ham_status_t
btree_traverse_tree(ham_page_t **page_ref, ham_s32_t *idxptr,
ham_db_t *db, ham_page_t *page, ham_key_t *key)
{
ham_status_t st;
ham_s32_t slot;
int_key_t *bte;
btree_node_t *node=ham_page_get_btree_node(page);
/*
* make sure that we're not in a leaf page, and that the
* page is not empty
*/
ham_assert(btree_node_get_count(node)>0, (0));
ham_assert(btree_node_get_ptr_left(node)!=0, (0));
st=btree_get_slot(db, page, key, &slot, 0);
if (st)
{
*page_ref = 0;
return st;
}
if (idxptr)
*idxptr=slot;
if (slot==-1)
{
st = db_fetch_page(page_ref, db, btree_node_get_ptr_left(node), 0);
ham_assert(st ? !*page_ref : 1, (0));
return st;
}
else
{
bte=btree_node_get_key(db, node, slot);
ham_assert(key_get_flags(bte)==0 ||
key_get_flags(bte)==KEY_IS_EXTENDED,
("invalid key flags 0x%x", key_get_flags(bte)));
st = db_fetch_page(page_ref, db, key_get_ptr(bte), 0);
ham_assert(st ? !*page_ref : 1, (0));
return st;
}
}
ham_status_t
btree_find_cursor(ham_btree_t *be, ham_bt_cursor_t *cursor,
ham_key_t *key, ham_record_t *record, ham_u32_t flags)
{
ham_status_t st;
ham_page_t *page = NULL;
btree_node_t *node = NULL;
int_key_t *entry;
ham_s32_t idx = -1;
ham_db_t *db=be_get_db(be);
find_hints_t hints = {flags, flags, 0, HAM_FALSE, HAM_FALSE, 1};
btree_find_get_hints(&hints, db, key);
if (hints.key_is_out_of_bounds) {
stats_update_find_fail_oob(db, &hints);
return HAM_KEY_NOT_FOUND;
}
if (hints.try_fast_track) {
/*
* see if we get a sure hit within this btree leaf; if not, revert to
* regular scan
*
* As this is a speed-improvement hint re-using recent material, the
* page should still sit in the cache, or we're using old info, which
* should be discarded.
*/
st = db_fetch_page(&page, db, hints.leaf_page_addr, DB_ONLY_FROM_CACHE);
ham_assert(st ? !page : 1, (0));
if (st)
return st;
if (page) {
node=ham_page_get_btree_node(page);
ham_assert(btree_node_is_leaf(node), (0));
ham_assert(btree_node_get_count(node) >= 3, (0)); /* edges + middle match */
idx = btree_node_search_by_key(db, page, key, hints.flags);
/*
* if we didn't hit a match OR a match at either edge, FAIL.
* A match at one of the edges is very risky, as this can also
* signal a match far away from the current node, so we need
* the full tree traversal then.
*/
if (idx <= 0 || idx >= btree_node_get_count(node) - 1) {
idx = -1;
}
/*
* else: we landed in the middle of the node, so we don't need to
* traverse the entire tree now.
*/
}
/* Reset any errors which may have been collected during the hinting
* phase -- this is done by setting 'idx = -1' above as that effectively
* clears the possible error code stored in there when (idx < -1)
*/
}
if (idx == -1) {
/* get the address of the root page */
if (!btree_get_rootpage(be)) {
stats_update_find_fail(db, &hints);
return HAM_KEY_NOT_FOUND;
}
/* load the root page */
st=db_fetch_page(&page, db, btree_get_rootpage(be), 0);
ham_assert(st ? !page : 1, (0));
if (!page) {
ham_assert(st, (0));
stats_update_find_fail(db, &hints);
return st ? st : HAM_INTERNAL_ERROR;
}
/* now traverse the root to the leaf nodes, till we find a leaf */
node=ham_page_get_btree_node(page);
if (!btree_node_is_leaf(node)) {
/* signal 'don't care' when we have multiple pages; we resolve
this once we've got a hit further down */
if (hints.flags & (HAM_FIND_LT_MATCH | HAM_FIND_GT_MATCH))
hints.flags |= (HAM_FIND_LT_MATCH | HAM_FIND_GT_MATCH);
for (;;) {
hints.cost++;
st=btree_traverse_tree(&page, 0, db, page, key);
if (!page) {
stats_update_find_fail(db, &hints);
return st ? st : HAM_KEY_NOT_FOUND;
}
node=ham_page_get_btree_node(page);
if (btree_node_is_leaf(node))
break;
}
}
/* check the leaf page for the key */
idx=btree_node_search_by_key(db, page, key, hints.flags);
if (idx < -1) {
stats_update_find_fail(db, &hints);
return (ham_status_t)idx;
}
} /* end of regular search */
/*
* When we are performing an approximate match, the worst case
* scenario is where we've picked the wrong side of the fence
* while sitting at a page/node boundary: that's what this
* next piece of code resolves:
*
* essentially it moves one record forwards or backward when
* the flags tell us this is mandatory and we're not yet in the proper
* position yet.
*
* The whole trick works, because the code above detects when
* we need to traverse a multi-page btree -- where this worst-case
* scenario can happen -- and adjusted the flags to accept
* both LT and GT approximate matches so that btree_node_search_by_key()
* will be hard pressed to return a 'key not found' signal (idx==-1),
* instead delivering the nearest LT or GT match; all we need to
* do now is ensure we've got the right one and if not,
* shift by one.
*/
if (idx >= 0) {
if ((ham_key_get_intflags(key) & KEY_IS_APPROXIMATE)
&& (hints.original_flags
& (HAM_FIND_LT_MATCH | HAM_FIND_GT_MATCH))
!= (HAM_FIND_LT_MATCH | HAM_FIND_GT_MATCH)) {
if ((ham_key_get_intflags(key) & KEY_IS_GT)
&& (hints.original_flags & HAM_FIND_LT_MATCH)) {
/*
* if the index-1 is still in the page, just decrement the
* index
*/
if (idx > 0) {
idx--;
}
else {
/*
* otherwise load the left sibling page
*/
if (!btree_node_get_left(node)) {
stats_update_find_fail(db, &hints);
ham_assert(node == ham_page_get_btree_node(page), (0));
stats_update_any_bound(db, page, key, hints.original_flags, -1);
return HAM_KEY_NOT_FOUND;
}
hints.cost++;
st = db_fetch_page(&page, db, btree_node_get_left(node), 0);
ham_assert(st ? !page : 1, (0));
if (!page) {
ham_assert(st, (0));
stats_update_find_fail(db, &hints);
return st ? st : HAM_INTERNAL_ERROR;
}
node = ham_page_get_btree_node(page);
idx = btree_node_get_count(node) - 1;
}
ham_key_set_intflags(key, (ham_key_get_intflags(key)
& ~KEY_IS_APPROXIMATE) | KEY_IS_LT);
}
else if ((ham_key_get_intflags(key) & KEY_IS_LT)
&& (hints.original_flags & HAM_FIND_GT_MATCH)) {
/*
* if the index+1 is still in the page, just increment the
* index
*/
if (idx + 1 < btree_node_get_count(node)) {
idx++;
}
else {
/*
* otherwise load the right sibling page
*/
if (!btree_node_get_right(node))
{
stats_update_find_fail(db, &hints);
ham_assert(node == ham_page_get_btree_node(page), (0));
stats_update_any_bound(db, page, key, hints.original_flags, -1);
return HAM_KEY_NOT_FOUND;
}
hints.cost++;
st = db_fetch_page(&page, db,
btree_node_get_right(node), 0);
if (!page) {
ham_assert(st, (0));
stats_update_find_fail(db, &hints);
return st ? st : HAM_INTERNAL_ERROR;
}
node = ham_page_get_btree_node(page);
idx = 0;
}
ham_key_set_intflags(key, (ham_key_get_intflags(key)
& ~KEY_IS_APPROXIMATE) | KEY_IS_GT);
}
}
else if (!(ham_key_get_intflags(key) & KEY_IS_APPROXIMATE)
&& !(hints.original_flags & HAM_FIND_EXACT_MATCH)
&& (hints.original_flags != 0)) {
/*
* 'true GT/LT' has been added @ 2009/07/18 to complete
* the EQ/LEQ/GEQ/LT/GT functionality;
*
* 'true LT/GT' is simply an extension upon the already existing
* LEQ/GEQ logic just above; all we do here is move one record
* up/down as it just happens that we get an exact ('equal')
* match here.
*
* The fact that the LT/GT constants share their bits with the
* LEQ/GEQ flags so that LEQ==(LT|EXACT) and GEQ==(GT|EXACT)
* ensures that we can restrict our work to a simple adjustment
* right here; everything else has already been taken of by the
* LEQ/GEQ logic in the section above when the key has been
* flagged with the KEY_IS_APPROXIMATE flag.
*/
if (hints.original_flags & HAM_FIND_LT_MATCH)
{
/*
* if the index-1 is still in the page, just decrement the
* index
*/
if (idx > 0)
{
idx--;
ham_key_set_intflags(key, (ham_key_get_intflags(key)
& ~KEY_IS_APPROXIMATE) | KEY_IS_LT);
}
else
{
/*
* otherwise load the left sibling page
*/
if (!btree_node_get_left(node))
{
/* when an error is otherwise unavoidable, see if
we have an escape route through GT? */
if (hints.original_flags & HAM_FIND_GT_MATCH)
{
/*
* if the index+1 is still in the page, just
* increment the index
*/
if (idx + 1 < btree_node_get_count(node))
{
idx++;
}
else
{
/*
* otherwise load the right sibling page
*/
if (!btree_node_get_right(node))
{
stats_update_find_fail(db, &hints);
ham_assert(node == ham_page_get_btree_node(page), (0));
stats_update_any_bound(db, page, key, hints.original_flags, -1);
return HAM_KEY_NOT_FOUND;
}
hints.cost++;
st = db_fetch_page(&page, db,
btree_node_get_right(node), 0);
if (!page)
{
ham_assert(st, (0));
stats_update_find_fail(db, &hints);
return st ? st : HAM_INTERNAL_ERROR;
}
node = ham_page_get_btree_node(page);
idx = 0;
}
ham_key_set_intflags(key, (ham_key_get_intflags(key) &
~KEY_IS_APPROXIMATE) | KEY_IS_GT);
}
else
{
stats_update_find_fail(db, &hints);
ham_assert(node == ham_page_get_btree_node(page), (0));
stats_update_any_bound(db, page, key, hints.original_flags, -1);
return HAM_KEY_NOT_FOUND;
}
}
else
{
hints.cost++;
st = db_fetch_page(&page, db,
btree_node_get_left(node), 0);
if (!page)
{
ham_assert(st, (0));
stats_update_find_fail(db, &hints);
return st ? st : HAM_INTERNAL_ERROR;
}
node = ham_page_get_btree_node(page);
idx = btree_node_get_count(node) - 1;
ham_key_set_intflags(key, (ham_key_get_intflags(key)
& ~KEY_IS_APPROXIMATE) | KEY_IS_LT);
}
}
}
else if (hints.original_flags & HAM_FIND_GT_MATCH)
{
/*
* if the index+1 is still in the page, just increment the
* index
*/
if (idx + 1 < btree_node_get_count(node))
{
idx++;
}
else
{
/*
* otherwise load the right sibling page
*/
if (!btree_node_get_right(node))
{
stats_update_find_fail(db, &hints);
ham_assert(node == ham_page_get_btree_node(page), (0));
stats_update_any_bound(db, page, key, hints.original_flags, -1);
return HAM_KEY_NOT_FOUND;
}
hints.cost++;
st = db_fetch_page(&page, db,
btree_node_get_right(node), 0);
if (!page)
{
ham_assert(st, (0));
stats_update_find_fail(db, &hints);
return st ? st : HAM_INTERNAL_ERROR;
}
node = ham_page_get_btree_node(page);
idx = 0;
}
ham_key_set_intflags(key, (ham_key_get_intflags(key)
& ~KEY_IS_APPROXIMATE) | KEY_IS_GT);
}
}
}
if (idx<0) {
stats_update_find_fail(db, &hints);
ham_assert(node, (0));
ham_assert(page, (0));
ham_assert(node == ham_page_get_btree_node(page), (0));
stats_update_any_bound(db, page, key, hints.original_flags, -1);
return HAM_KEY_NOT_FOUND;
}
/* load the entry, and store record ID and key flags */
entry=btree_node_get_key(db, node, idx);
/* set the cursor-position to this key */
if (cursor) {
ham_assert(!(bt_cursor_get_flags(cursor)&BT_CURSOR_FLAG_UNCOUPLED),
("coupling an uncoupled cursor, but need a nil-cursor"));
ham_assert(!(bt_cursor_get_flags(cursor)&BT_CURSOR_FLAG_COUPLED),
("coupling a coupled cursor, but need a nil-cursor"));
page_add_cursor(page, (ham_cursor_t *)cursor);
bt_cursor_set_flags(cursor,
bt_cursor_get_flags(cursor)|BT_CURSOR_FLAG_COUPLED);
bt_cursor_set_coupled_page(cursor, page);
bt_cursor_set_coupled_index(cursor, idx);
}
/*
* during util_read_key and util_read_record, new pages might be needed,
* and the page at which we're pointing could be moved out of memory;
* that would mean that the cursor would be uncoupled, and we're losing
* the 'entry'-pointer. therefore we 'lock' the page by incrementing
* the reference counter
*/
page_add_ref(page);
ham_assert(btree_node_is_leaf(node), ("iterator points to internal node"));
/* no need to load the key if we have an exact match: */
if (key && (ham_key_get_intflags(key) & KEY_IS_APPROXIMATE))
{
ham_status_t st=util_read_key(db, entry, key);
if (st)
{
page_release_ref(page);
stats_update_find_fail(db, &hints);
return (st);
}
}
if (record)
{
ham_status_t st;
record->_intflags=key_get_flags(entry);
record->_rid=key_get_ptr(entry);
st=util_read_record(db, record, flags);
if (st)
{
page_release_ref(page);
stats_update_find_fail(db, &hints);
return (st);
}
}
page_release_ref(page);
stats_update_find(db, page, &hints);
ham_assert(node == ham_page_get_btree_node(page), (0));
stats_update_any_bound(db, page, key, hints.original_flags, idx);
return (0);
}
static ham_status_t
__insert_split(ham_page_t *page, ham_key_t *key,
ham_offset_t rid, insert_scratchpad_t *scratchpad,
insert_hints_t *hints)
{
int cmp;
ham_status_t st;
ham_page_t *newpage, *oldsib;
int_key_t *nbte, *obte;
btree_node_t *nbtp, *obtp, *sbtp;
ham_size_t count, keysize;
ham_db_t *db=page_get_owner(page);
ham_env_t *env = db_get_env(db);
ham_key_t pivotkey, oldkey;
ham_offset_t pivotrid;
ham_u16_t pivot;
ham_bool_t pivot_at_end=HAM_FALSE;
ham_assert(page_get_owner(page), (0));
ham_assert(device_get_env(page_get_device(page))
== db_get_env(page_get_owner(page)), (0));
ham_assert(hints->force_append == HAM_FALSE, (0));
keysize=db_get_keysize(db);
/*
* allocate a new page
*/
hints->cost++;
st=db_alloc_page(&newpage, db, PAGE_TYPE_B_INDEX, 0);
ham_assert(st ? page == NULL : 1, (0));
ham_assert(!st ? page != NULL : 1, (0));
if (st)
return st;
ham_assert(page_get_owner(newpage), (""));
/* clear the node header */
memset(page_get_payload(newpage), 0, sizeof(btree_node_t));
stats_page_is_nuked(db, page, HAM_TRUE);
/*
* move half of the key/rid-tuples to the new page
*
* !! recno: keys are sorted; we do a "lazy split"
*/
nbtp=ham_page_get_btree_node(newpage);
nbte=btree_node_get_key(db, nbtp, 0);
obtp=ham_page_get_btree_node(page);
obte=btree_node_get_key(db, obtp, 0);
count=btree_node_get_count(obtp);
/*
* for databases with sequential access (this includes recno databases):
* do not split in the middle, but at the very end of the page
*
* if this page is the right-most page in the index, and this key is
* inserted at the very end, then we select the same pivot as for
* sequential access
*/
if (db_get_data_access_mode(db)&HAM_DAM_SEQUENTIAL_INSERT)
pivot_at_end=HAM_TRUE;
else if (btree_node_get_right(obtp)==0) {
cmp=key_compare_pub_to_int(db, page, key, btree_node_get_count(obtp)-1);
if (cmp>0)
pivot_at_end=HAM_TRUE;
}
/*
* internal pages set the count of the new page to count-pivot-1 (because
* the pivot element will become ptr_left of the new page).
* by using pivot=count-2 we make sure that at least 1 element will remain
* in the new node.
*/
if (pivot_at_end) {
pivot=count-2;
}
else {
pivot=count/2;
}
/*
* uncouple all cursors
*/
st=bt_uncouple_all_cursors(page, pivot);
if (st)
return (st);
/*
* if we split a leaf, we'll insert the pivot element in the leaf
* page, too. in internal nodes, we don't insert it, but propagate
* it to the parent node only.
*/
if (btree_node_is_leaf(obtp)) {
hints->cost += stats_memmove_cost((db_get_int_key_header_size()+keysize)*(count-pivot));
memcpy((char *)nbte,
((char *)obte)+(db_get_int_key_header_size()+keysize)*pivot,
(db_get_int_key_header_size()+keysize)*(count-pivot));
}
else {
hints->cost += stats_memmove_cost((db_get_int_key_header_size()+keysize)*(count-pivot-1));
memcpy((char *)nbte,
((char *)obte)+(db_get_int_key_header_size()+keysize)*(pivot+1),
(db_get_int_key_header_size()+keysize)*(count-pivot-1));
}
/*
* store the pivot element, we'll need it later to propagate it
* to the parent page
*/
nbte=btree_node_get_key(db, obtp, pivot);
memset(&pivotkey, 0, sizeof(pivotkey));
memset(&oldkey, 0, sizeof(oldkey));
oldkey.data=key_get_key(nbte);
oldkey.size=key_get_size(nbte);
oldkey._flags=key_get_flags(nbte);
st = util_copy_key(db, &oldkey, &pivotkey);
if (st)
{
(void)db_free_page(newpage, DB_MOVE_TO_FREELIST);
goto fail_dramatically;
}
pivotrid=page_get_self(newpage);
/*
* adjust the page count
*/
if (btree_node_is_leaf(obtp)) {
btree_node_set_count(obtp, pivot);
btree_node_set_count(nbtp, count-pivot);
}
else {
btree_node_set_count(obtp, pivot);
btree_node_set_count(nbtp, count-pivot-1);
}
/*
* if we're in an internal page: fix the ptr_left of the new page
* (it points to the ptr of the pivot key)
*/
if (!btree_node_is_leaf(obtp)) {
/*
* nbte still contains the pivot key
*/
btree_node_set_ptr_left(nbtp, key_get_ptr(nbte));
}
/*
* insert the new element
*/
hints->cost++;
cmp=key_compare_pub_to_int(db, page, key, pivot);
if (cmp < -1)
{
st = (ham_status_t)cmp;
goto fail_dramatically;
}
if (cmp>=0)
st=__insert_nosplit(newpage, key, rid,
scratchpad->record, scratchpad->cursor, hints);
else
st=__insert_nosplit(page, key, rid,
scratchpad->record, scratchpad->cursor, hints);
if (st)
{
goto fail_dramatically;
}
scratchpad->cursor=0; /* don't overwrite cursor if __insert_nosplit
is called again */
/*
* fix the double-linked list of pages, and mark the pages as dirty
*/
if (btree_node_get_right(obtp))
{
st=db_fetch_page(&oldsib, db, btree_node_get_right(obtp), 0);
if (st)
goto fail_dramatically;
}
else
{
oldsib=0;
}
if (oldsib) {
st=ham_log_add_page_before(oldsib);
if (st)
goto fail_dramatically;
}
btree_node_set_left (nbtp, page_get_self(page));
btree_node_set_right(nbtp, btree_node_get_right(obtp));
btree_node_set_right(obtp, page_get_self(newpage));
if (oldsib) {
sbtp=ham_page_get_btree_node(oldsib);
btree_node_set_left(sbtp, page_get_self(newpage));
page_set_dirty(oldsib, env);
}
page_set_dirty(newpage, env);
page_set_dirty(page, env);
/*
* propagate the pivot key to the parent page
*/
ham_assert(!(scratchpad->key.flags & HAM_KEY_USER_ALLOC), (0));
if (scratchpad->key.data)
allocator_free(env_get_allocator(env), scratchpad->key.data);
scratchpad->key=pivotkey;
scratchpad->rid=pivotrid;
ham_assert(!(scratchpad->key.flags & HAM_KEY_USER_ALLOC), (0));
return (SPLIT);
fail_dramatically:
ham_assert(!(pivotkey.flags & HAM_KEY_USER_ALLOC), (0));
if (pivotkey.data)
allocator_free(env_get_allocator(env), pivotkey.data);
return st;
}
static ham_status_t
__insert_cursor(ham_btree_t *be, ham_key_t *key, ham_record_t *record,
ham_bt_cursor_t *cursor, insert_hints_t *hints)
{
ham_status_t st;
ham_page_t *root;
ham_db_t *db=be_get_db(be);
ham_env_t *env = db_get_env(db);
insert_scratchpad_t scratchpad;
ham_assert(hints->force_append == HAM_FALSE, (0));
ham_assert(hints->force_prepend == HAM_FALSE, (0));
/*
* initialize the scratchpad
*/
memset(&scratchpad, 0, sizeof(scratchpad));
scratchpad.be=be;
scratchpad.record=record;
scratchpad.cursor=cursor;
/*
* get the root-page...
*/
ham_assert(btree_get_rootpage(be)!=0, ("btree has no root page"));
st=db_fetch_page(&root, db, btree_get_rootpage(be), 0);
ham_assert(st ? root == NULL : 1, (0));
if (st)
return st;
/*
* ... and start the recursion
*/
st=__insert_recursive(root, key, 0, &scratchpad, hints);
/*
* if the root page was split, we have to create a new
* root page.
*/
if (st==SPLIT) {
ham_page_t *newroot;
btree_node_t *node;
/*
* the root-page will be changed...
*/
st=ham_log_add_page_before(root);
if (st)
return (st);
/*
* allocate a new root page
*/
st=db_alloc_page(&newroot, db, PAGE_TYPE_B_ROOT, 0);
ham_assert(st ? newroot == NULL : 1, (0));
if (st)
return (st);
ham_assert(page_get_owner(newroot), (""));
/* clear the node header */
memset(page_get_payload(newroot), 0, sizeof(btree_node_t));
stats_page_is_nuked(db, root, HAM_TRUE);
/*
* insert the pivot element and the ptr_left
*/
node=ham_page_get_btree_node(newroot);
btree_node_set_ptr_left(node, btree_get_rootpage(be));
st=__insert_nosplit(newroot, &scratchpad.key,
scratchpad.rid, scratchpad.record, scratchpad.cursor,
hints);
ham_assert(!(scratchpad.key.flags & HAM_KEY_USER_ALLOC), (0));
scratchpad.cursor=0; /* don't overwrite cursor if __insert_nosplit
is called again */
if (st) {
ham_assert(!(scratchpad.key.flags & HAM_KEY_USER_ALLOC), (0));
if (scratchpad.key.data)
allocator_free(env_get_allocator(env), scratchpad.key.data);
return (st);
}
/*
* set the new root page
*
* !!
* do NOT delete the old root page - it's still in use!
*
* also don't forget to flush the backend - otherwise the header
* page of the database will not contain the updated information.
* The backend is flushed when the database is closed, but if
* recovery is enabled then the flush here is critical.
*/
btree_set_rootpage(be, page_get_self(newroot));
be_set_dirty(be, HAM_TRUE);
be->_fun_flush(be);
/*
* As we re-purpose a page, we will reset its pagecounter
* as well to signal its first use as the new type assigned
* here.
*/
if (env_get_cache(env) && (page_get_type(root)!=PAGE_TYPE_B_INDEX))
cache_update_page_access_counter(root, env_get_cache(env), 0);
page_set_type(root, PAGE_TYPE_B_INDEX);
page_set_dirty(root, env);
page_set_dirty(newroot, env);
/* the root page was modified (btree_set_rootpage) - make sure that
* it's logged */
if (env_get_rt_flags(env)&HAM_ENABLE_RECOVERY) {
st=txn_add_page(env_get_txn(env), env_get_header_page(env),
HAM_TRUE);
if (st)
return (st);
}
}
/*
* release the scratchpad-memory and return to caller
*/
ham_assert(!(scratchpad.key.flags & HAM_KEY_USER_ALLOC), (0));
if (scratchpad.key.data)
allocator_free(env_get_allocator(env), scratchpad.key.data);
return (st);
}
static ham_status_t
__append_key(ham_btree_t *be, ham_key_t *key, ham_record_t *record,
ham_bt_cursor_t *cursor, insert_hints_t *hints)
{
ham_status_t st=0;
ham_page_t *page;
btree_node_t *node;
ham_db_t *db;
#ifdef HAM_DEBUG
if (cursor && !bt_cursor_is_nil(cursor)) {
ham_assert(be_get_db(be) == bt_cursor_get_db(cursor), (0));
}
#endif
db = be_get_db(be);
/*
* see if we get this btree leaf; if not, revert to regular scan
*
* As this is a speed-improvement hint re-using recent material, the page
* should still sit in the cache, or we're using old info, which should be
* discarded.
*/
st = db_fetch_page(&page, db, hints->leaf_page_addr, DB_ONLY_FROM_CACHE);
if (st)
return st;
if (!page) {
hints->force_append = HAM_FALSE;
hints->force_prepend = HAM_FALSE;
return (__insert_cursor(be, key, record, cursor, hints));
}
page_add_ref(page);
node=ham_page_get_btree_node(page);
ham_assert(btree_node_is_leaf(node), ("iterator points to internal node"));
/*
* if the page is already full OR this page is not the right-most page
* when we APPEND or the left-most node when we PREPEND
* OR the new key is not the highest key: perform a normal insert
*/
if ((hints->force_append && btree_node_get_right(node))
|| (hints->force_prepend && btree_node_get_left(node))
|| btree_node_get_count(node) >= btree_get_maxkeys(be)) {
page_release_ref(page);
hints->force_append = HAM_FALSE;
hints->force_prepend = HAM_FALSE;
return (__insert_cursor(be, key, record, cursor, hints));
}
/*
* if the page is not empty: check if we append the key at the end / start
* (depending on force_append/force_prepend),
* or if it's actually inserted in the middle (when neither force_append
* or force_prepend is specified: that'd be SEQUENTIAL insertion
* hinting somewhere in the middle of the total key range.
*/
if (btree_node_get_count(node)!=0) {
int cmp_hi;
int cmp_lo;
hints->cost++;
if (!hints->force_prepend) {
cmp_hi = key_compare_pub_to_int(db, page, key,
btree_node_get_count(node)-1);
/* key is in the middle */
if (cmp_hi < -1) {
page_release_ref(page);
return (ham_status_t)cmp_hi;
}
/* key is at the end */
if (cmp_hi > 0) {
if (btree_node_get_right(node)) {
/* not at top end of the btree, so we can't do the
* fast track */
page_release_ref(page);
//hints->flags &= ~HAM_HINT_APPEND;
hints->force_append = HAM_FALSE;
hints->force_prepend = HAM_FALSE;
return (__insert_cursor(be, key, record, cursor, hints));
}
hints->force_append = HAM_TRUE;
hints->force_prepend = HAM_FALSE;
}
}
else { /* hints->force_prepend is true */
/* not bigger than the right-most node while we
* were trying to APPEND */
cmp_hi = -1;
}
if (!hints->force_append) {
cmp_lo = key_compare_pub_to_int(db, page, key, 0);
/* in the middle range */
if (cmp_lo < -1) {
page_release_ref(page);
return ((ham_status_t)cmp_lo);
}
/* key is at the start of page */
if (cmp_lo < 0) {
if (btree_node_get_left(node)) {
/* not at bottom end of the btree, so we can't
* do the fast track */
page_release_ref(page);
//hints->flags &= ~HAM_HINT_PREPEND;
hints->force_append = HAM_FALSE;
hints->force_prepend = HAM_FALSE;
return (__insert_cursor(be, key, record, cursor, hints));
}
hints->force_append = HAM_FALSE;
hints->force_prepend = HAM_TRUE;
}
}
else { /* hints->force_prepend is true */
/* not smaller than the left-most node while we were
* trying to PREPEND */
cmp_lo = +1;
}
/* handle inserts in the middle range */
if (cmp_lo >= 0 && cmp_hi <= 0) {
/*
* Depending on where we are in the btree, the current key either
* is going to end up in the middle of the given node/page,
* OR the given key is out of range of the given leaf node.
*/
if (hints->force_append || hints->force_prepend) {
/*
* when prepend or append is FORCED, we are expected to
* add keys ONLY at the beginning or end of the btree
* key range. Clearly the current key does not fit that
* criterium.
*/
page_release_ref(page);
//hints->flags &= ~HAM_HINT_PREPEND;
hints->force_append = HAM_FALSE;
hints->force_prepend = HAM_FALSE;
return (__insert_cursor(be, key, record, cursor, hints));
}
/*
* we discovered that the key must be inserted in the middle
* of the current leaf.
*
* It does not matter whether the current leaf is at the start or
* end of the btree range; as we need to add the key in the middle
* of the current leaf, that info alone is enough to continue with
* the fast track insert operation.
*/
ham_assert(!hints->force_prepend && !hints->force_append, (0));
}
ham_assert((hints->force_prepend + hints->force_append) < 2,
("Either APPEND or PREPEND flag MAY be set, but not both"));
}
else { /* empty page: force insertion in slot 0 */
hints->force_append = HAM_FALSE;
hints->force_prepend = HAM_TRUE;
}
/*
* the page will be changed - write it to the log (if a log exists)
*/
st=ham_log_add_page_before(page);
if (st) {
page_release_ref(page);
return (st);
}
/*
* OK - we're really appending/prepending the new key.
*/
ham_assert(hints->force_append || hints->force_prepend, (0));
st=__insert_nosplit(page, key, 0, record, cursor, hints);
page_release_ref(page);
return (st);
}
