rc_t btree_impl::_ux_adopt_foster_core (btree_page_h &parent, btree_page_h &child,
const w_keystr_t &new_child_key)
{
w_assert1 (g_xct()->is_single_log_sys_xct());
w_assert1 (parent.is_fixed());
w_assert1 (parent.latch_mode() == LATCH_EX);
w_assert1 (parent.is_node());
w_assert1 (child.is_fixed());
w_assert1 (child.latch_mode() == LATCH_EX);
w_assert0 (child.get_foster() != 0);
PageID new_child_pid = child.get_foster();
if (smlevel_0::bf->is_swizzled_pointer(new_child_pid)) {
smlevel_0::bf->unswizzle(parent.get_generic_page(),
GeneralRecordIds::FOSTER_CHILD, true, &new_child_pid);
}
w_assert1(!smlevel_0::bf->is_swizzled_pointer(new_child_pid));
lsn_t child_emlsn = child.get_foster_emlsn();
W_DO(log_btree_foster_adopt (parent, child, new_child_pid, child_emlsn, new_child_key));
_ux_adopt_foster_apply_parent (parent, new_child_pid, child_emlsn, new_child_key);
_ux_adopt_foster_apply_child (child);
// Switch parent of newly adopted child
// CS TODO: I'm not sure we can do this because we don't hold a latch on new_child_pid
smlevel_0::bf->switch_parent(new_child_pid, parent.get_generic_page());
w_assert3(parent.is_consistent(true, true));
w_assert3(child.is_consistent(true, true));
return RCOK;
}
rc_t btree_impl::_sx_adopt_foster_all_core (
btree_page_h &parent, bool is_root, bool recursive)
{
// TODO this should use the improved tree-walk-through
// See jira ticket:60 "Tree walk-through without more than 2 pages latched" (originally trac ticket:62)
w_assert1 (xct()->is_sys_xct());
w_assert1 (parent.is_fixed());
w_assert1 (parent.latch_mode() == LATCH_EX);
if (parent.is_node()) {
w_assert1(parent.pid0());
W_DO(_sx_adopt_foster_sweep(parent));
if (recursive) {
// also adopt at all children recursively
for (int i = -1; i < parent.nrecs(); ++i) {
btree_page_h child;
PageID shpid_opaqueptr = i == -1 ? parent.get_foster_opaqueptr() : parent.child_opaqueptr(i);
W_DO(child.fix_nonroot(parent, shpid_opaqueptr, LATCH_EX));
W_DO(_sx_adopt_foster_all_core(child, false, true));
}
}
}
// after all adopts, if this parent is the root and has foster,
// let's grow the tree
if (is_root && parent.get_foster()) {
W_DO(_sx_grow_tree(parent));
W_DO(_sx_adopt_foster_sweep(parent));
}
w_assert3(parent.is_consistent(true, true));
return RCOK;
}
rc_t btree_impl::_ux_assure_fence_low_entry(btree_page_h &leaf)
{
w_assert1(leaf.is_fixed());
w_assert1(leaf.latch_mode() == LATCH_EX);
if (!leaf.is_leaf()) {
// locks are taken only for leaf-page entries. this case isn't an issue
return RCOK;
}
w_keystr_t fence_low;
leaf.copy_fence_low_key(fence_low);
bool needs_to_create = false;
if (leaf.nrecs() == 0) {
if (leaf.compare_with_fence_high(fence_low) == 0) {
// low==high happens only during page split. In that case, no one can have a lock
// in the page being created. No need to assure the record.
return RCOK;
}
needs_to_create = true;
} else {
w_keystr_t first_key;
leaf.get_key(0, first_key);
w_assert1(fence_low.compare(first_key) <= 0); // can't be fence_low>first_key
if (fence_low.compare(first_key) < 0) {
// fence-low doesn't exist as an entry!
needs_to_create = true;
}
}
if (needs_to_create) {
W_DO(_sx_reserve_ghost(leaf, fence_low, 0)); // no data is needed
}
return RCOK;
}
rc_t btree_impl::_sx_opportunistic_adopt_foster (btree_page_h &parent,
btree_page_h &child, bool &pushedup,
const bool from_recovery)
{
w_assert1 (parent.is_fixed());
w_assert1 (parent.is_node());
w_assert1 (child.is_fixed());
pushedup = false;
// let's try upgrading parent to EX latch. This highly likely fails in high-load situation,
// so let's do it here to avoid system transaction creation cost.
// we start from parent because EX latch on child is assured to be available in this order
if (!parent.upgrade_latch_conditional()) {
DBGOUT1(<< "opportunistic_adopt gave it up because of parent. "
<< parent.pid() << ". do nothing.");
increase_ex_need(parent.pid()); // give a hint to subsequent accesses
return RCOK;
}
rc_t btree_impl::_ux_norec_alloc_core(btree_page_h &page, PageID &new_page_id) {
// This is called only in REDO-only SSX, so no compensation logging. Just apply.
w_assert1 (xct()->is_single_log_sys_xct());
w_assert1 (page.latch_mode() == LATCH_EX);
W_DO(smlevel_0::vol->alloc_a_page(new_page_id));
btree_page_h new_page;
w_rc_t rc;
rc = new_page.fix_nonroot(page, new_page_id, LATCH_EX, false, true);
if (rc.is_error()) {
// if failed for any reason, we release the allocated page.
W_DO(smlevel_0::vol ->deallocate_page(new_page_id));
return rc;
}
// The new page has an empty key range; parent's high to high.
w_keystr_t fence, chain_high;
page.copy_fence_high_key(fence);
bool was_right_most = (page.get_chain_fence_high_length() == 0);
page.copy_chain_fence_high_key(chain_high);
if (was_right_most) {
// this means there was no chain or the page was the right-most of it.
// (so its high=high of chain)
// upon the first foster split, we start setting the chain-high.
page.copy_fence_high_key(chain_high);
}
#if W_DEBUG_LEVEL >= 3
lsn_t old_lsn = page.get_page_lsn();
#endif //W_DEBUG_LEVEL
W_DO(log_btree_norec_alloc(page, new_page, new_page_id, fence, chain_high));
DBGOUT3(<< "btree_impl::_ux_norec_alloc_core, fence=" << fence << ", old-LSN="
<< old_lsn << ", new-LSN=" << page.get_page_lsn() << ", PID=" << new_page_id);
// initialize as an empty child:
new_page.format_steal(page.get_page_lsn(), new_page_id, page.store(),
page.root(), page.level(), 0, lsn_t::null,
page.get_foster_opaqueptr(), page.get_foster_emlsn(),
fence, fence, chain_high, false);
page.accept_empty_child(page.get_page_lsn(), new_page_id, false /*not from redo*/);
// in this operation, the log contains everything we need to recover without any
// write-order-dependency. So, no registration for WOD.
w_assert3(new_page.is_consistent(true, true));
w_assert1(new_page.is_fixed());
w_assert1(new_page.latch_mode() == LATCH_EX);
w_assert3(page.is_consistent(true, true));
w_assert1(page.is_fixed());
return RCOK;
}
rc_t bt_cursor_t::_advance_one_slot(btree_page_h &p, bool &eof)
{
w_assert1(p.is_fixed());
w_assert1(_slot <= p.nrecs());
if(_forward) {
++_slot;
} else {
--_slot;
}
eof = false;
// keep following the next page.
// because we might see empty pages to skip consecutively!
while (true) {
bool time2move = _forward ? (_slot >= p.nrecs()) : _slot < 0;
if (time2move) {
// Move to right(left) sibling
bool reached_end = _forward ? p.is_fence_high_supremum() : p.is_fence_low_infimum();
if (reached_end) {
eof = true;
return RCOK;
}
// now, use fence keys to tell where the neighboring page exists
w_keystr_t neighboring_fence;
btree_impl::traverse_mode_t traverse_mode;
bool only_low_fence_exact_match = false;
if (_forward) {
p.copy_fence_high_key(neighboring_fence);
traverse_mode = btree_impl::t_fence_low_match;
int d = _upper.compare(neighboring_fence);
if (d < 0 || (d == 0 && !_upper_inclusive)) {
eof = true;
return RCOK;
}
if (d == 0 && _upper_inclusive) {
// we will check the next page, but the only
// possible matching is an entry with
// the low-fence..
only_low_fence_exact_match = true;
}
} else {
// if we are going backwards, the current page had
// low = [current-fence-low], high = [current-fence-high]
// and the previous page should have
// low = [?], high = [current-fence-low].
p.copy_fence_low_key(neighboring_fence);
// let's find a page which has this value as high-fence
traverse_mode = btree_impl::t_fence_high_match;
int d = _lower.compare(neighboring_fence);
if (d >= 0) {
eof = true;
return RCOK;
}
}
p.unfix();
// take lock for the fence key
if (_needs_lock) {
lockid_t lid (_store, (const unsigned char*) neighboring_fence.buffer_as_keystr(), neighboring_fence.get_length_as_keystr());
okvl_mode lock_mode;
if (only_low_fence_exact_match) {
lock_mode = _ex_lock ? ALL_X_GAP_N: ALL_S_GAP_N;
} else {
lock_mode = _ex_lock ? ALL_X_GAP_X : ALL_S_GAP_S;
}
// we can unconditionally request lock because we already released latch
W_DO(ss_m::lm->lock(lid.hash(), lock_mode, true, true, true));
}
// TODO this part should check if we find an exact match of fence keys.
// because we unlatch above, it's possible to not find exact match.
// in that case, we should change the traverse_mode to fence_contains and continue
W_DO(btree_impl::_ux_traverse(_store, neighboring_fence, traverse_mode, LATCH_SH, p));
_slot = _forward ? 0 : p.nrecs() - 1;
_set_current_page(p);
continue;
}
// take lock on the next key.
// NOTE: until we get locks, we aren't sure the key really becomes
// the next key. So, we use the temporary variable _tmp_next_key_buf.
const okvl_mode *mode = NULL;
{
p.get_key(_slot, _tmp_next_key_buf);
if (_forward) {
int d = _tmp_next_key_buf.compare(_upper);
if (d < 0) {
mode = _ex_lock ? &ALL_X_GAP_X : &ALL_S_GAP_S;
} else if (d == 0 && _upper_inclusive) {
mode = _ex_lock ? &ALL_X_GAP_N : &ALL_S_GAP_N;
} else {
eof = true;
mode = &ALL_N_GAP_N;
}
} else {
int d = _tmp_next_key_buf.compare(_lower);
if (d > 0) {
mode = _ex_lock ? &ALL_X_GAP_X : &ALL_S_GAP_S;
} else if (d == 0 && _lower_inclusive) {
mode = _ex_lock ? &ALL_X_GAP_X : &ALL_S_GAP_S;
} else {
eof = true;
mode = _ex_lock ? &ALL_N_GAP_X : &ALL_N_GAP_S;
}
}
}
if (_needs_lock && !mode->is_empty()) {
rc_t rc = btree_impl::_ux_lock_key (_store, p, _tmp_next_key_buf,
LATCH_SH, *mode, false);
if (rc.is_error()) {
if (rc.err_num() == eLOCKRETRY) {
W_DO(_check_page_update(p));
continue;
} else {
return rc;
}
}
}
// okay, now we are sure the _tmp_next_key_buf is the key we want to use
_key = _tmp_next_key_buf;
return RCOK; // found a record! (or eof)
}
return RCOK;
}
