static void remove_index(struct cursor *cursor, struct chopped_index_info *cii)
{
int level = cursor->level;
struct bnode *node = level_node(cursor, level);
struct chopped_index_info *ciil = &cii[level];
/* Collect chopped index in this node for logging later */
if (!ciil->count)
ciil->start = be64_to_cpu((cursor->path[level].next - 1)->key);
ciil->count++;
/* Remove an index */
bnode_remove_index(node, cursor->path[level].next - 1, 1);
--(cursor->path[level].next);
mark_buffer_rollup_non(cursor->path[level].buffer);
/*
* Climb up to common parent and update separating key.
*
* What if index is now empty? (no deleted key)
*
* Then some key above is going to be deleted and used to set sep
* Climb the cursor while at first entry, bail out at root find the
* node with the old sep, set it to deleted key
*/
/* There is no separator for last entry or root node */
if (!level || cursor_level_finished(cursor))
return;
/* If removed index was not first entry, no change to separator */
if (cursor->path[level].next != node->entries)
return;
adjust_parent_sep(cursor, level - 1, cursor->path[level].next->key);
}
/*
* Cursor advance for btree traverse.
* < 0 - error
* 0 - Finished traverse
* 1 - Reached leaf
*/
static int cursor_advance(struct cursor *cursor)
{
int ret;
do {
if (!cursor_advance_up(cursor))
return 0;
} while (cursor_level_finished(cursor));
do {
ret = cursor_advance_down(cursor);
if (ret < 0)
return ret;
} while (ret);
return 1;
}
/*
* Cursor advance for btree traverse.
* < 0 - error
* 0 - Finished traverse
* 1 - Reached leaf
*/
static int cursor_advance(struct cursor *cursor)
{
if(DEBUG_MODE_K==1)
{
printf("\t\t\t\t%25s[K] %25s %4d #in\n",__FILE__,__func__,__LINE__);
}
int ret;
do {
if (!cursor_advance_up(cursor))
return 0;
} while (cursor_level_finished(cursor));
do {
ret = cursor_advance_down(cursor);
if (ret < 0)
return ret;
} while (ret);
return 1;
}
/*
* This is range deletion. So, instead of adjusting balance of the
* space on sibling nodes for each change, this just removes the range
* and merges from right to left even if it is not same parent.
*
* +--------------- (A, B, C)--------------------+
* | | |
* +-- (AA, AB, AC) -+ +- (BA, BB, BC) -+ + (CA, CB, CC) +
* | | | | | | | | |
* (AAA,AAB)(ABA,ABB)(ACA,ACB) (BAA,BAB)(BBA)(BCA,BCB) (CAA)(CBA,CBB)(CCA)
*
* [less : A, AA, AAA, AAB, AB, ABA, ABB, AC, ACA, ACB, B, BA ... : greater]
*
* If we merged from cousin (or re-distributed), we may have to update
* the index until common parent. (e.g. removed (ACB), then merged
* from (BAA,BAB) to (ACA), we have to adjust B in root node to BB)
*
* See, adjust_parent_sep().
*
* FIXME: no re-distribute. so, we don't guarantee above than 50%
* space efficiency. And if range is end of key (truncate() case), we
* don't need to merge, and adjust_parent_sep().
*
* FIXME2: we may want to split chop work for each step. instead of
* blocking for a long time.
*/
int btree_chop(struct btree *btree, tuxkey_t start, u64 len)
{
if(DEBUG_MODE_K==1)
{
printf("\t\t\t\t%25s[K] %25s %4d #in\n",__FILE__,__func__,__LINE__);
}
struct sb *sb = btree->sb;
struct btree_ops *ops = btree->ops;
struct buffer_head **prev, *leafprev = NULL;
struct chopped_index_info *cii;
struct cursor *cursor;
tuxkey_t limit;
int ret, done = 0;
if (!has_root(btree))
return 0;
/* Chop all range if len >= TUXKEY_LIMIT */
limit = (len >= TUXKEY_LIMIT) ? TUXKEY_LIMIT : start + len;
prev = malloc(sizeof(*prev) * btree->root.depth);
if (prev == NULL)
return -ENOMEM;
memset(prev, 0, sizeof(*prev) * btree->root.depth);
cii = malloc(sizeof(*cii) * btree->root.depth);
if (cii == NULL) {
ret = -ENOMEM;
goto error_cii;
}
memset(cii, 0, sizeof(*cii) * btree->root.depth);
cursor = alloc_cursor(btree, 0);
if (!cursor) {
ret = -ENOMEM;
goto error_alloc_cursor;
}
down_write(&btree->lock);
ret = btree_probe(cursor, start);
if (ret)
goto error_btree_probe;
/* Walk leaves */
while (1) {
struct buffer_head *leafbuf;
tuxkey_t this_key;
/*
* FIXME: If leaf was merged and freed later, we don't
* need to redirect leaf and leaf_chop()
*/
if ((ret = cursor_redirect(cursor)))
goto out;
leafbuf = cursor_pop(cursor);
/* Adjust start and len for this leaf */
this_key = cursor_level_this_key(cursor);
if (start < this_key) {
if (limit < TUXKEY_LIMIT)
len -= this_key - start;
start = this_key;
}
ret = ops->leaf_chop(btree, start, len, bufdata(leafbuf));
if (ret) {
if (ret < 0) {
blockput(leafbuf);
goto out;
}
mark_buffer_dirty_non(leafbuf);
}
/* Try to merge this leaf with prev */
if (leafprev) {
if (try_leaf_merge(btree, leafprev, leafbuf)) {
trace(">>> can merge leaf %p into leaf %p", leafbuf, leafprev);
remove_index(cursor, cii);
mark_buffer_dirty_non(leafprev);
blockput_free(sb, leafbuf);
goto keep_prev_leaf;
}
blockput(leafprev);
}
leafprev = leafbuf;
keep_prev_leaf:
if (cursor_level_next_key(cursor) >= limit)
done = 1;
/* Pop and try to merge finished nodes */
while (done || cursor_level_finished(cursor)) {
struct buffer_head *buf;
int level = cursor->level;
struct chopped_index_info *ciil = &cii[level];
/* Get merge src buffer, and go parent level */
buf = cursor_pop(cursor);
/*
* Logging chopped indexes
* FIXME: If node is freed later (e.g. merged),
* we dont't need to log this
*/
if (ciil->count) {
log_bnode_del(sb, bufindex(buf), ciil->start,
ciil->count);
}
memset(ciil, 0, sizeof(*ciil));
/* Try to merge node with prev */
if (prev[level]) {
assert(level);
if (try_bnode_merge(sb, prev[level], buf)) {
trace(">>> can merge node %p into node %p", buf, prev[level]);
remove_index(cursor, cii);
mark_buffer_unify_non(prev[level]);
blockput_free_unify(sb, buf);
goto keep_prev_node;
}
blockput(prev[level]);
}
prev[level] = buf;
keep_prev_node:
if (!level)
goto chop_root;
}
/* Push back down to leaf level */
do {
ret = cursor_advance_down(cursor);
if (ret < 0)
goto out;
} while (ret);
}
chop_root:
/* Remove depth if possible */
while (btree->root.depth > 1 && bcount(bufdata(prev[0])) == 1) {
trace("drop btree level");
btree->root.block = bufindex(prev[1]);
btree->root.depth--;
tux3_mark_btree_dirty(btree);
/*
* We know prev[0] is redirected and dirty. So, in
* here, we can just cancel bnode_redirect by bfree(),
* instead of defered_bfree()
* FIXME: we can optimize freeing bnode without
* bnode_redirect, and if we did, this is not true.
*/
bfree(sb, bufindex(prev[0]), 1);
log_bnode_free(sb, bufindex(prev[0]));
blockput_free_unify(sb, prev[0]);
vecmove(prev, prev + 1, btree->root.depth);
}
ret = 0;
out:
if (leafprev)
blockput(leafprev);
for (int i = 0; i < btree->root.depth; i++) {
if (prev[i])
blockput(prev[i]);
}
release_cursor(cursor);
error_btree_probe:
up_write(&btree->lock);
free_cursor(cursor);
error_alloc_cursor:
free(cii);
error_cii:
free(prev);
return ret;
}