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 btree_impl::_sx_adopt_foster (btree_page_h &parent, btree_page_h &child) {
w_keystr_t new_child_key;
child.copy_fence_high_key(new_child_key);
W_DO(_sx_split_if_needed(parent, new_child_key));
// Now, another SSX to move the pointer
sys_xct_section_t sxs(true);
W_DO(sxs.check_error_on_start());
rc_t ret = _ux_adopt_foster_core(parent, child, new_child_key);
W_DO (sxs.end_sys_xct (ret));
DBG(<< "Adopted " << child.pid() << " into " << parent.pid());
return ret;
}
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;
}
