Пример #1
0
void	mu_swap_root(glist *gl_ptr, int *root_swap_statistic_ptr)
{
	sgmnt_data_ptr_t	csd;
	sgmnt_addrs		*csa;
	node_local_ptr_t	cnl;
	srch_hist		*dir_hist_ptr, *gvt_hist_ptr;
	gv_namehead		*save_targ;
	block_id		root_blk_id, child_blk_id, free_blk_id;
	sm_uc_ptr_t		root_blk_ptr, child_blk_ptr;
	kill_set		kill_set_list;
	trans_num		curr_tn, ret_tn;
	int			level, root_blk_lvl;
	block_id		save_root;
	boolean_t		tn_aborted;
	unsigned int		lcl_t_tries;
	enum cdb_sc		status;
	DCL_THREADGBL_ACCESS;

	SETUP_THREADGBL_ACCESS;
	assert(mu_reorg_process);
	gv_target = gl_ptr->gvt;
	gv_target->root = 0;		/* reset root so we recompute it in DO_OP_GVNAME below */
	gv_target->clue.end = 0;	/* reset clue since reorg action on later globals might have invalidated it */
	reorg_gv_target->gvname.var_name = gv_target->gvname.var_name;	/* needed by SAVE_ROOTSRCH_ENTRY_STATE */
	dir_hist_ptr = gv_target->alt_hist;
	gvt_hist_ptr = &(gv_target->hist);
	inctn_opcode = inctn_invalid_op;
	DO_OP_GVNAME(gl_ptr);
		/* sets gv_target/gv_currkey/gv_cur_region/cs_addrs/cs_data to correspond to <globalname,reg> in gl_ptr */
	csa = cs_addrs;
	cnl = csa->nl;
	csd = cs_data;	/* Be careful to keep csd up to date. With MM, cs_data can change, and
			 * dereferencing an older copy can result in a SIG-11.
			 */
	if (gv_cur_region->read_only)
		return;	/* Cannot proceed for read-only data files */
	if (0 == gv_target->root)
	{	/* Global does not exist (online rollback). No problem. */
		gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_GBLNOEXIST, 2, GNAME(gl_ptr).len, GNAME(gl_ptr).addr);
		return;
	}
	if (dba_mm == csd->acc_meth)
		/* return for now without doing any swapping operation because later mu_truncate
		 * is going to issue the MUTRUNCNOTBG message.
		 */
		return;
	SET_GV_ALTKEY_TO_GBLNAME_FROM_GV_CURRKEY;		/* set up gv_altkey to be just the gblname */
	/* ------------ Swap root block of global variable tree --------- */
	t_begin(ERR_MUREORGFAIL, UPDTRNS_DB_UPDATED_MASK);
	for (;;)
	{
		curr_tn = csa->ti->curr_tn;
		kill_set_list.used = 0;
		save_root = gv_target->root;
		gv_target->root = csa->dir_tree->root;
		gv_target->clue.end = 0;
		if (cdb_sc_normal != (status = gvcst_search(gv_altkey, dir_hist_ptr)))
		{	/* Assign directory tree path to dir_hist_ptr */
			assert(t_tries < CDB_STAGNATE);
			gv_target->root = save_root;
			t_retry(status);
			continue;
		}
		gv_target->root = save_root;
		gv_target->clue.end = 0;
		if (cdb_sc_normal != (gvcst_search(gv_currkey, NULL)))
		{	/* Assign global variable tree path to gvt_hist_ptr */
			assert(t_tries < CDB_STAGNATE);
			t_retry(status);
			continue;
		}
		/* We've already search the directory tree in op_gvname/t_retry and obtained gv_target->root.
		 * Should restart with gvtrootmod2 if they don't agree. gvcst_root_search is the final arbiter.
		 * Really need that for debug info and also should assert(gv_currkey is global name).
		 */
		root_blk_lvl = gvt_hist_ptr->depth;
		assert(root_blk_lvl > 0);
		root_blk_ptr = gvt_hist_ptr->h[root_blk_lvl].buffaddr;
		root_blk_id = gvt_hist_ptr->h[root_blk_lvl].blk_num;
		assert((CDB_STAGNATE > t_tries) || (gv_target->root == gvt_hist_ptr->h[root_blk_lvl].blk_num));
		free_blk_id = swap_root_or_directory_block(0, root_blk_lvl, dir_hist_ptr, root_blk_id,
				root_blk_ptr, &kill_set_list, curr_tn);
		if (RETRY_SWAP == free_blk_id)
			continue;
		else if (ABORT_SWAP == free_blk_id)
			break;
		update_trans = UPDTRNS_DB_UPDATED_MASK;
		inctn_opcode = inctn_mu_reorg;
		assert(1 == kill_set_list.used);
		need_kip_incr = TRUE;
		if (!csa->now_crit)
			WAIT_ON_INHIBIT_KILLS(cnl, MAXWAIT2KILL);
		DEBUG_ONLY(lcl_t_tries = t_tries);
		TREF(in_mu_swap_root_state) = MUSWP_INCR_ROOT_CYCLE;
		assert(!TREF(in_gvcst_redo_root_search));
		if ((trans_num)0 == (ret_tn = t_end(gvt_hist_ptr, dir_hist_ptr, TN_NOT_SPECIFIED)))
		{
			TREF(in_mu_swap_root_state) = MUSWP_NONE;
			need_kip_incr = FALSE;
			assert(NULL == kip_csa);
			ABORT_TRANS_IF_GBL_EXIST_NOMORE(lcl_t_tries, tn_aborted);
			if (tn_aborted)
			{	/* It is not an error if the global (that once existed) doesn't exist anymore (due to ROLLBACK) */
				gtm_putmsg_csa(CSA_ARG(csa) VARLSTCNT(4) ERR_GBLNOEXIST, 2, GNAME(gl_ptr).len, GNAME(gl_ptr).addr);
				return;
			}
			continue;
		}
		TREF(in_mu_swap_root_state) = MUSWP_NONE;
		/* Note that this particular process's csa->root_search_cycle is now behind cnl->root_search_cycle.
		 * This forces a cdb_sc_gvtrootmod2 restart in gvcst_bmp_mark_free below.
		 */
		assert(cnl->root_search_cycle > csa->root_search_cycle);
		gvcst_kill_sort(&kill_set_list);
		GVCST_BMP_MARK_FREE(&kill_set_list, ret_tn, inctn_mu_reorg, inctn_bmp_mark_free_mu_reorg, inctn_opcode, csa);
		DECR_KIP(csd, csa, kip_csa);
		*root_swap_statistic_ptr += 1;
		break;
	}
	/* ------------ Swap blocks in branch of directory tree --------- */
	for (level = 0; level <= MAX_BT_DEPTH; level++)
	{
		t_begin(ERR_MUREORGFAIL, UPDTRNS_DB_UPDATED_MASK);
		for (;;)
		{
			curr_tn = csa->ti->curr_tn;
			kill_set_list.used = 0;
			save_root = gv_target->root;
			gv_target->root = csa->dir_tree->root;
			gv_target->clue.end = 0;
			if (cdb_sc_normal != (status = gvcst_search(gv_altkey, dir_hist_ptr)))
			{	/* assign branch path of directory tree into dir_hist_ptr */
				assert(t_tries < CDB_STAGNATE);
				gv_target->root = save_root;
				t_retry(status);
				continue;
			}
			gv_target->root = save_root;
			gv_target->clue.end = 0;
			if (level >= dir_hist_ptr->depth)
			{	/* done */
				t_abort(gv_cur_region, csa);
				return;
			}
			child_blk_ptr = dir_hist_ptr->h[level].buffaddr;
			child_blk_id = dir_hist_ptr->h[level].blk_num;
			assert(csa->dir_tree->root != child_blk_id);
			free_blk_id = swap_root_or_directory_block(level + 1, level, dir_hist_ptr, child_blk_id,
					child_blk_ptr, &kill_set_list, curr_tn);
			if (level == 0)
				/* set level as 1 to mark this kill set is for level-0 block in directory tree.
				 * The kill-set level later will be used in gvcst_bmp_markfree to assign a special value to
				 * cw_set_element, which will be eventually used by t_end to write the block to snapshot
				 */
				kill_set_list.blk[kill_set_list.used - 1].level = 1;
			if (RETRY_SWAP == free_blk_id)
				continue;
			else if (ABORT_SWAP == free_blk_id)
				break;
			update_trans = UPDTRNS_DB_UPDATED_MASK;
			inctn_opcode = inctn_mu_reorg;
			assert(1 == kill_set_list.used);
			need_kip_incr = TRUE;
			if (!csa->now_crit)
				WAIT_ON_INHIBIT_KILLS(cnl, MAXWAIT2KILL);
			DEBUG_ONLY(lcl_t_tries = t_tries);
			TREF(in_mu_swap_root_state) = MUSWP_DIRECTORY_SWAP;
			if ((trans_num)0 == (ret_tn = t_end(dir_hist_ptr, NULL, TN_NOT_SPECIFIED)))
			{
				TREF(in_mu_swap_root_state) = MUSWP_NONE;
				need_kip_incr = FALSE;
				assert(NULL == kip_csa);
				continue;
			}
			TREF(in_mu_swap_root_state) = MUSWP_NONE;
			gvcst_kill_sort(&kill_set_list);
			TREF(in_mu_swap_root_state) = MUSWP_FREE_BLK;
			GVCST_BMP_MARK_FREE(&kill_set_list, ret_tn, inctn_mu_reorg, inctn_bmp_mark_free_mu_reorg,
					inctn_opcode, csa);
			TREF(in_mu_swap_root_state) = MUSWP_NONE;
			DECR_KIP(csd, csa, kip_csa);
			break;
		}
	}
	return;
}
Пример #2
0
/****************************************************************
Input Parameter:
	gn = Global name
	exclude_glist_ptr = list of globals in EXCLUDE option
	index_fill_factor = index blocks' fill factor
	data_fill_factor = data blocks' fill factor
Input/Output Parameter:
	resume = resume flag
	reorg_op = What operations to do (coalesce or, swap or, split) [Default is all]
			[Only for debugging]
 ****************************************************************/
boolean_t mu_reorg(mval *gn, glist *exclude_glist_ptr, boolean_t *resume, int index_fill_factor, int data_fill_factor, int reorg_op)
{
	boolean_t		end_of_tree = FALSE, complete_merge, detailed_log;
	int			rec_size;
	/*
	 *
	 * "level" is the level of the working block.
	 * "pre_order_successor_level" is pre_order successor level except in the case
	 * where we are in a left-most descent of the tree
	 * in which case pre_order_successor_level will be the maximum height of that subtree
	 * until we reach the leaf level block .
	 * In other words, pre_order_successor_level and level variable controls the iterative pre-order traversal.
	 * We start reorg from the (root_level - 1) to 0. That is, level = pre_order_successor_level:-1:0.
	 */
	int			pre_order_successor_level, level;
	static block_id		dest_blk_id = 0;
	int			tkeysize;
	int			blks_killed, blks_processed, blks_reused, blks_coalesced, blks_split, blks_swapped,
				count, file_extended, lvls_reduced;
	int			d_max_fill, i_max_fill, blk_size, cur_blk_size, max_fill, toler, d_toler, i_toler;
	int			cnt1, cnt2;
	kill_set		kill_set_list;
	sm_uc_ptr_t		rPtr1;
	enum cdb_sc		status;
	srch_hist		*rtsib_hist;
	jnl_buffer_ptr_t	jbp;
	trans_num		ret_tn;

	error_def(ERR_MUREORGFAIL);
	error_def(ERR_DBRDONLY);
	error_def(ERR_GBLNOEXIST);
	error_def(ERR_MAXBTLEVEL);

	t_err = ERR_MUREORGFAIL;
	kill_set_tail = &kill_set_list;
	/* Initialization for current global */
	op_gvname(VARLSTCNT(1) gn);
	/* Cannot proceed for read-only data files */
	if (gv_cur_region->read_only)
	{
		gtm_putmsg(VARLSTCNT(4) ERR_DBRDONLY, 2, DB_LEN_STR(gv_cur_region));
		return FALSE;
	}
	dest_blk_id = cs_addrs->reorg_last_dest;
	inctn_opcode = inctn_mu_reorg;

	/* If resume option is present, then reorg_restart_key should be not null.
	 * Skip all globals until we are in the region for that global.
	 * Get the reorg_restart_key and reorg_restart_block from database header and restart from there.
	 */
	if (*resume && 0 != cs_data->reorg_restart_key[0])
	{
		/* resume from last key reorged in GVT */
		GET_KEY_LEN(tkeysize, &cs_data->reorg_restart_key[0]);
		memcpy(gv_currkey->base, cs_data->reorg_restart_key, tkeysize);
		gv_currkey->end = tkeysize - 1;
		dest_blk_id = cs_data->reorg_restart_block;
 		if (0 == memcmp(cs_data->reorg_restart_key, gn->str.addr, gn->str.len))
			/* Going to resume from current global, so it resumed and make it false */
			*resume = FALSE;
	} else
	{
		/* start from the left most leaf */
		memcpy(&gv_currkey->base[0], gn->str.addr, gn->str.len);
		gv_currkey->base[gn->str.len] = gv_currkey->base[gn->str.len + 1] = 0;
		gv_currkey->end = gn->str.len + 1;
	}
	if (*resume)
	{
		util_out_print("REORG cannot be resumed from this point, Skipping this global...", FLUSH);
		memcpy(&gv_currkey->base[0], gn->str.addr, gn->str.len);
		gv_currkey->base[gn->str.len] = gv_currkey->base[gn->str.len + 1] = 0;
		gv_currkey->end = gn->str.len + 1;
		return TRUE;
	}
 	memcpy(&gv_currkey_next_reorg->base[0], &gv_currkey->base[0], gv_currkey->end + 1);
	gv_currkey_next_reorg->end =  gv_currkey->end;
	if (2 > dest_blk_id)
		dest_blk_id = 2; /* we know that first block is bitmap and next one is directory tree root */
	file_extended = cs_data->trans_hist.total_blks;
	blk_size = cs_data->blk_size;
	d_max_fill = (double)data_fill_factor * blk_size / 100.0 - cs_data->reserved_bytes;
	i_max_fill = (double)index_fill_factor * blk_size / 100.0 - cs_data->reserved_bytes;
	d_toler = (double) DATA_FILL_TOLERANCE * blk_size / 100.0;
	i_toler = (double) INDEX_FILL_TOLERANCE * blk_size / 100.0;
	blks_killed = blks_processed = blks_reused = lvls_reduced = blks_coalesced = blks_split = blks_swapped = 0;
	pre_order_successor_level = level = MAX_BT_DEPTH + 1; /* Just some high value to initialize */

	/* --- more detailed debugging information --- */
	if (detailed_log = reorg_op & DETAIL)
		util_out_print("STARTING to work on global ^!AD from region !AD", TRUE,
			gn->str.len, gn->str.addr, REG_LEN_STR(gv_cur_region));

	/* In each iteration of MAIN loop, a working block is processed for a GVT */
	for (; ;)	/* ================ START MAIN LOOP ================ */
	{
		/* If right sibling is completely merged with the working block, do not swap the working block
		 * with its final destination block. Continue trying next right sibling. Swap only at the end.
		 */
		complete_merge = TRUE;
		while(complete_merge)	/* === START WHILE COMPLETE_MERGE === */
		{
			if (mu_ctrlc_occurred || mu_ctrly_occurred)
			{
				cs_data->reorg_restart_block = dest_blk_id;
				memcpy(&cs_data->reorg_restart_key[0], &gv_currkey->base[0], gv_currkey->end + 1);
				return FALSE;
			}
			complete_merge = FALSE;
			blks_processed++;
			t_begin(ERR_MUREORGFAIL, UPDTRNS_DB_UPDATED_MASK);
			/* Folllowing for loop is to handle concurrency retry for split/coalesce */
			for (; ;)		/* === SPLIT-COALESCE LOOP STARTS === */
			{
				gv_target->clue.end = 0;
				/* search gv_currkey and get the result in gv_target */
				if ((status = gvcst_search(gv_currkey, NULL)) != cdb_sc_normal)
				{
					assert(CDB_STAGNATE > t_tries);
					t_retry(status);
					continue;
				} else if (gv_currkey->end + 1 != gv_target->hist.h[0].curr_rec.match)
                                {
					if (SIZEOF(blk_hdr) == ((blk_hdr_ptr_t)gv_target->hist.h[0].buffaddr)->bsiz
						&& 1 == gv_target->hist.depth)
					{
						if (cs_addrs->now_crit)
						{
							t_abort(gv_cur_region, cs_addrs); /* do crit and other cleanup */
							gtm_putmsg(VARLSTCNT(4) ERR_GBLNOEXIST, 2, gn->str.len, gn->str.addr);
							reorg_finish(dest_blk_id, blks_processed, blks_killed, blks_reused,
								file_extended, lvls_reduced,
								blks_coalesced, blks_split, blks_swapped);
							return TRUE; /* It is not an error that global was killed */
						} else
						{
							assert(CDB_STAGNATE > t_tries);
							t_retry(status);
							continue;
						}
					}
                                }
				if (gv_target->hist.depth <= level)
				{
					/* Will come here
					 * 	1) first iteration of the for loop (since level == MAX_BT_DEPTH + 1) or,
					 *	2) tree depth decreased for mu_reduce_level or, M-kill
					 */
					pre_order_successor_level = gv_target->hist.depth - 1;
					if (MAX_BT_DEPTH + 1 != level)
					{
						/* break the loop when tree depth decreased (case 2) */
						level = pre_order_successor_level;
						break;
					}
					level = pre_order_successor_level;
				}
				max_fill = (0 == level)? d_max_fill : i_max_fill;
				toler = (0 == level)? d_toler:i_toler;
				cur_blk_size =  ((blk_hdr_ptr_t)(gv_target->hist.h[level].buffaddr))->bsiz;
				if (cur_blk_size > max_fill + toler && 0 == (reorg_op & NOSPLIT)) /* SPLIT BLOCK */
				{
					cnt1 = cnt2 = 0;
					/* history of current working block is in gv_target */
					status = mu_split(level, i_max_fill, d_max_fill, &cnt1, &cnt2);
					if (cdb_sc_maxlvl == status)
					{
						gtm_putmsg(VARLSTCNT(4) ERR_MAXBTLEVEL, 2, gn->str.len, gn->str.addr);
						reorg_finish(dest_blk_id, blks_processed, blks_killed, blks_reused,
							file_extended, lvls_reduced, blks_coalesced, blks_split, blks_swapped);
						return FALSE;
					} else if (cdb_sc_normal == status)
					{
						if ((trans_num)0 == (ret_tn = t_end(&(gv_target->hist), NULL, TN_NOT_SPECIFIED)))
						{
							need_kip_incr = FALSE;
							continue;
						}
						if (detailed_log)
							log_detailed_log("SPL", &(gv_target->hist), NULL, level, NULL, ret_tn);
						blks_reused += cnt1;
						lvls_reduced -= cnt2;
						blks_split++;
						break;
					} else if (cdb_sc_oprnotneeded == status)
					{	/* undo any update_array/cw_set changes and DROP THRU to mu_clsce */
						cw_set_depth = 0;
						CHECK_AND_RESET_UPDATE_ARRAY;	/* reset update_array_ptr to update_array */
						assert(0 == cw_map_depth); /* mu_swap_blk (that changes cw_map_depth) comes later */
					} else
					{
						assert(CDB_STAGNATE > t_tries);
						t_retry(status);
						continue;
					}
				} /* end if SPLIT BLOCK */
				/* We are here because, mu_split() was not called or, split was not done or, not required */
				rtsib_hist = gv_target->alt_hist;
				status = gvcst_rtsib(rtsib_hist, level);
				if (cdb_sc_normal != status && cdb_sc_endtree != status)
				{
					assert(CDB_STAGNATE > t_tries);
					t_retry(status);
					continue;
				}
				if (cdb_sc_endtree == status)
				{
					if (0 == level)
						end_of_tree = TRUE;
					break;
				} else if (0 == level)
					pre_order_successor_level = rtsib_hist->depth - 1;
				/* COALESCE WITH RTSIB */
				kill_set_list.used = 0;
				if (cur_blk_size < max_fill - toler && 0 == (reorg_op & NOCOALESCE))
				{
					/* histories are sent in &gv_target->hist and gv_target->alt_hist */
					status = mu_clsce(level, i_max_fill, d_max_fill, &kill_set_list, &complete_merge);
					if (cdb_sc_normal == status)
					{
						if (level) /* delete lower elements of array, t_end might confuse */
						{
							memmove(&rtsib_hist->h[0], &rtsib_hist->h[level],
								SIZEOF(srch_blk_status)*(rtsib_hist->depth - level + 2));
							rtsib_hist->depth = rtsib_hist->depth - level;
						}
						if (0 < kill_set_list.used)     /* increase kill_in_prog */
						{
							need_kip_incr = TRUE;
							if (!cs_addrs->now_crit)	/* Do not sleep while holding crit */
								WAIT_ON_INHIBIT_KILLS(cs_addrs->nl, MAXWAIT2KILL);
						}
						if ((trans_num)0 == (ret_tn = t_end(&(gv_target->hist), rtsib_hist,
							TN_NOT_SPECIFIED)))
						{
							need_kip_incr = FALSE;
							assert(NULL == kip_csa);
							if (level)
							{	/* reinitialize level member in rtsib_hist srch_blk_status' */
								for (count = 0; count < MAX_BT_DEPTH; count++)
									rtsib_hist->h[count].level = count;
							}
							continue;
						}
						if (level)
						{	/* reinitialize level member in rtsib_hist srch_blk_status' */
							for (count = 0; count < MAX_BT_DEPTH; count++)
								rtsib_hist->h[count].level = count;
						}
						if (detailed_log)
							log_detailed_log("CLS", &(gv_target->hist), rtsib_hist, level,
								NULL, ret_tn);
						assert(0 < kill_set_list.used || (NULL == kip_csa));
						if (0 < kill_set_list.used)     /* decrease kill_in_prog */
						{
							gvcst_kill_sort(&kill_set_list);
							GVCST_BMP_MARK_FREE(&kill_set_list, ret_tn, inctn_mu_reorg,
									inctn_bmp_mark_free_mu_reorg, inctn_opcode, cs_addrs)
							DECR_KIP(cs_data, cs_addrs, kip_csa);
							if (detailed_log)
								log_detailed_log("KIL", &(gv_target->hist), NULL, level,
									&kill_set_list, ret_tn);
							blks_killed += kill_set_list.used;
						}
						blks_coalesced++;
						break;
					} else if (cdb_sc_oprnotneeded == status)
					{	/* undo any update_array/cw_set changes and DROP THRU to t_end */
						cw_set_depth = 0;
						CHECK_AND_RESET_UPDATE_ARRAY;	/* reset update_array_ptr to update_array */
						assert(0 == cw_map_depth); /* mu_swap_blk (that changes cw_map_depth) comes later */
					} else
					{
						assert(CDB_STAGNATE > t_tries);
						t_retry(status);
						continue;
					}
				} /* end if try coalesce */
				if (0 == level)
				{
					/* Note: In data block level:
					 *      if split is successful or,
					 *	if coalesce is successful without a complete merge of rtsib,
					 *	then gv_currkey_next_reorg is already set from the called function.
					 *	if split or, coalesce do a retry or,
					 *	if coalesce is successful with a complete merge then
					 *	gv_currkey will not be changed.
					 * If split or, coalesce is not successful or, not needed then
					 *	here gv_currkey_next_reorg will be set from right sibling
					 */
					cw_set_depth = cw_map_depth = 0;
					GET_KEY_LEN(tkeysize, rtsib_hist->h[0].buffaddr + SIZEOF(blk_hdr) + SIZEOF(rec_hdr));
					if (2 < tkeysize && MAX_KEY_SZ >= tkeysize)
					{
						memcpy(&(gv_currkey_next_reorg->base[0]), rtsib_hist->h[0].buffaddr
							+ SIZEOF(blk_hdr) +SIZEOF(rec_hdr), tkeysize);
						gv_currkey_next_reorg->end = tkeysize - 1;
						inctn_opcode = inctn_invalid_op; /* temporary reset; satisfy an assert in t_end() */
						assert(UPDTRNS_DB_UPDATED_MASK == update_trans);
						update_trans = 0; /* tell t_end, this is no longer an update transaction */
						if ((trans_num)0 == (ret_tn = t_end(rtsib_hist, NULL, TN_NOT_SPECIFIED)))
						{
							need_kip_incr = FALSE;
							inctn_opcode = inctn_mu_reorg;	/* reset inctn_opcode to its default */
							update_trans = UPDTRNS_DB_UPDATED_MASK;/* reset update_trans to old value */
							assert(NULL == kip_csa);
							continue;
						}
						/* There is no need to reset update_trans in case of a successful "t_end" call.
						 * This is because before the next call to "t_end" we should have a call to
						 * "t_begin" which will reset update_trans anyways.
						 */
						inctn_opcode = inctn_mu_reorg;	/* reset inctn_opcode to its default */
						if (detailed_log)
							log_detailed_log("NOU", rtsib_hist, NULL, level, NULL, ret_tn);
					} else
					{
						assert(CDB_STAGNATE > t_tries);
						t_retry(status);
						continue;
					}
				} /* end if (0 == level) */
				break;
			}/* === SPLIT-COALESCE LOOP END === */
			t_abort(gv_cur_region, cs_addrs);	/* do crit and other cleanup */
		}/* === START WHILE COMPLETE_MERGE === */

		if (mu_ctrlc_occurred || mu_ctrly_occurred)
		{
			cs_data->reorg_restart_block = dest_blk_id;
			memcpy(&cs_data->reorg_restart_key[0], &gv_currkey->base[0], gv_currkey->end+1);
			return FALSE;
		}
		/* Now swap the working block */
		if (0 == (reorg_op & NOSWAP))
		{
			t_begin(ERR_MUREORGFAIL, UPDTRNS_DB_UPDATED_MASK);
			/* Following loop is to handle concurrency retry for swap */
			for (; ;)	/* === START OF SWAP LOOP === */
			{
				kill_set_list.used = 0;
				gv_target->clue.end = 0;
				/* search gv_currkey and get the result in gv_target */
				if ((status = gvcst_search(gv_currkey, NULL)) != cdb_sc_normal)
				{
					assert(CDB_STAGNATE > t_tries);
					t_retry(status);
					continue;
				} else if (gv_currkey->end + 1 != gv_target->hist.h[0].curr_rec.match)
                                {
					if (SIZEOF(blk_hdr) == ((blk_hdr_ptr_t)gv_target->hist.h[0].buffaddr)->bsiz
						&& 1 == gv_target->hist.depth)
					{
						if (cs_addrs->now_crit)
						{
							t_abort(gv_cur_region, cs_addrs); /* do crit and other cleanup */
							gtm_putmsg(VARLSTCNT(4) ERR_GBLNOEXIST, 2, gn->str.len, gn->str.addr);
							reorg_finish(dest_blk_id, blks_processed, blks_killed, blks_reused,
								file_extended, lvls_reduced,
								blks_coalesced, blks_split, blks_swapped);
							return TRUE; /* It is not an error that global was killed */
						} else
						{
							assert(CDB_STAGNATE > t_tries);
							t_retry(status);
							continue;
						}
					}
                                }
				if (gv_target->hist.depth <= level)
					break;
				/* swap working block with appropriate dest_blk_id block.
				   Historys are sent as gv_target->hist and reorg_gv_target->hist */
				mu_reorg_in_swap_blk = TRUE;
				status = mu_swap_blk(level, &dest_blk_id, &kill_set_list, exclude_glist_ptr);
				mu_reorg_in_swap_blk = FALSE;
				if (cdb_sc_oprnotneeded == status)
				{
					if (cs_data->trans_hist.total_blks <= dest_blk_id)
					{
						util_out_print("REORG may be incomplete for this global.", TRUE);
						reorg_finish(dest_blk_id, blks_processed, blks_killed, blks_reused,
							file_extended, lvls_reduced, blks_coalesced, blks_split, blks_swapped);
						return TRUE;
					}
				} else if (cdb_sc_normal == status)
				{
					if (0 < kill_set_list.used)
					{
						need_kip_incr = TRUE;
						if (!cs_addrs->now_crit)	/* Do not sleep while holding crit */
							WAIT_ON_INHIBIT_KILLS(cs_addrs->nl, MAXWAIT2KILL);
						/* second history not needed, because,
						   we are reusing a free block, which does not need history */
						if ((trans_num)0 == (ret_tn = t_end(&(gv_target->hist), NULL, TN_NOT_SPECIFIED)))
						{
							need_kip_incr = FALSE;
							assert(NULL == kip_csa);
							DECR_BLK_NUM(dest_blk_id);
							continue;
						}
						if (detailed_log)
							log_detailed_log("SWA", &(gv_target->hist), NULL, level, NULL, ret_tn);
						gvcst_kill_sort(&kill_set_list);
						GVCST_BMP_MARK_FREE(&kill_set_list, ret_tn, inctn_mu_reorg,
								inctn_bmp_mark_free_mu_reorg, inctn_opcode, cs_addrs)
						DECR_KIP(cs_data, cs_addrs, kip_csa);
						if (detailed_log)
							log_detailed_log("KIL", &(gv_target->hist), NULL, level,
								&kill_set_list, ret_tn);
						blks_reused += kill_set_list.used;
						blks_killed += kill_set_list.used;
					}
					/* gv_target->hist is for working block's history, and
					   reorg_gv_target->hist is for destinition block's history.
					   Note: gv_target and reorg_gv_target can be part of different GVT.  */
					else if ((trans_num)0 == (ret_tn = t_end(&(gv_target->hist), &(reorg_gv_target->hist),
						TN_NOT_SPECIFIED)))
					{
						need_kip_incr = FALSE;
						assert(NULL == kip_csa);
						DECR_BLK_NUM(dest_blk_id);
						continue;
					}
					if ((0 >= kill_set_list.used) && detailed_log)
						log_detailed_log("SWA", &(gv_target->hist), &(reorg_gv_target->hist),
							level, NULL, ret_tn);
					blks_swapped++;
					if (reorg_op & SWAPHIST)
						util_out_print("Dest !SL From !SL", TRUE, dest_blk_id,
							gv_target->hist.h[level].blk_num);
				} else
				{
					assert(CDB_STAGNATE > t_tries);
					t_retry(status);
					continue;
				}
				break;
			}	/* === END OF SWAP LOOP === */
			t_abort(gv_cur_region, cs_addrs);	/* do crit and other cleanup */
		}
		if (mu_ctrlc_occurred || mu_ctrly_occurred)
		{
			cs_data->reorg_restart_block = dest_blk_id;
			memcpy(&cs_data->reorg_restart_key[0], &gv_currkey->base[0], gv_currkey->end + 1);
			return FALSE;
		}
		if (end_of_tree)
			break;
		if (0 < level)
			level--; /* Order of reorg is root towards leaf */
		else
		{
			level = pre_order_successor_level;
			memcpy(&gv_currkey->base[0], &gv_currkey_next_reorg->base[0], gv_currkey_next_reorg->end + 1);
			gv_currkey->end =  gv_currkey_next_reorg->end;
			cs_data->reorg_restart_block = dest_blk_id;
			memcpy(&cs_data->reorg_restart_key[0], &gv_currkey->base[0], gv_currkey->end + 1);
		}
	}		/* ================ END MAIN LOOP ================ */

	/* =========== START REDUCE LEVEL ============== */
	memcpy(&gv_currkey->base[0], gn->str.addr, gn->str.len);
	gv_currkey->base[gn->str.len] = gv_currkey->base[gn->str.len + 1] = 0;
	gv_currkey->end = gn->str.len + 1;
	for (;;)	/* Reduce level continues until it fails to reduce */
	{
		t_begin(ERR_MUREORGFAIL, UPDTRNS_DB_UPDATED_MASK);
		cnt1 = 0;
		for (; ;) 	/* main reduce level loop starts */
		{
			kill_set_list.used = 0;
			gv_target->clue.end = 0;
			/* search gv_currkey and get the result in gv_target */
			if ((status = gvcst_search(gv_currkey, NULL)) != cdb_sc_normal)
			{
				assert(CDB_STAGNATE > t_tries);
				t_retry(status);
				continue;
			} else if (gv_currkey->end + 1 != gv_target->hist.h[0].curr_rec.match)
			{
				if (SIZEOF(blk_hdr) == ((blk_hdr_ptr_t)gv_target->hist.h[0].buffaddr)->bsiz
					&& 1 == gv_target->hist.depth)
				{
					if (cs_addrs->now_crit)
					{
						t_abort(gv_cur_region, cs_addrs);	/* do crit and other cleanup */
						gtm_putmsg(VARLSTCNT(4) ERR_GBLNOEXIST, 2, gn->str.len, gn->str.addr);
						reorg_finish(dest_blk_id, blks_processed, blks_killed, blks_reused,
							file_extended, lvls_reduced, blks_coalesced, blks_split, blks_swapped);
						return TRUE; /* It is not an error that global was killed */
					} else
					{
						assert(CDB_STAGNATE > t_tries);
						t_retry(status);
						continue;
					}
				}
			}
			if (gv_target->hist.depth <= level)
				break;
			/* History is passed in gv_target->hist */
			status = mu_reduce_level(&kill_set_list);
			if (cdb_sc_oprnotneeded != status && cdb_sc_normal != status)
			{
				assert(CDB_STAGNATE > t_tries);
				t_retry(status);
				continue;
			} else if (cdb_sc_normal == status)
			{
				assert(0 < kill_set_list.used);
				need_kip_incr = TRUE;
				if (!cs_addrs->now_crit)	/* Do not sleep while holding crit */
					WAIT_ON_INHIBIT_KILLS(cs_addrs->nl, MAXWAIT2KILL);
				if ((trans_num)0 == (ret_tn = t_end(&(gv_target->hist), NULL, TN_NOT_SPECIFIED)))
				{
					need_kip_incr = FALSE;
					assert(NULL == kip_csa);
					continue;
				}
				if (detailed_log)
					log_detailed_log("RDL", &(gv_target->hist), NULL, level, NULL, ret_tn);
				gvcst_kill_sort(&kill_set_list);
				GVCST_BMP_MARK_FREE(&kill_set_list, ret_tn, inctn_mu_reorg,
						inctn_bmp_mark_free_mu_reorg, inctn_opcode, cs_addrs)
				DECR_KIP(cs_data, cs_addrs, kip_csa);
				if (detailed_log)
					log_detailed_log("KIL", &(gv_target->hist), NULL, level, &kill_set_list, ret_tn);
				blks_reused += kill_set_list.used;
				blks_killed += kill_set_list.used;
				cnt1 = 1;
				lvls_reduced++;
			}
			break;
		} 		/* main reduce level loop ends */
		t_abort(gv_cur_region, cs_addrs); /* do crit and other cleanup */
		if (0 == cnt1)
			break;
	}
	/* =========== END REDUCE LEVEL ===========*/

	reorg_finish(dest_blk_id, blks_processed, blks_killed, blks_reused,
		file_extended, lvls_reduced, blks_coalesced, blks_split, blks_swapped);
	return TRUE;

} /* end mu_reorg() */