/*
* 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;
}
/*
* Recursively redirect non-dirty buffers on path to modify leaf.
*
* Redirect order is from root to leaf. Otherwise, blocks of path will
* be allocated by reverse order.
*
* FIXME: We can allocate/copy blocks before change common ancestor
* (before changing common ancestor, changes are not visible for
* reader). With this, we may be able to reduce locking time.
*/
int cursor_redirect(struct cursor *cursor)
{
if(DEBUG_MODE_K==1)
{
printf("\t\t\t\t%25s[K] %25s %4d #in\n",__FILE__,__func__,__LINE__);
}
struct btree *btree = cursor->btree;
struct sb *sb = btree->sb;
int level;
for (level = 0; level <= btree->root.depth; level++) {
struct buffer_head *buffer, *clone;
block_t parent, oldblock, newblock;
struct index_entry *entry;
int redirect, is_leaf = (level == btree->root.depth);
buffer = cursor->path[level].buffer;
/* If buffer needs to redirect to dirty, redirect it */
if (is_leaf)
redirect = leaf_need_redirect(sb, buffer);
else
redirect = bnode_need_redirect(sb, buffer);
/* No need to redirect */
if (!redirect)
continue;
/* Redirect buffer before changing */
clone = new_block(btree);
if (IS_ERR(clone))
return PTR_ERR(clone);
oldblock = bufindex(buffer);
newblock = bufindex(clone);
trace("redirect %Lx to %Lx", oldblock, newblock);
level_redirect_blockput(cursor, level, clone);
if (is_leaf) {
/* This is leaf buffer */
mark_buffer_dirty_atomic(clone);
log_leaf_redirect(sb, oldblock, newblock);
defer_bfree(&sb->defree, oldblock, 1);
} else {
/* This is bnode buffer */
mark_buffer_unify_atomic(clone);
log_bnode_redirect(sb, oldblock, newblock);
defer_bfree(&sb->deunify, oldblock, 1);
}
trace("update parent");
if (!level) {
/* Update pointer in btree->root */
trace("redirect root");
assert(oldblock == btree->root.block);
btree->root.block = newblock;
tux3_mark_btree_dirty(btree);
continue;
}
/* Update entry on parent for the redirected block */
parent = bufindex(cursor->path[level - 1].buffer);
entry = cursor->path[level - 1].next - 1;
entry->block = cpu_to_be64(newblock);
log_bnode_update(sb, parent, newblock, be64_to_cpu(entry->key));
}
cursor_check(cursor);
return 0;
}
/*
* Insert new leaf to next cursor position.
* keep == 1: keep current cursor position.
* keep == 0, set cursor position to new leaf.
*/
static int insert_leaf(struct cursor *cursor, tuxkey_t childkey, struct buffer_head *leafbuf, int keep)
{
if(DEBUG_MODE_K==1)
{
printf("\t\t\t\t%25s[K] %25s %4d #in\n",__FILE__,__func__,__LINE__);
}
struct btree *btree = cursor->btree;
struct sb *sb = btree->sb;
int level = btree->root.depth;
block_t childblock = bufindex(leafbuf);
if (keep)
blockput(leafbuf);
else {
cursor_pop_blockput(cursor);
cursor_push(cursor, leafbuf, NULL);
}
while (level--) {
struct path_level *at = &cursor->path[level];
struct buffer_head *parentbuf = at->buffer;
struct bnode *parent = bufdata(parentbuf);
/* insert and exit if not full */
if (bcount(parent) < btree->sb->entries_per_node) {
bnode_add_index(parent, at->next, childblock, childkey);
if (!keep)
at->next++;
log_bnode_add(sb, bufindex(parentbuf), childblock, childkey);
mark_buffer_unify_non(parentbuf);
cursor_check(cursor);
return 0;
}
/* split a full index node */
struct buffer_head *newbuf = new_node(btree);
if (IS_ERR(newbuf))
return PTR_ERR(newbuf);
struct bnode *newnode = bufdata(newbuf);
unsigned half = bcount(parent) / 2;
u64 newkey = be64_to_cpu(parent->entries[half].key);
bnode_split(parent, half, newnode);
log_bnode_split(sb, bufindex(parentbuf), half, bufindex(newbuf));
/* if the cursor is in the new node, use that as the parent */
int child_is_left = at->next <= parent->entries + half;
if (!child_is_left) {
struct index_entry *newnext;
mark_buffer_unify_non(parentbuf);
newnext = newnode->entries + (at->next - &parent->entries[half]);
get_bh(newbuf);
level_replace_blockput(cursor, level, newbuf, newnext);
parentbuf = newbuf;
parent = newnode;
} else
mark_buffer_unify_non(newbuf);
bnode_add_index(parent, at->next, childblock, childkey);
if (!keep)
at->next++;
log_bnode_add(sb, bufindex(parentbuf), childblock, childkey);
mark_buffer_unify_non(parentbuf);
childkey = newkey;
childblock = bufindex(newbuf);
blockput(newbuf);
/*
* If child is in left bnode, we should keep the
* cursor position to child, otherwise adjust cursor
* to new bnode.
*/
keep = child_is_left;
}
/* Make new root bnode */
trace("add tree level");
struct buffer_head *newbuf = new_node(btree);
if (IS_ERR(newbuf))
return PTR_ERR(newbuf);
struct bnode *newroot = bufdata(newbuf);
block_t newrootblock = bufindex(newbuf);
block_t oldrootblock = btree->root.block;
int left_node = bufindex(cursor->path[0].buffer) != childblock;
bnode_init_root(newroot, 2, oldrootblock, childblock, childkey);
cursor_root_add(cursor, newbuf, newroot->entries + 1 + !left_node);
log_bnode_root(sb, newrootblock, 2, oldrootblock, childblock, childkey);
/* Change btree to point the new root */
btree->root.block = newrootblock;
btree->root.depth++;
mark_buffer_unify_non(newbuf);
tux3_mark_btree_dirty(btree);
cursor_check(cursor);
return 0;
}
int alloc_empty_btree(struct btree *btree)
{
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 buffer_head *rootbuf = new_node(btree);
if (IS_ERR(rootbuf))
goto error;
struct buffer_head *leafbuf = new_leaf(btree);
if (IS_ERR(leafbuf))
goto error_leafbuf;
assert(!has_root(btree));
struct bnode *rootnode = bufdata(rootbuf);
block_t rootblock = bufindex(rootbuf);
block_t leafblock = bufindex(leafbuf);
trace("root at %Lx", rootblock);
trace("leaf at %Lx", leafblock);
bnode_init_root(rootnode, 1, leafblock, 0, 0);
log_bnode_root(sb, rootblock, 1, leafblock, 0, 0);
log_balloc(sb, leafblock, 1);
mark_buffer_unify_non(rootbuf);
blockput(rootbuf);
mark_buffer_dirty_non(leafbuf);
blockput(leafbuf);
btree->root = (struct root){ .block = rootblock, .depth = 1 };
tux3_mark_btree_dirty(btree);
return 0;
error_leafbuf:
(btree->ops->bfree)(sb, bufindex(rootbuf), 1);
blockput(rootbuf);
rootbuf = leafbuf;
error:
return PTR_ERR(rootbuf);
}
/* FIXME: right? and this should be done by btree_chop()? */
int free_empty_btree(struct btree *btree)
{
if(DEBUG_MODE_K==1)
{
printf("\t\t\t\t%25s[K] %25s %4d #in\n",__FILE__,__func__,__LINE__);
}
struct btree_ops *ops = btree->ops;
if (!has_root(btree))
return 0;
assert(btree->root.depth == 1);
struct sb *sb = btree->sb;
struct buffer_head *rootbuf = vol_bread(sb, btree->root.block);
if (!rootbuf)
return -EIO;
assert(bnode_sniff(bufdata(rootbuf)));
/* Make btree has no root */
btree->root = no_root;
tux3_mark_btree_dirty(btree);
struct bnode *rootnode = bufdata(rootbuf);
assert(bcount(rootnode) == 1);
block_t leaf = be64_to_cpu(rootnode->entries[0].block);
struct buffer_head *leafbuf = vol_find_get_block(sb, leaf);
if (leafbuf && !leaf_need_redirect(sb, leafbuf)) {
/*
* This is redirected leaf. So, in here, we can just
* cancel leaf_redirect by bfree(), instead of
* defered_bfree().
*/
bfree(sb, leaf, 1);
log_leaf_free(sb, leaf);
assert(ops->leaf_can_free(btree, bufdata(leafbuf)));
blockput_free(sb, leafbuf);
} else {
defer_bfree(&sb->defree, leaf, 1);
log_bfree(sb, leaf, 1);
if (leafbuf) {
assert(ops->leaf_can_free(btree, bufdata(leafbuf)));
blockput(leafbuf);
}
}
if (!bnode_need_redirect(sb, rootbuf)) {
/*
* This is redirected bnode. So, in here, we can just
* cancel bnode_redirect by bfree(), instead of
* defered_bfree().
*/
bfree(sb, bufindex(rootbuf), 1);
log_bnode_free(sb, bufindex(rootbuf));
blockput_free_unify(sb, rootbuf);
} else {
defer_bfree(&sb->deunify, bufindex(rootbuf), 1);
log_bfree_on_unify(sb, bufindex(rootbuf), 1);
blockput(rootbuf);
}
return 0;
}
int replay_bnode_redirect(struct replay *rp, block_t oldblock, block_t newblock)
{
if(DEBUG_MODE_K==1)
{
printf("\t\t\t\t%25s[K] %25s %4d #in\n",__FILE__,__func__,__LINE__);
}
struct sb *sb = rp->sb;
struct buffer_head *newbuf, *oldbuf;
int err = 0;
newbuf = vol_getblk(sb, newblock);
if (!newbuf) {
err = -ENOMEM; /* FIXME: error code */
goto error;
}
oldbuf = vol_bread(sb, oldblock);
if (!oldbuf) {
err = -EIO; /* FIXME: error code */
goto error_put_newbuf;
}
assert(bnode_sniff(bufdata(oldbuf)));
memcpy(bufdata(newbuf), bufdata(oldbuf), bufsize(newbuf));
mark_buffer_unify_atomic(newbuf);
blockput(oldbuf);
error_put_newbuf:
blockput(newbuf);
error:
return err;
}
int replay_bnode_root(struct replay *rp, block_t root, unsigned count,
block_t left, block_t right, tuxkey_t rkey)
{
if(DEBUG_MODE_K==1)
{
printf("\t\t\t\t%25s[K] %25s %4d #in\n",__FILE__,__func__,__LINE__);
}
struct sb *sb = rp->sb;
struct buffer_head *rootbuf;
rootbuf = vol_getblk(sb, root);
if (!rootbuf)
return -ENOMEM;
bnode_buffer_init(rootbuf);
bnode_init_root(bufdata(rootbuf), count, left, right, rkey);
mark_buffer_unify_atomic(rootbuf);
blockput(rootbuf);
return 0;
}
/*
* Before this replay, replay should already dirty the buffer of src.
* (e.g. by redirect)
*/
int replay_bnode_split(struct replay *rp, block_t src, unsigned pos,
block_t dst)
{
if(DEBUG_MODE_K==1)
{
printf("\t\t\t\t%25s[K] %25s %4d #in\n",__FILE__,__func__,__LINE__);
}
struct sb *sb = rp->sb;
struct buffer_head *srcbuf, *dstbuf;
int err = 0;
srcbuf = vol_getblk(sb, src);
if (!srcbuf) {
err = -ENOMEM; /* FIXME: error code */
goto error;
}
dstbuf = vol_getblk(sb, dst);
if (!dstbuf) {
err = -ENOMEM; /* FIXME: error code */
goto error_put_srcbuf;
}
bnode_buffer_init(dstbuf);
bnode_split(bufdata(srcbuf), pos, bufdata(dstbuf));
mark_buffer_unify_non(srcbuf);
mark_buffer_unify_atomic(dstbuf);
blockput(dstbuf);
error_put_srcbuf:
blockput(srcbuf);
error:
return err;
}
/*
* Before this replay, replay should already dirty the buffer of bnodeblock.
* (e.g. by redirect)
*/
static int replay_bnode_change(struct sb *sb, block_t bnodeblock,
u64 val1, u64 val2,
void (*change)(struct bnode *, u64, u64))
{
if(DEBUG_MODE_K==1)
{
printf("\t\t\t\t%25s[K] %25s %4d #in\n",__FILE__,__func__,__LINE__);
}
struct buffer_head *bnodebuf;
bnodebuf = vol_getblk(sb, bnodeblock);
if (!bnodebuf)
return -ENOMEM; /* FIXME: error code */
struct bnode *bnode = bufdata(bnodebuf);
change(bnode, val1, val2);
mark_buffer_unify_non(bnodebuf);
blockput(bnodebuf);
return 0;
}
static void add_func(struct bnode *bnode, u64 child, u64 key)
{
if(DEBUG_MODE_K==1)
{
printf("\t\t\t\t%25s[K] %25s %4d #in\n",__FILE__,__func__,__LINE__);
}
struct index_entry *entry = bnode_lookup(bnode, key) + 1;
bnode_add_index(bnode, entry, child, key);
}
int replay_bnode_add(struct replay *rp, block_t parent, block_t child,
tuxkey_t key)
{
if(DEBUG_MODE_K==1)
{
printf("\t\t\t\t%25s[K] %25s %4d #in\n",__FILE__,__func__,__LINE__);
}
return replay_bnode_change(rp->sb, parent, child, key, add_func);
}
static void update_func(struct bnode *bnode, u64 child, u64 key)
{
if(DEBUG_MODE_K==1)
{
printf("\t\t\t\t%25s[K] %25s %4d #in\n",__FILE__,__func__,__LINE__);
}
struct index_entry *entry = bnode_lookup(bnode, key);
assert(be64_to_cpu(entry->key) == key);
entry->block = cpu_to_be64(child);
}
int replay_bnode_update(struct replay *rp, block_t parent, block_t child,
tuxkey_t key)
{
if(DEBUG_MODE_K==1)
{
printf("\t\t\t\t%25s[K] %25s %4d #in\n",__FILE__,__func__,__LINE__);
}
return replay_bnode_change(rp->sb, parent, child, key, update_func);
}
int replay_bnode_merge(struct replay *rp, block_t src, block_t dst)
{
if(DEBUG_MODE_K==1)
{
printf("\t\t\t\t%25s[K] %25s %4d #in\n",__FILE__,__func__,__LINE__);
}
struct sb *sb = rp->sb;
struct buffer_head *srcbuf, *dstbuf;
int err = 0, ret;
srcbuf = vol_getblk(sb, src);
if (!srcbuf) {
err = -ENOMEM; /* FIXME: error code */
goto error;
}
dstbuf = vol_getblk(sb, dst);
if (!dstbuf) {
err = -ENOMEM; /* FIXME: error code */
goto error_put_srcbuf;
}
ret = bnode_merge_nodes(sb, bufdata(dstbuf), bufdata(srcbuf));
assert(ret == 1);
mark_buffer_unify_non(dstbuf);
mark_buffer_unify_non(srcbuf);
blockput(dstbuf);
error_put_srcbuf:
blockput(srcbuf);
error:
return err;
}
static void del_func(struct bnode *bnode, u64 key, u64 count)
{
if(DEBUG_MODE_K==1)
{
printf("\t\t\t\t%25s[K] %25s %4d #in\n",__FILE__,__func__,__LINE__);
}
struct index_entry *entry = bnode_lookup(bnode, key);
assert(be64_to_cpu(entry->key) == key);
bnode_remove_index(bnode, entry, count);
}
int replay_bnode_del(struct replay *rp, block_t bnode, tuxkey_t key,
unsigned count)
{
if(DEBUG_MODE_K==1)
{
printf("\t\t\t\t%25s[K] %25s %4d #in\n",__FILE__,__func__,__LINE__);
}
return replay_bnode_change(rp->sb, bnode, key, count, del_func);
}
static void adjust_func(struct bnode *bnode, u64 from, u64 to)
{
if(DEBUG_MODE_K==1)
{
printf("\t\t\t\t%25s[K] %25s %4d #in\n",__FILE__,__func__,__LINE__);
}
struct index_entry *entry = bnode_lookup(bnode, from);
assert(be64_to_cpu(entry->key) == from);
entry->key = cpu_to_be64(to);
}
int replay_bnode_adjust(struct replay *rp, block_t bnode, tuxkey_t from,
tuxkey_t to)
{
if(DEBUG_MODE_K==1)
{
printf("\t\t\t\t%25s[K] %25s %4d #in\n",__FILE__,__func__,__LINE__);
}
return replay_bnode_change(rp->sb, bnode, from, to, adjust_func);
}
int cursor_redirect(struct cursor *cursor)
{
struct btree *btree = cursor->btree;
unsigned level = btree->root.depth;
struct sb *sb = btree->sb;
block_t uninitialized_var(child);
while (1) {
struct buffer_head *buffer;
block_t uninitialized_var(oldblock);
block_t uninitialized_var(newblock);
int redirect, is_leaf = (level == btree->root.depth);
buffer = cursor->path[level].buffer;
/* If buffer needs to redirect to dirty, redirect it */
if (is_leaf)
redirect = leaf_need_redirect(sb, buffer);
else
redirect = bnode_need_redirect(sb, buffer);
if (redirect) {
/* Redirect buffer before changing */
struct buffer_head *clone = new_block(btree);
if (IS_ERR(clone))
return PTR_ERR(clone);
oldblock = bufindex(buffer);
newblock = bufindex(clone);
trace("redirect %Lx to %Lx", oldblock, newblock);
level_redirect_blockput(cursor, level, clone);
if (is_leaf) {
/* This is leaf buffer */
mark_buffer_dirty_atomic(clone);
log_leaf_redirect(sb, oldblock, newblock);
defer_bfree(&sb->defree, oldblock, 1);
goto parent_level;
}
/* This is bnode buffer */
mark_buffer_rollup_atomic(clone);
log_bnode_redirect(sb, oldblock, newblock);
defer_bfree(&sb->derollup, oldblock, 1);
} else {
if (is_leaf) {
/* This is leaf buffer */
goto parent_level;
}
}
/* Update entry for the redirected child block */
trace("update parent");
block_t block = bufindex(cursor->path[level].buffer);
struct index_entry *entry = cursor->path[level].next - 1;
entry->block = cpu_to_be64(child);
log_bnode_update(sb, block, child, be64_to_cpu(entry->key));
parent_level:
/* If it is already redirected, ancestor is also redirected */
if (!redirect) {
cursor_check(cursor);
return 0;
}
if (!level--) {
trace("redirect root");
assert(oldblock == btree->root.block);
btree->root.block = newblock;
tux3_mark_btree_dirty(btree);
cursor_check(cursor);
return 0;
}
child = newblock;
}
}
