2 * Copyright 2000 by Hans Reiser, licensing governed by reiserfs/README
12 ** directory_part_size
16 ** are_leaves_removable
20 ** is_left_neighbor_in_cache
24 ** can_node_be_removed
26 ** dc_check_balance_internal
27 ** dc_check_balance_leaf
37 #include <linux/time.h>
38 #include <linux/string.h>
39 #include <linux/reiserfs_fs.h>
40 #include <linux/buffer_head.h>
42 /* To make any changes in the tree we find a node, that contains item
43 to be changed/deleted or position in the node we insert a new item
44 to. We call this node S. To do balancing we need to decide what we
45 will shift to left/right neighbor, or to a new node, where new item
46 will be etc. To make this analysis simpler we build virtual
47 node. Virtual node is an array of items, that will replace items of
48 node S. (For instance if we are going to delete an item, virtual
49 node does not contain it). Virtual node keeps information about
50 item sizes and types, mergeability of first and last items, sizes
51 of all entries in directory item. We use this array of items when
52 calculating what we can shift to neighbors and how many nodes we
53 have to have if we do not any shiftings, if we shift to left/right
54 neighbor or to both. */
56 /* taking item number in virtual node, returns number of item, that it has in source buffer */
57 static inline int old_item_num(int new_num, int affected_item_num, int mode)
59 if (mode == M_PASTE || mode == M_CUT || new_num < affected_item_num)
62 if (mode == M_INSERT) {
65 "vs-8005: for INSERT mode and item number of inserted item");
70 RFALSE(mode != M_DELETE,
71 "vs-8010: old_item_num: mode must be M_DELETE (mode = \'%c\'",
77 static void create_virtual_node(struct tree_balance *tb, int h)
80 struct virtual_node *vn = tb->tb_vn;
82 struct buffer_head *Sh; /* this comes from tb->S[h] */
84 Sh = PATH_H_PBUFFER(tb->tb_path, h);
86 /* size of changed node */
88 MAX_CHILD_SIZE(Sh) - B_FREE_SPACE(Sh) + tb->insert_size[h];
90 /* for internal nodes array if virtual items is not created */
92 vn->vn_nr_item = (vn->vn_size - DC_SIZE) / (DC_SIZE + KEY_SIZE);
96 /* number of items in virtual node */
98 B_NR_ITEMS(Sh) + ((vn->vn_mode == M_INSERT) ? 1 : 0) -
99 ((vn->vn_mode == M_DELETE) ? 1 : 0);
101 /* first virtual item */
102 vn->vn_vi = (struct virtual_item *)(tb->tb_vn + 1);
103 memset(vn->vn_vi, 0, vn->vn_nr_item * sizeof(struct virtual_item));
104 vn->vn_free_ptr += vn->vn_nr_item * sizeof(struct virtual_item);
106 /* first item in the node */
107 ih = B_N_PITEM_HEAD(Sh, 0);
109 /* define the mergeability for 0-th item (if it is not being deleted) */
110 if (op_is_left_mergeable(&(ih->ih_key), Sh->b_size)
111 && (vn->vn_mode != M_DELETE || vn->vn_affected_item_num))
112 vn->vn_vi[0].vi_type |= VI_TYPE_LEFT_MERGEABLE;
114 /* go through all items those remain in the virtual node (except for the new (inserted) one) */
115 for (new_num = 0; new_num < vn->vn_nr_item; new_num++) {
117 struct virtual_item *vi = vn->vn_vi + new_num;
119 ((new_num != vn->vn_affected_item_num) ? 0 : 1);
121 if (is_affected && vn->vn_mode == M_INSERT)
124 /* get item number in source node */
125 j = old_item_num(new_num, vn->vn_affected_item_num,
128 vi->vi_item_len += ih_item_len(ih + j) + IH_SIZE;
130 vi->vi_item = B_I_PITEM(Sh, ih + j);
131 vi->vi_uarea = vn->vn_free_ptr;
133 // FIXME: there is no check, that item operation did not
134 // consume too much memory
136 op_create_vi(vn, vi, is_affected, tb->insert_size[0]);
137 if (tb->vn_buf + tb->vn_buf_size < vn->vn_free_ptr)
138 reiserfs_panic(tb->tb_sb, "vs-8030",
139 "virtual node space consumed");
142 /* this is not being changed */
145 if (vn->vn_mode == M_PASTE || vn->vn_mode == M_CUT) {
146 vn->vn_vi[new_num].vi_item_len += tb->insert_size[0];
147 vi->vi_new_data = vn->vn_data; // pointer to data which is going to be pasted
151 /* virtual inserted item is not defined yet */
152 if (vn->vn_mode == M_INSERT) {
153 struct virtual_item *vi = vn->vn_vi + vn->vn_affected_item_num;
155 RFALSE(vn->vn_ins_ih == NULL,
156 "vs-8040: item header of inserted item is not specified");
157 vi->vi_item_len = tb->insert_size[0];
158 vi->vi_ih = vn->vn_ins_ih;
159 vi->vi_item = vn->vn_data;
160 vi->vi_uarea = vn->vn_free_ptr;
162 op_create_vi(vn, vi, 0 /*not pasted or cut */ ,
166 /* set right merge flag we take right delimiting key and check whether it is a mergeable item */
168 struct reiserfs_key *key;
170 key = B_N_PDELIM_KEY(tb->CFR[0], tb->rkey[0]);
171 if (op_is_left_mergeable(key, Sh->b_size)
172 && (vn->vn_mode != M_DELETE
173 || vn->vn_affected_item_num != B_NR_ITEMS(Sh) - 1))
174 vn->vn_vi[vn->vn_nr_item - 1].vi_type |=
175 VI_TYPE_RIGHT_MERGEABLE;
177 #ifdef CONFIG_REISERFS_CHECK
178 if (op_is_left_mergeable(key, Sh->b_size) &&
179 !(vn->vn_mode != M_DELETE
180 || vn->vn_affected_item_num != B_NR_ITEMS(Sh) - 1)) {
181 /* we delete last item and it could be merged with right neighbor's first item */
184 && is_direntry_le_ih(B_N_PITEM_HEAD(Sh, 0))
185 && I_ENTRY_COUNT(B_N_PITEM_HEAD(Sh, 0)) == 1)) {
186 /* node contains more than 1 item, or item is not directory item, or this item contains more than 1 entry */
187 print_block(Sh, 0, -1, -1);
188 reiserfs_panic(tb->tb_sb, "vs-8045",
189 "rdkey %k, affected item==%d "
190 "(mode==%c) Must be %c",
191 key, vn->vn_affected_item_num,
192 vn->vn_mode, M_DELETE);
200 /* using virtual node check, how many items can be shifted to left
202 static void check_left(struct tree_balance *tb, int h, int cur_free)
205 struct virtual_node *vn = tb->tb_vn;
206 struct virtual_item *vi;
209 RFALSE(cur_free < 0, "vs-8050: cur_free (%d) < 0", cur_free);
213 tb->lnum[h] = cur_free / (DC_SIZE + KEY_SIZE);
219 if (!cur_free || !vn->vn_nr_item) {
220 /* no free space or nothing to move */
226 RFALSE(!PATH_H_PPARENT(tb->tb_path, 0),
227 "vs-8055: parent does not exist or invalid");
230 if ((unsigned int)cur_free >=
232 ((vi->vi_type & VI_TYPE_LEFT_MERGEABLE) ? IH_SIZE : 0))) {
233 /* all contents of S[0] fits into L[0] */
235 RFALSE(vn->vn_mode == M_INSERT || vn->vn_mode == M_PASTE,
236 "vs-8055: invalid mode or balance condition failed");
238 tb->lnum[0] = vn->vn_nr_item;
243 d_size = 0, ih_size = IH_SIZE;
245 /* first item may be merge with last item in left neighbor */
246 if (vi->vi_type & VI_TYPE_LEFT_MERGEABLE)
247 d_size = -((int)IH_SIZE), ih_size = 0;
250 for (i = 0; i < vn->vn_nr_item;
251 i++, ih_size = IH_SIZE, d_size = 0, vi++) {
252 d_size += vi->vi_item_len;
253 if (cur_free >= d_size) {
254 /* the item can be shifted entirely */
260 /* the item cannot be shifted entirely, try to split it */
261 /* check whether L[0] can hold ih and at least one byte of the item body */
262 if (cur_free <= ih_size) {
263 /* cannot shift even a part of the current item */
269 tb->lbytes = op_check_left(vi, cur_free, 0, 0);
270 if (tb->lbytes != -1)
271 /* count partially shifted item */
280 /* using virtual node check, how many items can be shifted to right
282 static void check_right(struct tree_balance *tb, int h, int cur_free)
285 struct virtual_node *vn = tb->tb_vn;
286 struct virtual_item *vi;
289 RFALSE(cur_free < 0, "vs-8070: cur_free < 0");
293 tb->rnum[h] = cur_free / (DC_SIZE + KEY_SIZE);
299 if (!cur_free || !vn->vn_nr_item) {
306 RFALSE(!PATH_H_PPARENT(tb->tb_path, 0),
307 "vs-8075: parent does not exist or invalid");
309 vi = vn->vn_vi + vn->vn_nr_item - 1;
310 if ((unsigned int)cur_free >=
312 ((vi->vi_type & VI_TYPE_RIGHT_MERGEABLE) ? IH_SIZE : 0))) {
313 /* all contents of S[0] fits into R[0] */
315 RFALSE(vn->vn_mode == M_INSERT || vn->vn_mode == M_PASTE,
316 "vs-8080: invalid mode or balance condition failed");
318 tb->rnum[h] = vn->vn_nr_item;
323 d_size = 0, ih_size = IH_SIZE;
325 /* last item may be merge with first item in right neighbor */
326 if (vi->vi_type & VI_TYPE_RIGHT_MERGEABLE)
327 d_size = -(int)IH_SIZE, ih_size = 0;
330 for (i = vn->vn_nr_item - 1; i >= 0;
331 i--, d_size = 0, ih_size = IH_SIZE, vi--) {
332 d_size += vi->vi_item_len;
333 if (cur_free >= d_size) {
334 /* the item can be shifted entirely */
340 /* check whether R[0] can hold ih and at least one byte of the item body */
341 if (cur_free <= ih_size) { /* cannot shift even a part of the current item */
346 /* R[0] can hold the header of the item and at least one byte of its body */
347 cur_free -= ih_size; /* cur_free is still > 0 */
349 tb->rbytes = op_check_right(vi, cur_free);
350 if (tb->rbytes != -1)
351 /* count partially shifted item */
361 * from - number of items, which are shifted to left neighbor entirely
362 * to - number of item, which are shifted to right neighbor entirely
363 * from_bytes - number of bytes of boundary item (or directory entries) which are shifted to left neighbor
364 * to_bytes - number of bytes of boundary item (or directory entries) which are shifted to right neighbor */
365 static int get_num_ver(int mode, struct tree_balance *tb, int h,
366 int from, int from_bytes,
367 int to, int to_bytes, short *snum012, int flow)
373 struct virtual_node *vn = tb->tb_vn;
374 // struct virtual_item * vi;
376 int total_node_size, max_node_size, current_item_size;
378 int start_item, /* position of item we start filling node from */
379 end_item, /* position of item we finish filling node by */
380 start_bytes, /* number of first bytes (entries for directory) of start_item-th item
381 we do not include into node that is being filled */
382 end_bytes; /* number of last bytes (entries for directory) of end_item-th item
383 we do node include into node that is being filled */
384 int split_item_positions[2]; /* these are positions in virtual item of
385 items, that are split between S[0] and
386 S1new and S1new and S2new */
388 split_item_positions[0] = -1;
389 split_item_positions[1] = -1;
391 /* We only create additional nodes if we are in insert or paste mode
392 or we are in replace mode at the internal level. If h is 0 and
393 the mode is M_REPLACE then in fix_nodes we change the mode to
394 paste or insert before we get here in the code. */
395 RFALSE(tb->insert_size[h] < 0 || (mode != M_INSERT && mode != M_PASTE),
396 "vs-8100: insert_size < 0 in overflow");
398 max_node_size = MAX_CHILD_SIZE(PATH_H_PBUFFER(tb->tb_path, h));
400 /* snum012 [0-2] - number of items, that lay
401 to S[0], first new node and second new node */
402 snum012[3] = -1; /* s1bytes */
403 snum012[4] = -1; /* s2bytes */
407 i = ((to - from) * (KEY_SIZE + DC_SIZE) + DC_SIZE);
408 if (i == max_node_size)
410 return (i / max_node_size + 1);
416 cur_free = max_node_size;
418 // start from 'from'-th item
420 // skip its first 'start_bytes' units
421 start_bytes = ((from_bytes != -1) ? from_bytes : 0);
423 // last included item is the 'end_item'-th one
424 end_item = vn->vn_nr_item - to - 1;
425 // do not count last 'end_bytes' units of 'end_item'-th item
426 end_bytes = (to_bytes != -1) ? to_bytes : 0;
428 /* go through all item beginning from the start_item-th item and ending by
429 the end_item-th item. Do not count first 'start_bytes' units of
430 'start_item'-th item and last 'end_bytes' of 'end_item'-th item */
432 for (i = start_item; i <= end_item; i++) {
433 struct virtual_item *vi = vn->vn_vi + i;
434 int skip_from_end = ((i == end_item) ? end_bytes : 0);
436 RFALSE(needed_nodes > 3, "vs-8105: too many nodes are needed");
438 /* get size of current item */
439 current_item_size = vi->vi_item_len;
441 /* do not take in calculation head part (from_bytes) of from-th item */
443 op_part_size(vi, 0 /*from start */ , start_bytes);
445 /* do not take in calculation tail part of last item */
447 op_part_size(vi, 1 /*from end */ , skip_from_end);
449 /* if item fits into current node entierly */
450 if (total_node_size + current_item_size <= max_node_size) {
451 snum012[needed_nodes - 1]++;
452 total_node_size += current_item_size;
457 if (current_item_size > max_node_size) {
458 /* virtual item length is longer, than max size of item in
459 a node. It is impossible for direct item */
460 RFALSE(is_direct_le_ih(vi->vi_ih),
462 "direct item length is %d. It can not be longer than %d",
463 current_item_size, max_node_size);
464 /* we will try to split it */
469 /* as we do not split items, take new node and continue */
475 // calculate number of item units which fit into node being
480 free_space = max_node_size - total_node_size - IH_SIZE;
482 op_check_left(vi, free_space, start_bytes,
485 /* nothing fits into current node, take new node and continue */
486 needed_nodes++, i--, total_node_size = 0;
491 /* something fits into the current node */
492 //if (snum012[3] != -1 || needed_nodes != 1)
493 // reiserfs_panic (tb->tb_sb, "vs-8115: get_num_ver: too many nodes required");
494 //snum012[needed_nodes - 1 + 3] = op_unit_num (vi) - start_bytes - units;
495 start_bytes += units;
496 snum012[needed_nodes - 1 + 3] = units;
498 if (needed_nodes > 2)
499 reiserfs_warning(tb->tb_sb, "vs-8111",
500 "split_item_position is out of range");
501 snum012[needed_nodes - 1]++;
502 split_item_positions[needed_nodes - 1] = i;
504 /* continue from the same item with start_bytes != -1 */
510 // sum012[4] (if it is not -1) contains number of units of which
511 // are to be in S1new, snum012[3] - to be in S0. They are supposed
512 // to be S1bytes and S2bytes correspondingly, so recalculate
513 if (snum012[4] > 0) {
515 int bytes_to_r, bytes_to_l;
518 split_item_num = split_item_positions[1];
520 ((from == split_item_num
521 && from_bytes != -1) ? from_bytes : 0);
523 ((end_item == split_item_num
524 && end_bytes != -1) ? end_bytes : 0);
526 ((split_item_positions[0] ==
527 split_item_positions[1]) ? snum012[3] : 0);
531 op_unit_num(&vn->vn_vi[split_item_num]) - snum012[4] -
532 bytes_to_r - bytes_to_l - bytes_to_S1new;
534 if (vn->vn_vi[split_item_num].vi_index != TYPE_DIRENTRY &&
535 vn->vn_vi[split_item_num].vi_index != TYPE_INDIRECT)
536 reiserfs_warning(tb->tb_sb, "vs-8115",
537 "not directory or indirect item");
540 /* now we know S2bytes, calculate S1bytes */
541 if (snum012[3] > 0) {
543 int bytes_to_r, bytes_to_l;
546 split_item_num = split_item_positions[0];
548 ((from == split_item_num
549 && from_bytes != -1) ? from_bytes : 0);
551 ((end_item == split_item_num
552 && end_bytes != -1) ? end_bytes : 0);
554 ((split_item_positions[0] == split_item_positions[1]
555 && snum012[4] != -1) ? snum012[4] : 0);
559 op_unit_num(&vn->vn_vi[split_item_num]) - snum012[3] -
560 bytes_to_r - bytes_to_l - bytes_to_S2new;
566 #ifdef CONFIG_REISERFS_CHECK
567 extern struct tree_balance *cur_tb;
570 /* Set parameters for balancing.
571 * Performs write of results of analysis of balancing into structure tb,
572 * where it will later be used by the functions that actually do the balancing.
574 * tb tree_balance structure;
575 * h current level of the node;
576 * lnum number of items from S[h] that must be shifted to L[h];
577 * rnum number of items from S[h] that must be shifted to R[h];
578 * blk_num number of blocks that S[h] will be splitted into;
579 * s012 number of items that fall into splitted nodes.
580 * lbytes number of bytes which flow to the left neighbor from the item that is not
581 * not shifted entirely
582 * rbytes number of bytes which flow to the right neighbor from the item that is not
583 * not shifted entirely
584 * s1bytes number of bytes which flow to the first new node when S[0] splits (this number is contained in s012 array)
587 static void set_parameters(struct tree_balance *tb, int h, int lnum,
588 int rnum, int blk_num, short *s012, int lb, int rb)
593 tb->blknum[h] = blk_num;
595 if (h == 0) { /* only for leaf level */
598 tb->s1num = *s012++, tb->s2num = *s012++;
599 tb->s1bytes = *s012++;
605 PROC_INFO_ADD(tb->tb_sb, lnum[h], lnum);
606 PROC_INFO_ADD(tb->tb_sb, rnum[h], rnum);
608 PROC_INFO_ADD(tb->tb_sb, lbytes[h], lb);
609 PROC_INFO_ADD(tb->tb_sb, rbytes[h], rb);
612 /* check, does node disappear if we shift tb->lnum[0] items to left
613 neighbor and tb->rnum[0] to the right one. */
614 static int is_leaf_removable(struct tree_balance *tb)
616 struct virtual_node *vn = tb->tb_vn;
617 int to_left, to_right;
621 /* number of items, that will be shifted to left (right) neighbor
623 to_left = tb->lnum[0] - ((tb->lbytes != -1) ? 1 : 0);
624 to_right = tb->rnum[0] - ((tb->rbytes != -1) ? 1 : 0);
625 remain_items = vn->vn_nr_item;
627 /* how many items remain in S[0] after shiftings to neighbors */
628 remain_items -= (to_left + to_right);
630 if (remain_items < 1) {
631 /* all content of node can be shifted to neighbors */
632 set_parameters(tb, 0, to_left, vn->vn_nr_item - to_left, 0,
637 if (remain_items > 1 || tb->lbytes == -1 || tb->rbytes == -1)
638 /* S[0] is not removable */
641 /* check, whether we can divide 1 remaining item between neighbors */
643 /* get size of remaining item (in item units) */
644 size = op_unit_num(&(vn->vn_vi[to_left]));
646 if (tb->lbytes + tb->rbytes >= size) {
647 set_parameters(tb, 0, to_left + 1, to_right + 1, 0, NULL,
655 /* check whether L, S, R can be joined in one node */
656 static int are_leaves_removable(struct tree_balance *tb, int lfree, int rfree)
658 struct virtual_node *vn = tb->tb_vn;
660 struct buffer_head *S0;
662 S0 = PATH_H_PBUFFER(tb->tb_path, 0);
665 if (vn->vn_nr_item) {
666 if (vn->vn_vi[0].vi_type & VI_TYPE_LEFT_MERGEABLE)
669 if (vn->vn_vi[vn->vn_nr_item - 1].
670 vi_type & VI_TYPE_RIGHT_MERGEABLE)
673 /* there was only one item and it will be deleted */
674 struct item_head *ih;
676 RFALSE(B_NR_ITEMS(S0) != 1,
677 "vs-8125: item number must be 1: it is %d",
680 ih = B_N_PITEM_HEAD(S0, 0);
682 && !comp_short_le_keys(&(ih->ih_key),
683 B_N_PDELIM_KEY(tb->CFR[0],
685 if (is_direntry_le_ih(ih)) {
686 /* Directory must be in correct state here: that is
687 somewhere at the left side should exist first directory
688 item. But the item being deleted can not be that first
689 one because its right neighbor is item of the same
690 directory. (But first item always gets deleted in last
691 turn). So, neighbors of deleted item can be merged, so
692 we can save ih_size */
695 /* we might check that left neighbor exists and is of the
697 RFALSE(le_ih_k_offset(ih) == DOT_OFFSET,
698 "vs-8130: first directory item can not be removed until directory is not empty");
703 if (MAX_CHILD_SIZE(S0) + vn->vn_size <= rfree + lfree + ih_size) {
704 set_parameters(tb, 0, -1, -1, -1, NULL, -1, -1);
705 PROC_INFO_INC(tb->tb_sb, leaves_removable);
712 /* when we do not split item, lnum and rnum are numbers of entire items */
713 #define SET_PAR_SHIFT_LEFT \
718 to_l = (MAX_NR_KEY(Sh)+1 - lpar + vn->vn_nr_item + 1) / 2 -\
719 (MAX_NR_KEY(Sh) + 1 - lpar);\
721 set_parameters (tb, h, to_l, 0, lnver, NULL, -1, -1);\
725 if (lset==LEFT_SHIFT_FLOW)\
726 set_parameters (tb, h, lpar, 0, lnver, snum012+lset,\
729 set_parameters (tb, h, lpar - (tb->lbytes!=-1), 0, lnver, snum012+lset,\
733 #define SET_PAR_SHIFT_RIGHT \
738 to_r = (MAX_NR_KEY(Sh)+1 - rpar + vn->vn_nr_item + 1) / 2 - (MAX_NR_KEY(Sh) + 1 - rpar);\
740 set_parameters (tb, h, 0, to_r, rnver, NULL, -1, -1);\
744 if (rset==RIGHT_SHIFT_FLOW)\
745 set_parameters (tb, h, 0, rpar, rnver, snum012+rset,\
748 set_parameters (tb, h, 0, rpar - (tb->rbytes!=-1), rnver, snum012+rset,\
752 static void free_buffers_in_tb(struct tree_balance *p_s_tb)
756 pathrelse(p_s_tb->tb_path);
758 for (n_counter = 0; n_counter < MAX_HEIGHT; n_counter++) {
759 brelse(p_s_tb->L[n_counter]);
760 brelse(p_s_tb->R[n_counter]);
761 brelse(p_s_tb->FL[n_counter]);
762 brelse(p_s_tb->FR[n_counter]);
763 brelse(p_s_tb->CFL[n_counter]);
764 brelse(p_s_tb->CFR[n_counter]);
766 p_s_tb->L[n_counter] = NULL;
767 p_s_tb->R[n_counter] = NULL;
768 p_s_tb->FL[n_counter] = NULL;
769 p_s_tb->FR[n_counter] = NULL;
770 p_s_tb->CFL[n_counter] = NULL;
771 p_s_tb->CFR[n_counter] = NULL;
775 /* Get new buffers for storing new nodes that are created while balancing.
776 * Returns: SCHEDULE_OCCURRED - schedule occurred while the function worked;
777 * CARRY_ON - schedule didn't occur while the function worked;
778 * NO_DISK_SPACE - no disk space.
780 /* The function is NOT SCHEDULE-SAFE! */
781 static int get_empty_nodes(struct tree_balance *p_s_tb, int n_h)
783 struct buffer_head *p_s_new_bh,
784 *p_s_Sh = PATH_H_PBUFFER(p_s_tb->tb_path, n_h);
785 b_blocknr_t *p_n_blocknr, a_n_blocknrs[MAX_AMOUNT_NEEDED] = { 0, };
786 int n_counter, n_number_of_freeblk, n_amount_needed, /* number of needed empty blocks */
788 struct super_block *sb = p_s_tb->tb_sb;
790 /* number_of_freeblk is the number of empty blocks which have been
791 acquired for use by the balancing algorithm minus the number of
792 empty blocks used in the previous levels of the analysis,
793 number_of_freeblk = tb->cur_blknum can be non-zero if a schedule occurs
794 after empty blocks are acquired, and the balancing analysis is
795 then restarted, amount_needed is the number needed by this level
796 (n_h) of the balancing analysis.
798 Note that for systems with many processes writing, it would be
799 more layout optimal to calculate the total number needed by all
800 levels and then to run reiserfs_new_blocks to get all of them at once. */
802 /* Initiate number_of_freeblk to the amount acquired prior to the restart of
803 the analysis or 0 if not restarted, then subtract the amount needed
804 by all of the levels of the tree below n_h. */
805 /* blknum includes S[n_h], so we subtract 1 in this calculation */
806 for (n_counter = 0, n_number_of_freeblk = p_s_tb->cur_blknum;
807 n_counter < n_h; n_counter++)
808 n_number_of_freeblk -=
809 (p_s_tb->blknum[n_counter]) ? (p_s_tb->blknum[n_counter] -
812 /* Allocate missing empty blocks. */
813 /* if p_s_Sh == 0 then we are getting a new root */
814 n_amount_needed = (p_s_Sh) ? (p_s_tb->blknum[n_h] - 1) : 1;
815 /* Amount_needed = the amount that we need more than the amount that we have. */
816 if (n_amount_needed > n_number_of_freeblk)
817 n_amount_needed -= n_number_of_freeblk;
818 else /* If we have enough already then there is nothing to do. */
821 /* No need to check quota - is not allocated for blocks used for formatted nodes */
822 if (reiserfs_new_form_blocknrs(p_s_tb, a_n_blocknrs,
823 n_amount_needed) == NO_DISK_SPACE)
824 return NO_DISK_SPACE;
826 /* for each blocknumber we just got, get a buffer and stick it on FEB */
827 for (p_n_blocknr = a_n_blocknrs, n_counter = 0;
828 n_counter < n_amount_needed; p_n_blocknr++, n_counter++) {
830 RFALSE(!*p_n_blocknr,
831 "PAP-8135: reiserfs_new_blocknrs failed when got new blocks");
833 p_s_new_bh = sb_getblk(sb, *p_n_blocknr);
834 RFALSE(buffer_dirty(p_s_new_bh) ||
835 buffer_journaled(p_s_new_bh) ||
836 buffer_journal_dirty(p_s_new_bh),
837 "PAP-8140: journlaled or dirty buffer %b for the new block",
840 /* Put empty buffers into the array. */
841 RFALSE(p_s_tb->FEB[p_s_tb->cur_blknum],
842 "PAP-8141: busy slot for new buffer");
844 set_buffer_journal_new(p_s_new_bh);
845 p_s_tb->FEB[p_s_tb->cur_blknum++] = p_s_new_bh;
848 if (n_retval == CARRY_ON && FILESYSTEM_CHANGED_TB(p_s_tb))
849 n_retval = REPEAT_SEARCH;
854 /* Get free space of the left neighbor, which is stored in the parent
855 * node of the left neighbor. */
856 static int get_lfree(struct tree_balance *tb, int h)
858 struct buffer_head *l, *f;
861 if ((f = PATH_H_PPARENT(tb->tb_path, h)) == NULL ||
862 (l = tb->FL[h]) == NULL)
866 order = PATH_H_B_ITEM_ORDER(tb->tb_path, h) - 1;
868 order = B_NR_ITEMS(l);
872 return (MAX_CHILD_SIZE(f) - dc_size(B_N_CHILD(f, order)));
875 /* Get free space of the right neighbor,
876 * which is stored in the parent node of the right neighbor.
878 static int get_rfree(struct tree_balance *tb, int h)
880 struct buffer_head *r, *f;
883 if ((f = PATH_H_PPARENT(tb->tb_path, h)) == NULL ||
884 (r = tb->FR[h]) == NULL)
888 order = PATH_H_B_ITEM_ORDER(tb->tb_path, h) + 1;
894 return (MAX_CHILD_SIZE(f) - dc_size(B_N_CHILD(f, order)));
898 /* Check whether left neighbor is in memory. */
899 static int is_left_neighbor_in_cache(struct tree_balance *p_s_tb, int n_h)
901 struct buffer_head *p_s_father, *left;
902 struct super_block *sb = p_s_tb->tb_sb;
903 b_blocknr_t n_left_neighbor_blocknr;
904 int n_left_neighbor_position;
906 if (!p_s_tb->FL[n_h]) /* Father of the left neighbor does not exist. */
909 /* Calculate father of the node to be balanced. */
910 p_s_father = PATH_H_PBUFFER(p_s_tb->tb_path, n_h + 1);
912 RFALSE(!p_s_father ||
913 !B_IS_IN_TREE(p_s_father) ||
914 !B_IS_IN_TREE(p_s_tb->FL[n_h]) ||
915 !buffer_uptodate(p_s_father) ||
916 !buffer_uptodate(p_s_tb->FL[n_h]),
917 "vs-8165: F[h] (%b) or FL[h] (%b) is invalid",
918 p_s_father, p_s_tb->FL[n_h]);
920 /* Get position of the pointer to the left neighbor into the left father. */
921 n_left_neighbor_position = (p_s_father == p_s_tb->FL[n_h]) ?
922 p_s_tb->lkey[n_h] : B_NR_ITEMS(p_s_tb->FL[n_h]);
923 /* Get left neighbor block number. */
924 n_left_neighbor_blocknr =
925 B_N_CHILD_NUM(p_s_tb->FL[n_h], n_left_neighbor_position);
926 /* Look for the left neighbor in the cache. */
927 if ((left = sb_find_get_block(sb, n_left_neighbor_blocknr))) {
929 RFALSE(buffer_uptodate(left) && !B_IS_IN_TREE(left),
930 "vs-8170: left neighbor (%b %z) is not in the tree",
939 #define LEFT_PARENTS 'l'
940 #define RIGHT_PARENTS 'r'
942 static void decrement_key(struct cpu_key *p_s_key)
944 // call item specific function for this key
945 item_ops[cpu_key_k_type(p_s_key)]->decrement_key(p_s_key);
948 /* Calculate far left/right parent of the left/right neighbor of the current node, that
949 * is calculate the left/right (FL[h]/FR[h]) neighbor of the parent F[h].
950 * Calculate left/right common parent of the current node and L[h]/R[h].
951 * Calculate left/right delimiting key position.
952 * Returns: PATH_INCORRECT - path in the tree is not correct;
953 SCHEDULE_OCCURRED - schedule occurred while the function worked;
954 * CARRY_ON - schedule didn't occur while the function worked;
956 static int get_far_parent(struct tree_balance *p_s_tb,
958 struct buffer_head **pp_s_father,
959 struct buffer_head **pp_s_com_father, char c_lr_par)
961 struct buffer_head *p_s_parent;
962 INITIALIZE_PATH(s_path_to_neighbor_father);
963 struct treepath *p_s_path = p_s_tb->tb_path;
964 struct cpu_key s_lr_father_key;
966 n_position = INT_MAX,
967 n_first_last_position = 0,
968 n_path_offset = PATH_H_PATH_OFFSET(p_s_path, n_h);
970 /* Starting from F[n_h] go upwards in the tree, and look for the common
971 ancestor of F[n_h], and its neighbor l/r, that should be obtained. */
973 n_counter = n_path_offset;
975 RFALSE(n_counter < FIRST_PATH_ELEMENT_OFFSET,
976 "PAP-8180: invalid path length");
978 for (; n_counter > FIRST_PATH_ELEMENT_OFFSET; n_counter--) {
979 /* Check whether parent of the current buffer in the path is really parent in the tree. */
981 (p_s_parent = PATH_OFFSET_PBUFFER(p_s_path, n_counter - 1)))
982 return REPEAT_SEARCH;
983 /* Check whether position in the parent is correct. */
985 PATH_OFFSET_POSITION(p_s_path,
987 B_NR_ITEMS(p_s_parent))
988 return REPEAT_SEARCH;
989 /* Check whether parent at the path really points to the child. */
990 if (B_N_CHILD_NUM(p_s_parent, n_position) !=
991 PATH_OFFSET_PBUFFER(p_s_path, n_counter)->b_blocknr)
992 return REPEAT_SEARCH;
993 /* Return delimiting key if position in the parent is not equal to first/last one. */
994 if (c_lr_par == RIGHT_PARENTS)
995 n_first_last_position = B_NR_ITEMS(p_s_parent);
996 if (n_position != n_first_last_position) {
997 *pp_s_com_father = p_s_parent;
998 get_bh(*pp_s_com_father);
999 /*(*pp_s_com_father = p_s_parent)->b_count++; */
1004 /* if we are in the root of the tree, then there is no common father */
1005 if (n_counter == FIRST_PATH_ELEMENT_OFFSET) {
1006 /* Check whether first buffer in the path is the root of the tree. */
1007 if (PATH_OFFSET_PBUFFER
1009 FIRST_PATH_ELEMENT_OFFSET)->b_blocknr ==
1010 SB_ROOT_BLOCK(p_s_tb->tb_sb)) {
1011 *pp_s_father = *pp_s_com_father = NULL;
1014 return REPEAT_SEARCH;
1017 RFALSE(B_LEVEL(*pp_s_com_father) <= DISK_LEAF_NODE_LEVEL,
1018 "PAP-8185: (%b %z) level too small",
1019 *pp_s_com_father, *pp_s_com_father);
1021 /* Check whether the common parent is locked. */
1023 if (buffer_locked(*pp_s_com_father)) {
1024 __wait_on_buffer(*pp_s_com_father);
1025 if (FILESYSTEM_CHANGED_TB(p_s_tb)) {
1026 brelse(*pp_s_com_father);
1027 return REPEAT_SEARCH;
1031 /* So, we got common parent of the current node and its left/right neighbor.
1032 Now we are geting the parent of the left/right neighbor. */
1034 /* Form key to get parent of the left/right neighbor. */
1035 le_key2cpu_key(&s_lr_father_key,
1036 B_N_PDELIM_KEY(*pp_s_com_father,
1038 LEFT_PARENTS) ? (p_s_tb->lkey[n_h - 1] =
1040 1) : (p_s_tb->rkey[n_h -
1044 if (c_lr_par == LEFT_PARENTS)
1045 decrement_key(&s_lr_father_key);
1048 (p_s_tb->tb_sb, &s_lr_father_key, &s_path_to_neighbor_father,
1049 n_h + 1) == IO_ERROR)
1053 if (FILESYSTEM_CHANGED_TB(p_s_tb)) {
1054 pathrelse(&s_path_to_neighbor_father);
1055 brelse(*pp_s_com_father);
1056 return REPEAT_SEARCH;
1059 *pp_s_father = PATH_PLAST_BUFFER(&s_path_to_neighbor_father);
1061 RFALSE(B_LEVEL(*pp_s_father) != n_h + 1,
1062 "PAP-8190: (%b %z) level too small", *pp_s_father, *pp_s_father);
1063 RFALSE(s_path_to_neighbor_father.path_length <
1064 FIRST_PATH_ELEMENT_OFFSET, "PAP-8192: path length is too small");
1066 s_path_to_neighbor_father.path_length--;
1067 pathrelse(&s_path_to_neighbor_father);
1071 /* Get parents of neighbors of node in the path(S[n_path_offset]) and common parents of
1072 * S[n_path_offset] and L[n_path_offset]/R[n_path_offset]: F[n_path_offset], FL[n_path_offset],
1073 * FR[n_path_offset], CFL[n_path_offset], CFR[n_path_offset].
1074 * Calculate numbers of left and right delimiting keys position: lkey[n_path_offset], rkey[n_path_offset].
1075 * Returns: SCHEDULE_OCCURRED - schedule occurred while the function worked;
1076 * CARRY_ON - schedule didn't occur while the function worked;
1078 static int get_parents(struct tree_balance *p_s_tb, int n_h)
1080 struct treepath *p_s_path = p_s_tb->tb_path;
1083 n_path_offset = PATH_H_PATH_OFFSET(p_s_tb->tb_path, n_h);
1084 struct buffer_head *p_s_curf, *p_s_curcf;
1086 /* Current node is the root of the tree or will be root of the tree */
1087 if (n_path_offset <= FIRST_PATH_ELEMENT_OFFSET) {
1088 /* The root can not have parents.
1089 Release nodes which previously were obtained as parents of the current node neighbors. */
1090 brelse(p_s_tb->FL[n_h]);
1091 brelse(p_s_tb->CFL[n_h]);
1092 brelse(p_s_tb->FR[n_h]);
1093 brelse(p_s_tb->CFR[n_h]);
1094 p_s_tb->FL[n_h] = p_s_tb->CFL[n_h] = p_s_tb->FR[n_h] =
1095 p_s_tb->CFR[n_h] = NULL;
1099 /* Get parent FL[n_path_offset] of L[n_path_offset]. */
1100 if ((n_position = PATH_OFFSET_POSITION(p_s_path, n_path_offset - 1))) {
1101 /* Current node is not the first child of its parent. */
1102 /*(p_s_curf = p_s_curcf = PATH_OFFSET_PBUFFER(p_s_path, n_path_offset - 1))->b_count += 2; */
1103 p_s_curf = p_s_curcf =
1104 PATH_OFFSET_PBUFFER(p_s_path, n_path_offset - 1);
1107 p_s_tb->lkey[n_h] = n_position - 1;
1109 /* Calculate current parent of L[n_path_offset], which is the left neighbor of the current node.
1110 Calculate current common parent of L[n_path_offset] and the current node. Note that
1111 CFL[n_path_offset] not equal FL[n_path_offset] and CFL[n_path_offset] not equal F[n_path_offset].
1112 Calculate lkey[n_path_offset]. */
1113 if ((n_ret_value = get_far_parent(p_s_tb, n_h + 1, &p_s_curf,
1115 LEFT_PARENTS)) != CARRY_ON)
1119 brelse(p_s_tb->FL[n_h]);
1120 p_s_tb->FL[n_h] = p_s_curf; /* New initialization of FL[n_h]. */
1121 brelse(p_s_tb->CFL[n_h]);
1122 p_s_tb->CFL[n_h] = p_s_curcf; /* New initialization of CFL[n_h]. */
1124 RFALSE((p_s_curf && !B_IS_IN_TREE(p_s_curf)) ||
1125 (p_s_curcf && !B_IS_IN_TREE(p_s_curcf)),
1126 "PAP-8195: FL (%b) or CFL (%b) is invalid", p_s_curf, p_s_curcf);
1128 /* Get parent FR[n_h] of R[n_h]. */
1130 /* Current node is the last child of F[n_h]. FR[n_h] != F[n_h]. */
1131 if (n_position == B_NR_ITEMS(PATH_H_PBUFFER(p_s_path, n_h + 1))) {
1132 /* Calculate current parent of R[n_h], which is the right neighbor of F[n_h].
1133 Calculate current common parent of R[n_h] and current node. Note that CFR[n_h]
1134 not equal FR[n_path_offset] and CFR[n_h] not equal F[n_h]. */
1136 get_far_parent(p_s_tb, n_h + 1, &p_s_curf, &p_s_curcf,
1137 RIGHT_PARENTS)) != CARRY_ON)
1140 /* Current node is not the last child of its parent F[n_h]. */
1141 /*(p_s_curf = p_s_curcf = PATH_OFFSET_PBUFFER(p_s_path, n_path_offset - 1))->b_count += 2; */
1142 p_s_curf = p_s_curcf =
1143 PATH_OFFSET_PBUFFER(p_s_path, n_path_offset - 1);
1146 p_s_tb->rkey[n_h] = n_position;
1149 brelse(p_s_tb->FR[n_h]);
1150 p_s_tb->FR[n_h] = p_s_curf; /* New initialization of FR[n_path_offset]. */
1152 brelse(p_s_tb->CFR[n_h]);
1153 p_s_tb->CFR[n_h] = p_s_curcf; /* New initialization of CFR[n_path_offset]. */
1155 RFALSE((p_s_curf && !B_IS_IN_TREE(p_s_curf)) ||
1156 (p_s_curcf && !B_IS_IN_TREE(p_s_curcf)),
1157 "PAP-8205: FR (%b) or CFR (%b) is invalid", p_s_curf, p_s_curcf);
1162 /* it is possible to remove node as result of shiftings to
1163 neighbors even when we insert or paste item. */
1164 static inline int can_node_be_removed(int mode, int lfree, int sfree, int rfree,
1165 struct tree_balance *tb, int h)
1167 struct buffer_head *Sh = PATH_H_PBUFFER(tb->tb_path, h);
1168 int levbytes = tb->insert_size[h];
1169 struct item_head *ih;
1170 struct reiserfs_key *r_key = NULL;
1172 ih = B_N_PITEM_HEAD(Sh, 0);
1174 r_key = B_N_PDELIM_KEY(tb->CFR[h], tb->rkey[h]);
1176 if (lfree + rfree + sfree < MAX_CHILD_SIZE(Sh) + levbytes
1177 /* shifting may merge items which might save space */
1180 && op_is_left_mergeable(&(ih->ih_key), Sh->b_size)) ? IH_SIZE : 0)
1183 && op_is_left_mergeable(r_key, Sh->b_size)) ? IH_SIZE : 0)
1184 + ((h) ? KEY_SIZE : 0)) {
1185 /* node can not be removed */
1186 if (sfree >= levbytes) { /* new item fits into node S[h] without any shifting */
1190 ((mode == M_INSERT) ? 1 : 0);
1191 set_parameters(tb, h, 0, 0, 1, NULL, -1, -1);
1192 return NO_BALANCING_NEEDED;
1195 PROC_INFO_INC(tb->tb_sb, can_node_be_removed[h]);
1196 return !NO_BALANCING_NEEDED;
1199 /* Check whether current node S[h] is balanced when increasing its size by
1200 * Inserting or Pasting.
1201 * Calculate parameters for balancing for current level h.
1203 * tb tree_balance structure;
1204 * h current level of the node;
1205 * inum item number in S[h];
1206 * mode i - insert, p - paste;
1207 * Returns: 1 - schedule occurred;
1208 * 0 - balancing for higher levels needed;
1209 * -1 - no balancing for higher levels needed;
1210 * -2 - no disk space.
1212 /* ip means Inserting or Pasting */
1213 static int ip_check_balance(struct tree_balance *tb, int h)
1215 struct virtual_node *vn = tb->tb_vn;
1216 int levbytes, /* Number of bytes that must be inserted into (value
1217 is negative if bytes are deleted) buffer which
1218 contains node being balanced. The mnemonic is
1219 that the attempted change in node space used level
1220 is levbytes bytes. */
1223 int lfree, sfree, rfree /* free space in L, S and R */ ;
1225 /* nver is short for number of vertixes, and lnver is the number if
1226 we shift to the left, rnver is the number if we shift to the
1227 right, and lrnver is the number if we shift in both directions.
1228 The goal is to minimize first the number of vertixes, and second,
1229 the number of vertixes whose contents are changed by shifting,
1230 and third the number of uncached vertixes whose contents are
1231 changed by shifting and must be read from disk. */
1232 int nver, lnver, rnver, lrnver;
1234 /* used at leaf level only, S0 = S[0] is the node being balanced,
1235 sInum [ I = 0,1,2 ] is the number of items that will
1236 remain in node SI after balancing. S1 and S2 are new
1237 nodes that might be created. */
1239 /* we perform 8 calls to get_num_ver(). For each call we calculate five parameters.
1240 where 4th parameter is s1bytes and 5th - s2bytes
1242 short snum012[40] = { 0, }; /* s0num, s1num, s2num for 8 cases
1243 0,1 - do not shift and do not shift but bottle
1244 2 - shift only whole item to left
1245 3 - shift to left and bottle as much as possible
1246 4,5 - shift to right (whole items and as much as possible
1247 6,7 - shift to both directions (whole items and as much as possible)
1250 /* Sh is the node whose balance is currently being checked */
1251 struct buffer_head *Sh;
1253 Sh = PATH_H_PBUFFER(tb->tb_path, h);
1254 levbytes = tb->insert_size[h];
1256 /* Calculate balance parameters for creating new root. */
1259 reiserfs_panic(tb->tb_sb, "vs-8210",
1260 "S[0] can not be 0");
1261 switch (n_ret_value = get_empty_nodes(tb, h)) {
1263 set_parameters(tb, h, 0, 0, 1, NULL, -1, -1);
1264 return NO_BALANCING_NEEDED; /* no balancing for higher levels needed */
1270 reiserfs_panic(tb->tb_sb, "vs-8215", "incorrect "
1271 "return value of get_empty_nodes");
1275 if ((n_ret_value = get_parents(tb, h)) != CARRY_ON) /* get parents of S[h] neighbors. */
1278 sfree = B_FREE_SPACE(Sh);
1280 /* get free space of neighbors */
1281 rfree = get_rfree(tb, h);
1282 lfree = get_lfree(tb, h);
1284 if (can_node_be_removed(vn->vn_mode, lfree, sfree, rfree, tb, h) ==
1285 NO_BALANCING_NEEDED)
1286 /* and new item fits into node S[h] without any shifting */
1287 return NO_BALANCING_NEEDED;
1289 create_virtual_node(tb, h);
1292 determine maximal number of items we can shift to the left neighbor (in tb structure)
1293 and the maximal number of bytes that can flow to the left neighbor
1294 from the left most liquid item that cannot be shifted from S[0] entirely (returned value)
1296 check_left(tb, h, lfree);
1299 determine maximal number of items we can shift to the right neighbor (in tb structure)
1300 and the maximal number of bytes that can flow to the right neighbor
1301 from the right most liquid item that cannot be shifted from S[0] entirely (returned value)
1303 check_right(tb, h, rfree);
1305 /* all contents of internal node S[h] can be moved into its
1306 neighbors, S[h] will be removed after balancing */
1307 if (h && (tb->rnum[h] + tb->lnum[h] >= vn->vn_nr_item + 1)) {
1310 /* Since we are working on internal nodes, and our internal
1311 nodes have fixed size entries, then we can balance by the
1312 number of items rather than the space they consume. In this
1313 routine we set the left node equal to the right node,
1314 allowing a difference of less than or equal to 1 child
1317 ((MAX_NR_KEY(Sh) << 1) + 2 - tb->lnum[h] - tb->rnum[h] +
1318 vn->vn_nr_item + 1) / 2 - (MAX_NR_KEY(Sh) + 1 -
1320 set_parameters(tb, h, vn->vn_nr_item + 1 - to_r, to_r, 0, NULL,
1325 /* this checks balance condition, that any two neighboring nodes can not fit in one node */
1327 (tb->lnum[h] >= vn->vn_nr_item + 1 ||
1328 tb->rnum[h] >= vn->vn_nr_item + 1),
1329 "vs-8220: tree is not balanced on internal level");
1330 RFALSE(!h && ((tb->lnum[h] >= vn->vn_nr_item && (tb->lbytes == -1)) ||
1331 (tb->rnum[h] >= vn->vn_nr_item && (tb->rbytes == -1))),
1332 "vs-8225: tree is not balanced on leaf level");
1334 /* all contents of S[0] can be moved into its neighbors
1335 S[0] will be removed after balancing. */
1336 if (!h && is_leaf_removable(tb))
1339 /* why do we perform this check here rather than earlier??
1340 Answer: we can win 1 node in some cases above. Moreover we
1341 checked it above, when we checked, that S[0] is not removable
1343 if (sfree >= levbytes) { /* new item fits into node S[h] without any shifting */
1345 tb->s0num = vn->vn_nr_item;
1346 set_parameters(tb, h, 0, 0, 1, NULL, -1, -1);
1347 return NO_BALANCING_NEEDED;
1351 int lpar, rpar, nset, lset, rset, lrset;
1353 * regular overflowing of the node
1356 /* get_num_ver works in 2 modes (FLOW & NO_FLOW)
1357 lpar, rpar - number of items we can shift to left/right neighbor (including splitting item)
1358 nset, lset, rset, lrset - shows, whether flowing items give better packing
1361 #define NO_FLOW 0 /* do not any splitting */
1363 /* we choose one the following */
1364 #define NOTHING_SHIFT_NO_FLOW 0
1365 #define NOTHING_SHIFT_FLOW 5
1366 #define LEFT_SHIFT_NO_FLOW 10
1367 #define LEFT_SHIFT_FLOW 15
1368 #define RIGHT_SHIFT_NO_FLOW 20
1369 #define RIGHT_SHIFT_FLOW 25
1370 #define LR_SHIFT_NO_FLOW 30
1371 #define LR_SHIFT_FLOW 35
1376 /* calculate number of blocks S[h] must be split into when
1377 nothing is shifted to the neighbors,
1378 as well as number of items in each part of the split node (s012 numbers),
1379 and number of bytes (s1bytes) of the shared drop which flow to S1 if any */
1380 nset = NOTHING_SHIFT_NO_FLOW;
1381 nver = get_num_ver(vn->vn_mode, tb, h,
1382 0, -1, h ? vn->vn_nr_item : 0, -1,
1388 /* note, that in this case we try to bottle between S[0] and S1 (S1 - the first new node) */
1389 nver1 = get_num_ver(vn->vn_mode, tb, h,
1391 snum012 + NOTHING_SHIFT_FLOW, FLOW);
1393 nset = NOTHING_SHIFT_FLOW, nver = nver1;
1396 /* calculate number of blocks S[h] must be split into when
1397 l_shift_num first items and l_shift_bytes of the right most
1398 liquid item to be shifted are shifted to the left neighbor,
1399 as well as number of items in each part of the splitted node (s012 numbers),
1400 and number of bytes (s1bytes) of the shared drop which flow to S1 if any
1402 lset = LEFT_SHIFT_NO_FLOW;
1403 lnver = get_num_ver(vn->vn_mode, tb, h,
1404 lpar - ((h || tb->lbytes == -1) ? 0 : 1),
1405 -1, h ? vn->vn_nr_item : 0, -1,
1406 snum012 + LEFT_SHIFT_NO_FLOW, NO_FLOW);
1410 lnver1 = get_num_ver(vn->vn_mode, tb, h,
1412 ((tb->lbytes != -1) ? 1 : 0),
1414 snum012 + LEFT_SHIFT_FLOW, FLOW);
1416 lset = LEFT_SHIFT_FLOW, lnver = lnver1;
1419 /* calculate number of blocks S[h] must be split into when
1420 r_shift_num first items and r_shift_bytes of the left most
1421 liquid item to be shifted are shifted to the right neighbor,
1422 as well as number of items in each part of the splitted node (s012 numbers),
1423 and number of bytes (s1bytes) of the shared drop which flow to S1 if any
1425 rset = RIGHT_SHIFT_NO_FLOW;
1426 rnver = get_num_ver(vn->vn_mode, tb, h,
1428 h ? (vn->vn_nr_item - rpar) : (rpar -
1433 snum012 + RIGHT_SHIFT_NO_FLOW, NO_FLOW);
1437 rnver1 = get_num_ver(vn->vn_mode, tb, h,
1440 ((tb->rbytes != -1) ? 1 : 0)),
1442 snum012 + RIGHT_SHIFT_FLOW, FLOW);
1445 rset = RIGHT_SHIFT_FLOW, rnver = rnver1;
1448 /* calculate number of blocks S[h] must be split into when
1449 items are shifted in both directions,
1450 as well as number of items in each part of the splitted node (s012 numbers),
1451 and number of bytes (s1bytes) of the shared drop which flow to S1 if any
1453 lrset = LR_SHIFT_NO_FLOW;
1454 lrnver = get_num_ver(vn->vn_mode, tb, h,
1455 lpar - ((h || tb->lbytes == -1) ? 0 : 1),
1457 h ? (vn->vn_nr_item - rpar) : (rpar -
1462 snum012 + LR_SHIFT_NO_FLOW, NO_FLOW);
1466 lrnver1 = get_num_ver(vn->vn_mode, tb, h,
1468 ((tb->lbytes != -1) ? 1 : 0),
1471 ((tb->rbytes != -1) ? 1 : 0)),
1473 snum012 + LR_SHIFT_FLOW, FLOW);
1474 if (lrnver > lrnver1)
1475 lrset = LR_SHIFT_FLOW, lrnver = lrnver1;
1478 /* Our general shifting strategy is:
1479 1) to minimized number of new nodes;
1480 2) to minimized number of neighbors involved in shifting;
1481 3) to minimized number of disk reads; */
1483 /* we can win TWO or ONE nodes by shifting in both directions */
1484 if (lrnver < lnver && lrnver < rnver) {
1486 (tb->lnum[h] != 1 ||
1488 lrnver != 1 || rnver != 2 || lnver != 2
1489 || h != 1), "vs-8230: bad h");
1490 if (lrset == LR_SHIFT_FLOW)
1491 set_parameters(tb, h, tb->lnum[h], tb->rnum[h],
1492 lrnver, snum012 + lrset,
1493 tb->lbytes, tb->rbytes);
1495 set_parameters(tb, h,
1497 ((tb->lbytes == -1) ? 0 : 1),
1499 ((tb->rbytes == -1) ? 0 : 1),
1500 lrnver, snum012 + lrset, -1, -1);
1505 /* if shifting doesn't lead to better packing then don't shift */
1506 if (nver == lrnver) {
1507 set_parameters(tb, h, 0, 0, nver, snum012 + nset, -1,
1512 /* now we know that for better packing shifting in only one
1513 direction either to the left or to the right is required */
1515 /* if shifting to the left is better than shifting to the right */
1516 if (lnver < rnver) {
1521 /* if shifting to the right is better than shifting to the left */
1522 if (lnver > rnver) {
1523 SET_PAR_SHIFT_RIGHT;
1527 /* now shifting in either direction gives the same number
1528 of nodes and we can make use of the cached neighbors */
1529 if (is_left_neighbor_in_cache(tb, h)) {
1534 /* shift to the right independently on whether the right neighbor in cache or not */
1535 SET_PAR_SHIFT_RIGHT;
1540 /* Check whether current node S[h] is balanced when Decreasing its size by
1541 * Deleting or Cutting for INTERNAL node of S+tree.
1542 * Calculate parameters for balancing for current level h.
1544 * tb tree_balance structure;
1545 * h current level of the node;
1546 * inum item number in S[h];
1547 * mode i - insert, p - paste;
1548 * Returns: 1 - schedule occurred;
1549 * 0 - balancing for higher levels needed;
1550 * -1 - no balancing for higher levels needed;
1551 * -2 - no disk space.
1553 * Note: Items of internal nodes have fixed size, so the balance condition for
1554 * the internal part of S+tree is as for the B-trees.
1556 static int dc_check_balance_internal(struct tree_balance *tb, int h)
1558 struct virtual_node *vn = tb->tb_vn;
1560 /* Sh is the node whose balance is currently being checked,
1561 and Fh is its father. */
1562 struct buffer_head *Sh, *Fh;
1563 int maxsize, n_ret_value;
1564 int lfree, rfree /* free space in L and R */ ;
1566 Sh = PATH_H_PBUFFER(tb->tb_path, h);
1567 Fh = PATH_H_PPARENT(tb->tb_path, h);
1569 maxsize = MAX_CHILD_SIZE(Sh);
1571 /* using tb->insert_size[h], which is negative in this case, create_virtual_node calculates: */
1572 /* new_nr_item = number of items node would have if operation is */
1573 /* performed without balancing (new_nr_item); */
1574 create_virtual_node(tb, h);
1576 if (!Fh) { /* S[h] is the root. */
1577 if (vn->vn_nr_item > 0) {
1578 set_parameters(tb, h, 0, 0, 1, NULL, -1, -1);
1579 return NO_BALANCING_NEEDED; /* no balancing for higher levels needed */
1581 /* new_nr_item == 0.
1582 * Current root will be deleted resulting in
1583 * decrementing the tree height. */
1584 set_parameters(tb, h, 0, 0, 0, NULL, -1, -1);
1588 if ((n_ret_value = get_parents(tb, h)) != CARRY_ON)
1591 /* get free space of neighbors */
1592 rfree = get_rfree(tb, h);
1593 lfree = get_lfree(tb, h);
1595 /* determine maximal number of items we can fit into neighbors */
1596 check_left(tb, h, lfree);
1597 check_right(tb, h, rfree);
1599 if (vn->vn_nr_item >= MIN_NR_KEY(Sh)) { /* Balance condition for the internal node is valid.
1600 * In this case we balance only if it leads to better packing. */
1601 if (vn->vn_nr_item == MIN_NR_KEY(Sh)) { /* Here we join S[h] with one of its neighbors,
1602 * which is impossible with greater values of new_nr_item. */
1603 if (tb->lnum[h] >= vn->vn_nr_item + 1) {
1604 /* All contents of S[h] can be moved to L[h]. */
1610 PATH_H_B_ITEM_ORDER(tb->tb_path,
1612 0) ? B_NR_ITEMS(tb->FL[h]) : n - 1;
1613 n = dc_size(B_N_CHILD(tb->FL[h], order_L)) /
1614 (DC_SIZE + KEY_SIZE);
1615 set_parameters(tb, h, -n - 1, 0, 0, NULL, -1,
1620 if (tb->rnum[h] >= vn->vn_nr_item + 1) {
1621 /* All contents of S[h] can be moved to R[h]. */
1627 PATH_H_B_ITEM_ORDER(tb->tb_path,
1629 B_NR_ITEMS(Fh)) ? 0 : n + 1;
1630 n = dc_size(B_N_CHILD(tb->FR[h], order_R)) /
1631 (DC_SIZE + KEY_SIZE);
1632 set_parameters(tb, h, 0, -n - 1, 0, NULL, -1,
1638 if (tb->rnum[h] + tb->lnum[h] >= vn->vn_nr_item + 1) {
1639 /* All contents of S[h] can be moved to the neighbors (L[h] & R[h]). */
1643 ((MAX_NR_KEY(Sh) << 1) + 2 - tb->lnum[h] -
1644 tb->rnum[h] + vn->vn_nr_item + 1) / 2 -
1645 (MAX_NR_KEY(Sh) + 1 - tb->rnum[h]);
1646 set_parameters(tb, h, vn->vn_nr_item + 1 - to_r, to_r,
1651 /* Balancing does not lead to better packing. */
1652 set_parameters(tb, h, 0, 0, 1, NULL, -1, -1);
1653 return NO_BALANCING_NEEDED;
1656 /* Current node contain insufficient number of items. Balancing is required. */
1657 /* Check whether we can merge S[h] with left neighbor. */
1658 if (tb->lnum[h] >= vn->vn_nr_item + 1)
1659 if (is_left_neighbor_in_cache(tb, h)
1660 || tb->rnum[h] < vn->vn_nr_item + 1 || !tb->FR[h]) {
1666 PATH_H_B_ITEM_ORDER(tb->tb_path,
1668 0) ? B_NR_ITEMS(tb->FL[h]) : n - 1;
1669 n = dc_size(B_N_CHILD(tb->FL[h], order_L)) / (DC_SIZE +
1671 set_parameters(tb, h, -n - 1, 0, 0, NULL, -1, -1);
1675 /* Check whether we can merge S[h] with right neighbor. */
1676 if (tb->rnum[h] >= vn->vn_nr_item + 1) {
1682 PATH_H_B_ITEM_ORDER(tb->tb_path,
1683 h)) == B_NR_ITEMS(Fh)) ? 0 : (n + 1);
1684 n = dc_size(B_N_CHILD(tb->FR[h], order_R)) / (DC_SIZE +
1686 set_parameters(tb, h, 0, -n - 1, 0, NULL, -1, -1);
1690 /* All contents of S[h] can be moved to the neighbors (L[h] & R[h]). */
1691 if (tb->rnum[h] + tb->lnum[h] >= vn->vn_nr_item + 1) {
1695 ((MAX_NR_KEY(Sh) << 1) + 2 - tb->lnum[h] - tb->rnum[h] +
1696 vn->vn_nr_item + 1) / 2 - (MAX_NR_KEY(Sh) + 1 -
1698 set_parameters(tb, h, vn->vn_nr_item + 1 - to_r, to_r, 0, NULL,
1703 /* For internal nodes try to borrow item from a neighbor */
1704 RFALSE(!tb->FL[h] && !tb->FR[h], "vs-8235: trying to borrow for root");
1706 /* Borrow one or two items from caching neighbor */
1707 if (is_left_neighbor_in_cache(tb, h) || !tb->FR[h]) {
1711 (MAX_NR_KEY(Sh) + 1 - tb->lnum[h] + vn->vn_nr_item +
1712 1) / 2 - (vn->vn_nr_item + 1);
1713 set_parameters(tb, h, -from_l, 0, 1, NULL, -1, -1);
1717 set_parameters(tb, h, 0,
1718 -((MAX_NR_KEY(Sh) + 1 - tb->rnum[h] + vn->vn_nr_item +
1719 1) / 2 - (vn->vn_nr_item + 1)), 1, NULL, -1, -1);
1723 /* Check whether current node S[h] is balanced when Decreasing its size by
1724 * Deleting or Truncating for LEAF node of S+tree.
1725 * Calculate parameters for balancing for current level h.
1727 * tb tree_balance structure;
1728 * h current level of the node;
1729 * inum item number in S[h];
1730 * mode i - insert, p - paste;
1731 * Returns: 1 - schedule occurred;
1732 * 0 - balancing for higher levels needed;
1733 * -1 - no balancing for higher levels needed;
1734 * -2 - no disk space.
1736 static int dc_check_balance_leaf(struct tree_balance *tb, int h)
1738 struct virtual_node *vn = tb->tb_vn;
1740 /* Number of bytes that must be deleted from
1741 (value is negative if bytes are deleted) buffer which
1742 contains node being balanced. The mnemonic is that the
1743 attempted change in node space used level is levbytes bytes. */
1745 /* the maximal item size */
1746 int maxsize, n_ret_value;
1747 /* S0 is the node whose balance is currently being checked,
1748 and F0 is its father. */
1749 struct buffer_head *S0, *F0;
1750 int lfree, rfree /* free space in L and R */ ;
1752 S0 = PATH_H_PBUFFER(tb->tb_path, 0);
1753 F0 = PATH_H_PPARENT(tb->tb_path, 0);
1755 levbytes = tb->insert_size[h];
1757 maxsize = MAX_CHILD_SIZE(S0); /* maximal possible size of an item */
1759 if (!F0) { /* S[0] is the root now. */
1761 RFALSE(-levbytes >= maxsize - B_FREE_SPACE(S0),
1762 "vs-8240: attempt to create empty buffer tree");
1764 set_parameters(tb, h, 0, 0, 1, NULL, -1, -1);
1765 return NO_BALANCING_NEEDED;
1768 if ((n_ret_value = get_parents(tb, h)) != CARRY_ON)
1771 /* get free space of neighbors */
1772 rfree = get_rfree(tb, h);
1773 lfree = get_lfree(tb, h);
1775 create_virtual_node(tb, h);
1777 /* if 3 leaves can be merge to one, set parameters and return */
1778 if (are_leaves_removable(tb, lfree, rfree))
1781 /* determine maximal number of items we can shift to the left/right neighbor
1782 and the maximal number of bytes that can flow to the left/right neighbor
1783 from the left/right most liquid item that cannot be shifted from S[0] entirely
1785 check_left(tb, h, lfree);
1786 check_right(tb, h, rfree);
1788 /* check whether we can merge S with left neighbor. */
1789 if (tb->lnum[0] >= vn->vn_nr_item && tb->lbytes == -1)
1790 if (is_left_neighbor_in_cache(tb, h) || ((tb->rnum[0] - ((tb->rbytes == -1) ? 0 : 1)) < vn->vn_nr_item) || /* S can not be merged with R */
1794 "vs-8245: dc_check_balance_leaf: FL[h] must exist");
1796 /* set parameter to merge S[0] with its left neighbor */
1797 set_parameters(tb, h, -1, 0, 0, NULL, -1, -1);
1801 /* check whether we can merge S[0] with right neighbor. */
1802 if (tb->rnum[0] >= vn->vn_nr_item && tb->rbytes == -1) {
1803 set_parameters(tb, h, 0, -1, 0, NULL, -1, -1);
1807 /* All contents of S[0] can be moved to the neighbors (L[0] & R[0]). Set parameters and return */
1808 if (is_leaf_removable(tb))
1811 /* Balancing is not required. */
1812 tb->s0num = vn->vn_nr_item;
1813 set_parameters(tb, h, 0, 0, 1, NULL, -1, -1);
1814 return NO_BALANCING_NEEDED;
1817 /* Check whether current node S[h] is balanced when Decreasing its size by
1818 * Deleting or Cutting.
1819 * Calculate parameters for balancing for current level h.
1821 * tb tree_balance structure;
1822 * h current level of the node;
1823 * inum item number in S[h];
1824 * mode d - delete, c - cut.
1825 * Returns: 1 - schedule occurred;
1826 * 0 - balancing for higher levels needed;
1827 * -1 - no balancing for higher levels needed;
1828 * -2 - no disk space.
1830 static int dc_check_balance(struct tree_balance *tb, int h)
1832 RFALSE(!(PATH_H_PBUFFER(tb->tb_path, h)),
1833 "vs-8250: S is not initialized");
1836 return dc_check_balance_internal(tb, h);
1838 return dc_check_balance_leaf(tb, h);
1841 /* Check whether current node S[h] is balanced.
1842 * Calculate parameters for balancing for current level h.
1845 * tb tree_balance structure:
1847 * tb is a large structure that must be read about in the header file
1848 * at the same time as this procedure if the reader is to successfully
1849 * understand this procedure
1851 * h current level of the node;
1852 * inum item number in S[h];
1853 * mode i - insert, p - paste, d - delete, c - cut.
1854 * Returns: 1 - schedule occurred;
1855 * 0 - balancing for higher levels needed;
1856 * -1 - no balancing for higher levels needed;
1857 * -2 - no disk space.
1859 static int check_balance(int mode,
1860 struct tree_balance *tb,
1864 struct item_head *ins_ih, const void *data)
1866 struct virtual_node *vn;
1868 vn = tb->tb_vn = (struct virtual_node *)(tb->vn_buf);
1869 vn->vn_free_ptr = (char *)(tb->tb_vn + 1);
1871 vn->vn_affected_item_num = inum;
1872 vn->vn_pos_in_item = pos_in_item;
1873 vn->vn_ins_ih = ins_ih;
1876 RFALSE(mode == M_INSERT && !vn->vn_ins_ih,
1877 "vs-8255: ins_ih can not be 0 in insert mode");
1879 if (tb->insert_size[h] > 0)
1880 /* Calculate balance parameters when size of node is increasing. */
1881 return ip_check_balance(tb, h);
1883 /* Calculate balance parameters when size of node is decreasing. */
1884 return dc_check_balance(tb, h);
1887 /* Check whether parent at the path is the really parent of the current node.*/
1888 static int get_direct_parent(struct tree_balance *p_s_tb, int n_h)
1890 struct buffer_head *bh;
1891 struct treepath *p_s_path = p_s_tb->tb_path;
1893 n_path_offset = PATH_H_PATH_OFFSET(p_s_tb->tb_path, n_h);
1895 /* We are in the root or in the new root. */
1896 if (n_path_offset <= FIRST_PATH_ELEMENT_OFFSET) {
1898 RFALSE(n_path_offset < FIRST_PATH_ELEMENT_OFFSET - 1,
1899 "PAP-8260: invalid offset in the path");
1901 if (PATH_OFFSET_PBUFFER(p_s_path, FIRST_PATH_ELEMENT_OFFSET)->
1902 b_blocknr == SB_ROOT_BLOCK(p_s_tb->tb_sb)) {
1903 /* Root is not changed. */
1904 PATH_OFFSET_PBUFFER(p_s_path, n_path_offset - 1) = NULL;
1905 PATH_OFFSET_POSITION(p_s_path, n_path_offset - 1) = 0;
1908 return REPEAT_SEARCH; /* Root is changed and we must recalculate the path. */
1912 (bh = PATH_OFFSET_PBUFFER(p_s_path, n_path_offset - 1)))
1913 return REPEAT_SEARCH; /* Parent in the path is not in the tree. */
1916 PATH_OFFSET_POSITION(p_s_path,
1917 n_path_offset - 1)) > B_NR_ITEMS(bh))
1918 return REPEAT_SEARCH;
1920 if (B_N_CHILD_NUM(bh, n_position) !=
1921 PATH_OFFSET_PBUFFER(p_s_path, n_path_offset)->b_blocknr)
1922 /* Parent in the path is not parent of the current node in the tree. */
1923 return REPEAT_SEARCH;
1925 if (buffer_locked(bh)) {
1926 __wait_on_buffer(bh);
1927 if (FILESYSTEM_CHANGED_TB(p_s_tb))
1928 return REPEAT_SEARCH;
1931 return CARRY_ON; /* Parent in the path is unlocked and really parent of the current node. */
1934 /* Using lnum[n_h] and rnum[n_h] we should determine what neighbors
1936 * need in order to balance S[n_h], and get them if necessary.
1937 * Returns: SCHEDULE_OCCURRED - schedule occurred while the function worked;
1938 * CARRY_ON - schedule didn't occur while the function worked;
1940 static int get_neighbors(struct tree_balance *p_s_tb, int n_h)
1942 int n_child_position,
1943 n_path_offset = PATH_H_PATH_OFFSET(p_s_tb->tb_path, n_h + 1);
1944 unsigned long n_son_number;
1945 struct super_block *sb = p_s_tb->tb_sb;
1946 struct buffer_head *bh;
1948 PROC_INFO_INC(sb, get_neighbors[n_h]);
1950 if (p_s_tb->lnum[n_h]) {
1951 /* We need left neighbor to balance S[n_h]. */
1952 PROC_INFO_INC(sb, need_l_neighbor[n_h]);
1953 bh = PATH_OFFSET_PBUFFER(p_s_tb->tb_path, n_path_offset);
1955 RFALSE(bh == p_s_tb->FL[n_h] &&
1956 !PATH_OFFSET_POSITION(p_s_tb->tb_path, n_path_offset),
1957 "PAP-8270: invalid position in the parent");
1961 p_s_tb->FL[n_h]) ? p_s_tb->lkey[n_h] : B_NR_ITEMS(p_s_tb->
1963 n_son_number = B_N_CHILD_NUM(p_s_tb->FL[n_h], n_child_position);
1964 bh = sb_bread(sb, n_son_number);
1967 if (FILESYSTEM_CHANGED_TB(p_s_tb)) {
1969 PROC_INFO_INC(sb, get_neighbors_restart[n_h]);
1970 return REPEAT_SEARCH;
1973 RFALSE(!B_IS_IN_TREE(p_s_tb->FL[n_h]) ||
1974 n_child_position > B_NR_ITEMS(p_s_tb->FL[n_h]) ||
1975 B_N_CHILD_NUM(p_s_tb->FL[n_h], n_child_position) !=
1976 bh->b_blocknr, "PAP-8275: invalid parent");
1977 RFALSE(!B_IS_IN_TREE(bh), "PAP-8280: invalid child");
1980 MAX_CHILD_SIZE(bh) -
1981 dc_size(B_N_CHILD(p_s_tb->FL[0], n_child_position)),
1982 "PAP-8290: invalid child size of left neighbor");
1984 brelse(p_s_tb->L[n_h]);
1985 p_s_tb->L[n_h] = bh;
1988 if (p_s_tb->rnum[n_h]) { /* We need right neighbor to balance S[n_path_offset]. */
1989 PROC_INFO_INC(sb, need_r_neighbor[n_h]);
1990 bh = PATH_OFFSET_PBUFFER(p_s_tb->tb_path, n_path_offset);
1992 RFALSE(bh == p_s_tb->FR[n_h] &&
1993 PATH_OFFSET_POSITION(p_s_tb->tb_path,
1996 "PAP-8295: invalid position in the parent");
1999 (bh == p_s_tb->FR[n_h]) ? p_s_tb->rkey[n_h] + 1 : 0;
2000 n_son_number = B_N_CHILD_NUM(p_s_tb->FR[n_h], n_child_position);
2001 bh = sb_bread(sb, n_son_number);
2004 if (FILESYSTEM_CHANGED_TB(p_s_tb)) {
2006 PROC_INFO_INC(sb, get_neighbors_restart[n_h]);
2007 return REPEAT_SEARCH;
2009 brelse(p_s_tb->R[n_h]);
2010 p_s_tb->R[n_h] = bh;
2013 && B_FREE_SPACE(bh) !=
2014 MAX_CHILD_SIZE(bh) -
2015 dc_size(B_N_CHILD(p_s_tb->FR[0], n_child_position)),
2016 "PAP-8300: invalid child size of right neighbor (%d != %d - %d)",
2017 B_FREE_SPACE(bh), MAX_CHILD_SIZE(bh),
2018 dc_size(B_N_CHILD(p_s_tb->FR[0], n_child_position)));
2024 static int get_virtual_node_size(struct super_block *sb, struct buffer_head *bh)
2026 int max_num_of_items;
2027 int max_num_of_entries;
2028 unsigned long blocksize = sb->s_blocksize;
2030 #define MIN_NAME_LEN 1
2032 max_num_of_items = (blocksize - BLKH_SIZE) / (IH_SIZE + MIN_ITEM_LEN);
2033 max_num_of_entries = (blocksize - BLKH_SIZE - IH_SIZE) /
2034 (DEH_SIZE + MIN_NAME_LEN);
2036 return sizeof(struct virtual_node) +
2037 max(max_num_of_items * sizeof(struct virtual_item),
2038 sizeof(struct virtual_item) + sizeof(struct direntry_uarea) +
2039 (max_num_of_entries - 1) * sizeof(__u16));
2042 /* maybe we should fail balancing we are going to perform when kmalloc
2043 fails several times. But now it will loop until kmalloc gets
2045 static int get_mem_for_virtual_node(struct tree_balance *tb)
2051 size = get_virtual_node_size(tb->tb_sb, PATH_PLAST_BUFFER(tb->tb_path));
2053 if (size > tb->vn_buf_size) {
2054 /* we have to allocate more memory for virtual node */
2056 /* free memory allocated before */
2058 /* this is not needed if kfree is atomic */
2062 /* virtual node requires now more memory */
2063 tb->vn_buf_size = size;
2065 /* get memory for virtual item */
2066 buf = kmalloc(size, GFP_ATOMIC | __GFP_NOWARN);
2068 /* getting memory with GFP_KERNEL priority may involve
2069 balancing now (due to indirect_to_direct conversion on
2070 dcache shrinking). So, release path and collected
2072 free_buffers_in_tb(tb);
2073 buf = kmalloc(size, GFP_NOFS);
2075 tb->vn_buf_size = 0;
2079 return REPEAT_SEARCH;
2085 if (check_fs && FILESYSTEM_CHANGED_TB(tb))
2086 return REPEAT_SEARCH;
2091 #ifdef CONFIG_REISERFS_CHECK
2092 static void tb_buffer_sanity_check(struct super_block *sb,
2093 struct buffer_head *bh,
2094 const char *descr, int level)
2097 if (atomic_read(&(bh->b_count)) <= 0)
2099 reiserfs_panic(sb, "jmacd-1", "negative or zero "
2100 "reference counter for buffer %s[%d] "
2101 "(%b)", descr, level, bh);
2103 if (!buffer_uptodate(bh))
2104 reiserfs_panic(sb, "jmacd-2", "buffer is not up "
2105 "to date %s[%d] (%b)",
2108 if (!B_IS_IN_TREE(bh))
2109 reiserfs_panic(sb, "jmacd-3", "buffer is not "
2110 "in tree %s[%d] (%b)",
2113 if (bh->b_bdev != sb->s_bdev)
2114 reiserfs_panic(sb, "jmacd-4", "buffer has wrong "
2115 "device %s[%d] (%b)",
2118 if (bh->b_size != sb->s_blocksize)
2119 reiserfs_panic(sb, "jmacd-5", "buffer has wrong "
2120 "blocksize %s[%d] (%b)",
2123 if (bh->b_blocknr > SB_BLOCK_COUNT(sb))
2124 reiserfs_panic(sb, "jmacd-6", "buffer block "
2125 "number too high %s[%d] (%b)",
2130 static void tb_buffer_sanity_check(struct super_block *sb,
2131 struct buffer_head *bh,
2132 const char *descr, int level)
2137 static int clear_all_dirty_bits(struct super_block *s, struct buffer_head *bh)
2139 return reiserfs_prepare_for_journal(s, bh, 0);
2142 static int wait_tb_buffers_until_unlocked(struct tree_balance *p_s_tb)
2144 struct buffer_head *locked;
2145 #ifdef CONFIG_REISERFS_CHECK
2146 int repeat_counter = 0;
2154 for (i = p_s_tb->tb_path->path_length;
2155 !locked && i > ILLEGAL_PATH_ELEMENT_OFFSET; i--) {
2156 if (PATH_OFFSET_PBUFFER(p_s_tb->tb_path, i)) {
2157 /* if I understand correctly, we can only be sure the last buffer
2158 ** in the path is in the tree --clm
2160 #ifdef CONFIG_REISERFS_CHECK
2161 if (PATH_PLAST_BUFFER(p_s_tb->tb_path) ==
2162 PATH_OFFSET_PBUFFER(p_s_tb->tb_path, i)) {
2163 tb_buffer_sanity_check(p_s_tb->tb_sb,
2171 if (!clear_all_dirty_bits(p_s_tb->tb_sb,
2176 PATH_OFFSET_PBUFFER(p_s_tb->tb_path,
2182 for (i = 0; !locked && i < MAX_HEIGHT && p_s_tb->insert_size[i];
2185 if (p_s_tb->lnum[i]) {
2188 tb_buffer_sanity_check(p_s_tb->tb_sb,
2191 if (!clear_all_dirty_bits
2192 (p_s_tb->tb_sb, p_s_tb->L[i]))
2193 locked = p_s_tb->L[i];
2196 if (!locked && p_s_tb->FL[i]) {
2197 tb_buffer_sanity_check(p_s_tb->tb_sb,
2200 if (!clear_all_dirty_bits
2201 (p_s_tb->tb_sb, p_s_tb->FL[i]))
2202 locked = p_s_tb->FL[i];
2205 if (!locked && p_s_tb->CFL[i]) {
2206 tb_buffer_sanity_check(p_s_tb->tb_sb,
2209 if (!clear_all_dirty_bits
2210 (p_s_tb->tb_sb, p_s_tb->CFL[i]))
2211 locked = p_s_tb->CFL[i];
2216 if (!locked && (p_s_tb->rnum[i])) {
2219 tb_buffer_sanity_check(p_s_tb->tb_sb,
2222 if (!clear_all_dirty_bits
2223 (p_s_tb->tb_sb, p_s_tb->R[i]))
2224 locked = p_s_tb->R[i];
2227 if (!locked && p_s_tb->FR[i]) {
2228 tb_buffer_sanity_check(p_s_tb->tb_sb,
2231 if (!clear_all_dirty_bits
2232 (p_s_tb->tb_sb, p_s_tb->FR[i]))
2233 locked = p_s_tb->FR[i];
2236 if (!locked && p_s_tb->CFR[i]) {
2237 tb_buffer_sanity_check(p_s_tb->tb_sb,
2240 if (!clear_all_dirty_bits
2241 (p_s_tb->tb_sb, p_s_tb->CFR[i]))
2242 locked = p_s_tb->CFR[i];
2246 /* as far as I can tell, this is not required. The FEB list seems
2247 ** to be full of newly allocated nodes, which will never be locked,
2248 ** dirty, or anything else.
2249 ** To be safe, I'm putting in the checks and waits in. For the moment,
2250 ** they are needed to keep the code in journal.c from complaining
2251 ** about the buffer. That code is inside CONFIG_REISERFS_CHECK as well.
2254 for (i = 0; !locked && i < MAX_FEB_SIZE; i++) {
2255 if (p_s_tb->FEB[i]) {
2256 if (!clear_all_dirty_bits
2257 (p_s_tb->tb_sb, p_s_tb->FEB[i]))
2258 locked = p_s_tb->FEB[i];
2263 #ifdef CONFIG_REISERFS_CHECK
2265 if ((repeat_counter % 10000) == 0) {
2266 reiserfs_warning(p_s_tb->tb_sb, "reiserfs-8200",
2267 "too many iterations waiting "
2268 "for buffer to unlock "
2271 /* Don't loop forever. Try to recover from possible error. */
2273 return (FILESYSTEM_CHANGED_TB(p_s_tb)) ?
2274 REPEAT_SEARCH : CARRY_ON;
2277 __wait_on_buffer(locked);
2278 if (FILESYSTEM_CHANGED_TB(p_s_tb)) {
2279 return REPEAT_SEARCH;
2288 /* Prepare for balancing, that is
2289 * get all necessary parents, and neighbors;
2290 * analyze what and where should be moved;
2291 * get sufficient number of new nodes;
2292 * Balancing will start only after all resources will be collected at a time.
2294 * When ported to SMP kernels, only at the last moment after all needed nodes
2295 * are collected in cache, will the resources be locked using the usual
2296 * textbook ordered lock acquisition algorithms. Note that ensuring that
2297 * this code neither write locks what it does not need to write lock nor locks out of order
2298 * will be a pain in the butt that could have been avoided. Grumble grumble. -Hans
2300 * fix is meant in the sense of render unchanging
2302 * Latency might be improved by first gathering a list of what buffers are needed
2303 * and then getting as many of them in parallel as possible? -Hans
2306 * op_mode i - insert, d - delete, c - cut (truncate), p - paste (append)
2307 * tb tree_balance structure;
2308 * inum item number in S[h];
2309 * pos_in_item - comment this if you can
2310 * ins_ih & ins_sd are used when inserting
2311 * Returns: 1 - schedule occurred while the function worked;
2312 * 0 - schedule didn't occur while the function worked;
2313 * -1 - if no_disk_space
2316 int fix_nodes(int n_op_mode, struct tree_balance *p_s_tb, struct item_head *p_s_ins_ih, // item head of item being inserted
2317 const void *data // inserted item or data to be pasted
2320 int n_ret_value, n_h, n_item_num = PATH_LAST_POSITION(p_s_tb->tb_path);
2323 /* we set wait_tb_buffers_run when we have to restore any dirty bits cleared
2324 ** during wait_tb_buffers_run
2326 int wait_tb_buffers_run = 0;
2327 struct buffer_head *p_s_tbS0 = PATH_PLAST_BUFFER(p_s_tb->tb_path);
2329 ++REISERFS_SB(p_s_tb->tb_sb)->s_fix_nodes;
2331 n_pos_in_item = p_s_tb->tb_path->pos_in_item;
2333 p_s_tb->fs_gen = get_generation(p_s_tb->tb_sb);
2335 /* we prepare and log the super here so it will already be in the
2336 ** transaction when do_balance needs to change it.
2337 ** This way do_balance won't have to schedule when trying to prepare
2338 ** the super for logging
2340 reiserfs_prepare_for_journal(p_s_tb->tb_sb,
2341 SB_BUFFER_WITH_SB(p_s_tb->tb_sb), 1);
2342 journal_mark_dirty(p_s_tb->transaction_handle, p_s_tb->tb_sb,
2343 SB_BUFFER_WITH_SB(p_s_tb->tb_sb));
2344 if (FILESYSTEM_CHANGED_TB(p_s_tb))
2345 return REPEAT_SEARCH;
2347 /* if it possible in indirect_to_direct conversion */
2348 if (buffer_locked(p_s_tbS0)) {
2349 __wait_on_buffer(p_s_tbS0);
2350 if (FILESYSTEM_CHANGED_TB(p_s_tb))
2351 return REPEAT_SEARCH;
2353 #ifdef CONFIG_REISERFS_CHECK
2355 print_cur_tb("fix_nodes");
2356 reiserfs_panic(p_s_tb->tb_sb, "PAP-8305",
2357 "there is pending do_balance");
2360 if (!buffer_uptodate(p_s_tbS0) || !B_IS_IN_TREE(p_s_tbS0)) {
2361 reiserfs_panic(p_s_tb->tb_sb, "PAP-8320", "S[0] (%b %z) is "
2362 "not uptodate at the beginning of fix_nodes "
2363 "or not in tree (mode %c)",
2364 p_s_tbS0, p_s_tbS0, n_op_mode);
2367 /* Check parameters. */
2368 switch (n_op_mode) {
2370 if (n_item_num <= 0 || n_item_num > B_NR_ITEMS(p_s_tbS0))
2371 reiserfs_panic(p_s_tb->tb_sb, "PAP-8330", "Incorrect "
2372 "item number %d (in S0 - %d) in case "
2373 "of insert", n_item_num,
2374 B_NR_ITEMS(p_s_tbS0));
2379 if (n_item_num < 0 || n_item_num >= B_NR_ITEMS(p_s_tbS0)) {
2380 print_block(p_s_tbS0, 0, -1, -1);
2381 reiserfs_panic(p_s_tb->tb_sb, "PAP-8335", "Incorrect "
2382 "item number(%d); mode = %c "
2384 n_item_num, n_op_mode,
2385 p_s_tb->insert_size[0]);
2389 reiserfs_panic(p_s_tb->tb_sb, "PAP-8340", "Incorrect mode "
2394 if (get_mem_for_virtual_node(p_s_tb) == REPEAT_SEARCH)
2395 // FIXME: maybe -ENOMEM when tb->vn_buf == 0? Now just repeat
2396 return REPEAT_SEARCH;
2398 /* Starting from the leaf level; for all levels n_h of the tree. */
2399 for (n_h = 0; n_h < MAX_HEIGHT && p_s_tb->insert_size[n_h]; n_h++) {
2400 if ((n_ret_value = get_direct_parent(p_s_tb, n_h)) != CARRY_ON) {
2405 check_balance(n_op_mode, p_s_tb, n_h, n_item_num,
2406 n_pos_in_item, p_s_ins_ih,
2407 data)) != CARRY_ON) {
2408 if (n_ret_value == NO_BALANCING_NEEDED) {
2409 /* No balancing for higher levels needed. */
2411 get_neighbors(p_s_tb, n_h)) != CARRY_ON) {
2414 if (n_h != MAX_HEIGHT - 1)
2415 p_s_tb->insert_size[n_h + 1] = 0;
2416 /* ok, analysis and resource gathering are complete */
2422 if ((n_ret_value = get_neighbors(p_s_tb, n_h)) != CARRY_ON) {
2426 if ((n_ret_value = get_empty_nodes(p_s_tb, n_h)) != CARRY_ON) {
2427 goto repeat; /* No disk space, or schedule occurred and
2428 analysis may be invalid and needs to be redone. */
2431 if (!PATH_H_PBUFFER(p_s_tb->tb_path, n_h)) {
2432 /* We have a positive insert size but no nodes exist on this
2433 level, this means that we are creating a new root. */
2435 RFALSE(p_s_tb->blknum[n_h] != 1,
2436 "PAP-8350: creating new empty root");
2438 if (n_h < MAX_HEIGHT - 1)
2439 p_s_tb->insert_size[n_h + 1] = 0;
2440 } else if (!PATH_H_PBUFFER(p_s_tb->tb_path, n_h + 1)) {
2441 if (p_s_tb->blknum[n_h] > 1) {
2442 /* The tree needs to be grown, so this node S[n_h]
2443 which is the root node is split into two nodes,
2444 and a new node (S[n_h+1]) will be created to
2445 become the root node. */
2447 RFALSE(n_h == MAX_HEIGHT - 1,
2448 "PAP-8355: attempt to create too high of a tree");
2450 p_s_tb->insert_size[n_h + 1] =
2452 KEY_SIZE) * (p_s_tb->blknum[n_h] - 1) +
2454 } else if (n_h < MAX_HEIGHT - 1)
2455 p_s_tb->insert_size[n_h + 1] = 0;
2457 p_s_tb->insert_size[n_h + 1] =
2458 (DC_SIZE + KEY_SIZE) * (p_s_tb->blknum[n_h] - 1);
2461 if ((n_ret_value = wait_tb_buffers_until_unlocked(p_s_tb)) == CARRY_ON) {
2462 if (FILESYSTEM_CHANGED_TB(p_s_tb)) {
2463 wait_tb_buffers_run = 1;
2464 n_ret_value = REPEAT_SEARCH;
2470 wait_tb_buffers_run = 1;
2475 // fix_nodes was unable to perform its calculation due to
2476 // filesystem got changed under us, lack of free disk space or i/o
2477 // failure. If the first is the case - the search will be
2478 // repeated. For now - free all resources acquired so far except
2479 // for the new allocated nodes
2483 /* Release path buffers. */
2484 if (wait_tb_buffers_run) {
2485 pathrelse_and_restore(p_s_tb->tb_sb, p_s_tb->tb_path);
2487 pathrelse(p_s_tb->tb_path);
2489 /* brelse all resources collected for balancing */
2490 for (i = 0; i < MAX_HEIGHT; i++) {
2491 if (wait_tb_buffers_run) {
2492 reiserfs_restore_prepared_buffer(p_s_tb->tb_sb,
2494 reiserfs_restore_prepared_buffer(p_s_tb->tb_sb,
2496 reiserfs_restore_prepared_buffer(p_s_tb->tb_sb,
2498 reiserfs_restore_prepared_buffer(p_s_tb->tb_sb,
2500 reiserfs_restore_prepared_buffer(p_s_tb->tb_sb,
2503 reiserfs_restore_prepared_buffer(p_s_tb->tb_sb,
2508 brelse(p_s_tb->L[i]);
2509 brelse(p_s_tb->R[i]);
2510 brelse(p_s_tb->FL[i]);
2511 brelse(p_s_tb->FR[i]);
2512 brelse(p_s_tb->CFL[i]);
2513 brelse(p_s_tb->CFR[i]);
2515 p_s_tb->L[i] = NULL;
2516 p_s_tb->R[i] = NULL;
2517 p_s_tb->FL[i] = NULL;
2518 p_s_tb->FR[i] = NULL;
2519 p_s_tb->CFL[i] = NULL;
2520 p_s_tb->CFR[i] = NULL;
2523 if (wait_tb_buffers_run) {
2524 for (i = 0; i < MAX_FEB_SIZE; i++) {
2525 if (p_s_tb->FEB[i]) {
2526 reiserfs_restore_prepared_buffer
2527 (p_s_tb->tb_sb, p_s_tb->FEB[i]);
2536 /* Anatoly will probably forgive me renaming p_s_tb to tb. I just
2537 wanted to make lines shorter */
2538 void unfix_nodes(struct tree_balance *tb)
2542 /* Release path buffers. */
2543 pathrelse_and_restore(tb->tb_sb, tb->tb_path);
2545 /* brelse all resources collected for balancing */
2546 for (i = 0; i < MAX_HEIGHT; i++) {
2547 reiserfs_restore_prepared_buffer(tb->tb_sb, tb->L[i]);
2548 reiserfs_restore_prepared_buffer(tb->tb_sb, tb->R[i]);
2549 reiserfs_restore_prepared_buffer(tb->tb_sb, tb->FL[i]);
2550 reiserfs_restore_prepared_buffer(tb->tb_sb, tb->FR[i]);
2551 reiserfs_restore_prepared_buffer(tb->tb_sb, tb->CFL[i]);
2552 reiserfs_restore_prepared_buffer(tb->tb_sb, tb->CFR[i]);
2562 /* deal with list of allocated (used and unused) nodes */
2563 for (i = 0; i < MAX_FEB_SIZE; i++) {
2565 b_blocknr_t blocknr = tb->FEB[i]->b_blocknr;
2566 /* de-allocated block which was not used by balancing and
2567 bforget about buffer for it */
2569 reiserfs_free_block(tb->transaction_handle, NULL,
2573 /* release used as new nodes including a new root */
2574 brelse(tb->used[i]);