tb->rnum[0] == n) &&
 265                        (tb->lbytes != -1 || tb->rbytes != -1),
 266                        "PAP-12055: bad rbytes (%d)/lbytes (%d) parameters when items are not split",
 267                        tb->rbytes, tb->lbytes);
 268                 RFALSE((tb->lnum[0] + tb->rnum[0] == n + 1) &&
 269                        (tb->lbytes < 1 || tb->rbytes != -1),
 270                        "PAP-12060: bad rbytes (%d)/lbytes (%d) parameters when items are split",
 271                        tb->rbytes, tb->lbytes);
 272
 273                 leaf_shift_left(tb, tb->lnum[0], tb->lbytes);
 274                 leaf_shift_right(tb, tb->rnum[0], tb->rbytes);
 275
 276                 reiserfs_invalidate_buffer(tb, tbS0);
 277
 278                 return 0;
 279         }
 280
 281         if (tb->rnum[0] == -1) {
 282                 /* all contents of R[0] and S[0] will be in R[0] */
 283                 leaf_shift_right(tb, n, -1);
 284                 reiserfs_invalidate_buffer(tb, tbS0);
 285                 return 0;
 286         }
 287
 288         RFALSE(tb->rnum[0],
 289                "PAP-12065: bad rnum parameter must be 0 (%d)", tb->rnum[0]);
 290         return 0;
 291 }
 292
 293 static void balance_leaf_insert_left(struct tree_balance *tb,
 294                                      struct item_head *ih, const char *body)
 295 {
 296         int ret_val;
 297         struct buffer_info bi;
 298         int n = B_NR_ITEMS(tb->L[0]);
 299
 300                                 if (tb->item_pos == tb->lnum[0] - 1 && tb->lbytes != -1) {
 301                                         /* part of new item falls into L[0] */
 302                                         int new_item_len;
 303                                         int version;
 304
 305                                         ret_val = leaf_shift_left(tb, tb->lnum[0] - 1, -1);
 306
 307                                         /* Calculate item length to insert to S[0] */
 308                                         new_item_len = ih_item_len(ih) - tb->lbytes;
 309                                         /* Calculate and check item length to insert to L[0] */
 310                                         put_ih_item_len(ih, ih_item_len(ih) - new_item_len);
 311
 312                                         RFALSE(ih_item_len(ih) <= 0,
 313                                                "PAP-12080: there is nothing to insert into L[0]: ih_item_len=%d",
 314                                                ih_item_len(ih));
 315
 316                                         /* Insert new item into L[0] */
 317                                         buffer_info_init_left(tb, &bi);
 318                                         leaf_insert_into_buf(&bi,
 319                                                         n + tb->item_pos - ret_val, ih, body,
 320                                                         tb->zeroes_num > ih_item_len(ih) ? ih_item_len(ih) : tb->zeroes_num);
 321
 322                                         version = ih_version(ih);
 323
 324                                         /* Calculate key component, item length and body to insert into S[0] */
 325                                         set_le_ih_k_offset(ih, le_ih_k_offset(ih) +
 326                                                         (tb->lbytes << (is_indirect_le_ih(ih) ? tb->tb_sb->s_blocksize_bits - UNFM_P_SHIFT : 0)));
 327
 328                                         put_ih_item_len(ih, new_item_len);
 329                                         if (tb->lbytes > tb->zeroes_num) {
 330                                                 body += (tb->lbytes - tb->zeroes_num);
 331                                                 tb->zeroes_num = 0;
 332                                         } else
 333                                                 tb->zeroes_num -= tb->lbytes;
 334
 335                                         RFALSE(ih_item_len(ih) <= 0,
 336                                                "PAP-12085: there is nothing to insert into S[0]: ih_item_len=%d",
 337                                                ih_item_len(ih));
 338                                 } else {
 339                                         /* new item in whole falls into L[0] */
 340                                         /* Shift lnum[0]-1 items to L[0] */
 341                                         ret_val = leaf_shift_left(tb, tb->lnum[0] - 1, tb->lbytes);
 342                                         /* Insert new item into L[0] */
 343                                         buffer_info_init_left(tb, &bi);
 344                                         leaf_insert_into_buf(&bi, n + tb->item_pos - ret_val, ih, body, tb->zeroes_num);
 345                                         tb->insert_size[0] = 0;
 346                                         tb->zeroes_num = 0;
 347                                 }
 348
 349 }
 350
 351 static void balance_leaf_paste_left(struct tree_balance *tb,
 352                                     struct item_head *ih, const char *body)
 353 {
 354         struct buffer_head *tbS0 = PATH_PLAST_BUFFER(tb->tb_path);
 355         int ret_val;
 356         struct buffer_info bi;
 357         int n = B_NR_ITEMS(tb->L[0]);
 358
 359                                 if (tb->item_pos == tb->lnum[0] - 1 && tb->lbytes != -1) {
 360                                         /* we must shift the part of the appended item */
 361                                         if (is_direntry_le_ih(item_head(tbS0, tb->item_pos))) {
 362
 363                                                 RFALSE(tb->zeroes_num,
 364                                                        "PAP-12090: invalid parameter in case of a directory");
 365                                                 /* directory item */
 366                                                 if (tb->lbytes > tb->pos_in_item) {
 367                                                         /* new directory entry falls into L[0] */
 368                                                         struct item_head *pasted;
 369                                                         int l_pos_in_item = tb->pos_in_item;
 370
 371                                                         /* Shift lnum[0] - 1 items in whole. Shift lbytes - 1 entries from given directory item */
 372                                                         ret_val = leaf_shift_left(tb, tb->lnum[0], tb->lbytes-1);
 373                                                         if (ret_val && !tb->item_pos) {
 374                                                                 pasted = item_head(tb->L[0], B_NR_ITEMS(tb->L[0]) - 1);
 375                                                                 l_pos_in_item += ih_entry_count(pasted) - (tb->lbytes -1);
 376                                                         }
 377
 378                                                         /* Append given directory entry to directory item */
 379                                                         buffer_info_init_left(tb, &bi);
 380                                                         leaf_paste_in_buffer(&bi, n + tb->item_pos - ret_val, l_pos_in_item, tb->insert_size[0], body, tb->zeroes_num);
 381
 382                                                         /* previous string prepared space for pasting new entry, following string pastes this entry */
 383
 384                                                         /* when we have merge directory item, pos_in_item has been changed too */
 385
 386                                                         /* paste new directory entry. 1 is entry number */
 387                                                         leaf_paste_entries(&bi, n + tb->item_pos - ret_val, l_pos_in_item,
 388                                                                            1, (struct reiserfs_de_head *) body,
 389                                                                            body + DEH_SIZE, tb->insert_size[0]);
 390                                                         tb->insert_size[0] = 0;
 391                                                 } else {
 392                                                         /* new directory item doesn't fall into L[0] */
 393                                                         /* Shift lnum[0]-1 items in whole. Shift lbytes directory entries from directory item number lnum[0] */
 394                                                         leaf_shift_left(tb, tb->lnum[0], tb->lbytes);
 395                                                 }
 396                                                 /* Calculate new position to append in item body */
 397                                                 tb->pos_in_item -= tb->lbytes;
 398                                         } else {
 399                                                 /* regular object */
 400                                                 RFALSE(tb->lbytes <= 0, "PAP-12095: there is nothing to shift to L[0]. lbytes=%d", tb->lbytes);
 401                                                 RFALSE(tb->pos_in_item != ih_item_len(item_head(tbS0, tb->item_pos)),
 402                                                        "PAP-12100: incorrect position to paste: item_len=%d, pos_in_item=%d",
 403                                                        ih_item_len(item_head(tbS0, tb->item_pos)), tb->pos_in_item);
 404
 405                                                 if (tb->lbytes >= tb->pos_in_item) {
 406                                                         /* appended item will be in L[0] in whole */
 407                                                         int l_n;
 408
 409                                                         /* this bytes number must be appended to the last item of L[h] */
 410                                                         l_n = tb->lbytes - tb->pos_in_item;
 411
 412                                                         /* Calculate new insert_size[0] */
 413                                                         tb->insert_size[0] -= l_n;
 414
 415                                                         RFALSE(tb->insert_size[0] <= 0,
 416                                                                "PAP-12105: there is nothing to paste into L[0]. insert_size=%d",
 417                                                                tb->insert_size[0]);
 418                                                         ret_val = leaf_shift_left(tb, tb->lnum[0], ih_item_len
 419                                                                             (item_head(tbS0, tb->item_pos)));
 420                                                         /* Append to body of item in L[0] */
 421                                                         buffer_info_init_left(tb, &bi);
 422                                                         leaf_paste_in_buffer
 423                                                             (&bi, n + tb->item_pos - ret_val, ih_item_len
 424                                                              (item_head(tb->L[0], n + tb->item_pos - ret_val)),
 425                                                              l_n, body,
 426                                                              tb->zeroes_num > l_n ? l_n : tb->zeroes_num);
 427                                                         /* 0-th item in S0 can be only of DIRECT type when l_n != 0 */
 428                                                         {
 429                                                                 int version;
 430                                                                 int temp_l = l_n;
 431
 432                                                                 RFALSE(ih_item_len(item_head(tbS0, 0)),
 433                                                                      "PAP-12106: item length must be 0");
 434                                                                 RFALSE(comp_short_le_keys(leaf_key(tbS0, 0), leaf_key
 435                                                                       (tb->L[0], n + tb->item_pos - ret_val)),
 436                                                                      "PAP-12107: items must be of the same file");
 437                                                                 if (is_indirect_le_ih(item_head(tb->L[0], n + tb->item_pos - ret_val))) {
 438                                                                         temp_l = l_n << (tb->tb_sb-> s_blocksize_bits - UNFM_P_SHIFT);
 439                                                                 }
 440                                                                 /* update key of first item in S0 */
 441                                                                 version = ih_version(item_head(tbS0, 0));
 442                                                                 set_le_key_k_offset(version, leaf_key(tbS0, 0),
 443                                                                      le_key_k_offset(version,leaf_key(tbS0, 0)) + temp_l);
 444                                                                 /* update left delimiting key */
 445                                                                 set_le_key_k_offset(version, internal_key(tb->CFL[0], tb->lkey[0]),
 446                                                                      le_key_k_offset(version, internal_key(tb->CFL[0], tb->lkey[0])) + temp_l);
 447                                                         }
 448
 449                                                         /* Calculate new body, position in item and insert_size[0] */
 450                                                         if (l_n > tb->zeroes_num) {
 451                                                                 body += (l_n - tb->zeroes_num);
 452                                                                 tb->zeroes_num = 0;
 453                                                         } else
 454                                                                 tb->zeroes_num -= l_n;
 455                                                         tb->pos_in_item = 0;
 456
 457                                                         RFALSE(comp_short_le_keys(leaf_key(tbS0, 0), leaf_key(tb->L[0], B_NR_ITEMS(tb->L[0]) - 1))
 458                                                              || !op_is_left_mergeable(leaf_key(tbS0, 0), tbS0->b_size)
 459                                                              || !op_is_left_mergeable(internal_key(tb->CFL[0], tb->lkey[0]), tbS0->b_size),
 460                                                              "PAP-12120: item must be merge-able with left neighboring item");
 461                                                 } else {        /* only part of the appended item will be in L[0] */
 462
 463                                                         /* Calculate position in item for append in S[0] */
 464                                                         tb->pos_in_item -= tb->lbytes;
 465
 466                                                         RFALSE(tb->pos_in_item <= 0, "PAP-12125: no place for paste. pos_in_item=%d", tb->pos_in_item);
 467
 468                                                         /* Shift lnum[0] - 1 items in whole. Shift lbytes - 1 byte from item number lnum[0] */
 469                                                         leaf_shift_left(tb, tb->lnum[0], tb->lbytes);
 470                                                 }
 471                                         }
 472                                 } else {        /* appended item will be in L[0] in whole */
 473
 474                                         struct item_head *pasted;
 475
 476                                         if (!tb->item_pos && op_is_left_mergeable(leaf_key(tbS0, 0), tbS0->b_size)) {   /* if we paste into first item of S[0] and it is left mergable */
 477                                                 /* then increment pos_in_item by the size of the last item in L[0] */
 478                                                 pasted = item_head(tb->L[0], n - 1);
 479                                                 if (is_direntry_le_ih(pasted))
 480                                                         tb->pos_in_item += ih_entry_count(pasted);
 481                                                 else
 482                                                         tb->pos_in_item += ih_item_len(pasted);
 483                                         }
 484
 485                                         /* Shift lnum[0] - 1 items in whole. Shift lbytes - 1 byte from item number lnum[0] */
 486                                         ret_val = leaf_shift_left(tb, tb->lnum[0], tb->lbytes);
 487                                         /* Append to body of item in L[0] */
 488                                         buffer_info_init_left(tb, &bi);
 489                                         leaf_paste_in_buffer(&bi, n + tb->item_pos - ret_val,
 490                                                              tb->pos_in_item,
 491                                                              tb->insert_size[0],
 492                                                              body, tb->zeroes_num);
 493
 494                                         /* if appended item is directory, paste entry */
 495                                         pasted = item_head(tb->L[0], n + tb->item_pos - ret_val);
 496                                         if (is_direntry_le_ih(pasted))
 497                                                 leaf_paste_entries(&bi, n + tb->item_pos - ret_val,
 498                                                                    tb->pos_in_item, 1,
 499                                                                    (struct reiserfs_de_head *) body,
 500                                                                    body + DEH_SIZE,
 501                                                                    tb->insert_size[0]);
 502                                         /* if appended item is indirect item, put unformatted node into un list */
 503                                         if (is_indirect_le_ih(pasted))
 504                                                 set_ih_free_space(pasted, 0);
 505                                         tb->insert_size[0] = 0;
 506                                         tb->zeroes_num = 0;
 507                                 }
 508
 509 }
 510
 511 /* Shift lnum[0] items from S[0] to the left neighbor L[0] */
 512 static void balance_leaf_left(struct tree_balance *tb, struct item_head *ih,
 513                               const char *body, int flag)
 514 {
 515         if (tb->lnum[0] <= 0)
 516                 return;
 517
 518         /* new item or it part falls to L[0], shift it too */
 519         if (tb->item_pos < tb->lnum[0]) {
 520                 BUG_ON(flag != M_INSERT && flag != M_PASTE);
 521
 522                 if (flag == M_INSERT)
 523                         balance_leaf_insert_left(tb, ih, body);
 524                 else /* M_PASTE */
 525                         balance_leaf_paste_left(tb, ih, body);
 526         } else
 527                 /* new item doesn't fall into L[0] */
 528                 leaf_shift_left(tb, tb->lnum[0], tb->lbytes);
 529 }
 530
 531
 532 static void balance_leaf_insert_right(struct tree_balance *tb,
 533                                       struct item_head *ih, const char *body)
 534 {
 535
 536         struct buffer_head *tbS0 = PATH_PLAST_BUFFER(tb->tb_path);
 537         int n = B_NR_ITEMS(tbS0);
 538         struct buffer_info bi;
 539         int ret_val;
 540                         if (n - tb->rnum[0] < tb->item_pos) {   /* new item or its part falls to R[0] */
 541                                 if (tb->item_pos == n - tb->rnum[0] + 1 && tb->rbytes != -1) {  /* part of new item falls into R[0] */
 542                                         loff_t old_key_comp, old_len, r_zeroes_number;
 543                                         const char *r_body;
 544                                         int version;
 545                                         loff_t offset;
 546
 547                                         leaf_shift_right(tb, tb->rnum[0] - 1, -1);
 548
 549                                         version = ih_version(ih);
 550                                         /* Remember key component and item length */
 551                                         old_key_comp = le_ih_k_offset(ih);
 552                                         old_len = ih_item_len(ih);
 553
 554                                         /* Calculate key component and item length to insert into R[0] */
 555                                         offset = le_ih_k_offset(ih) + ((old_len - tb->rbytes) << (is_indirect_le_ih(ih) ? tb->tb_sb->s_blocksize_bits - UNFM_P_SHIFT : 0));
 556                                         set_le_ih_k_offset(ih, offset);
 557                                         put_ih_item_len(ih, tb->rbytes);
 558                                         /* Insert part of the item into R[0] */
 559                                         buffer_info_init_right(tb, &bi);
 560                                         if ((old_len - tb->rbytes) > tb->zeroes_num) {
 561                                                 r_zeroes_number = 0;
 562                                                 r_body = body + (old_len - tb->rbytes) - tb->zeroes_num;
 563                                         } else {
 564                                                 r_body = body;
 565                                                 r_zeroes_number = tb->zeroes_num - (old_len - tb->rbytes);
 566                                                 tb->zeroes_num -= r_zeroes_number;
 567                                         }
 568
 569                                         leaf_insert_into_buf(&bi, 0, ih, r_body,
 570                                                              r_zeroes_number);
 571
 572                                         /* Replace right delimiting key by first key in R[0] */
 573                                         replace_key(tb, tb->CFR[0], tb->rkey[0],
 574                                                     tb->R[0], 0);
 575
 576                                         /* Calculate key component and item length to insert into S[0] */
 577                                         set_le_ih_k_offset(ih, old_key_comp);
 578                                         put_ih_item_len(ih, old_len - tb->rbytes);
 579
 580                                         tb->insert_size[0] -= tb->rbytes;
 581
 582                                 } else {        /* whole new item falls into R[0] */
 583
 584                                         /* Shift rnum[0]-1 items to R[0] */
 585                                         ret_val = leaf_shift_right(tb, tb->rnum[0] - 1, tb->rbytes);
 586                                         /* Insert new item into R[0] */
 587                                         buffer_info_init_right(tb, &bi);
 588                                         leaf_insert_into_buf(&bi, tb->item_pos - n + tb->rnum[0] - 1,
 589                                                              ih, body, tb->zeroes_num);
 590
 591                                         if (tb->item_pos - n + tb->rnum[0] - 1 == 0) {
 592                                                 replace_key(tb, tb->CFR[0],
 593                                                             tb->rkey[0],
 594                                                             tb->R[0], 0);
 595
 596                                         }
 597                                         tb->zeroes_num = tb->insert_size[0] = 0;
 598                                 }
 599                         } else {        /* new item or part of it doesn't fall into R[0] */
 600
 601                                 leaf_shift_right(tb, tb->rnum[0], tb->rbytes);
 602                         }
 603
 604 }
 605
 606 static void balance_leaf_paste_right(struct tree_balance *tb,
 607                                      struct item_head *ih, const char *body)
 608 {
 609         struct buffer_head *tbS0 = PATH_PLAST_BUFFER(tb->tb_path);
 610         int n = B_NR_ITEMS(tbS0);
 611         struct buffer_info bi;
 612         int ret_val;
 613
 614                         if (n - tb->rnum[0] <= tb->item_pos) {  /* pasted item or part of it falls to R[0] */
 615                                 if (tb->item_pos == n - tb->rnum[0] && tb->rbytes != -1) {      /* we must shift the part of the appended item */
 616                                         if (is_direntry_le_ih(item_head(tbS0, tb->item_pos))) { /* we append to directory item */
 617                                                 int entry_count;
 618
 619                                                 RFALSE(tb->zeroes_num,
 620                                                        "PAP-12145: invalid parameter in case of a directory");
 621                                                 entry_count = ih_entry_count(item_head
 622                                                                   (tbS0, tb->item_pos));
 623                                                 if (entry_count - tb->rbytes <
 624                                                     tb->pos_in_item)
 625                                                         /* new directory entry falls into R[0] */
 626                                                 {
 627                                                         int paste_entry_position;
 628
 629                                                         RFALSE(tb->rbytes - 1 >= entry_count || !tb-> insert_size[0],
 630                                                                "PAP-12150: no enough of entries to shift to R[0]: rbytes=%d, entry_count=%d",
 631                                                                tb->rbytes, entry_count);
 632                                                         /* Shift rnum[0]-1 items in whole. Shift rbytes-1 directory entries from directory item number rnum[0] */
 633                                                         leaf_shift_right(tb, tb->rnum[0], tb->rbytes - 1);
 634                                                         /* Paste given directory entry to directory item */
 635                                                         paste_entry_position = tb->pos_in_item - entry_count + tb->rbytes - 1;
 636                                                         buffer_info_init_right(tb, &bi);
 637                                                         leaf_paste_in_buffer(&bi, 0, paste_entry_position, tb->insert_size[0], body, tb->zeroes_num);
 638                                                         /* paste entry */
 639                                                         leaf_paste_entries(&bi, 0, paste_entry_position, 1,
 640                                                                            (struct reiserfs_de_head *) body,
 641                                                                            body + DEH_SIZE, tb->insert_size[0]);
 642
 643                                                         if (paste_entry_position == 0) {
 644                                                                 /* change delimiting keys */
 645                                                                 replace_key(tb, tb->CFR[0], tb->rkey[0], tb->R[0],0);
 646                                                         }
 647
 648                                                         tb->insert_size[0] = 0;
 649                                                         tb->pos_in_item++;
 650                                                 } else {        /* new directory entry doesn't fall into R[0] */
 651
 652                                                         leaf_shift_right(tb, tb->rnum[0], tb->rbytes);
 653                                                 }
 654                                         } else {        /* regular object */
 655
 656                                                 int n_shift, n_rem, r_zeroes_number;
 657                                                 const char *r_body;
 658
 659                                                 /* Calculate number of bytes which must be shifted from appended item */
 660                                                 if ((n_shift = tb->rbytes - tb->insert_size[0]) < 0)
 661                                                         n_shift = 0;
 662
 663                                                 RFALSE(tb->pos_in_item != ih_item_len
 664                                                        (item_head(tbS0, tb->item_pos)),
 665                                                        "PAP-12155: invalid position to paste. ih_item_len=%d, pos_in_item=%d",
 666                                                        tb->pos_in_item, ih_item_len
 667                                                        (item_head(tbS0, tb->item_pos)));
 668
 669                                                 leaf_shift_right(tb, tb->rnum[0], n_shift);
 670                                                 /* Calculate number of bytes which must remain in body after appending to R[0] */
 671                                                 if ((n_rem = tb->insert_size[0] - tb->rbytes) < 0)
 672                                                         n_rem = 0;
 673
 674                                                 {
 675                                                         int version;
 676                                                         unsigned long temp_rem = n_rem;
 677
 678                                                         version = ih_version(item_head(tb->R[0], 0));
 679                                                         if (is_indirect_le_key(version, leaf_key(tb->R[0], 0))) {
 680                                                                 temp_rem = n_rem << (tb->tb_sb->s_blocksize_bits - UNFM_P_SHIFT);
 681                                                         }
 682                                                         set_le_key_k_offset(version, leaf_key(tb->R[0], 0),
 683                                                              le_key_k_offset(version, leaf_key(tb->R[0], 0)) + temp_rem);
 684                                                         set_le_key_k_offset(version, internal_key(tb->CFR[0], tb->rkey[0]),
 685                                                              le_key_k_offset(version, internal_key(tb->CFR[0], tb->rkey[0])) + temp_rem);
 686                                                 }
 687 /*                k_offset (leaf_key(tb->R[0],0)) += n_rem;
 688                   k_offset (internal_key(tb->CFR[0],tb->rkey[0])) += n_rem;*/
 689                                                 do_balance_mark_internal_dirty(tb, tb->CFR[0], 0);
 690
 691                                                 /* Append part of body into R[0] */
 692                                                 buffer_info_init_right(tb, &bi);
 693                                                 if (n_rem > tb->zeroes_num) {
 694                                                         r_zeroes_number = 0;
 695                                                         r_body = body + n_rem - tb->zeroes_num;
 696                                                 } else {
 697                                                         r_body = body;
 698                                                         r_zeroes_number = tb->zeroes_num - n_rem;
 699                                                         tb->zeroes_num -= r_zeroes_number;
 700                                                 }
 701
 702                                                 leaf_paste_in_buffer(&bi, 0, n_shift,
 703                                                                      tb->insert_size[0] - n_rem,
 704                                                                      r_body, r_zeroes_number);
 705
 706                                                 if (is_indirect_le_ih(item_head(tb->R[0], 0))) {
 707 #if 0
 708                                                         RFALSE(n_rem,
 709                                                                "PAP-12160: paste more than one unformatted node pointer");
 710 #endif
 711                                                         set_ih_free_space(item_head(tb->R[0], 0), 0);
 712                                                 }
 713                                                 tb->insert_size[0] = n_rem;
 714                                                 if (!n_rem)
 715                                                         tb->pos_in_item++;
 716                                         }
 717                                 } else {        /* pasted item in whole falls into R[0] */
 718
 719                                         struct item_head *pasted;
 720
 721                                         ret_val = leaf_shift_right(tb, tb->rnum[0], tb->rbytes);
 722                                         /* append item in R[0] */
 723                                         if (tb->pos_in_item >= 0) {
 724                                                 buffer_info_init_right(tb, &bi);
 725                                                 leaf_paste_in_buffer(&bi, tb->item_pos - n + tb->rnum[0], tb->pos_in_item,
 726                                                                      tb->insert_size[0], body, tb->zeroes_num);
 727                                         }
 728
 729                                         /* paste new entry, if item is directory item */
 730                                         pasted = item_head(tb->R[0], tb->item_pos - n + tb->rnum[0]);
 731                                         if (is_direntry_le_ih(pasted) && tb->pos_in_item >= 0) {
 732                                                 leaf_paste_entries(&bi, tb->item_pos - n + tb->rnum[0],
 733                                                                    tb->pos_in_item, 1,
 734                                                                    (struct reiserfs_de_head *) body,
 735                                                                    body + DEH_SIZE, tb->insert_size[0]);
 736                                                 if (!tb->pos_in_item) {
 737
 738                                                         RFALSE(tb->item_pos - n + tb->rnum[0],
 739                                                                "PAP-12165: directory item must be first item of node when pasting is in 0th position");
 740
 741                                                         /* update delimiting keys */
 742                                                         replace_key(tb, tb->CFR[0], tb->rkey[0], tb->R[0], 0);
 743                                                 }
 744                                         }
 745
 746                                         if (is_indirect_le_ih(pasted))
 747                                                 set_ih_free_space(pasted, 0);
 748                                         tb->zeroes_num = tb->insert_size[0] = 0;
 749                                 }
 750                         } else {        /* new item doesn't fall into R[0] */
 751
 752                                 leaf_shift_right(tb, tb->rnum[0], tb->rbytes);
 753                         }
 754
 755 }
 756
 757 /* shift rnum[0] items from S[0] to the right neighbor R[0] */
 758 static void balance_leaf_right(struct tree_balance *tb, struct item_head *ih,
 759                                const char *body, int flag)
 760 {
 761         if (tb->rnum[0] <= 0)
 762                 return;
 763
 764         BUG_ON(flag != M_INSERT && flag != M_PASTE);
 765
 766         if (flag == M_INSERT)
 767                 balance_leaf_insert_right(tb, ih, body);
 768         else /* M_PASTE */
 769                 balance_leaf_paste_right(tb, ih, body);
 770
 771 }
 772
 773 static void balance_leaf_new_nodes_insert(struct tree_balance *tb,
 774                                           struct item_head *ih,
 775                                           const char *body,
 776                                           struct item_head *insert_key,
 777                                           struct buffer_head **insert_ptr,
 778                                           int i)
 779 {
 780         struct buffer_head *tbS0 = PATH_PLAST_BUFFER(tb->tb_path);
 781         int n = B_NR_ITEMS(tbS0);
 782         struct buffer_info bi;
 783                         if (n - tb->snum[i] < tb->item_pos) {   /* new item or it's part falls to first new node S_new[i] */
 784                                 if (tb->item_pos == n - tb->snum[i] + 1 && tb->sbytes[i] != -1) {       /* part of new item falls into S_new[i] */
 785                                         int old_key_comp, old_len, r_zeroes_number;
 786                                         const char *r_body;
 787                                         int version;
 788
 789                                         /* Move snum[i]-1 items from S[0] to S_new[i] */
 790                                         leaf_move_items(LEAF_FROM_S_TO_SNEW, tb,
 791                                                         tb->snum[i] - 1, -1,
 792                                                         tb->S_new[i]);
 793                                         /* Remember key component and item length */
 794                                         version = ih_version(ih);
 795                                         old_key_comp = le_ih_k_offset(ih);
 796                                         old_len = ih_item_len(ih);
 797
 798                                         /* Calculate key component and item length to insert into S_new[i] */
 799                                         set_le_ih_k_offset(ih, le_ih_k_offset(ih) +
 800                                                            ((old_len - tb->sbytes[i]) << (is_indirect_le_ih(ih) ? tb->tb_sb->s_blocksize_bits - UNFM_P_SHIFT : 0)));
 801
 802                                         put_ih_item_len(ih, tb->sbytes[i]);
 803
 804                                         /* Insert part of the item into S_new[i] before 0-th item */
 805                                         buffer_info_init_bh(tb, &bi, tb->S_new[i]);
 806
 807                                         if ((old_len - tb->sbytes[i]) > tb->zeroes_num) {
 808                                                 r_zeroes_number = 0;
 809                                                 r_body = body + (old_len - tb->sbytes[i]) - tb->zeroes_num;
 810                                         } else {
 811                                                 r_body = body;
 812                                                 r_zeroes_number = tb->zeroes_num - (old_len - tb->sbytes[i]);
 813                                                 tb->zeroes_num -= r_zeroes_number;
 814                                         }
 815
 816                                         leaf_insert_into_buf(&bi, 0, ih, r_body, r_zeroes_number);
 817
 818                                         /* Calculate key component and item length to insert into S[i] */
 819                                         set_le_ih_k_offset(ih, old_key_comp);
 820                                         put_ih_item_len(ih, old_len - tb->sbytes[i]);
 821                                         tb->insert_size[0] -= tb->sbytes[i];
 822                                 } else {        /* whole new item falls into S_new[i] */
 823
 824                                         /* Shift snum[0] - 1 items to S_new[i] (sbytes[i] of split item) */
 825                                         leaf_move_items(LEAF_FROM_S_TO_SNEW, tb,
 826                                                         tb->snum[i] - 1, tb->sbytes[i], tb->S_new[i]);
 827
 828                                         /* Insert new item into S_new[i] */
 829                                         buffer_info_init_bh(tb, &bi, tb->S_new[i]);
 830                                         leaf_insert_into_buf(&bi, tb->item_pos - n + tb->snum[i] - 1,
 831                                                              ih, body, tb->zeroes_num);
 832
 833                                         tb->zeroes_num = tb->insert_size[0] = 0;
 834                                 }
 835                         }
 836
 837                         else {  /* new item or it part don't falls into S_new[i] */
 838
 839                                 leaf_move_items(LEAF_FROM_S_TO_SNEW, tb,
 840                                                 tb->snum[i], tb->sbytes[i], tb->S_new[i]);
 841                         }
 842 }
 843
 844 static void balance_leaf_new_nodes_paste(struct tree_balance *tb,
 845                                          struct item_head *ih,
 846                                          const char *body,
 847                                          struct item_head *insert_key,
 848                                          struct buffer_head **insert_ptr,
 849                                          int i)
 850 {
 851         struct buffer_head *tbS0 = PATH_PLAST_BUFFER(tb->tb_path);
 852         int n = B_NR_ITEMS(tbS0);
 853         struct buffer_info bi;
 854                         if (n - tb->snum[i] <= tb->item_pos) {  /* pasted item or part if it falls to S_new[i] */
 855                                 if (tb->item_pos == n - tb->snum[i] && tb->sbytes[i] != -1) {   /* we must shift part of the appended item */
 856                                         struct item_head *aux_ih;
 857
 858                                         RFALSE(ih, "PAP-12210: ih must be 0");
 859
 860                                         aux_ih = item_head(tbS0, tb->item_pos);
 861                                         if (is_direntry_le_ih(aux_ih)) {
 862                                                 /* we append to directory item */
 863
 864                                                 int entry_count;
 865
 866                                                 entry_count = ih_entry_count(aux_ih);
 867
 868                                                 if (entry_count - tb->sbytes[i] < tb->pos_in_item && tb->pos_in_item <= entry_count) {
 869                                                         /* new directory entry falls into S_new[i] */
 870
 871                                                         RFALSE(!tb->insert_size[0], "PAP-12215: insert_size is already 0");
 872                                                         RFALSE(tb->sbytes[i] - 1 >= entry_count,
 873                                                                "PAP-12220: there are no so much entries (%d), only %d",
 874                                                                tb->sbytes[i] - 1, entry_count);
 875
 876                                                         /* Shift snum[i]-1 items in whole. Shift sbytes[i] directory entries from directory item number snum[i] */
 877                                                         leaf_move_items(LEAF_FROM_S_TO_SNEW, tb, tb->snum[i], tb->sbytes[i] - 1, tb->S_new[i]);
 878                                                         /* Paste given directory entry to directory item */
 879                                                         buffer_info_init_bh(tb, &bi, tb->S_new[i]);
 880                                                         leaf_paste_in_buffer(&bi, 0, tb->pos_in_item - entry_count + tb->sbytes[i] - 1,
 881                                                              tb->insert_size[0], body, tb->zeroes_num);
 882                                                         /* paste new directory entry */
 883                                                         leaf_paste_entries(&bi, 0, tb->pos_in_item - entry_count + tb->sbytes[i] - 1, 1,
 884                                                                            (struct reiserfs_de_head *) body,
 885                                                                            body + DEH_SIZE, tb->insert_size[0]);
 886                                                         tb->insert_size[0] = 0;
 887                                                         tb->pos_in_item++;
 888                                                 } else {        /* new directory entry doesn't fall into S_new[i] */
 889                                                         leaf_move_items(LEAF_FROM_S_TO_SNEW, tb, tb->snum[i], tb->sbytes[i], tb->S_new[i]);
 890                                                 }
 891                                         } else {        /* regular object */
 892
 893                                                 int n_shift, n_rem, r_zeroes_number;
 894                                                 const char *r_body;
 895
 896                                                 RFALSE(tb->pos_in_item != ih_item_len(item_head(tbS0, tb->item_pos)) || tb->insert_size[0] <= 0,
 897                                                        "PAP-12225: item too short or insert_size <= 0");
 898
 899                                                 /* Calculate number of bytes which must be shifted from appended item */
 900                                                 n_shift = tb->sbytes[i] - tb->insert_size[0];
 901                                                 if (n_shift < 0)
 902                                                         n_shift = 0;
 903                                                 leaf_move_items(LEAF_FROM_S_TO_SNEW, tb, tb->snum[i], n_shift, tb->S_new[i]);
 904
 905                                                 /* Calculate number of bytes which must remain in body after append to S_new[i] */
 906                                                 n_rem = tb->insert_size[0] - tb->sbytes[i];
 907                                                 if (n_rem < 0)
 908                                                         n_rem = 0;
 909                                                 /* Append part of body into S_new[0] */
 910                                                 buffer_info_init_bh(tb, &bi, tb->S_new[i]);
 911                                                 if (n_rem > tb->zeroes_num) {
 912                                                         r_zeroes_number = 0;
 913                                                         r_body = body + n_rem - tb->zeroes_num;
 914                                                 } else {
 915                                                         r_body = body;
 916                                                         r_zeroes_number = tb->zeroes_num - n_rem;
 917                                                         tb->zeroes_num -= r_zeroes_number;
 918                                                 }
 919
 920                                                 leaf_paste_in_buffer(&bi, 0, n_shift,
 921                                                                      tb->insert_size[0] - n_rem,
 922                                                                      r_body, r_zeroes_number);
 923                                                 {
 924                                                         struct item_head *tmp;
 925
 926                                                         tmp = item_head(tb->S_new[i], 0);
 927                                                         if (is_indirect_le_ih
 928                                                             (tmp)) {
 929                                                                 set_ih_free_space(tmp, 0);
 930                                                                 set_le_ih_k_offset(tmp, le_ih_k_offset(tmp) + (n_rem << (tb->tb_sb->s_blocksize_bits - UNFM_P_SHIFT)));
 931                                                         } else {
 932                                                                 set_le_ih_k_offset(tmp, le_ih_k_offset(tmp) + n_rem);
 933                                                         }
 934                                                 }
 935
 936                                                 tb->insert_size[0] = n_rem;
 937                                                 if (!n_rem)
 938                                                         tb->pos_in_item++;
 939                                         }
 940                                 } else
 941                                         /* item falls wholly into S_new[i] */
 942                                 {
 943                                         int leaf_mi;
 944                                         struct item_head *pasted;
 945
 946 #ifdef CONFIG_REISERFS_CHECK
 947                                         struct item_head *ih_check = item_head(tbS0, tb->item_pos);
 948
 949                                         if (!is_direntry_le_ih(ih_check)
 950                                             && (tb->pos_in_item != ih_item_len(ih_check)
 951                                                 || tb->insert_size[0] <= 0))
 952                                                 reiserfs_panic(tb->tb_sb,
 953                                                              "PAP-12235",
 954                                                              "pos_in_item "
 955                                                              "must be equal "
 956                                                              "to ih_item_len");
 957 #endif                          /* CONFIG_REISERFS_CHECK */
 958
 959                                         leaf_mi = leaf_move_items(LEAF_FROM_S_TO_SNEW,
 960                                                             tb, tb->snum[i],
 961                                                             tb->sbytes[i],
 962                                                             tb->S_new[i]);
 963
 964                                         RFALSE(leaf_mi,
 965                                                "PAP-12240: unexpected value returned by leaf_move_items (%d)",
 966                                                leaf_mi);
 967
 968                                         /* paste into item */
 969                                         buffer_info_init_bh(tb, &bi, tb->S_new[i]);
 970                                         leaf_paste_in_buffer(&bi,
 971                                                              tb->item_pos - n + tb->snum[i],
 972                                                              tb->pos_in_item,
 973                                                              tb->insert_size[0],
 974                                                              body, tb->zeroes_num);
 975
 976                                         pasted = item_head(tb->S_new[i], tb->item_pos - n + tb->snum[i]);
 977                                         if (is_direntry_le_ih(pasted)) {
 978                                                 leaf_paste_entries(&bi,
 979                                                                    tb->item_pos - n + tb->snum[i],
 980                                                                    tb->pos_in_item, 1,
 981                                                                    (struct reiserfs_de_head *)body,
 982                                                                    body + DEH_SIZE,
 983                                                                    tb->insert_size[0]
 984                                                     );
 985                                         }
 986
 987                                         /* if we paste to indirect item update ih_free_space */
 988                                         if (is_indirect_le_ih(pasted))
 989                                                 set_ih_free_space(pasted, 0);
 990                                         tb->zeroes_num = tb->insert_size[0] = 0;
 991                                 }
 992                         }
 993
 994                         else {  /* pasted item doesn't fall into S_new[i] */
 995
 996                                 leaf_move_items(LEAF_FROM_S_TO_SNEW, tb,
 997                                                 tb->snum[i], tb->sbytes[i], tb->S_new[i]);
 998                         }
 999
1000 }
1001
1002 /* Fill new nodes that appear in place of S[0] */
1003 static void balance_leaf_new_nodes(struct tree_balance *tb,
1004                                    struct item_head *ih,
1005                                    const char *body,
1006                                    struct item_head *insert_key,
1007                                    struct buffer_head **insert_ptr,
1008                                    int flag)
1009 {
1010         int i;
1011         for (i = tb->blknum[0] - 2; i >= 0; i--) {
1012
1013                 RFALSE(!tb->snum[i],
1014                        "PAP-12200: snum[%d] == %d. Must be > 0", i,
1015                        tb->snum[i]);
1016
1017                 /* here we shift from S to S_new nodes */
1018
1019                 tb->S_new[i] = get_FEB(tb);
1020
1021                 /* initialized block type and tree level */
1022                 set_blkh_level(B_BLK_HEAD(tb->S_new[i]), DISK_LEAF_NODE_LEVEL);
1023
1024                 switch (flag) {
1025                 case M_INSERT:  /* insert item */
1026                         balance_leaf_new_nodes_insert(tb, ih, body, insert_key,
1027                                                       insert_ptr, i);
1028                         break;
1029
1030                 case M_PASTE:   /* append item */
1031                         balance_leaf_new_nodes_paste(tb, ih, body, insert_key,
1032                                                      insert_ptr, i);
1033                         break;
1034                 default:        /* cases d and t */
1035                         reiserfs_panic(tb->tb_sb, "PAP-12245",
1036                                        "blknum > 2: unexpected mode: %s(%d)",
1037                                        (flag == M_DELETE) ? "DELETE" : ((flag == M_CUT) ? "CUT" : "UNKNOWN"), flag);
1038                 }
1039
1040                 memcpy(insert_key + i, leaf_key(tb->S_new[i], 0), KEY_SIZE);
1041                 insert_ptr[i] = tb->S_new[i];
1042
1043                 RFALSE(!buffer_journaled(tb->S_new[i])
1044                        || buffer_journal_dirty(tb->S_new[i])
1045                        || buffer_dirty(tb->S_new[i]),
1046                        "PAP-12247: S_new[%d] : (%b)",
1047                        i, format_bh(tb->S_new[i]));
1048         }
1049 }
1050
1051 static void balance_leaf_finish_node_insert(struct tree_balance *tb,
1052                                             struct item_head *ih,
1053                                             const char *body)
1054 {
1055         struct buffer_head *tbS0 = PATH_PLAST_BUFFER(tb->tb_path);
1056         struct buffer_info bi;
1057                         buffer_info_init_tbS0(tb, &bi);
1058                         leaf_insert_into_buf(&bi, tb->item_pos, ih,
1059                                              body, tb->zeroes_num);
1060
1061                         /*
1062                          * If we insert the first key
1063                          * change the delimiting key
1064                          */
1065                         if (tb->item_pos == 0) {
1066                                 if (tb->CFL[0]) /* can be 0 in reiserfsck */
1067                                         replace_key(tb, tb->CFL[0], tb->lkey[0], tbS0, 0);
1068                         }
1069 }
1070
1071 static void balance_leaf_finish_node_paste(struct tree_balance *tb,
1072                                            struct item_head *ih,
1073                                            const char *body)
1074 {
1075         struct buffer_head *tbS0 = PATH_PLAST_BUFFER(tb->tb_path);
1076         struct buffer_info bi;
1077                                 struct item_head *pasted;
1078
1079                                 pasted = item_head(tbS0, tb->item_pos);
1080                                 /* when directory, may be new entry already pasted */
1081                                 if (is_direntry_le_ih(pasted)) {
1082                                         if (tb->pos_in_item >= 0 && tb->pos_in_item <= ih_entry_count(pasted)) {
1083
1084                                                 RFALSE(!tb->insert_size[0],
1085                                                        "PAP-12260: insert_size is 0 already");
1086
1087                                                 /* prepare space */
1088                                                 buffer_info_init_tbS0(tb, &bi);
1089                                                 leaf_paste_in_buffer(&bi, tb->item_pos, tb->pos_in_item,
1090                                                                      tb->insert_size[0], body,
1091                                                                      tb->zeroes_num);
1092
1093                                                 /* paste entry */
1094                                                 leaf_paste_entries(&bi, tb->item_pos, tb->pos_in_item, 1,
1095                                                                    (struct reiserfs_de_head *)body,
1096                                                                    body + DEH_SIZE,
1097                                                                    tb->insert_size[0]);
1098                                                 if (!tb->item_pos && !tb->pos_in_item) {
1099                                                         RFALSE(!tb->CFL[0] || !tb->L[0],
1100                                                                "PAP-12270: CFL[0]/L[0] must be specified");
1101                                                         if (tb->CFL[0])
1102                                                                 replace_key(tb, tb->CFL[0], tb->lkey[0], tbS0, 0);
1103                                                 }
1104                                                 tb->insert_size[0] = 0;
1105                                         }
1106                                 } else {        /* regular object */
1107                                         if (tb->pos_in_item == ih_item_len(pasted)) {
1108
1109                                                 RFALSE(tb->insert_size[0] <= 0,
1110                                                        "PAP-12275: insert size must not be %d",
1111                                                        tb->insert_size[0]);
1112                                                 buffer_info_init_tbS0(tb, &bi);
1113                                                 leaf_paste_in_buffer(&bi, tb->item_pos, tb->pos_in_item,
1114                                                                      tb->insert_size[0], body, tb->zeroes_num);
1115
1116                                                 if (is_indirect_le_ih(pasted)) {
1117 #if 0
1118                                                         RFALSE(tb->
1119                                                                insert_size[0] !=
1120                                                                UNFM_P_SIZE,
1121                                                                "PAP-12280: insert_size for indirect item must be %d, not %d",
1122                                                                UNFM_P_SIZE,
1123                                                                tb->
1124                                                                insert_size[0]);
1125 #endif
1126                                                         set_ih_free_space(pasted, 0);
1127                                                 }
1128                                                 tb->insert_size[0] = 0;
1129                                         }
1130 #ifdef CONFIG_REISERFS_CHECK
1131                                         else {
1132                                                 if (tb->insert_size[0]) {
1133                                                         print_cur_tb("12285");
1134                                                         reiserfs_panic(tb->tb_sb,
1135                                                             "PAP-12285",
1136                                                             "insert_size "
1137                                                             "must be 0 "
1138                                                             "(%d)",
1139                                                             tb->insert_size[0]);
1140                                                 }
1141                                         }
1142 #endif                          /* CONFIG_REISERFS_CHECK */
1143
1144                                 }
1145 }
1146
1147 /*
1148  * if the affected item was not wholly shifted then we
1149  * perform all necessary operations on that part or whole
1150  * of the affected item which remains in S
1151  */
1152 static void balance_leaf_finish_node(struct tree_balance *tb,
1153                                       struct item_head *ih,
1154                                       const char *body, int flag)
1155 {
1156         /* if we must insert or append into buffer S[0] */
1157         if (0 <= tb->item_pos && tb->item_pos < tb->s0num) {
1158                 if (flag == M_INSERT)
1159                         balance_leaf_finish_node_insert(tb, ih, body);
1160                 else /* M_PASTE */
1161                         balance_leaf_finish_node_paste(tb, ih, body);
1162         }
1163 }
1164
1165 /**
1166  * balance_leaf - reiserfs tree balancing algorithm
1167  * @tb: tree balance state
1168  * @ih: item header of inserted item (little endian)
1169  * @body: body of inserted item or bytes to paste
1170  * @flag: i - insert, d - delete, c - cut, p - paste (see do_balance)
1171  * passed back:
1172  * @insert_key: key to insert new nodes
1173  * @insert_ptr: array of nodes to insert at the next level
1174  *
1175  * In our processing of one level we sometimes determine what must be
1176  * inserted into the next higher level.  This insertion consists of a
1177  * key or two keys and their corresponding pointers.
1178  */
1179 static int balance_leaf(struct tree_balance *tb, struct item_head *ih,
1180                         const char *body, int flag,
1181                         struct item_head *insert_key,
1182                         struct buffer_head **insert_ptr)
1183 {
1184         struct buffer_head *tbS0 = PATH_PLAST_BUFFER(tb->tb_path);
1185
1186         PROC_INFO_INC(tb->tb_sb, balance_at[0]);
1187
1188         /* Make balance in case insert_size[0] < 0 */
1189         if (tb->insert_size[0] < 0)
1190                 return balance_leaf_when_delete(tb, flag);
1191
1192         tb->item_pos = PATH_LAST_POSITION(tb->tb_path),
1193         tb->pos_in_item = tb->tb_path->pos_in_item,
1194         tb->zeroes_num = 0;
1195         if (flag == M_INSERT && !body)
1196                 tb->zeroes_num = ih_item_len(ih);
1197
1198         /*
1199          * for indirect item pos_in_item is measured in unformatted node
1200          * pointers. Recalculate to bytes
1201          */
1202         if (flag != M_INSERT
1203             && is_indirect_le_ih(item_head(tbS0, tb->item_pos)))
1204                 tb->pos_in_item *= UNFM_P_SIZE;
1205
1206         balance_leaf_left(tb, ih, body, flag);
1207
1208         /* tb->lnum[0] > 0 */
1209         /* Calculate new item position */
1210         tb->item_pos -= (tb->lnum[0] - ((tb->lbytes != -1) ? 1 : 0));
1211
1212         balance_leaf_right(tb, ih, body, flag);
1213
1214         /* tb->rnum[0] > 0 */
1215         RFALSE(tb->blknum[0] > 3,
1216                "PAP-12180: blknum can not be %d. It must be <= 3", tb->blknum[0]);
1217         RFALSE(tb->blknum[0] < 0,
1218                "PAP-12185: blknum can not be %d. It must be >= 0", tb->blknum[0]);
1219
1220         /*
1221          * if while adding to a node we discover that it is possible to split
1222          * it in two, and merge the left part into the left neighbor and the
1223          * right part into the right neighbor, eliminating the node
1224          */
1225         if (tb->blknum[0] == 0) {       /* node S[0] is empty now */
1226
1227                 RFALSE(!tb->lnum[0] || !tb->rnum[0],
1228                        "PAP-12190: lnum and rnum must not be zero");
1229                 /*
1230                  * if insertion was done before 0-th position in R[0], right
1231                  * delimiting key of the tb->L[0]'s and left delimiting key are
1232                  * not set correctly
1233                  */
1234                 if (tb->CFL[0]) {
1235                         if (!tb->CFR[0])
1236                                 reiserfs_panic(tb->tb_sb, "vs-12195",
1237                                                "CFR not initialized");
1238                         copy_key(internal_key(tb->CFL[0], tb->lkey[0]),
1239                                  internal_key(tb->CFR[0], tb->rkey[0]));
1240                         do_balance_mark_internal_dirty(tb, tb->CFL[0], 0);
1241                 }
1242
1243                 reiserfs_invalidate_buffer(tb, tbS0);
1244                 return 0;
1245         }
1246
1247         balance_leaf_new_nodes(tb, ih, body, insert_key, insert_ptr, flag);
1248
1249         balance_leaf_finish_node(tb, ih, body, flag);
1250
1251 #ifdef CONFIG_REISERFS_CHECK
1252         if (flag == M_PASTE && tb->insert_size[0]) {
1253                 print_cur_tb("12290");
1254                 reiserfs_panic(tb->tb_sb,
1255                                "PAP-12290", "insert_size is still not 0 (%d)",
1256                                tb->insert_size[0]);
1257         }
1258 #endif
1259
1260         /* Leaf level of the tree is balanced (end of balance_leaf) */
1261         return 0;
1262 }
1263
1264 /* Make empty node */
1265 void make_empty_node(struct buffer_info *bi)
1266 {
1267         struct block_head *blkh;
1268
1269         RFALSE(bi->bi_bh == NULL, "PAP-12295: pointer to the buffer is NULL");
1270
1271         blkh = B_BLK_HEAD(bi->bi_bh);
1272         set_blkh_nr_item(blkh, 0);
1273         set_blkh_free_space(blkh, MAX_CHILD_SIZE(bi->bi_bh));
1274
1275         if (bi->bi_parent)
1276                 B_N_CHILD(bi->bi_parent, bi->bi_position)->dc_size = 0; /* Endian safe if 0 */
1277 }
1278
1279 /* Get first empty buffer */
1280 struct buffer_head *get_FEB(struct tree_balance *tb)
1281 {
1282         int i;
1283         struct buffer_info bi;
1284
1285         for (i = 0; i < MAX_FEB_SIZE; i++)
1286                 if (tb->FEB[i] != NULL)
1287                         break;
1288
1289         if (i == MAX_FEB_SIZE)
1290                 reiserfs_panic(tb->tb_sb, "vs-12300", "FEB list is empty");
1291
1292         buffer_info_init_bh(tb, &bi, tb->FEB[i]);
1293         make_empty_node(&bi);
1294         set_buffer_uptodate(tb->FEB[i]);
1295         tb->used[i] = tb->FEB[i];
1296         tb->FEB[i] = NULL;
1297
1298         return tb->used[i];
1299 }
1300
1301 /* This is now used because reiserfs_free_block has to be able to schedule. */
1302 static void store_thrown(struct tree_balance *tb, struct buffer_head *bh)
1303 {
1304         int i;
1305
1306         if (buffer_dirty(bh))
1307                 reiserfs_warning(tb->tb_sb, "reiserfs-12320",
1308                                  "called with dirty buffer");
1309         for (i = 0; i < ARRAY_SIZE(tb->thrown); i++)
1310                 if (!tb->thrown[i]) {
1311                         tb->thrown[i] = bh;
1312                         get_bh(bh);     /* free_thrown puts this */
1313                         return;
1314                 }
1315         reiserfs_warning(tb->tb_sb, "reiserfs-12321",
1316                          "too many thrown buffers");
1317 }
1318
1319 static void free_thrown(struct tree_balance *tb)
1320 {
1321         int i;
1322         b_blocknr_t blocknr;
1323         for (i = 0; i < ARRAY_SIZE(tb->thrown); i++) {
1324                 if (tb->thrown[i]) {
1325                         blocknr = tb->thrown[i]->b_blocknr;
1326                         if (buffer_dirty(tb->thrown[i]))
1327                                 reiserfs_warning(tb->tb_sb, "reiserfs-12322",
1328                                                  "called with dirty buffer %d",
1329                                                  blocknr);
1330                         brelse(tb->thrown[i]);  /* incremented in store_thrown */
1331                         reiserfs_free_block(tb->transaction_handle, NULL,
1332                                             blocknr, 0);
1333                 }
1334         }
1335 }
1336
1337 void reiserfs_invalidate_buffer(struct tree_balance *tb, struct buffer_head *bh)
1338 {
1339         struct block_head *blkh;
1340         blkh = B_BLK_HEAD(bh);
1341         set_blkh_level(blkh, FREE_LEVEL);
1342         set_blkh_nr_item(blkh, 0);
1343
1344         clear_buffer_dirty(bh);
1345         store_thrown(tb, bh);
1346 }
1347
1348 /* Replace n_dest'th key in buffer dest by n_src'th key of buffer src.*/
1349 void replace_key(struct tree_balance *tb, struct buffer_head *dest, int n_dest,
1350                  struct buffer_head *src, int n_src)
1351 {
1352
1353         RFALSE(dest == NULL || src == NULL,
1354                "vs-12305: source or destination buffer is 0 (src=%p, dest=%p)",
1355                src, dest);
1356         RFALSE(!B_IS_KEYS_LEVEL(dest),
1357                "vs-12310: invalid level (%z) for destination buffer. dest must be leaf",
1358                dest);
1359         RFALSE(n_dest < 0 || n_src < 0,
1360                "vs-12315: src(%d) or dest(%d) key number < 0", n_src, n_dest);
1361         RFALSE(n_dest >= B_NR_ITEMS(dest) || n_src >= B_NR_ITEMS(src),
1362                "vs-12320: src(%d(%d)) or dest(%d(%d)) key number is too big",
1363                n_src, B_NR_ITEMS(src), n_dest, B_NR_ITEMS(dest));
1364
1365         if (B_IS_ITEMS_LEVEL(src))
1366                 /* source buffer contains leaf node */
1367                 memcpy(internal_key(dest, n_dest), item_head(src, n_src),
1368                        KEY_SIZE);
1369         else
1370                 memcpy(internal_key(dest, n_dest), internal_key(src, n_src),
1371                        KEY_SIZE);
1372
1373         do_balance_mark_internal_dirty(tb, dest, 0);
1374 }
1375
1376 int get_left_neighbor_position(struct tree_balance *tb, int h)
1377 {
1378         int Sh_position = PATH_H_POSITION(tb->tb_path, h + 1);
1379
1380         RFALSE(PATH_H_PPARENT(tb->tb_path, h) == NULL || tb->FL[h] == NULL,
1381                "vs-12325: FL[%d](%p) or F[%d](%p) does not exist",
1382                h, tb->FL[h], h, PATH_H_PPARENT(tb->tb_path, h));
1383
1384         if (Sh_position == 0)
1385                 return B_NR_ITEMS(tb->FL[h]);
1386         else
1387                 return Sh_position - 1;
1388 }
1389
1390 int get_right_neighbor_position(struct tree_balance *tb, int h)
1391 {
1392         int Sh_position = PATH_H_POSITION(tb->tb_path, h + 1);
1393
1394         RFALSE(PATH_H_PPARENT(tb->tb_path, h) == NULL || tb->FR[h] == NULL,
1395                "vs-12330: F[%d](%p) or FR[%d](%p) does not exist",
1396                h, PATH_H_PPARENT(tb->tb_path, h), h, tb->FR[h]);
1397
1398         if (Sh_position == B_NR_ITEMS(PATH_H_PPARENT(tb->tb_path, h)))
1399                 return 0;
1400         else
1401                 return Sh_position + 1;
1402 }
1403
1404 #ifdef CONFIG_REISERFS_CHECK
1405
1406 int is_reusable(struct super_block *s, b_blocknr_t block, int bit_value);
1407 static void check_internal_node(struct super_block *s, struct buffer_head *bh,
1408                                 char *mes)
1409 {
1410         struct disk_child *dc;
1411         int i;
1412
1413         RFALSE(!bh, "PAP-12336: bh == 0");
1414
1415         if (!bh || !B_IS_IN_TREE(bh))
1416                 return;
1417
1418         RFALSE(!buffer_dirty(bh) &&
1419                !(buffer_journaled(bh) || buffer_journal_dirty(bh)),
1420                "PAP-12337: buffer (%b) must be dirty", bh);
1421         dc = B_N_CHILD(bh, 0);
1422
1423         for (i = 0; i <= B_NR_ITEMS(bh); i++, dc++) {
1424                 if (!is_reusable(s, dc_block_number(dc), 1)) {
1425                         print_cur_tb(mes);
1426                         reiserfs_panic(s, "PAP-12338",
1427                                        "invalid child pointer %y in %b",
1428                                        dc, bh);
1429                 }
1430         }
1431 }
1432
1433 static int locked_or_not_in_tree(struct tree_balance *tb,
1434                                   struct buffer_head *bh, char *which)
1435 {
1436         if ((!buffer_journal_prepared(bh) && buffer_locked(bh)) ||
1437             !B_IS_IN_TREE(bh)) {
1438                 reiserfs_warning(tb->tb_sb, "vs-12339", "%s (%b)", which, bh);
1439                 return 1;
1440         }
1441         return 0;
1442 }
1443
1444 static int check_before_balancing(struct tree_balance *tb)
1445 {
1446         int retval = 0;
1447
1448         if (REISERFS_SB(tb->tb_sb)->cur_tb) {
1449                 reiserfs_panic(tb->tb_sb, "vs-12335", "suspect that schedule "
1450                                "occurred based on cur_tb not being null at "
1451                                "this point in code. do_balance cannot properly "
1452                                "handle concurrent tree accesses on a same "
1453                                "mount point.");
1454         }
1455
1456         /*
1457          * double check that buffers that we will modify are unlocked.
1458          * (fix_nodes should already have prepped all of these for us).
1459          */
1460         if (tb->lnum[0]) {
1461                 retval |= locked_or_not_in_tree(tb, tb->L[0], "L[0]");
1462                 retval |= locked_or_not_in_tree(tb, tb->FL[0], "FL[0]");
1463                 retval |= locked_or_not_in_tree(tb, tb->CFL[0], "CFL[0]");
1464                 check_leaf(tb->L[0]);
1465         }
1466         if (tb->rnum[0]) {
1467                 retval |= locked_or_not_in_tree(tb, tb->R[0], "R[0]");
1468                 retval |= locked_or_not_in_tree(tb, tb->FR[0], "FR[0]");
1469                 retval |= locked_or_not_in_tree(tb, tb->CFR[0], "CFR[0]");
1470                 check_leaf(tb->R[0]);
1471         }
1472         retval |= locked_or_not_in_tree(tb, PATH_PLAST_BUFFER(tb->tb_path),
1473                                         "S[0]");
1474         check_leaf(PATH_PLAST_BUFFER(tb->tb_path));
1475
1476         return retval;
1477 }
1478
1479 static void check_after_balance_leaf(struct tree_balance *tb)
1480 {
1481         if (tb->lnum[0]) {
1482                 if (B_FREE_SPACE(tb->L[0]) !=
1483                     MAX_CHILD_SIZE(tb->L[0]) -
1484                     dc_size(B_N_CHILD
1485                             (tb->FL[0], get_left_neighbor_position(tb, 0)))) {
1486                         print_cur_tb("12221");
1487                         reiserfs_panic(tb->tb_sb, "PAP-12355",
1488                                        "shift to left was incorrect");
1489                 }
1490         }
1491         if (tb->rnum[0]) {
1492                 if (B_FREE_SPACE(tb->R[0]) !=
1493                     MAX_CHILD_SIZE(tb->R[0]) -
1494                     dc_size(B_N_CHILD
1495                             (tb->FR[0], get_right_neighbor_position(tb, 0)))) {
1496                         print_cur_tb("12222");
1497                         reiserfs_panic(tb->tb_sb, "PAP-12360",
1498                                        "shift to right was incorrect");
1499                 }
1500         }
1501         if (PATH_H_PBUFFER(tb->tb_path, 1) &&
1502             (B_FREE_SPACE(PATH_H_PBUFFER(tb->tb_path, 0)) !=
1503              (MAX_CHILD_SIZE(PATH_H_PBUFFER(tb->tb_path, 0)) -
1504               dc_size(B_N_CHILD(PATH_H_PBUFFER(tb->tb_path, 1),
1505                                 PATH_H_POSITION(tb->tb_path, 1)))))) {
1506                 int left = B_FREE_SPACE(PATH_H_PBUFFER(tb->tb_path, 0));
1507                 int right = (MAX_CHILD_SIZE(PATH_H_PBUFFER(tb->tb_path, 0)) -
1508                              dc_size(B_N_CHILD(PATH_H_PBUFFER(tb->tb_path, 1),
1509                                                PATH_H_POSITION(tb->tb_path,
1510                                                                1))));
1511                 print_cur_tb("12223");
1512                 reiserfs_warning(tb->tb_sb, "reiserfs-12363",
1513                                  "B_FREE_SPACE (PATH_H_PBUFFER(tb->tb_path,0)) = %d; "
1514                                  "MAX_CHILD_SIZE (%d) - dc_size( %y, %d ) [%d] = %d",
1515                                  left,
1516                                  MAX_CHILD_SIZE(PATH_H_PBUFFER(tb->tb_path, 0)),
1517                                  PATH_H_PBUFFER(tb->tb_path, 1),
1518                                  PATH_H_POSITION(tb->tb_path, 1),
1519                                  dc_size(B_N_CHILD
1520                                          (PATH_H_PBUFFER(tb->tb_path, 1),
1521                                           PATH_H_POSITION(tb->tb_path, 1))),
1522                                  right);
1523                 reiserfs_panic(tb->tb_sb, "PAP-12365", "S is incorrect");
1524         }
1525 }
1526
1527 static void check_leaf_level(struct tree_balance *tb)
1528 {
1529         check_leaf(tb->L[0]);
1530         check_leaf(tb->R[0]);
1531         check_leaf(PATH_PLAST_BUFFER(tb->tb_path));
1532 }
1533
1534 static void check_internal_levels(struct tree_balance *tb)
1535 {
1536         int h;
1537
1538         /* check all internal nodes */
1539         for (h = 1; tb->insert_size[h]; h++) {
1540                 check_internal_node(tb->tb_sb, PATH_H_PBUFFER(tb->tb_path, h),
1541                                     "BAD BUFFER ON PATH");
1542                 if (tb->lnum[h])
1543                         check_internal_node(tb->tb_sb, tb->L[h], "BAD L");
1544                 if (tb->rnum[h])
1545                         check_internal_node(tb->tb_sb, tb->R[h], "BAD R");
1546         }
1547
1548 }
1549
1550 #endif
1551
1552 /*
1553  * Now we have all of the buffers that must be used in balancing of
1554  * the tree.  We rely on the assumption that schedule() will not occur
1555  * while do_balance works. ( Only interrupt handlers are acceptable.)
1556  * We balance the tree according to the analysis made before this,
1557  * using buffers already obtained.  For SMP support it will someday be
1558  * necessary to add ordered locking of tb.
1559  */
1560
1561 /*
1562  * Some interesting rules of balancing:
1563  * we delete a maximum of two nodes per level per balancing: we never
1564  * delete R, when we delete two of three nodes L, S, R then we move
1565  * them into R.
1566  *
1567  * we only delete L if we are deleting two nodes, if we delete only
1568  * one node we delete S
1569  *
1570  * if we shift leaves then we shift as much as we can: this is a
1571  * deliberate policy of extremism in node packing which results in
1572  * higher average utilization after repeated random balance operations
1573  * at the cost of more memory copies and more balancing as a result of
1574  * small insertions to full nodes.
1575  *
1576  * if we shift internal nodes we try to evenly balance the node
1577  * utilization, with consequent less balancing at the cost of lower
1578  * utilization.
1579  *
1580  * one could argue that the policy for directories in leaves should be
1581  * that of internal nodes, but we will wait until another day to
1582  * evaluate this....  It would be nice to someday measure and prove
1583  * these assumptions as to what is optimal....
1584  */
1585
1586 static inline void do_balance_starts(struct tree_balance *tb)
1587 {
1588         /* use print_cur_tb() to see initial state of struct tree_balance */
1589
1590         /* store_print_tb (tb); */
1591
1592         /* do not delete, just comment it out */
1593         /*
1594         print_tb(flag, PATH_LAST_POSITION(tb->tb_path),
1595                  tb->tb_path->pos_in_item, tb, "check");
1596         */
1597         RFALSE(check_before_balancing(tb), "PAP-12340: locked buffers in TB");
1598 #ifdef CONFIG_REISERFS_CHECK
1599         REISERFS_SB(tb->tb_sb)->cur_tb = tb;
1600 #endif
1601 }
1602
1603 static inline void do_balance_completed(struct tree_balance *tb)
1604 {
1605
1606 #ifdef CONFIG_REISERFS_CHECK
1607         check_leaf_level(tb);
1608         check_internal_levels(tb);
1609         REISERFS_SB(tb->tb_sb)->cur_tb = NULL;
1610 #endif
1611
1612         /*
1613          * reiserfs_free_block is no longer schedule safe.  So, we need to
1614          * put the buffers we want freed on the thrown list during do_balance,
1615          * and then free them now
1616          */
1617
1618         REISERFS_SB(tb->tb_sb)->s_do_balance++;
1619
1620         /* release all nodes hold to perform the balancing */
1621         unfix_nodes(tb);
1622
1623         free_thrown(tb);
1624 }
1625
1626 /*
1627  * do_balance - balance the tree
1628  *
1629  * @tb: tree_balance structure
1630  * @ih: item header of inserted item
1631  * @body: body of inserted item or bytes to paste
1632  * @flag: 'i' - insert, 'd' - delete, 'c' - cut, 'p' paste
1633  *
1634  * Cut means delete part of an item (includes removing an entry from a
1635  * directory).
1636  *
1637  * Delete means delete whole item.
1638  *
1639  * Insert means add a new item into the tree.
1640  *
1641  * Paste means to append to the end of an existing file or to
1642  * insert a directory entry.
1643  */
1644 void do_balance(struct tree_balance *tb, struct item_head *ih,
1645                 const char *body, int flag)
1646 {
1647         int child_pos;          /* position of a child node in its parent */
1648         int h;                  /* level of the tree being processed */
1649
1650         /*
1651          * in our processing of one level we sometimes determine what
1652          * must be inserted into the next higher level.  This insertion
1653          * consists of a key or two keys and their corresponding
1654          * pointers
1655          */
1656         struct item_head insert_key[2];
1657
1658         /* inserted node-ptrs for the next level */
1659         struct buffer_head *insert_ptr[2];
1660
1661         tb->tb_mode = flag;
1662         tb->need_balance_dirty = 0;
1663
1664         if (FILESYSTEM_CHANGED_TB(tb)) {
1665                 reiserfs_panic(tb->tb_sb, "clm-6000", "fs generation has "
1666                                "changed");
1667         }
1668         /* if we have no real work to do  */
1669         if (!tb->insert_size[0]) {
1670                 reiserfs_warning(tb->tb_sb, "PAP-12350",
1671                                  "insert_size == 0, mode == %c", flag);
1672                 unfix_nodes(tb);
1673                 return;
1674         }
1675
1676         atomic_inc(&fs_generation(tb->tb_sb));
1677         do_balance_starts(tb);
1678
1679         /*
1680          * balance_leaf returns 0 except if combining L R and S into
1681          * one node.  see balance_internal() for explanation of this
1682          * line of code.
1683          */
1684         child_pos = PATH_H_B_ITEM_ORDER(tb->tb_path, 0) +
1685             balance_leaf(tb, ih, body, flag, insert_key, insert_ptr);
1686
1687 #ifdef CONFIG_REISERFS_CHECK
1688         check_after_balance_leaf(tb);
1689 #endif
1690
1691         /* Balance internal level of the tree. */
1692         for (h = 1; h < MAX_HEIGHT && tb->insert_size[h]; h++)
1693                 child_pos =
1694                     balance_internal(tb, h, child_pos, insert_key, insert_ptr);
1695
1696         do_balance_completed(tb);
1697
1698 }