firebird-mirror/src/jrd/tra.cpp

/*
 *	PROGRAM:	JRD Access Method
 *	MODULE:		tra.cpp
 *	DESCRIPTION:	Transaction manager
 *
 * The contents of this file are subject to the Interbase Public
 * License Version 1.0 (the "License"); you may not use this file
 * except in compliance with the License. You may obtain a copy
 * of the License at http://www.Inprise.com/IPL.html
 *
 * Software distributed under the License is distributed on an
 * "AS IS" basis, WITHOUT WARRANTY OF ANY KIND, either express
 * or implied. See the License for the specific language governing
 * rights and limitations under the License.
 *
 * The Original Code was created by Inprise Corporation
 * and its predecessors. Portions created by Inprise Corporation are
 * Copyright (C) Inprise Corporation.
 *
 * All Rights Reserved.
 * Contributor(s): ______________________________________.
 * 2001.07.06 Sean Leyne - Code Cleanup, removed "#ifdef READONLY_DATABASE"
 *                         conditionals, as the engine now fully supports
 *                         readonly databases.
 * 2002.10.29 Nickolay Samofatov: Added support for savepoints
 */

#include "firebird.h"
#include <string.h>
#include "../jrd/common.h"
#include "../jrd/jrd.h"
#include "../jrd/tra.h"
#include "../jrd/ods.h"
#include "../jrd/pag.h"
#include "../jrd/lck.h"
#include "../jrd/ibase.h"
#include "../jrd/lls.h"
#include "../jrd/btr.h"
#include "../jrd/req.h"
#include "../jrd/exe.h"
#include "../jrd/extds/ExtDS.h"
#include "../jrd/rse.h"
#include "../jrd/intl_classes.h"
#include "../jrd/jrd_pwd.h"
#include "../jrd/ThreadStart.h"
#include "../jrd/UserManagement.h"
#include "../jrd/blb_proto.h"
#include "../jrd/cch_proto.h"
#include "../jrd/cmp_proto.h"
#include "../jrd/dfw_proto.h"
#include "../jrd/dpm_proto.h"
#include "../jrd/err_proto.h"
#include "../jrd/exe_proto.h"
#include "../jrd/ext_proto.h"
#include "../jrd/gds_proto.h"
#include "../jrd/isc_proto.h"
#include "../jrd/lck_proto.h"
#include "../jrd/met_proto.h"
#include "../jrd/mov_proto.h"
#include "../jrd/pag_proto.h"
#include "../jrd/rlck_proto.h"
#include "../jrd/thread_proto.h"
#include "../jrd/tpc_proto.h"
#include "../jrd/tra_proto.h"
#include "../jrd/vio_proto.h"
#include "../jrd/enc_proto.h"
#include "../jrd/jrd_proto.h"
#include "../common/classes/ClumpletWriter.h"
#include "../common/classes/TriState.h"
#include "../common/utils_proto.h"
#include "../lock/lock_proto.h"
#include "../dsql/dsql.h"
#include "../dsql/dsql_proto.h"
#include "../common/StatusArg.h"
#include "../jrd/trace/TraceManager.h"
#include "../jrd/trace/TraceJrdHelpers.h"


const int DYN_MSG_FAC	= 8;

using namespace Jrd;
using namespace Ods;
using namespace Firebird;

#ifdef GARBAGE_THREAD
#include "../jrd/isc_s_proto.h"
#endif

typedef Firebird::GenericMap<Firebird::Pair<Firebird::NonPooled<USHORT, UCHAR> > > RelationLockTypeMap;


#ifdef SUPERSERVER_V2
static SLONG bump_transaction_id(thread_db*, WIN *);
#else
static header_page* bump_transaction_id(thread_db*, WIN *);
#endif
static Lock* create_transaction_lock(thread_db* tdbb, void* object);
static void retain_context(thread_db*, jrd_tra*, bool, SSHORT);
#ifdef VMS
static void compute_oldest_retaining(thread_db*, jrd_tra*, bool);
#endif
static void expand_view_lock(thread_db* tdbb, jrd_tra*, jrd_rel*, UCHAR lock_type,
	const char* option_name, RelationLockTypeMap& lockmap, const int level);
static tx_inv_page* fetch_inventory_page(thread_db*, WIN *, SLONG, USHORT);
static const char* get_lockname_v3(const UCHAR lock);
static SLONG inventory_page(thread_db*, SLONG);
static SSHORT limbo_transaction(thread_db*, SLONG);
static void link_transaction(thread_db*, jrd_tra*);
static void restart_requests(thread_db*, jrd_tra*);
static void start_sweeper(thread_db*);
static THREAD_ENTRY_DECLARE sweep_database(THREAD_ENTRY_PARAM);
static void transaction_options(thread_db*, jrd_tra*, const UCHAR*, USHORT);
static jrd_tra* transaction_start(thread_db* tdbb, jrd_tra* temp);

static const UCHAR sweep_tpb[] =
{
	isc_tpb_version1, isc_tpb_read,
	isc_tpb_read_committed, isc_tpb_rec_version
};


void TRA_attach_request(Jrd::jrd_tra* transaction, Jrd::jrd_req* request)
{
	// When request finishes normally transaction reference is not cleared.
	// Then if afterwards request is restarted TRA_attach_request is called again.
	if (request->req_transaction)
	{
		if (request->req_transaction == transaction)
			return;
		TRA_detach_request(request);
	}

	fb_assert(request->req_transaction == NULL);
	fb_assert(request->req_tra_next == NULL);
	fb_assert(request->req_tra_prev == NULL);

	// Assign transaction reference
	request->req_transaction = transaction;

	// Add request to the doubly linked list
	if (transaction->tra_requests)
	{
		fb_assert(transaction->tra_requests->req_tra_prev == NULL);
		transaction->tra_requests->req_tra_prev = request;
		request->req_tra_next = transaction->tra_requests;
	}
	transaction->tra_requests = request;
}

void TRA_detach_request(Jrd::jrd_req* request)
{
	if (!request->req_transaction)
		return;

	// Remove request from the doubly linked list
	if (request->req_tra_next)
	{
		fb_assert(request->req_tra_next->req_tra_prev == request);
		request->req_tra_next->req_tra_prev = request->req_tra_prev;
	}

	if (request->req_tra_prev)
	{
		fb_assert(request->req_tra_prev->req_tra_next == request);
		request->req_tra_prev->req_tra_next = request->req_tra_next;
	}
	else
	{
		fb_assert(request->req_transaction->tra_requests == request);
		request->req_transaction->tra_requests = request->req_tra_next;
	}

	// Clear references
	request->req_transaction = NULL;
	request->req_tra_next = NULL;
	request->req_tra_prev = NULL;
}

bool TRA_active_transactions(thread_db* tdbb, Database* dbb)
{
/**************************************
 *
 *	T R A _ a c t i v e _ t r a n s a c t i o n s
 *
 **************************************
 *
 * Functional description
 *	Determine if any transactions are active.
 *	Return true is active transactions; otherwise
 *	return false if no active transactions.
 *
 **************************************/
	SET_TDBB(tdbb);

#ifndef VMS
	return ((LCK_query_data(tdbb, dbb->dbb_lock, LCK_tra, LCK_ANY)) ? true : false);
#else

	// Read header page and allocate transaction number.

	ULONG number, oldest, active;
#ifdef SUPERSERVER_V2
	number = dbb->dbb_next_transaction;
	oldest = dbb->dbb_oldest_transaction;
	active = MAX(dbb->dbb_oldest_active, dbb->dbb_oldest_transaction);
#else
	if (dbb->dbb_flags & DBB_read_only)
	{
		number = dbb->dbb_next_transaction;
		oldest = dbb->dbb_oldest_transaction;
		active = MAX(dbb->dbb_oldest_active, dbb->dbb_oldest_transaction);
	}
	else
	{
		WIN window(HEADER_PAGE_NUMBER);
		const header_page* header = (header_page*) CCH_FETCH(tdbb, &window, LCK_read, pag_header);
		number = header->hdr_next_transaction;
		oldest = header->hdr_oldest_transaction;
		active = MAX(header->hdr_oldest_active, header->hdr_oldest_transaction);
		CCH_RELEASE(tdbb, &window);
	}
#endif // SUPERSERVER_V2

	const ULONG base = oldest & ~TRA_MASK;
	const size_t length = (number - base + TRA_MASK) / 4;

	MemoryPool* const pool = dbb->dbb_permanent;
	Firebird::AutoPtr<jrd_tra> trans =
		FB_NEW(*pool) jrd_tra(pool, &dbb->dbb_memory_stats, NULL, NULL, length);

	// Build transaction bitmap to scan for active transactions.

	TRA_get_inventory(tdbb, trans->tra_transactions, base, number);

	Lock temp_lock;
	temp_lock.lck_dbb = dbb;
	temp_lock.lck_object = trans;
	temp_lock.lck_type = LCK_tra;
	temp_lock.lck_owner_handle = LCK_get_owner_handle(tdbb, temp_lock.lck_type);
	temp_lock.lck_parent = dbb->dbb_lock;
	temp_lock.lck_length = sizeof(SLONG);

	for (; active <= number; active++)
	{
		const ULONG byte = TRANS_OFFSET(active - base);
		const USHORT shift = TRANS_SHIFT(active);
		const USHORT state = (trans->tra_transactions[byte] >> shift) & TRA_MASK;
		if (state == tra_active)
		{
			temp_lock.lck_key.lck_long = active;
			if (!LCK_lock(tdbb, &temp_lock, LCK_read, LCK_NO_WAIT)) {
				return true;
			}
			LCK_release(tdbb, &temp_lock);
		}
	}

	return false;
#endif
}

void TRA_cleanup(thread_db* tdbb)
{
/**************************************
 *
 *	T R A _ c l e a n u p
 *
 **************************************
 *
 * Functional description
 *	TRA_cleanup is called at startup while an exclusive lock is
 *	held on the database.  Because we haven't started a transaction,
 *	and we have an exclusive lock on the db, any transactions marked
 *	as active on the transaction inventory pages are indeed dead.
 *	Mark them so.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* dbb = tdbb->getDatabase();
	CHECK_DBB(dbb);

	// Return without cleaning up the TIP's for a ReadOnly database
	if (dbb->dbb_flags & DBB_read_only)
		return;

	// First, make damn sure there are no outstanding transactions

	for (Attachment* attachment = dbb->dbb_attachments; attachment; attachment = attachment->att_next)
	{
		if (attachment->att_transactions)
			return;
	}

	const SLONG trans_per_tip = dbb->dbb_page_manager.transPerTIP;

	// Read header page and allocate transaction number.  Since
	// the transaction inventory page was initialized to zero, it
	// transaction is automatically marked active.

	WIN window(HEADER_PAGE_NUMBER);
	const header_page* header = (header_page*) CCH_FETCH(tdbb, &window, LCK_read, pag_header);
	const SLONG ceiling = header->hdr_next_transaction;
	const SLONG active = header->hdr_oldest_active;
	CCH_RELEASE(tdbb, &window);

	if (ceiling == 0)
		return;

	// Zip thru transactions from the "oldest active" to the next looking for
	// active transactions.  When one is found, declare it dead.

	const SLONG last = ceiling / trans_per_tip;
	SLONG number = active % trans_per_tip;
	SLONG limbo = 0;

	for (SLONG sequence = active / trans_per_tip; sequence <= last; sequence++, number = 0)
	{
		window.win_page = inventory_page(tdbb, sequence);
		tx_inv_page* tip = (tx_inv_page*) CCH_FETCH(tdbb, &window, LCK_write, pag_transactions);
		SLONG max = ceiling - (sequence * trans_per_tip);
		if (max > trans_per_tip)
			max = trans_per_tip - 1;
		for (; number <= max; number++)
		{
			const SLONG trans_offset = TRANS_OFFSET(number);
			UCHAR* byte = tip->tip_transactions + trans_offset;
			const SSHORT shift = TRANS_SHIFT(number);
			const SSHORT state = (*byte >> shift) & TRA_MASK;
			if (state == tra_limbo && limbo == 0)
				limbo = sequence * trans_per_tip + number;
			else if (state == tra_active)
			{
				CCH_MARK(tdbb, &window);
				*byte &= ~(TRA_MASK << shift);

				// hvlad: mark system transaction as committed
				if (sequence == 0 && number == 0) {
					*byte |= tra_committed << shift;
				}
				else {
					*byte |= tra_dead << shift;
				}
			}
		}
#ifdef SUPERSERVER_V2
		if (sequence == last)
		{
			CCH_MARK(tdbb, &window);
			for (; number < trans_per_tip; number++)
			{
				const SLONG trans_offset = TRANS_OFFSET(number);
				UCHAR* byte = tip->tip_transactions + trans_offset;
				const SSHORT shift = TRANS_SHIFT(number);
				*byte &= ~(TRA_MASK << shift);
				if (tip->tip_next)
					*byte |= tra_committed << shift;
				else
					*byte |= tra_active << shift;
			}
		}
#endif
		CCH_RELEASE(tdbb, &window);
	}

#ifdef SUPERSERVER_V2
	window.win_page = inventory_page(tdbb, last);
	tx_inv_page* tip = (tx_inv_page*) CCH_FETCH(tdbb, &window, LCK_write, pag_transactions);

	while (tip->tip_next)
	{
		CCH_RELEASE(tdbb, &window);
		window.win_page = inventory_page(tdbb, ++last);
		tip = (tx_inv_page*) CCH_FETCH(tdbb, &window, LCK_write, pag_transactions);
		CCH_MARK(tdbb, &window);
		for (number = 0; number < trans_per_tip; number++)
		{
			const SLONG trans_offset = TRANS_OFFSET(number);
			UCHAR* byte = tip->tip_transactions + trans_offset;
			const USHORT shift = TRANS_SHIFT(number);
			*byte &= ~(TRA_MASK << shift);
			if (tip->tip_next || !number)
				*byte |= tra_committed << shift;
			else
				*byte |= tra_active << shift;
		}

		if (!tip->tip_next)
			dbb->dbb_next_transaction = last * trans_per_tip;
	}

	CCH_RELEASE(tdbb, &window);
#endif
}


void TRA_commit(thread_db* tdbb, jrd_tra* transaction, const bool retaining_flag)
{
/**************************************
 *
 *	T R A _ c o m m i t
 *
 **************************************
 *
 * Functional description
 *	Commit a transaction.
 *
 **************************************/
	SET_TDBB(tdbb);

	TraceTransactionEnd trace(transaction, true, retaining_flag);

	EDS::Transaction::jrdTransactionEnd(tdbb, transaction, true, retaining_flag, false);

	// If this is a commit retaining, and no updates have been performed,
	// and no events have been posted (via stored procedures etc)
	// no-op the operation.

	if (retaining_flag && !(transaction->tra_flags & TRA_write || transaction->tra_deferred_job))
	{
		transaction->tra_flags &= ~TRA_prepared;
		// Get rid of all user savepoints
		while (transaction->tra_save_point && transaction->tra_save_point->sav_flags & SAV_user)
		{
			Savepoint* const next = transaction->tra_save_point->sav_next;
			transaction->tra_save_point->sav_next = NULL;
			VIO_verb_cleanup(tdbb, transaction);
			transaction->tra_save_point = next;
		}

		trace.finish(res_successful);
		return;
	}

	if (transaction->tra_flags & TRA_invalidated)
		ERR_post(Arg::Gds(isc_trans_invalid));

	Jrd::ContextPoolHolder context(tdbb, transaction->tra_pool);

	// Perform any meta data work deferred

	if (!(transaction->tra_flags & TRA_prepared))
		DFW_perform_work(tdbb, transaction);

	if (transaction->tra_flags & (TRA_prepare2 | TRA_reconnected))
		MET_update_transaction(tdbb, transaction, true);

	// Check in with external file system

	EXT_trans_commit(transaction);

#ifdef GARBAGE_THREAD
	// Flush pages if transaction logically modified data

	if (transaction->tra_flags & TRA_write)
#endif
		CCH_flush(tdbb, FLUSH_TRAN, transaction->tra_number);
#ifdef GARBAGE_THREAD
	else if (transaction->tra_flags & (TRA_prepare2 | TRA_reconnected))
	{
		// If the transaction only read data but is a member of a
		// multi-database transaction with a transaction description
		// message then flush RDB$TRANSACTIONS.

		CCH_flush(tdbb, FLUSH_SYSTEM, 0);
	}
#endif

	if (retaining_flag)
	{
		trace.finish(res_successful);
		retain_context(tdbb, transaction, true, tra_committed);
		return;
	}

	// Set the state on the inventory page to be committed

	TRA_set_state(tdbb, transaction, transaction->tra_number, tra_committed);

	// Perform any post commit work

	DFW_perform_post_commit_work(transaction);

	// notify any waiting locks that this transaction is committing;
	// there could be no lock if this transaction is being reconnected

	++transaction->tra_use_count;
	Lock* lock = transaction->tra_lock;
	if (lock && (lock->lck_logical < LCK_write))
		LCK_convert(tdbb, lock, LCK_write, LCK_WAIT);
	--transaction->tra_use_count;

	TRA_release_transaction(tdbb, transaction, &trace);
}


void TRA_extend_tip(thread_db* tdbb, ULONG sequence) //, WIN* precedence_window)
{
/**************************************
 *
 *	T R A _ e x t e n d _ t i p
 *
 **************************************
 *
 * Functional description
 *	Allocate and link in new TIP (transaction inventory page).
 *	This is called from TRA_start and from validate/repair.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* dbb = tdbb->getDatabase();
	CHECK_DBB(dbb);

	// Start by fetching prior transaction page, if any
	tx_inv_page* prior_tip = NULL;
	WIN prior_window(DB_PAGE_SPACE, -1);
	if (sequence) {
		prior_tip = fetch_inventory_page(tdbb, &prior_window, (SLONG) (sequence - 1), LCK_write);
	}

	// Allocate and format new page
	WIN window(DB_PAGE_SPACE, -1);
	tx_inv_page* tip = (tx_inv_page*) DPM_allocate(tdbb, &window);
	tip->tip_header.pag_type = pag_transactions;

	CCH_must_write(&window);
	CCH_RELEASE(tdbb, &window);

	// Release prior page

	if (sequence)
	{
		CCH_MARK_MUST_WRITE(tdbb, &prior_window);
		prior_tip->tip_next = window.win_page.getPageNum();
		CCH_RELEASE(tdbb, &prior_window);
	}

	// Link into internal data structures

	vcl* vector = dbb->dbb_t_pages =
		vcl::newVector(*dbb->dbb_permanent, dbb->dbb_t_pages, sequence + 1);
	(*vector)[sequence] = window.win_page.getPageNum();

	// Write into pages relation

	DPM_pages(tdbb, 0, pag_transactions, sequence, window.win_page.getPageNum());
}


int TRA_fetch_state(thread_db* tdbb, SLONG number)
{
/**************************************
 *
 *	T R A _ f e t c h _ s t a t e
 *
 **************************************
 *
 * Functional description
 *	Physically fetch the state of a given
 *	transaction on the transaction inventory
 *	page.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* dbb = tdbb->getDatabase();
	CHECK_DBB(dbb);

	// locate and fetch the proper TIP page

    const ULONG tip_number = (ULONG) number;
	const SLONG trans_per_tip = dbb->dbb_page_manager.transPerTIP;
	const ULONG tip_seq = tip_number / trans_per_tip;
	WIN window(DB_PAGE_SPACE, -1);
	const tx_inv_page* tip = fetch_inventory_page(tdbb, &window, tip_seq, LCK_read);

	// calculate the state of the desired transaction

	const ULONG byte = TRANS_OFFSET(tip_number % trans_per_tip);
	const USHORT shift = TRANS_SHIFT(tip_number);
	const USHORT state = (tip->tip_transactions[byte] >> shift) & TRA_MASK;

	CCH_RELEASE(tdbb, &window);

	return state;
}


void TRA_get_inventory(thread_db* tdbb, UCHAR* bit_vector, ULONG base, ULONG top)
{
/**************************************
 *
 *	T R A _ g e t _ i n v e n t o r y
 *
 **************************************
 *
 * Functional description
 *	Get an inventory of the state of all transactions
 *	between the base and top transactions passed.
 *	To get a consistent view of the transaction
 *	inventory (in case we ever implement sub-transactions),
 *	do handoffs to read the pages in order.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* dbb = tdbb->getDatabase();
	CHECK_DBB(dbb);

	const ULONG trans_per_tip = dbb->dbb_page_manager.transPerTIP;
	ULONG sequence = base / trans_per_tip;
	const ULONG last = top / trans_per_tip;

	// fetch the first inventory page

	WIN window(DB_PAGE_SPACE, -1);
	const tx_inv_page* tip = fetch_inventory_page(tdbb, &window, (SLONG) sequence++, LCK_read);

	// move the first page into the bit vector

	UCHAR* p = bit_vector;
	if (p)
	{
		ULONG l = base % trans_per_tip;
		const UCHAR* q = tip->tip_transactions + TRANS_OFFSET(l);
		l = TRANS_OFFSET(MIN((top + TRA_MASK - base), trans_per_tip - l));
		memcpy(p, q, l);
		p += l;
	}

	// move successive pages into the bit vector

	while (sequence <= last)
	{
		base = sequence * trans_per_tip;

		// release the read lock as we go, so that some one else can
		// commit without having to signal all other transactions.

		tip = (tx_inv_page*) CCH_HANDOFF(tdbb, &window, inventory_page(tdbb, sequence++),
							  LCK_read, pag_transactions);
		TPC_update_cache(tdbb, tip, sequence - 1);
		if (p)
		{
			const ULONG l = TRANS_OFFSET(MIN((top + TRA_MASK - base), trans_per_tip));
			memcpy(p, tip->tip_transactions, l);
			p += l;
		}
	}

	CCH_RELEASE(tdbb, &window);
}


int TRA_get_state(thread_db* tdbb, SLONG number)
{
/**************************************
 *
 *	T R A _ g e t _ s t a t e
 *
 **************************************
 *
 * Functional description
 *	Get the state of a given transaction on the
 *	transaction inventory page.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* dbb = tdbb->getDatabase();
	CHECK_DBB(dbb);

	if (dbb->dbb_tip_cache)
		return TPC_snapshot_state(tdbb, number);

	if (number && dbb->dbb_pc_transactions)
	{
		if (TRA_precommited(tdbb, number, number))
			return tra_precommitted;
	}

	return TRA_fetch_state(tdbb, number);
}


#ifdef SUPERSERVER_V2
void TRA_header_write(thread_db* tdbb, Database* dbb, SLONG number)
{
/**************************************
 *
 *	T R A _ h e a d e r _ w r i t e
 *
 **************************************
 *
 * Functional description
 *	Force transaction ID on header to disk.
 *	Do post fetch check of the transaction
 *	ID header write as a concurrent thread
 *	might have written the header page
 *	while blocked on the latch.
 *
 *	The idea is to amortize the cost of
 *	header page I/O across multiple transactions.
 *
 **************************************/
	SET_TDBB(tdbb);

	// If transaction number is already on disk just return.

	if (!number || dbb->dbb_last_header_write < number)
	{
		WIN window(HEADER_PAGE_NUMBER);
		header_page* header = (header_page*) CCH_FETCH(tdbb, &window, LCK_write, pag_header);

		if (header->hdr_next_transaction)
		{
			if (header->hdr_oldest_active > header->hdr_next_transaction)
				BUGCHECK(266);	//next transaction older than oldest active

			if (header->hdr_oldest_transaction > header->hdr_next_transaction)
				BUGCHECK(267);	// next transaction older than oldest transaction
		}

		// The header page might have been written while waiting
		// for the latch; perform a post fetch check and optimize
		// this case by not writing the page again.

		if (!number || dbb->dbb_last_header_write < number)
		{
			CCH_MARK_MUST_WRITE(tdbb, &window);
			if (dbb->dbb_next_transaction > header->hdr_next_transaction)
				header->hdr_next_transaction = dbb->dbb_next_transaction;

			if (dbb->dbb_oldest_active > header->hdr_oldest_active)
				header->hdr_oldest_active = dbb->dbb_oldest_active;

			if (dbb->dbb_oldest_transaction > header->hdr_oldest_transaction)
				header->hdr_oldest_transaction = dbb->dbb_oldest_transaction;

			if (dbb->dbb_oldest_snapshot > header->hdr_oldest_snapshot)
				header->hdr_oldest_snapshot = dbb->dbb_oldest_snapshot;
		}

		CCH_RELEASE(tdbb, &window);
	}
}
#endif


void TRA_init(Database* dbb)
{
/**************************************
 *
 *	T R A _ i n i t
 *
 **************************************
 *
 * Functional description
 *	"Start" the system transaction.
 *
 **************************************/
	CHECK_DBB(dbb);

	MemoryPool* const pool = dbb->dbb_permanent;
	jrd_tra* const trans = FB_NEW(*pool) jrd_tra(pool, &dbb->dbb_memory_stats, NULL, NULL);
	dbb->dbb_sys_trans = trans;
	trans->tra_flags |= TRA_system | TRA_ignore_limbo;
}


void TRA_invalidate(Database* database, ULONG mask)
{
/**************************************
 *
 *	T R A _ i n v a l i d a t e
 *
 **************************************
 *
 * Functional description
 *	Invalidate any active transactions that may have
 *	modified a page that couldn't be written.
 *
 **************************************/
	for (Attachment* attachment = database->dbb_attachments; attachment;
		attachment = attachment->att_next)
	{
		for (jrd_tra* transaction = attachment->att_transactions; transaction;
			transaction = transaction->tra_next)
		{
			const ULONG transaction_mask = 1L << (transaction->tra_number & (BITS_PER_LONG - 1));
			if (transaction_mask & mask && transaction->tra_flags & TRA_write)
				transaction->tra_flags |= TRA_invalidated;
		}
	}
}


void TRA_link_cursor(jrd_tra* transaction, dsql_req* cursor)
{
/**************************************
 *
 *	T R A _ l i n k _ c u r s o r
 *
 **************************************
 *
 * Functional description
 *	Add cursor to the list of open cursors belonging to this transaction.
 *
 **************************************/

	fb_assert(!transaction->tra_open_cursors.exist(cursor));
	transaction->tra_open_cursors.add(cursor);
}


void TRA_unlink_cursor(jrd_tra* transaction, dsql_req* cursor)
{
/**************************************
 *
 *	T R A _ u n l i n k _ c u r s o r
 *
 **************************************
 *
 * Functional description
 *	Remove cursor from the list of open cursors.
 *
 **************************************/

	size_t pos;
	if (transaction->tra_open_cursors.find(cursor, pos))
	{
		transaction->tra_open_cursors.remove(pos);
	}
}


void TRA_post_resources(thread_db* tdbb, jrd_tra* transaction, ResourceList& resources)
{
/**************************************
 *
 *	T R A _ p o s t _ r e s o u r c e s
 *
 **************************************
 *
 * Functional description
 *	Post interest in relation/procedure/collation existence to transaction.
 *	This guarantees that the relation/procedure/collation won't be dropped
 *	out from under the transaction.
 *
 **************************************/
	SET_TDBB(tdbb);

	Jrd::ContextPoolHolder context(tdbb, transaction->tra_pool);

	for (Resource* rsc = resources.begin(); rsc < resources.end(); rsc++)
	{
		if (rsc->rsc_type == Resource::rsc_relation ||
			rsc->rsc_type == Resource::rsc_procedure ||
			rsc->rsc_type == Resource::rsc_collation)
		{
			size_t i;
			if (!transaction->tra_resources.find(*rsc, i))
			{
				transaction->tra_resources.insert(i, *rsc);
				switch (rsc->rsc_type)
				{
				case Resource::rsc_relation:
					MET_post_existence(tdbb, rsc->rsc_rel);
					if (rsc->rsc_rel->rel_file) {
						EXT_tra_attach(rsc->rsc_rel->rel_file, transaction);
					}
					break;
				case Resource::rsc_procedure:
					rsc->rsc_prc->prc_use_count++;
#ifdef DEBUG_PROCS
					{
						char buffer[256];
						sprintf(buffer,
								"Called from TRA_post_resources():\n\t Incrementing use count of %s\n",
								rsc->rsc_prc->prc_name->c_str());
						JRD_print_procedure_info(tdbb, buffer);
					}
#endif
					break;
				case Resource::rsc_collation:
					rsc->rsc_coll->incUseCount(tdbb);
					break;
				default:   // shut up compiler warning
					break;
				}
			}
		}
	}
}


bool TRA_pc_active(thread_db* tdbb, SLONG number)
{
/**************************************
 *
 *	T R A _ p c _ a c t i v e
 *
 **************************************
 *
 * Functional description
 *	Returns whether a given precommitted transaction
 *  owned by some other guy active or not.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* dbb = tdbb->getDatabase();
	CHECK_DBB(dbb);

	Lock temp_lock;
	temp_lock.lck_dbb = dbb;
	temp_lock.lck_type = LCK_tra_pc;
	temp_lock.lck_owner_handle = LCK_get_owner_handle(tdbb, temp_lock.lck_type);
	temp_lock.lck_parent = dbb->dbb_lock;
	temp_lock.lck_length = sizeof(SLONG);
	temp_lock.lck_key.lck_long = number;

	// If we can't get a lock on the transaction, it must be active

	if (!LCK_lock(tdbb, &temp_lock, LCK_read, LCK_NO_WAIT))
	{
		fb_utils::init_status(tdbb->tdbb_status_vector);
		return true;
	}

	LCK_release(tdbb, &temp_lock);
	return false;
}


bool TRA_precommited(thread_db* tdbb, SLONG old_number, SLONG new_number)
{
/**************************************
 *
 *	T R A _ p r e c o m m i t e d	(s i c)
 *
 **************************************
 *
 * Functional description
 *	Maintain a vector of active precommitted
 *	transactions. If old_number <> new_number
 *	then swap old_number with new_number in
 *	the vector. If old_number equals new_number
 *	then test for that number's presence in
 *	the vector.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* dbb = tdbb->getDatabase();
	CHECK_DBB(dbb);

	vcl* vector = dbb->dbb_pc_transactions;
	if (!vector)
	{
		if (old_number == new_number)
			return false;
		vector = dbb->dbb_pc_transactions = vcl::newVector(*dbb->dbb_permanent, 1);
	}

	SLONG* zp = 0;
	for (vcl::iterator p = vector->begin(), end = vector->end(); p < end; ++p)
	{
		if (*p == old_number)
			return (*p = new_number) ? true : false;
		if (!zp && !*p)
			zp = &*p;
	}

	if (old_number == new_number || new_number == 0)
		return false;

	if (zp)
		*zp = new_number;
	else
	{
		vector->resize(vector->count() + 1);
		(*vector)[vector->count() - 1] = new_number;
	}

	return true;
}


void TRA_prepare(thread_db* tdbb, jrd_tra* transaction, USHORT length, const UCHAR* msg)
{
/**************************************
 *
 *	T R A _ p r e p a r e
 *
 **************************************
 *
 * Functional description
 *	Put a transaction into limbo.
 *
 **************************************/

	SET_TDBB(tdbb);

	if (transaction->tra_flags & TRA_prepared)
		return;

	if (transaction->tra_flags & TRA_invalidated)
		ERR_post(Arg::Gds(isc_trans_invalid));

	/* If there's a transaction description message, log it to RDB$TRANSACTION
	We should only log a message to RDB$TRANSACTION if there is a message
	to log (if the length = 0, we won't log the transaction in RDB$TRANSACTION)
	These messages are used to recover transactions in limbo.  The message indicates
	the action that is to be performed (hence, if nothing is getting logged, don't
	bother).
	*/

	/* Make sure that if msg is NULL there is no length.  The two
	should go hand in hand
	          msg == NULL || *msg == NULL
	*/
	fb_assert(!(!msg && length) || (msg && (!*msg && length)));

	if (msg && length)
	{
		MET_prepare(tdbb, transaction, length, msg);
		transaction->tra_flags |= TRA_prepare2;
	}

	// Check in with external file system

	EXT_trans_prepare(transaction);

	// Perform any meta data work deferred

	DFW_perform_work(tdbb, transaction);

#ifdef GARBAGE_THREAD
	// Flush pages if transaction logically modified data

	if (transaction->tra_flags & TRA_write)
#endif
		CCH_flush(tdbb, FLUSH_TRAN, transaction->tra_number);
#ifdef GARBAGE_THREAD
	else if (transaction->tra_flags & TRA_prepare2)
	{
		// If the transaction only read data but is a member of a
		// multi-database transaction with a transaction description
		// message then flush RDB$TRANSACTIONS.

		CCH_flush(tdbb, FLUSH_SYSTEM, 0);
	}
#endif

	// Set the state on the inventory page to be limbo

	transaction->tra_flags |= TRA_prepared;
	TRA_set_state(tdbb, transaction, transaction->tra_number, tra_limbo);
}


jrd_tra* TRA_reconnect(thread_db* tdbb, const UCHAR* id, USHORT length)
{
/**************************************
 *
 *	T R A _ r e c o n n e c t
 *
 **************************************
 *
 * Functional description
 *	Reconnect to a transaction in limbo.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* const dbb = tdbb->getDatabase();
	CHECK_DBB(dbb);
	Attachment* const attachment = tdbb->getAttachment();

	// Cannot work on limbo transactions for ReadOnly database
	if (dbb->dbb_flags & DBB_read_only)
		ERR_post(Arg::Gds(isc_read_only_database));

	const SLONG number = gds__vax_integer(id, length);
	if (number > dbb->dbb_next_transaction)
		PAG_header(tdbb, true);

	const UCHAR state = (number > dbb->dbb_next_transaction) ?
		255 : limbo_transaction(tdbb, number);

	if (state != tra_limbo)
	{
		USHORT message;

		switch (state)
		{
		case tra_active:
			message = 262;		// ACTIVE
			break;
		case tra_dead:
			message = 264;		// ROLLED BACK
			break;
		case tra_committed:
			message = 263;		// COMMITTED
			break;
		default:
			message = 265;		// ILL DEFINED
			break;
		}

		TEXT text[128];
		USHORT flags = 0;
		gds__msg_lookup(NULL, JRD_BUGCHK, message, sizeof(text), text, &flags);

		ERR_post(Arg::Gds(isc_no_recon) <<
				 Arg::Gds(isc_tra_state) << Arg::Num(number) << Arg::Str(text));
	}

	MemoryPool* const pool = dbb->createPool();
	Jrd::ContextPoolHolder context(tdbb, pool);
	jrd_tra* const trans = jrd_tra::create(pool, attachment, NULL);
	trans->tra_number = number;
	trans->tra_flags |= TRA_prepared | TRA_reconnected | TRA_write;

	link_transaction(tdbb, trans);

	return trans;
}


void TRA_release_transaction(thread_db* tdbb, jrd_tra* transaction, TraceTransactionEnd* trace)
{
/**************************************
 *
 *	T R A _ r e l e a s e _ t r a n s a c t i o n
 *
 **************************************
 *
 * Functional description
 *	Cleanup a transaction.  This is called by both COMMIT and
 *	ROLLBACK as well as code in JRD to get rid of remote
 *	transactions.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* dbb = tdbb->getDatabase();
	Attachment* attachment = tdbb->getAttachment();

	if (!transaction->tra_outer)
	{
		if (transaction->tra_blobs->getFirst())
		{
			while (true)
			{
				BlobIndex *current = &transaction->tra_blobs->current();
				if (current->bli_materialized)
				{
					if (!transaction->tra_blobs->getNext())
						break;
				}
				else
				{
					ULONG temp_id = current->bli_temp_id;
					BLB_cancel(tdbb, current->bli_blob_object);
					if (!transaction->tra_blobs->locate(Firebird::locGreat, temp_id))
						break;
				}
			}
		}

		while (transaction->tra_arrays)
			BLB_release_array(transaction->tra_arrays);
	}

	if (transaction->tra_pool)
	{
		// Iterate the doubly linked list of requests for transaction and null out the transaction references
		while (transaction->tra_requests)
			TRA_detach_request(transaction->tra_requests);
	}

	// Release interest in relation/procedure existence for transaction

	for (Resource* rsc = transaction->tra_resources.begin();
		rsc < transaction->tra_resources.end(); rsc++)
	{
		switch (rsc->rsc_type)
		{
		case Resource::rsc_relation:
			MET_release_existence(tdbb, rsc->rsc_rel);
			if (rsc->rsc_rel->rel_file) {
				EXT_tra_detach(rsc->rsc_rel->rel_file, transaction);
			}
			break;
		case Resource::rsc_procedure:
			CMP_decrement_prc_use_count(tdbb, rsc->rsc_prc);
			break;
		case Resource::rsc_collation:
			rsc->rsc_coll->decUseCount(tdbb);
			break;
		default:
			fb_assert(false);
		}
	}

	{ // scope
		vec<jrd_rel*>& rels = *dbb->dbb_relations;
		for (size_t i = 0; i < rels.count(); i++)
		{
			jrd_rel* relation = rels[i];
			if (relation && (relation->rel_flags & REL_temp_tran))
			{
				relation->delPages(tdbb, transaction->tra_number);
			}
		}

	} // end scope

	// Release the locks associated with the transaction

	vec<Lock*>* vector = transaction->tra_relation_locks;
	if (vector)
	{
		vec<Lock*>::iterator lock = vector->begin();
		for (ULONG i = 0; i < vector->count(); ++i, ++lock)
		{
			if (*lock)
				LCK_release(tdbb, *lock);
		}
	}

	++transaction->tra_use_count;
	if (transaction->tra_lock)
		LCK_release(tdbb, transaction->tra_lock);
	--transaction->tra_use_count;

	// release the sparse bit map used for commit retain transaction

	delete transaction->tra_commit_sub_trans;

	if (transaction->tra_flags & TRA_precommitted)
		TRA_precommited(tdbb, transaction->tra_number, 0);

	if (trace)
		trace->finish(res_successful);

	// Unlink the transaction from the database block

	for (jrd_tra** ptr = &attachment->att_transactions; *ptr; ptr = &(*ptr)->tra_next)
	{
		if (*ptr == transaction)
		{
			*ptr = transaction->tra_next;
			break;
		}
	}

	// Release transaction's under-modification-rpb list

	delete transaction->tra_rpblist;

	// Release the database snapshot, if any

	delete transaction->tra_db_snapshot;

	// Close all open DSQL cursors

	while (transaction->tra_open_cursors.getCount())
	{
		DSQL_free_statement(tdbb, transaction->tra_open_cursors.pop(), DSQL_close);
	}

	// Release the transaction and its pool

	jrd_tra::destroy(dbb, transaction);
}


void TRA_rollback(thread_db* tdbb, jrd_tra* transaction, const bool retaining_flag,
				  const bool force_flag)
{
/**************************************
 *
 *	T R A _ r o l l b a c k
 *
 **************************************
 *
 * Functional description
 *	Rollback a transaction.
 *
 **************************************/
	SET_TDBB(tdbb);

	TraceTransactionEnd trace(transaction, false, retaining_flag);

	EDS::Transaction::jrdTransactionEnd(tdbb, transaction, false, retaining_flag, false /*force_flag ?*/);

	Jrd::ContextPoolHolder context(tdbb, transaction->tra_pool);

	// Check in with external file system

	EXT_trans_rollback(transaction);

	if (transaction->tra_flags & (TRA_prepare2 | TRA_reconnected))
		MET_update_transaction(tdbb, transaction, false);

	// If force flag is true, get rid of all savepoints to mark the transaction as dead
	if (force_flag || transaction->tra_flags & TRA_invalidated)
	{
		// Free all savepoint data
		// We can do it in reverse order because nothing except simple deallocation
		// of memory is really done in VIO_verb_cleanup when we pass NULL as sav_next
		while (transaction->tra_save_point)
		{
			Savepoint* const next = transaction->tra_save_point->sav_next;
			transaction->tra_save_point->sav_next = NULL;
			VIO_verb_cleanup(tdbb, transaction);
			transaction->tra_save_point = next;
		}
	}
	else
	{
		VIO_temp_cleanup(tdbb, transaction);
	}

	//  Find out if there is a transaction savepoint we can use to rollback our transaction
	bool tran_sav = false;
	for (const Savepoint* temp = transaction->tra_save_point; temp; temp = temp->sav_next)
	{
		if (temp->sav_flags & SAV_trans_level)
		{
			tran_sav = true;
			break;
		}
	}

	// Measure transaction savepoint size if there is one. We'll use it for undo
	// only if it is small enough
	IPTR count = SAV_LARGE;
	if (tran_sav)
	{
		for (const Savepoint* temp = transaction->tra_save_point; temp; temp = temp->sav_next)
		{
		    count = VIO_savepoint_large(temp, count);
			if (count < 0)
				break;
		}
	}

	// We are going to use savepoint to undo transaction
	if (tran_sav && count > 0)
	{
		// Undo all user savepoints work
		while (transaction->tra_save_point->sav_flags & SAV_user)
		{
			++transaction->tra_save_point->sav_verb_count;	// cause undo
			VIO_verb_cleanup(tdbb, transaction);
		}
	}
	else
	{
		// Free all savepoint data
		// We can do it in reverse order because nothing except simple deallocation
		// of memory is really done in VIO_verb_cleanup when we pass NULL as sav_next
		while (transaction->tra_save_point && transaction->tra_save_point->sav_flags & SAV_user)
		{
			Savepoint* const next = transaction->tra_save_point->sav_next;
			transaction->tra_save_point->sav_next = NULL;
			VIO_verb_cleanup(tdbb, transaction);
			transaction->tra_save_point = next;
		}
		if (transaction->tra_save_point)
		{
			if (!(transaction->tra_save_point->sav_flags & SAV_trans_level))
				BUGCHECK(287);		// Too many savepoints
			// This transaction savepoint contains wrong data now. Clean it up
			VIO_verb_cleanup(tdbb, transaction);	// get rid of transaction savepoint
		}
	}

	SSHORT state = tra_dead;

	// Only transaction savepoint could be there
	if (transaction->tra_save_point)
	{
		if (!(transaction->tra_save_point->sav_flags & SAV_trans_level))
			BUGCHECK(287);		// Too many savepoints

		// Make sure that any error during savepoint undo is handled by marking
		// the transaction as dead.

		try {

			// In an attempt to avoid deadlocks, clear the precedence by writing
			// all dirty buffers for this transaction.

			if (transaction->tra_flags & TRA_write)
			{
				CCH_flush(tdbb, FLUSH_TRAN, transaction->tra_number);
				++transaction->tra_save_point->sav_verb_count;	// cause undo
				VIO_verb_cleanup(tdbb, transaction);
				CCH_flush(tdbb, FLUSH_TRAN, transaction->tra_number);
			}
			else
				VIO_verb_cleanup(tdbb, transaction);

			// All changes are undone, so we may mark the transaction
			// as committed
			state = tra_committed;
		}
		catch (const Firebird::Exception&)
		{
			// Prevent a bugcheck in TRA_set_state to cause a loop
			// Clear the error because the rollback will succeed.
			fb_utils::init_status(tdbb->tdbb_status_vector);
		}
	}
	else if (!(transaction->tra_flags & TRA_write))
	{
		// There were no changes within the transaction, so we may mark it
		// as committed
		state = tra_committed;
	}

	// If this is a rollback retain abort this transaction and start a new one.

	if (retaining_flag)
	{
		trace.finish(res_successful);
		retain_context(tdbb, transaction, false, state);
		return;
	}

	TRA_set_state(tdbb, transaction, transaction->tra_number, state);

	TRA_release_transaction(tdbb, transaction, &trace);
}


void TRA_set_state(thread_db* tdbb, jrd_tra* transaction, SLONG number, SSHORT state)
{
/**************************************
 *
 *	T R A _ s e t _ s t a t e
 *
 **************************************
 *
 * Functional description
 *	Set the state of a transaction in the inventory page.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* dbb = tdbb->getDatabase();
	CHECK_DBB(dbb);

	// If we're terminating ourselves and we've been precommitted then just return.

	if (transaction && transaction->tra_number == number && transaction->tra_flags & TRA_precommitted)
	{
		return;
	}

	// If it is a ReadOnly DB, set the new state in the TIP cache and return
	if ((dbb->dbb_flags & DBB_read_only) && dbb->dbb_tip_cache)
	{
		TPC_set_state(tdbb, number, state);
		return;
	}

	const ULONG trans_per_tip = dbb->dbb_page_manager.transPerTIP;
	const SLONG sequence = number / trans_per_tip;
	//trans_per_tip = dbb->dbb_page_manager.transPerTIP;
	const ULONG byte = TRANS_OFFSET(number % trans_per_tip);
	const SSHORT shift = TRANS_SHIFT(number);

	WIN window(DB_PAGE_SPACE, -1);
	tx_inv_page* tip = fetch_inventory_page(tdbb, &window, (SLONG) sequence, LCK_write);

#ifdef SUPERSERVER_V2
	CCH_MARK(tdbb, &window);
	const ULONG generation = tip->pag_generation;
#else
	CCH_MARK_MUST_WRITE(tdbb, &window);
#endif

	// set the state on the TIP page

	UCHAR* address = tip->tip_transactions + byte;
	*address &= ~(TRA_MASK << shift);
	*address |= state << shift;

	// set the new state in the TIP cache as well

	if (dbb->dbb_tip_cache)
		TPC_set_state(tdbb, number, state);

	CCH_RELEASE(tdbb, &window);

#ifdef SUPERSERVER_V2
	// Let the TIP be lazily updated for read-only queries.
	// To amortize write of TIP page for update transactions,
	// exit engine to allow other transactions to update the TIP
	// and use page generation to determine if page was written.

	if (transaction && !(transaction->tra_flags & TRA_write))
		return;

	{ //scope
		Database::Checkout dcoHolder(dbb);
		THREAD_YIELD();
	}
	tip = reinterpret_cast<tx_inv_page*>(CCH_FETCH(tdbb, &window, LCK_write, pag_transactions));
	if (generation == tip->pag_generation)
		CCH_MARK_MUST_WRITE(tdbb, &window);
	CCH_RELEASE(tdbb, &window);
#endif

}


int TRA_snapshot_state(thread_db* tdbb, const jrd_tra* trans, SLONG number)
{
/**************************************
 *
 *	T R A _ s n a p s h o t _ s t a t e
 *
 **************************************
 *
 * Functional description
 *	Get the state of a numbered transaction when a
 *	transaction started.
 *
 **************************************/

	SET_TDBB(tdbb);

	if (number && TRA_precommited(tdbb, number, number))
		return tra_precommitted;

	if (number == trans->tra_number)
		return tra_us;

	// If the transaction is older than the oldest
	// interesting transaction, it must be committed.

	if (number < trans->tra_oldest)
		return tra_committed;

	// If the transaction is the system transaction, it is considered committed.

	if (number == TRA_system_transaction)
		return tra_committed;

	// Look in the transaction cache for read committed transactions
	// fast, and the system transaction.  The system transaction can read
	// data from active transactions.

	if (trans->tra_flags & TRA_read_committed)
		return TPC_snapshot_state(tdbb, number);

	if (trans->tra_flags & TRA_system)
	{
		int state = TPC_snapshot_state(tdbb, number);
		if (state == tra_active)
			return tra_committed;

		return state;
	}

	// If the transaction is a commited sub-transction - do the easy lookup.

	if (trans->tra_commit_sub_trans && UInt32Bitmap::test(trans->tra_commit_sub_trans, number))
	{
		return tra_committed;
	}

	// If the transaction is younger than we are and we are not read committed
	// or the system transaction, the transaction must be considered active.

	if (number > trans->tra_top)
		return tra_active;

	return TRA_state(trans->tra_transactions.begin(), trans->tra_oldest, number);
}


jrd_tra* TRA_start(thread_db* tdbb, ULONG flags, SSHORT lock_timeout, Jrd::jrd_tra* outer)
{
/**************************************
 *
 *	T R A _ s t a r t
 *
 **************************************
 *
 * Functional description
 *	Start a user transaction.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* const dbb = tdbb->getDatabase();
	Attachment* const attachment = tdbb->getAttachment();

	if (dbb->dbb_ast_flags & DBB_shut_tran)
	{
		ERR_post(Arg::Gds(isc_shutinprog) << Arg::Str(attachment->att_filename));
	}

	// To handle the problems of relation locks, allocate a temporary
	// transaction block first, seize relation locks, then go ahead and
	// make up the real transaction block.
	MemoryPool* const pool = outer ? outer->getAutonomousPool() : dbb->createPool();
	Jrd::ContextPoolHolder context(tdbb, pool);
	jrd_tra* const temp = jrd_tra::create(pool, attachment, outer);

	temp->tra_flags = flags & TRA_OPTIONS_MASK;
	temp->tra_lock_timeout = lock_timeout;

	jrd_tra* transaction = NULL;
	try
	{
		transaction = transaction_start(tdbb, temp);
		delete temp;
	}
	catch (const Exception&)
	{
		jrd_tra::destroy(dbb, temp);
		throw;
	}

	if (attachment->att_trace_manager->needs().event_transaction_start)
	{
		TraceConnectionImpl conn(attachment);
		TraceTransactionImpl tran(transaction);
		attachment->att_trace_manager->event_transaction_start(&conn,
			&tran, 0, NULL, res_successful);
	}

	return transaction;
}


jrd_tra* TRA_start(thread_db* tdbb, int tpb_length, const UCHAR* tpb, Jrd::jrd_tra* outer)
{
/**************************************
 *
 *	T R A _ s t a r t
 *
 **************************************
 *
 * Functional description
 *	Start a user transaction.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* dbb = tdbb->getDatabase();
	Attachment* attachment = tdbb->getAttachment();

	if (dbb->dbb_ast_flags & DBB_shut_tran)
	{
		ERR_post(Arg::Gds(isc_shutinprog) << Arg::Str(attachment->att_filename));
	}

	// To handle the problems of relation locks, allocate a temporary
	// transaction block first, seize relation locks, then go ahead and
	// make up the real transaction block.
	MemoryPool* const pool = outer ? outer->getAutonomousPool() : dbb->createPool();
	Jrd::ContextPoolHolder context(tdbb, pool);
	jrd_tra* const temp = jrd_tra::create(pool, attachment, outer);

	jrd_tra* transaction = NULL;
	try
	{
		transaction_options(tdbb, temp, tpb, tpb_length);
		transaction = transaction_start(tdbb, temp);
		delete temp;
	}
	catch (const Exception&)
	{
		jrd_tra::destroy(dbb, temp);
		throw;
	}

	if (attachment->att_trace_manager->needs().event_transaction_start)
	{
		TraceConnectionImpl conn(attachment);
		TraceTransactionImpl tran(transaction);
		attachment->att_trace_manager->event_transaction_start(&conn,
			&tran, tpb_length, tpb, res_successful);
	}

	return transaction;
}


int TRA_state(const UCHAR* bit_vector, ULONG oldest, ULONG number)
{
/**************************************
 *
 *	T R A _ s t a t e
 *
 **************************************
 *
 * Functional description
 *	Get the state of a transaction from a cached
 *	bit vector.
 *	NOTE: This code is reproduced elsewhere in
 *	this module for speed.  If changes are made
 *	to this code make them in the replicated code also.
 *
 **************************************/
	const ULONG base = oldest & ~TRA_MASK;
	const ULONG byte = TRANS_OFFSET(number - base);
	const USHORT shift = TRANS_SHIFT(number);

	return (bit_vector[byte] >> shift) & TRA_MASK;
}


void TRA_sweep(thread_db* tdbb)
{
/**************************************
 *
 *	T R A _ s w e e p
 *
 **************************************
 *
 * Functional description
 *	Make a garbage collection pass thru database.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* const dbb = tdbb->getDatabase();
	CHECK_DBB(dbb);

	if (!dbb->allowSweepRun(tdbb))
	{
		dbb->clearSweepFlags(tdbb);
		return;
	}

	fb_assert(dbb->dbb_flags & DBB_sweep_in_progress);

	Attachment* const attachment = tdbb->getAttachment();

	jrd_tra* const tdbb_old_trans = tdbb->getTransaction();

	jrd_tra* transaction = NULL;

	// Clean up the temporary locks we've gotten in case anything goes wrong

	try {

	// Identify ourselves as a sweeper thread. This accomplishes two goals:
	// 1) Sweep transaction is started "precommitted" and
	// 2) Execution is throttled in JRD_reschedule() by
	// yielding the processor when our quantum expires.

	tdbb->tdbb_flags |= TDBB_sweeper;

	TraceSweepEvent traceSweep(tdbb);

	// Start a transaction to perform the sweep.
	// Save the transaction's oldest snapshot as it is refreshed
	// during the course of the database sweep. Since it is used
	// below to advance the OIT we must save it before it changes.

	transaction = TRA_start(tdbb, sizeof(sweep_tpb), sweep_tpb);

	SLONG transaction_oldest_active = transaction->tra_oldest_active;
	tdbb->setTransaction(transaction);

#ifdef GARBAGE_THREAD
	// The garbage collector runs asynchronously with respect to
	// our database sweep. This isn't good enough since we must
	// be absolutely certain that all dead transactions have been
	// swept from the database before advancing the OIT. Turn off
	// the "notify garbage collector" flag for the attachment and
	// synchronously perform the garbage collection ourselves.

	attachment->att_flags &= ~ATT_notify_gc;
#endif

	if (VIO_sweep(tdbb, transaction, &traceSweep))
	{
		// At this point, we know that no record versions belonging to dead
		// transactions remain anymore. However, there may still be limbo
		// transactions, so we need to find the oldest one between tra_oldest and tra_top.
		// As our transaction is read-committed (see sweep_tpb), we have to scan
		// the global TIP cache.

		const SLONG oldest_limbo =
			TPC_find_limbo(tdbb, transaction->tra_oldest, transaction->tra_top - 1);

		const SLONG active = oldest_limbo ? oldest_limbo : transaction->tra_top;

		// Flush page buffers to insure that no dangling records from
		// dead transactions are left on-disk. This must be done before
		// the OIT is advanced and the header page is written to disk.
		// If the header page was written before flushing the page buffers
		// and there was a server crash, the dead records would appear
		// committed since their TID would now be less than the OIT recorded
		// in the database.

		CCH_flush(tdbb, FLUSH_SWEEP, 0);

		WIN window(HEADER_PAGE_NUMBER);
		header_page* header = (header_page*) CCH_FETCH(tdbb, &window, LCK_write, pag_header);

		if (header->hdr_oldest_transaction < --transaction_oldest_active)
		{
			CCH_MARK_MUST_WRITE(tdbb, &window);
			header->hdr_oldest_transaction = MIN((ULONG) active, (ULONG) transaction_oldest_active);
		}

		traceSweep.update(header);

		CCH_RELEASE(tdbb, &window);

		traceSweep.finish();
	}

	TRA_commit(tdbb, transaction, false);

	tdbb->tdbb_flags &= ~TDBB_sweeper;
	tdbb->setTransaction(tdbb_old_trans);
	dbb->clearSweepFlags(tdbb);
	}	// try
	catch (const Firebird::Exception& ex)
	{
		iscLogException("Error during sweep:", ex);

		Firebird::stuff_exception(tdbb->tdbb_status_vector, ex);

		if (transaction)
		{
			try
			{
				TRA_commit(tdbb, transaction, false);
			}
			catch (const Firebird::Exception& ex2)
			{
				Firebird::stuff_exception(tdbb->tdbb_status_vector, ex2);
			}
		}

		tdbb->tdbb_flags &= ~TDBB_sweeper;
		tdbb->setTransaction(tdbb_old_trans);
		dbb->clearSweepFlags(tdbb);

		throw;
	}
}


int TRA_wait(thread_db* tdbb, jrd_tra* trans, SLONG number, jrd_tra::wait_t wait)
{
/**************************************
 *
 *	T R A _ w a i t
 *
 **************************************
 *
 * Functional description
 *	Wait for a given transaction to drop into a stable state (i.e. non-active)
 *	state.  To do this, we first wait on the transaction number.  When we
 *	are able to get the lock, the transaction is not longer bona fide
 *	active.  Next, we determine the state of the transaction from the
 *	transaction inventory page.  If either committed, dead, or limbo,
 *	we return the state.  If the transaction is still marked active,
 *	however, declare the transaction dead, and mark the transaction
 *	inventory page accordingly.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* dbb = tdbb->getDatabase();
	CHECK_DBB(dbb);

	// Create, wait on, and release lock on target transaction.  If
	// we can't get the lock due to deadlock

	if (wait != jrd_tra::tra_no_wait)
	{
		Lock temp_lock;
		temp_lock.lck_dbb = dbb;
		temp_lock.lck_type = LCK_tra;
		temp_lock.lck_owner_handle = LCK_get_owner_handle(tdbb, temp_lock.lck_type);
		temp_lock.lck_parent = dbb->dbb_lock;
		temp_lock.lck_length = sizeof(SLONG);
		temp_lock.lck_key.lck_long = number;

		const SSHORT timeout = (wait == jrd_tra::tra_wait) ? trans->getLockWait() : 0;

		if (!LCK_lock(tdbb, &temp_lock, LCK_read, timeout))
			return tra_active;

		LCK_release(tdbb, &temp_lock);
	}

	USHORT state = TRA_get_state(tdbb, number);

	if (wait != jrd_tra::tra_no_wait && state == tra_committed)
		return state;

	if (state == tra_precommitted)
		return state;

	// If the recorded state of the transaction is active, we know better.  If
	// it were active, he'd be alive now.  Mark him dead.

	if (state == tra_active)
	{
		state = tra_dead;
		TRA_set_state(tdbb, 0, number, tra_dead);
	}

	if (number > trans->tra_top)
		return state;

	// If the transaction disppeared into limbo, died, for constructively
	// died, tweak the transaction state snapshot to reflect the new state.
	// This is guarenteed safe.

	const ULONG byte = TRANS_OFFSET(number - (trans->tra_oldest & ~TRA_MASK));
	const USHORT shift = TRANS_SHIFT(number);

	if (trans->tra_flags & TRA_read_committed)
		TPC_set_state(tdbb, number, state);
	else
	{
		trans->tra_transactions[byte] &= ~(TRA_MASK << shift);
		trans->tra_transactions[byte] |= state << shift;
	}

	return state;
}


#ifdef SUPERSERVER_V2
static SLONG bump_transaction_id(thread_db* tdbb, WIN* window)
{
/**************************************
 *
 *	b u m p _ t r a n s a c t i o n _ i d
 *
 **************************************
 *
 * Functional description
 *	Fetch header and bump next transaction id.  If necessary,
 *	extend TIP.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* dbb = tdbb->getDatabase();
	CHECK_DBB(dbb);

	if (dbb->dbb_next_transaction >= MAX_TRA_NUMBER - 1)
	{
		CCH_RELEASE(tdbb, window);
		ERR_post(Arg::Gds(isc_imp_exc) <<
				 Arg::Gds(isc_tra_num_exc));
	}
	const SLONG number = ++dbb->dbb_next_transaction;

	// No need to write TID onto the TIP page, for a RO DB
	if (dbb->dbb_flags & DBB_read_only)
		return number;

	// If this is the first transaction on a TIP, allocate the TIP now.

	const bool new_tip = (number == 1 || (number % dbb->dbb_page_manager.transPerTIP) == 0);

	if (new_tip) {
		TRA_extend_tip(tdbb, (ULONG) (number / dbb->dbb_page_manager.transPerTIP)); //, window);
	}

	return number;
}
#else


static header_page* bump_transaction_id(thread_db* tdbb, WIN* window)
{
/**************************************
 *
 *	b u m p _ t r a n s a c t i o n _ i d
 *
 **************************************
 *
 * Functional description
 *	Fetch header and bump next transaction id.  If necessary,
 *	extend TIP.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* dbb = tdbb->getDatabase();
	CHECK_DBB(dbb);

	window->win_page = HEADER_PAGE_NUMBER;
	header_page* header = (header_page*) CCH_FETCH(tdbb, window, LCK_write, pag_header);

	// Before incrementing the next transaction Id, make sure the current one is valid
	if (header->hdr_next_transaction)
	{
		if (header->hdr_oldest_active > header->hdr_next_transaction)
			BUGCHECK(266);		//next transaction older than oldest active

		if (header->hdr_oldest_transaction > header->hdr_next_transaction)
			BUGCHECK(267);		// next transaction older than oldest transaction
	}

	if (header->hdr_next_transaction >= MAX_TRA_NUMBER - 1)
	{
		CCH_RELEASE(tdbb, window);
		ERR_post(Arg::Gds(isc_imp_exc) <<
				 Arg::Gds(isc_tra_num_exc));
	}
	const SLONG number = header->hdr_next_transaction + 1;

	// If this is the first transaction on a TIP, allocate the TIP now.

	const bool new_tip = (number == 1 || (number % dbb->dbb_page_manager.transPerTIP) == 0);

	if (new_tip) {
		TRA_extend_tip(tdbb, (ULONG) (number / dbb->dbb_page_manager.transPerTIP)); //, window);
	}

	// Extend, if necessary, has apparently succeeded.  Next, update header page

	CCH_MARK_MUST_WRITE(tdbb, window);
	header->hdr_next_transaction = number;

	if (dbb->dbb_oldest_active > header->hdr_oldest_active)
		header->hdr_oldest_active = dbb->dbb_oldest_active;

	if (dbb->dbb_oldest_transaction > header->hdr_oldest_transaction)
		header->hdr_oldest_transaction = dbb->dbb_oldest_transaction;

	if (dbb->dbb_oldest_snapshot > header->hdr_oldest_snapshot)
		header->hdr_oldest_snapshot = dbb->dbb_oldest_snapshot;

	return header;
}
#endif


static Lock* create_transaction_lock(thread_db* tdbb, void* object)
{
/**************************************
 *
 *	c r e a t e _ t r a n s a c t i o n _ l o c k
 *
 **************************************
 *
 * Functional description
 *	Allocate a transaction lock block.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* dbb = tdbb->getDatabase();

	Lock* lock = FB_NEW_RPT(*tdbb->getDefaultPool(), sizeof(SLONG)) Lock();
	lock->lck_type = LCK_tra;
	lock->lck_owner_handle = LCK_get_owner_handle(tdbb, lock->lck_type);
	lock->lck_length = sizeof(SLONG);

	lock->lck_dbb = dbb;
	lock->lck_parent = dbb->dbb_lock;
	lock->lck_object = object;

	return lock;
}


#ifdef VMS
static void compute_oldest_retaining(thread_db* tdbb, jrd_tra* transaction, const bool write_flag)
{
/**************************************
 *
 *	c o m p u t e _ o l d e s t _ r e t a i n i n g
 *
 **************************************
 *
 * Functional description
 *	Read the oldest active for all transactions
 *	younger than us up to the youngest retaining
 *	transaction. If an "older" oldest active is
 *	found, by all means use it. Write flag is true
 *	to write retaining lock and false to read it.
 *	The retaining lock holds the youngest commit
 *	retaining transaction.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* dbb = tdbb->getDatabase();
	CHECK_DBB(dbb);

	// Get a commit retaining lock, if not present.

	Lock* lock = dbb->dbb_retaining_lock;
	if (!lock)
	{
		lock = FB_NEW_RPT(*dbb->dbb_permanent, sizeof(SLONG)) Lock();
		lock->lck_dbb = dbb;
		lock->lck_type = LCK_retaining;
		lock->lck_owner_handle = LCK_get_owner_handle(tdbb, lock->lck_type);
		lock->lck_parent = dbb->dbb_lock;
		lock->lck_length = sizeof(SLONG);
		lock->lck_object = dbb;
		LCK_lock(tdbb, lock, LCK_SR, LCK_WAIT);
		dbb->dbb_retaining_lock = lock;
	}

	SLONG number = transaction->tra_number;

	// Writers must synchronize their lock update so that
	// an older retaining is not written over a younger retaining.
	// In any case, lock types have been selected so that
	// readers and writers don't interfere.

	SLONG youngest_retaining;

	if (write_flag)
	{
		LCK_convert(tdbb, lock, LCK_PW, LCK_WAIT);
		youngest_retaining = LOCK_read_data(lock->lck_id);
		if (number > youngest_retaining)
			LCK_write_data(lock, number);
		LCK_convert(tdbb, lock, LCK_SR, LCK_WAIT);
	}
	else
	{
		youngest_retaining = LOCK_read_data(lock->lck_id);
		if (number > youngest_retaining)
			return;

		// fill out a lock block, zeroing it out first
		Lock temp_lock;
		temp_lock.lck_dbb = dbb;
		temp_lock.lck_type = LCK_tra;
		temp_lock.lck_owner_handle = LCK_get_owner_handle(tdbb, temp_lock.lck_type);
		temp_lock.lck_parent = dbb->dbb_lock;
		temp_lock.lck_length = sizeof(SLONG);
		temp_lock.lck_object = transaction;

		while (number < youngest_retaining)
		{
			temp_lock.lck_key.lck_long = ++number;
			const SLONG data = LCK_read_data(&temp_lock);
			if (data && data < transaction->tra_oldest_active)
				transaction->tra_oldest_active = data;
		}
	}
}
#endif


static void expand_view_lock(thread_db* tdbb, jrd_tra* transaction, jrd_rel* relation,
	UCHAR lock_type, const char* option_name, RelationLockTypeMap& lockmap, const int level)
{
/**************************************
 *
 *	e x p a n d _ v i e w _ l o c k
 *
 **************************************
 *
 * Functional description
 *	A view in a RESERVING will lead to all tables in the
 *	view being locked.
 *  Some checks only apply when the user reserved directly the table or view.
 *
 **************************************/
	SET_TDBB(tdbb);

	if (level == 30)
	{
		ERR_post(Arg::Gds(isc_bad_tpb_content) <<
				 Arg::Gds(isc_tpb_reserv_max_recursion) << Arg::Num(30));
	}

	const char* const relation_name = relation->rel_name.c_str();

	// LCK_none < LCK_SR < LCK_PR < LCK_SW < LCK_EX
	UCHAR oldlock;
	const bool found = lockmap.get(relation->rel_id, oldlock);

	if (found && oldlock > lock_type)
	{
		const char* newname = get_lockname_v3(lock_type);
		const char* oldname = get_lockname_v3(oldlock);

		if (level)
		{
			lock_type = oldlock; // Preserve the old, more powerful lock.
			ERR_post_warning(Arg::Warning(isc_tpb_reserv_stronger_wng) << Arg::Str(relation_name) <<
																		  Arg::Str(oldname) <<
																		  Arg::Str(newname));
		}
		else
		{
			ERR_post(Arg::Gds(isc_bad_tpb_content) <<
					 Arg::Gds(isc_tpb_reserv_stronger) << Arg::Str(relation_name) <<
														  Arg::Str(oldname) <<
														  Arg::Str(newname));
		}
	}

	if (level == 0)
	{
		fb_assert(!relation->rel_view_rse && !relation->rel_view_contexts.getCount());
		// Reject explicit attempts to take locks on virtual tables.
		if (relation->isVirtual())
		{
			ERR_post(Arg::Gds(isc_bad_tpb_content) <<
					 Arg::Gds(isc_tpb_reserv_virtualtbl) << Arg::Str(relation_name));
		}

		// Reject explicit attempts to take locks on system tables, but RDB$ADMIN role
		// can do that for whatever is needed.
		if (relation->isSystem() && !tdbb->getAttachment()->locksmith())
		{
			ERR_post(Arg::Gds(isc_bad_tpb_content) <<
		    		 Arg::Gds(isc_tpb_reserv_systbl) << Arg::Str(relation_name));
		}

		if (relation->isTemporary() && (lock_type == LCK_PR || lock_type == LCK_EX))
		{
			ERR_post(Arg::Gds(isc_bad_tpb_content) <<
					 Arg::Gds(isc_tpb_reserv_temptbl) << Arg::Str(get_lockname_v3(LCK_PR)) <<
					 									 Arg::Str(get_lockname_v3(LCK_EX)) <<
														 Arg::Str(relation_name));
		}
	}
	else
	{
		fb_assert(relation->rel_view_rse && relation->rel_view_contexts.getCount());
		// Ignore implicit attempts to take locks on special tables through views.
		if (relation->isVirtual() || relation->isSystem())
			return;

		// We can't propagate a view's LCK_PR or LCK_EX to a temporary table.
		if (relation->isTemporary())
		{
			switch (lock_type)
			{
			case LCK_PR:
				lock_type = LCK_SR;
				break;
			case LCK_EX:
				lock_type = LCK_SW;
				break;
			}
		}
	}

	// set up the lock on the relation/view
	Lock* lock = RLCK_transaction_relation_lock(tdbb, transaction, relation);
	lock->lck_logical = lock_type;

	if (!found)
		*lockmap.put(relation->rel_id) = lock_type;

	const ViewContexts& ctx = relation->rel_view_contexts;

	for (size_t i = 0; i < ctx.getCount(); ++i)
	{
		if (!ctx[i]->vcx_is_relation)
			continue;

		jrd_rel* base_rel = MET_lookup_relation(tdbb, ctx[i]->vcx_relation_name);
		if (!base_rel)
		{
			// should be a BUGCHECK
			ERR_post(Arg::Gds(isc_bad_tpb_content) <<
					 Arg::Gds(isc_tpb_reserv_baserelnotfound) << Arg::Str(ctx[i]->vcx_relation_name) <<
																 Arg::Str(relation_name) <<
																 Arg::Str(option_name));
		}

		// force a scan to read view information
		MET_scan_relation(tdbb, base_rel);

		expand_view_lock(tdbb, transaction, base_rel, lock_type, option_name, lockmap, level + 1);
	}
}


static tx_inv_page* fetch_inventory_page(thread_db* tdbb,
										 WIN* window,
										 SLONG sequence,
										 USHORT lock_level)
{
/**************************************
 *
 *	f e t c h _ i n v e n t o r y _ p a g e
 *
 **************************************
 *
 * Functional description
 *	Fetch a transaction inventory page.
 *	Use the opportunity to cache the info
 *	in the TIP cache.
 *
 **************************************/
	SET_TDBB(tdbb);

	window->win_page = inventory_page(tdbb, sequence);
	tx_inv_page* tip = (tx_inv_page*) CCH_FETCH(tdbb, window, lock_level, pag_transactions);

	TPC_update_cache(tdbb, tip, sequence);

	return tip;
}


static const char* get_lockname_v3(const UCHAR lock)
{
/**************************************
 *
 * g e t _ l o c k n a m e _ v 3
 *
 **************************************
 *
 * Functional description
 *	Get the lock mnemonic, given its binary value.
 *	This is for TPB versions 1 & 3.
 *
 **************************************/
	const char* typestr = "unknown";
	switch (lock)
	{
	case LCK_none:
	case LCK_SR:
		typestr = "isc_tpb_lock_read, isc_tpb_shared";
		break;
	case LCK_PR:
		typestr = "isc_tpb_lock_read, isc_tpb_protected/isc_tpb_exclusive";
		break;
	case LCK_SW:
		typestr = "isc_tpb_lock_write, isc_tpb_shared";
		break;
	case LCK_EX:
		typestr = "isc_tpb_lock_write, isc_tpb_protected/isc_tpb_exclusive";
		break;
	}
	return typestr;
}


static SLONG inventory_page(thread_db* tdbb, SLONG sequence)
{
/**************************************
 *
 *	i n v e n t o r y _ p a g e
 *
 **************************************
 *
 * Functional description
 *	Get the physical page number of the n-th transaction inventory
 *	page. If not found, try to reconstruct using sibling pointer
 *	from last known TIP page.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* dbb = tdbb->getDatabase();
	CHECK_DBB(dbb);

	WIN window(DB_PAGE_SPACE, -1);
	vcl* vector = dbb->dbb_t_pages;
	while (!vector || sequence >= (SLONG) vector->count())
	{
		DPM_scan_pages(tdbb);
		if ((vector = dbb->dbb_t_pages) && sequence < (SLONG) vector->count())
			break;
		if (!vector)
			BUGCHECK(165);		// msg 165 cannot find tip page
		window.win_page = (*vector)[vector->count() - 1];
		tx_inv_page* tip = (tx_inv_page*) CCH_FETCH(tdbb, &window, LCK_read, pag_transactions);
		const SLONG next = tip->tip_next;
		CCH_RELEASE(tdbb, &window);
		if (!(window.win_page = next))
			BUGCHECK(165);		// msg 165 cannot find tip page
		// Type check it
		tip = (tx_inv_page*) CCH_FETCH(tdbb, &window, LCK_read, pag_transactions);
		CCH_RELEASE(tdbb, &window);
		DPM_pages(tdbb, 0, pag_transactions, vector->count(), window.win_page.getPageNum());
	}

	return (*vector)[sequence];
}


static SSHORT limbo_transaction(thread_db* tdbb, SLONG id)
{
/**************************************
 *
 *	l i m b o _ t r a n s a c t i o n
 *
 **************************************
 *
 * Functional description
 *
 *	limbo_state is called when reconnecting
 *	to an existing transaction to assure that
 *	the transaction is actually in limbo.
 *	It returns the transaction state.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* dbb = tdbb->getDatabase();
	CHECK_DBB(dbb);

	const SLONG trans_per_tip = dbb->dbb_page_manager.transPerTIP;

	const SLONG page = id / trans_per_tip;
	const SLONG number = id % trans_per_tip;

	WIN window(DB_PAGE_SPACE, -1);
	const tx_inv_page* tip = fetch_inventory_page(tdbb, &window, page, LCK_write);

	const SLONG trans_offset = TRANS_OFFSET(number);
	const UCHAR* byte = tip->tip_transactions + trans_offset;
	const SSHORT shift = TRANS_SHIFT(number);
	const SSHORT state = (*byte >> shift) & TRA_MASK;
	CCH_RELEASE(tdbb, &window);

	return state;
}


static void link_transaction(thread_db* tdbb, jrd_tra* transaction)
{
/**************************************
 *
 *	l i n k _ t r a n s a c t i o n
 *
 **************************************
 *
 * Functional description
 *	Link transaction block into database attachment.
 *
 **************************************/
	SET_TDBB(tdbb);

	Attachment* attachment = tdbb->getAttachment();
	transaction->tra_next = attachment->att_transactions;
	attachment->att_transactions = transaction;
}


static void restart_requests(thread_db* tdbb, jrd_tra* trans)
{
/**************************************
 *
 *	r e s t a r t _ r e q u e s t s
 *
 **************************************
 *
 * Functional description
 *	Restart all requests in the current
 *	attachment to utilize the passed
 *	transaction.
 *
 **************************************/
	SET_TDBB(tdbb);
	for (jrd_req* request = trans->tra_attachment->att_requests; request;
		request = request->req_request)
	{
		if (request->req_transaction)
		{
			EXE_unwind(tdbb, request);
			EXE_start(tdbb, request, trans);
		}

		// now take care of any other request levels;
		// start at level 1 since level 0 was just handled

		vec<jrd_req*>* vector = request->req_sub_requests;
		if (vector)
		{
			for (USHORT level = 1; level < vector->count(); level++)
			{
				jrd_req* clone = (*vector)[level];
				if (clone && clone->req_transaction)
				{
					EXE_unwind(tdbb, clone);
					EXE_start(tdbb, clone, trans);
				}
			}
		}
	}
}


static void retain_context(thread_db* tdbb, jrd_tra* transaction, bool commit, SSHORT state)
{
/**************************************
 *
 *	r e t a i n _ c o n t e x t
 *
 **************************************
 *
 * Functional description
 *	If 'commit' flag is true, commit the transaction,
 *	else rollback the transaction.
 *
 *	Commit/rollback a transaction while preserving the
 *	context, in particular, its snapshot. The
 *	trick is to insure that the transaction's
 *	oldest active is seen by other transactions
 *	simultaneously starting up.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* dbb = tdbb->getDatabase();
	CHECK_DBB(dbb);

	// The new transaction needs to remember the 'commit-retained' transaction
	// because it must see the operations of the 'commit-retained' transaction and
	// its snapshot doesn't contain these operations.

	if (commit) {
		SBM_SET(tdbb->getDefaultPool(), &transaction->tra_commit_sub_trans, transaction->tra_number);
	}

	// Create a new transaction lock, inheriting oldest active from transaction being committed.

	WIN window(DB_PAGE_SPACE, -1);
	SLONG new_number;
#ifdef SUPERSERVER_V2
	new_number = bump_transaction_id(tdbb, &window);
#else
	if (dbb->dbb_flags & DBB_read_only)
		new_number = dbb->dbb_next_transaction + dbb->generateTransactionId(tdbb);
	else
	{
		const header_page* header = bump_transaction_id(tdbb, &window);
		new_number = header->hdr_next_transaction;
	}
#endif

	Lock* new_lock = 0;
	Lock* old_lock = transaction->tra_lock;
	if (old_lock)
	{
		new_lock = create_transaction_lock(tdbb, transaction);
		new_lock->lck_key.lck_long = new_number;
		new_lock->lck_data = transaction->tra_lock->lck_data;

		if (!LCK_lock(tdbb, new_lock, LCK_write, LCK_WAIT))
		{
#ifndef SUPERSERVER_V2
			if (!(dbb->dbb_flags & DBB_read_only))
				CCH_RELEASE(tdbb, &window);
#endif
			ERR_post(Arg::Gds(isc_lock_conflict));
		}
	}

#ifndef SUPERSERVER_V2
	if (!(dbb->dbb_flags & DBB_read_only))
		CCH_RELEASE(tdbb, &window);
#endif

	// Update database notion of the youngest commit retaining
	// transaction before committing the first transaction. This
	// secures the original snapshot by insuring the oldest active
	// is seen by other transactions.

	const SLONG old_number = transaction->tra_number;
#ifdef VMS
	transaction->tra_number = new_number;
	compute_oldest_retaining(tdbb, transaction, true);
	transaction->tra_number = old_number;
#endif

	if (!(dbb->dbb_flags & DBB_read_only))
	{
		// Set the state on the inventory page
		TRA_set_state(tdbb, transaction, old_number, state);
	}
	transaction->tra_number = new_number;

	// Release transaction lock since it isn't needed
	// anymore and the new one is already in place.

	if (old_lock)
	{
		++transaction->tra_use_count;
		LCK_release(tdbb, old_lock);
		transaction->tra_lock = new_lock;
		--transaction->tra_use_count;
		delete old_lock;
	}

	// Perform any post commit work OR delete entries from deferred list

	if (commit)
		DFW_perform_post_commit_work(transaction);
	else
		DFW_delete_deferred(transaction, -1);

	transaction->tra_flags &= ~(TRA_write | TRA_prepared);

	// We have to mimic a TRA_commit and a TRA_start while reusing the
	// 'transaction' control block: get rid of the transaction-level
	// savepoint and possibly start a new transaction-level savepoint.

	// Get rid of all user savepoints
	// Why we can do this in reverse order described in commit method
	while (transaction->tra_save_point && transaction->tra_save_point->sav_flags & SAV_user)
	{
		Savepoint* const next = transaction->tra_save_point->sav_next;
		transaction->tra_save_point->sav_next = NULL;
		VIO_verb_cleanup(tdbb, transaction);
		transaction->tra_save_point = next;
	}

	if (transaction->tra_save_point)
	{
		if (!(transaction->tra_save_point->sav_flags & SAV_trans_level))
			BUGCHECK(287);		// Too many savepoints
		VIO_verb_cleanup(tdbb, transaction);	// get rid of transaction savepoint
		VIO_start_save_point(tdbb, transaction);	// start new savepoint
		transaction->tra_save_point->sav_flags |= SAV_trans_level;
	}

	if (transaction->tra_flags & TRA_precommitted)
	{
		if (!(dbb->dbb_flags & DBB_read_only))
		{
			transaction->tra_flags &= ~TRA_precommitted;
			TRA_set_state(tdbb, transaction, new_number, tra_committed);
			transaction->tra_flags |= TRA_precommitted;
		}

		TRA_precommited(tdbb, old_number, new_number);
	}
}


static void start_sweeper(thread_db* tdbb)
{
/**************************************
 *
 *	s t a r t _ s w e e p e r
 *
 **************************************
 *
 * Functional description
 *	Start a thread to sweep the database.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* const dbb = tdbb->getDatabase();

	if (!dbb->allowSweepThread(tdbb))
		return;

	// allocate space for the string and a null at the end
	const char* pszFilename = tdbb->getAttachment()->att_filename.c_str();

	char* database = (char*) gds__alloc(strlen(pszFilename) + 1);

	if (database)
	{
		strcpy(database, pszFilename);
		if (!gds__thread_start(sweep_database, database, THREAD_medium, 0, 0))
			return;

		gds__free(database);
		ERR_log(0, 0, "cannot start sweep thread");
	}
	else
	{
		ERR_log(0, 0, "cannot start sweep thread, Out of Memory");
	}
	dbb->clearSweepFlags(tdbb);
}


static THREAD_ENTRY_DECLARE sweep_database(THREAD_ENTRY_PARAM database)
{
/**************************************
 *
 *	s w e e p _ d a t a b a s e
 *
 **************************************
 *
 * Functional description
 *	Sweep database.
 *
 **************************************/
	Firebird::ClumpletWriter dpb(Firebird::ClumpletReader::Tagged, MAX_DPB_SIZE, isc_dpb_version1);

	dpb.insertByte(isc_dpb_sweep, isc_dpb_records);
	// sometimes security database is also to be swept
	dpb.insertByte(isc_dpb_gsec_attach, 1);
	// use trusted authentication to attach database
	const char* szAuthenticator = "sweeper";
	dpb.insertString(isc_dpb_trusted_auth, szAuthenticator, strlen(szAuthenticator));

	ISC_STATUS_ARRAY status_vector = {0};
	isc_db_handle db_handle = 0;

	isc_attach_database(status_vector, 0, (const char*) database,
						&db_handle, dpb.getBufferLength(),
						reinterpret_cast<const char*>(dpb.getBuffer()));

	if (db_handle)
	{
		isc_detach_database(status_vector, &db_handle);
	}

	gds__free(database);
	return 0;
}


static void transaction_options(thread_db* tdbb,
								jrd_tra* transaction,
								const UCHAR* tpb, USHORT tpb_length)
{
/**************************************
 *
 *	t r a n s a c t i o n _ o p t i o n s
 *
 **************************************
 *
 * Functional description
 *	Process transaction options.
 *
 **************************************/
	SET_TDBB(tdbb);

	if (!tpb_length)
		return;

	const UCHAR* const end = tpb + tpb_length;

	if (*tpb != isc_tpb_version3 && *tpb != isc_tpb_version1)
		ERR_post(Arg::Gds(isc_bad_tpb_form) <<
				 Arg::Gds(isc_wrotpbver));

	RelationLockTypeMap	lockmap;

	TriState wait, lock_timeout;
	TriState isolation, read_only, rec_version;
	bool anylock_write = false;

	++tpb;

	while (tpb < end)
	{
		const USHORT op = *tpb++;
		switch (op)
		{
		case isc_tpb_consistency:
			if (!isolation.assignOnce(true))
				ERR_post(Arg::Gds(isc_bad_tpb_content) <<
						 Arg::Gds(isc_tpb_multiple_txn_isolation));

			transaction->tra_flags |= TRA_degree3;
			transaction->tra_flags &= ~TRA_read_committed;
			break;

		case isc_tpb_concurrency:
			if (!isolation.assignOnce(true))
				ERR_post(Arg::Gds(isc_bad_tpb_content) <<
						 Arg::Gds(isc_tpb_multiple_txn_isolation));

			transaction->tra_flags &= ~TRA_degree3;
			transaction->tra_flags &= ~TRA_read_committed;
			break;

		case isc_tpb_read_committed:
			if (!isolation.assignOnce(true))
				ERR_post(Arg::Gds(isc_bad_tpb_content) <<
						 Arg::Gds(isc_tpb_multiple_txn_isolation));

			transaction->tra_flags &= ~TRA_degree3;
			transaction->tra_flags |= TRA_read_committed;
			break;

		case isc_tpb_shared:
			ERR_post(Arg::Gds(isc_bad_tpb_content) <<
					 Arg::Gds(isc_tpb_reserv_before_table) << Arg::Str("isc_tpb_shared"));
			break;

		case isc_tpb_protected:
			ERR_post(Arg::Gds(isc_bad_tpb_content) <<
					 Arg::Gds(isc_tpb_reserv_before_table) << Arg::Str("isc_tpb_protected"));
			break;

		case isc_tpb_exclusive:
			ERR_post(Arg::Gds(isc_bad_tpb_content) <<
					 Arg::Gds(isc_tpb_reserv_before_table) << Arg::Str("isc_tpb_exclusive"));
			break;

		case isc_tpb_wait:
			if (!wait.assignOnce(true))
			{
				if (!wait.asBool())
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_conflicting_options) << Arg::Str("isc_tpb_wait") <<
																	  Arg::Str("isc_tpb_nowait"));
				}
				else
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_multiple_spec) << Arg::Str("isc_tpb_wait"));
				}
			}
			break;

		case isc_tpb_rec_version:
			if (isolation.isAssigned() && !(transaction->tra_flags & TRA_read_committed))
			{
				ERR_post(Arg::Gds(isc_bad_tpb_content) <<
						 Arg::Gds(isc_tpb_option_without_rc) << Arg::Str("isc_tpb_rec_version"));
			}

			if (!rec_version.assignOnce(true))
			{
				ERR_post(Arg::Gds(isc_bad_tpb_content) <<
						 Arg::Gds(isc_tpb_multiple_spec) << Arg::Str("isc_tpb_rec_version"));
			}

			transaction->tra_flags |= TRA_rec_version;
			break;

		case isc_tpb_no_rec_version:
			if (isolation.isAssigned() && !(transaction->tra_flags & TRA_read_committed))
			{
				ERR_post(Arg::Gds(isc_bad_tpb_content) <<
						 Arg::Gds(isc_tpb_option_without_rc) << Arg::Str("isc_tpb_no_rec_version"));
			}

			if (!rec_version.assignOnce(false))
			{
				ERR_post(Arg::Gds(isc_bad_tpb_content) <<
						 Arg::Gds(isc_tpb_multiple_spec) << Arg::Str("isc_tpb_no_rec_version"));
			}

			transaction->tra_flags &= ~TRA_rec_version;
			break;

		case isc_tpb_nowait:
			if (lock_timeout.asBool())
			{
				ERR_post(Arg::Gds(isc_bad_tpb_content) <<
						 Arg::Gds(isc_tpb_conflicting_options) << Arg::Str("isc_tpb_nowait") <<
																  Arg::Str("isc_tpb_lock_timeout"));
			}

			if (!wait.assignOnce(false))
			{
				if (wait.asBool())
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_conflicting_options) << Arg::Str("isc_tpb_nowait") <<
																	  Arg::Str("isc_tpb_wait"));
				}
				else
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_multiple_spec) << Arg::Str("isc_tpb_nowait"));
				}
			}

			transaction->tra_lock_timeout = 0;
			break;

		case isc_tpb_read:
			if (!read_only.assignOnce(true))
			{
				if (!read_only.asBool())
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_conflicting_options) << Arg::Str("isc_tpb_read") <<
																	  Arg::Str("isc_tpb_write"));
				}
				else
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_multiple_spec) << Arg::Str("isc_tpb_read"));
				}
			}

			// Cannot set the whole txn to R/O if we already saw a R/W table reservation.
			if (anylock_write)
			{
				ERR_post(Arg::Gds(isc_bad_tpb_content) <<
						 Arg::Gds(isc_tpb_readtxn_after_writelock));
			}

			transaction->tra_flags |= TRA_readonly;
			break;

		case isc_tpb_write:
			if (!read_only.assignOnce(false))
			{
				if (read_only.asBool())
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_conflicting_options) << Arg::Str("isc_tpb_write") <<
																	  Arg::Str("isc_tpb_read"));
				}
				else
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_multiple_spec) << Arg::Str("isc_tpb_write"));
				}
			}

			transaction->tra_flags &= ~TRA_readonly;
			break;

		case isc_tpb_ignore_limbo:
			transaction->tra_flags |= TRA_ignore_limbo;
			break;

		case isc_tpb_no_auto_undo:
			transaction->tra_flags |= TRA_no_auto_undo;
			break;

		case isc_tpb_lock_write:
			// Cannot set a R/W table reservation if the whole txn is R/O.
			if (read_only.asBool())
			{
				ERR_post(Arg::Gds(isc_bad_tpb_content) <<
						 Arg::Gds(isc_tpb_writelock_after_readtxn));
			}
			anylock_write = true;
			// fall into
		case isc_tpb_lock_read:
			{
				const char* option_name = (op == isc_tpb_lock_read) ?
					"isc_tpb_lock_read" : "isc_tpb_lock_write";

				// Do we have space for the identifier length?
				if (tpb >= end)
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_reserv_missing_tlen) << Arg::Str(option_name));
				}

				const USHORT len = *tpb++;
				if (len > MAX_SQL_IDENTIFIER_LEN)
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_reserv_long_tlen) << Arg::Num(len) <<
																   Arg::Str(option_name));
				}

				if (!len)
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_reserv_null_tlen) << Arg::Str(option_name));
				}

				// Does the identifier length surpasses the remaining of the TPB?
				if (tpb >= end)
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_reserv_missing_tname) << Arg::Num(len) <<
							 										   Arg::Str(option_name));
				}

				if (end - tpb < len)
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_reserv_corrup_tlen) << Arg::Num(len) <<
							 										 Arg::Str(option_name));
				}

				const Firebird::MetaName name(reinterpret_cast<const char*>(tpb), len);
				tpb += len;
				jrd_rel* relation = MET_lookup_relation(tdbb, name);
				if (!relation)
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_reserv_relnotfound) << Arg::Str(name) <<
																	 Arg::Str(option_name));
				}

				// force a scan to read view information
				MET_scan_relation(tdbb, relation);

				UCHAR lock_type = (op == isc_tpb_lock_read) ? LCK_none : LCK_SW;
				if (tpb < end)
				{
					switch (*tpb)
					{
					case isc_tpb_shared:
						++tpb;
						break;
					case isc_tpb_protected:
					case isc_tpb_exclusive:
						++tpb;
						lock_type = (lock_type == LCK_SW) ? LCK_EX : LCK_PR;
						break;
					// We'll assume table reservation doesn't make the concurrency type mandatory.
					//default:
					//    ERR_post(isc-arg-end);
					}
				}

				expand_view_lock(tdbb, transaction, relation, lock_type, option_name, lockmap, 0);
			}
			break;

		case isc_tpb_verb_time:
		case isc_tpb_commit_time:
			{
				const char* option_name = (op == isc_tpb_verb_time) ?
					"isc_tpb_verb_time" : "isc_tpb_commit_time";
				// Harmless for now even if formally invalid.
				if (tpb >= end)
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_missing_len) << Arg::Str(option_name));
				}

				const USHORT len = *tpb++;

				if (tpb >= end && len > 0)
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_missing_value) << Arg::Num(len) << Arg::Str(option_name));
				}

				if (end - tpb < len)
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_corrupt_len) << Arg::Num(len) << Arg::Str(option_name));
				}

				tpb += len;
			}
			break;

		case isc_tpb_autocommit:
			transaction->tra_flags |= TRA_autocommit;
			break;

		case isc_tpb_restart_requests:
			transaction->tra_flags |= TRA_restart_requests;
			break;

		case isc_tpb_lock_timeout:
			{
				if (wait.isAssigned() && !wait.asBool())
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_conflicting_options) << Arg::Str("isc_tpb_lock_timeout") <<
																	  Arg::Str("isc_tpb_nowait"));
				}

				if (!lock_timeout.assignOnce(true))
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_multiple_spec) << Arg::Str("isc_tpb_lock_timeout"));
				}

				// Do we have space for the identifier length?
				if (tpb >= end)
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_missing_len) << Arg::Str("isc_tpb_lock_timeout"));
				}

				const USHORT len = *tpb++;

				// Does the encoded number's length surpasses the remaining of the TPB?
				if (tpb >= end)
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_missing_value) << Arg::Num(len) <<
																Arg::Str("isc_tpb_lock_timeout"));
				}

				if (end - tpb < len)
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_corrupt_len) << Arg::Num(len) <<
															  Arg::Str("isc_tpb_lock_timeout"));
				}

				if (!len)
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_null_len) << Arg::Str("isc_tpb_lock_timeout"));
				}

				if (len > sizeof(ULONG))
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_overflow_len) << Arg::Num(len) << Arg::Str("isc_tpb_lock_timeout"));
				}

				const SLONG value = gds__vax_integer(tpb, len);

				if (value <= 0 || value > MAX_SSHORT)
				{
					ERR_post(Arg::Gds(isc_bad_tpb_content) <<
							 Arg::Gds(isc_tpb_invalid_value) << Arg::Num(value) << Arg::Str("isc_tpb_lock_timeout"));
				}

				transaction->tra_lock_timeout = (SSHORT) value;

				tpb += len;
			}
			break;

		default:
			ERR_post(Arg::Gds(isc_bad_tpb_form));
		}
	}

	if (rec_version.isAssigned() && !(transaction->tra_flags & TRA_read_committed))
	{
		if (rec_version.asBool())
		{
			ERR_post(Arg::Gds(isc_bad_tpb_content) <<
					 Arg::Gds(isc_tpb_option_without_rc) << Arg::Str("isc_tpb_rec_version"));
		}
		else
		{
			ERR_post(Arg::Gds(isc_bad_tpb_content) <<
					 Arg::Gds(isc_tpb_option_without_rc) << Arg::Str("isc_tpb_no_rec_version"));
		}
	}


	// If there aren't any relation locks to seize, we're done.

	vec<Lock*>* vector = transaction->tra_relation_locks;
	if (!vector)
		return;

	// Try to seize all relation locks. If any can't be seized, release all and try again.

	for (ULONG id = 0; id < vector->count(); id++)
	{
		Lock* lock = (*vector)[id];
		if (!lock)
			continue;
		USHORT level = lock->lck_logical;
		if (level == LCK_none || LCK_lock(tdbb, lock, level, transaction->getLockWait()))
		{
			continue;
		}
		for (ULONG l = 0; l < id; l++)
		{
			if ( (lock = (*vector)[l]) )
			{
				level = lock->lck_logical;
				LCK_release(tdbb, lock);
				lock->lck_logical = level;
			}
		}
		id = 0;
		ERR_punt();
	}
}


static jrd_tra* transaction_start(thread_db* tdbb, jrd_tra* temp)
{
/**************************************
 *
 *	t r a n s a c t i o n _ s t a r t
 *
 **************************************
 *
 * Functional description
 *	Start a transaction.
 *
 **************************************/
	SET_TDBB(tdbb);
	Database* const dbb = tdbb->getDatabase();
	Attachment* const attachment = tdbb->getAttachment();
	WIN window(DB_PAGE_SPACE, -1);

	Lock* lock = create_transaction_lock(tdbb, temp);

	// Read header page and allocate transaction number.  Since
	// the transaction inventory page was initialized to zero, it
	// transaction is automatically marked active.

	ULONG oldest, number, active, oldest_active, oldest_snapshot;

#ifdef SUPERSERVER_V2
	number = bump_transaction_id(tdbb, &window);
	oldest = dbb->dbb_oldest_transaction;
	active = MAX(dbb->dbb_oldest_active, dbb->dbb_oldest_transaction);
	oldest_active = dbb->dbb_oldest_active;
	oldest_snapshot = dbb->dbb_oldest_snapshot;

#else // SUPERSERVER_V2
	if (dbb->dbb_flags & DBB_read_only)
	{
		number = dbb->dbb_next_transaction + dbb->generateTransactionId(tdbb);
		oldest = dbb->dbb_oldest_transaction;
		oldest_active = dbb->dbb_oldest_active;
		oldest_snapshot = dbb->dbb_oldest_snapshot;
	}
	else
	{
		const header_page* header = bump_transaction_id(tdbb, &window);
		number = header->hdr_next_transaction;
		oldest = header->hdr_oldest_transaction;
		oldest_active = header->hdr_oldest_active;
		oldest_snapshot = header->hdr_oldest_snapshot;
	}

	// oldest (OIT) > oldest_active (OAT) if OIT was advanced by sweep
	// and no transactions was started after the sweep starts
	active = MAX(oldest_active, oldest);

#endif // SUPERSERVER_V2

	// Allocate pool and transactions block.  Since, by policy,
	// all transactions older than the oldest are either committed
	// or cleaned up, they can be all considered as committed.  To
	// make everything simpler, round down the oldest to a multiple
	// of four, which puts the transaction on a byte boundary.

	ULONG base = oldest & ~TRA_MASK;

	const size_t length = (temp->tra_flags & TRA_read_committed) ? 0 : (number - base + TRA_MASK) / 4;

	MemoryPool* const pool = tdbb->getDefaultPool();
	jrd_tra* const trans = jrd_tra::create(pool, attachment, temp->tra_outer, length);

	fb_assert(trans->tra_pool == temp->tra_pool);
	trans->tra_relation_locks = temp->tra_relation_locks;
	trans->tra_lock_timeout = temp->tra_lock_timeout;
	trans->tra_flags = temp->tra_flags;
	trans->tra_number = number;
	trans->tra_top = number;
	trans->tra_oldest = oldest;
	trans->tra_oldest_active = active;

	trans->tra_lock = lock;
	lock->lck_key.lck_long = number;

	// Put the TID of the oldest active transaction (from the header page)
	// in the new transaction's lock.
	// hvlad: it is important to put transaction number for read-committed
	// transaction instead of oldest active to correctly calculate new oldest
	// active value (look at call to LCK_query_data below which will take into
	// account this new lock too)

	lock->lck_data = (trans->tra_flags & TRA_read_committed) ? number : active;
	lock->lck_object = trans;

	if (!LCK_lock(tdbb, lock, LCK_write, LCK_WAIT))
	{
#ifndef SUPERSERVER_V2
		if (!(dbb->dbb_flags & DBB_read_only))
			CCH_RELEASE(tdbb, &window);
#endif
		jrd_tra::destroy(dbb, trans);
		ERR_post(Arg::Gds(isc_lock_conflict));
	}

	// Link the transaction to the attachment block before releasing
	// header page for handling signals.

	link_transaction(tdbb, trans);

#ifndef SUPERSERVER_V2
	if (!(dbb->dbb_flags & DBB_read_only))
		CCH_RELEASE(tdbb, &window);
#endif

	if (dbb->dbb_flags & DBB_read_only)
	{
		// Set transaction flags to TRA_precommitted, TRA_readonly
		trans->tra_flags |= (TRA_readonly | TRA_precommitted);
	}

	// Next, take a snapshot of all transactions between the oldest interesting
	// transaction and the current.  Don't bother to get a snapshot for
	// read-committed transactions; they use the snapshot off the dbb block
	// since they need to know what is currently committed.

	if (trans->tra_flags & TRA_read_committed)
		TPC_initialize_tpc(tdbb, number);
	else
		TRA_get_inventory(tdbb, trans->tra_transactions.begin(), base, number);

	// Next task is to find the oldest active transaction on the system.  This
	// is needed for garbage collection.  Things are made ever so slightly
	// more complicated by the fact that existing transaction may have oldest
	// actives older than they are.

	Lock temp_lock;
	temp_lock.lck_dbb = dbb;
	temp_lock.lck_object = trans;
	temp_lock.lck_type = LCK_tra;
	temp_lock.lck_owner_handle = LCK_get_owner_handle(tdbb, temp_lock.lck_type);
	temp_lock.lck_parent = dbb->dbb_lock;
	temp_lock.lck_length = sizeof(SLONG);

	trans->tra_oldest_active = number;
	base = oldest & ~TRA_MASK;
	oldest_active = number;
	bool cleanup = !(number % TRA_ACTIVE_CLEANUP);
	USHORT oldest_state;

	for (; active < number; active++)
	{
		if (trans->tra_flags & TRA_read_committed)
			oldest_state = TPC_cache_state(tdbb, active);
		else
		{
			const ULONG byte = TRANS_OFFSET(active - base);
			const USHORT shift = TRANS_SHIFT(active);
			oldest_state = (trans->tra_transactions[byte] >> shift) & TRA_MASK;
		}
		if (oldest_state == tra_active)
		{
			temp_lock.lck_key.lck_long = active;
			SLONG data = LCK_read_data(tdbb, &temp_lock);
			if (!data)
			{
				if (cleanup)
				{
					if (TRA_wait(tdbb, trans, active, jrd_tra::tra_no_wait) == tra_committed)
						cleanup = false;
					continue;
				}

				data = active;
			}

			oldest_active = MIN(oldest_active, active);

			// Find the oldest record version that cannot be garbage collected yet
			// by taking the minimum of all all versions needed by all active transactions.

			if (data < trans->tra_oldest_active)
				trans->tra_oldest_active = data;

			// If the lock data for any active transaction matches a previously
			// computed value then there is no need to continue. There can't be
			// an older lock data in the remaining active transactions.

			if (trans->tra_oldest_active == (SLONG) oldest_snapshot)
				break;
#ifndef VMS
			// Query the minimum lock data for all active transaction locks.
			// This will be the oldest active snapshot used for regulating garbage collection.

			data = LCK_query_data(tdbb, dbb->dbb_lock, LCK_tra, LCK_MIN);
			if (data && data < trans->tra_oldest_active)
				trans->tra_oldest_active = data;
			break;
#endif
		}
	}

	// Calculate oldest active and oldest snapshot numbers looking at current
	// attachment's transactions only. Calculated values are used to determine
	// garbage collection threshold for attachment-local data such as temporary
	// tables (GTTs)

	trans->tra_att_oldest_active = number;
	SLONG att_oldest_active = number;
	SLONG att_oldest_snapshot = number;
	for (jrd_tra* tx_att = attachment->att_transactions; tx_att; tx_att = tx_att->tra_next)
	{
		att_oldest_active = MIN(att_oldest_active, tx_att->tra_number);
		att_oldest_snapshot = MIN(att_oldest_snapshot, tx_att->tra_att_oldest_active);
	}
	trans->tra_att_oldest_active = (trans->tra_flags & TRA_read_committed) ? number : att_oldest_active;
	if (attachment->att_oldest_snapshot < att_oldest_snapshot)
		attachment->att_oldest_snapshot = att_oldest_snapshot;

	// Put the TID of the oldest active transaction (just calculated)
	// in the new transaction's lock.
	// hvlad: for read-committed transaction put tra_number to prevent
	// unnecessary blocking of garbage collection by read-committed
	// transactions

	const ULONG lck_data = (trans->tra_flags & TRA_read_committed) ? number : oldest_active;

	if (lock->lck_data != (SLONG) lck_data)
		LCK_write_data(tdbb, lock, lck_data);

	// Scan commit retaining transactions which have started after us but which
	// want to preserve an oldest active from an already committed transaction.
	// If a previously computed oldest snapshot was matched then there's no
	// need to worry about commit retaining transactions.

#ifdef VMS
	if (trans->tra_oldest_active != oldest_snapshot)
		compute_oldest_retaining(tdbb, trans, false);
#endif

	// Finally, scan transactions looking for the oldest interesting transaction -- the oldest
	// non-commited transaction.  This will not be updated immediately, but saved until the
	// next update access to the header page

	oldest_state = tra_committed;

	for (oldest = trans->tra_oldest; oldest < number; oldest++)
	{
		if (trans->tra_flags & TRA_read_committed)
			oldest_state = TPC_cache_state(tdbb, oldest);
		else
		{
			const ULONG byte = TRANS_OFFSET(oldest - base);
			const USHORT shift = TRANS_SHIFT(oldest);
			oldest_state = (trans->tra_transactions[byte] >> shift) & TRA_MASK;
		}

		if (oldest_state != tra_committed && oldest_state != tra_precommitted)
			break;
	}

	if (--oldest > (ULONG) dbb->dbb_oldest_transaction)
		dbb->dbb_oldest_transaction = oldest;

	if (oldest_active > (ULONG) dbb->dbb_oldest_active)
		dbb->dbb_oldest_active = oldest_active;

	if (trans->tra_oldest_active > dbb->dbb_oldest_snapshot)
	{
		dbb->dbb_oldest_snapshot = trans->tra_oldest_active;

#if defined(GARBAGE_THREAD)
		if (!(dbb->dbb_flags & DBB_gc_active) && (dbb->dbb_flags & DBB_gc_background))
		{
			dbb->dbb_flags |= DBB_gc_pending;
			dbb->dbb_gc_sem.release();
		}
#endif
	}

	// If the transaction block is getting out of hand, force a sweep

	if (dbb->dbb_sweep_interval &&
		(trans->tra_oldest_active > oldest) &&
		(trans->tra_oldest_active - oldest > dbb->dbb_sweep_interval) &&
		oldest_state != tra_limbo)
	{
		start_sweeper(tdbb);
	}

	// Check in with external file system

	EXT_trans_start(trans);

	// Start a 'transaction-level' savepoint, unless this is the
	// system transaction, or unless the transactions doesn't want
	// a savepoint to be started.  This savepoint will be used to
	// undo the transaction if it rolls back.

	if ((trans != dbb->dbb_sys_trans) && !(trans->tra_flags & TRA_no_auto_undo))
	{
		VIO_start_save_point(tdbb, trans);
		trans->tra_save_point->sav_flags |= SAV_trans_level;
	}

	// if the user asked us to restart all requests in this attachment,
	// do so now using the new transaction

	if (trans->tra_flags & TRA_restart_requests)
		restart_requests(tdbb, trans);

	// If the transaction is read-only and read committed, it can be
	// precommitted because it can't modify any records and doesn't
	// need a snapshot preserved. This transaction type can run
	// forever without impacting garbage collection or causing
	// transaction bitmap growth.

	if (trans->tra_flags & TRA_readonly && trans->tra_flags & TRA_read_committed)
	{
		TRA_set_state(tdbb, trans, trans->tra_number, tra_committed);
		LCK_release(tdbb, lock);

		lock->lck_type = LCK_tra_pc;
		lock->lck_owner_handle = LCK_get_owner_handle(tdbb, lock->lck_type);
		lock->lck_data = 0;
		if (!LCK_lock(tdbb, lock, LCK_write, LCK_WAIT))
		{
			jrd_tra::destroy(dbb, trans);
			ERR_post(Arg::Gds(isc_lock_conflict));
		}
		
		trans->tra_flags |= TRA_precommitted;
	}

	if (trans->tra_flags & TRA_precommitted)
		TRA_precommited(tdbb, 0, trans->tra_number);

	return trans;
}


jrd_tra::~jrd_tra()
{
	delete tra_undo_record;
	delete tra_undo_space;
	delete tra_user_management;

	if (!tra_outer)
	{
		delete tra_blob_space;
	}
	else
	{
		fb_assert(!tra_arrays);
	}

	DFW_delete_deferred(this, -1);

	if (tra_autonomous_pool)
	{
		MemoryPool::deletePool(tra_autonomous_pool);
	}
}


UserManagement* jrd_tra::getUserManagement()
{
	if (!tra_user_management)
	{
		tra_user_management = FB_NEW(*tra_pool) UserManagement(this);
	}
	return tra_user_management;
}


jrd_tra* jrd_tra::getOuter()
{
	jrd_tra* tra = this;

	while (tra->tra_outer)
	{
		tra = tra->tra_outer;
	}

	return tra;
}


MemoryPool* jrd_tra::getAutonomousPool()
{
	if (!tra_autonomous_pool)
	{
		MemoryPool* pool = tra_pool;
		jrd_tra* outer = tra_outer;
		while (outer)
		{
			pool = outer->tra_pool;
			outer = outer->tra_outer;
		}
		tra_autonomous_pool = MemoryPool::createPool(pool, tra_memory_stats);
		tra_autonomous_cnt = 0;
	}

	return tra_autonomous_pool;
}


void jrd_tra::releaseAutonomousPool(MemoryPool* toRelease)
{
	fb_assert(tra_autonomous_pool == toRelease);
	if (++tra_autonomous_cnt > TRA_AUTONOMOUS_PER_POOL)
	{
		MemoryPool::deletePool(tra_autonomous_pool);
		tra_autonomous_pool = NULL;
	}
}


/// class TraceSweepEvent

TraceSweepEvent::TraceSweepEvent(thread_db* tdbb) :
  m_request(tdbb->getDefaultPool(), NULL)
{
	m_tdbb = tdbb;

	WIN window(HEADER_PAGE_NUMBER);
	Ods::header_page *header = (Ods::header_page*) CCH_FETCH(m_tdbb, &window, LCK_read, pag_header);

	m_sweep_info.update(header);
	CCH_RELEASE(m_tdbb, &window);

	Attachment* att = m_tdbb->getAttachment();

	gds__log("Sweep is started by %s\n"
		"\tDatabase \"%s\" \n"
		"\tOIT %"SLONGFORMAT", OAT %"SLONGFORMAT", OST %"SLONGFORMAT", Next %"SLONGFORMAT,
		att->att_user->usr_user_name.c_str(),
		att->att_filename.c_str(),
		m_sweep_info.getOIT(),
		m_sweep_info.getOAT(),
		m_sweep_info.getOST(),
		m_sweep_info.getNext());

	TraceManager* trace_mgr = att->att_trace_manager;

	m_need_trace = trace_mgr->needs().event_sweep;

	if (!m_need_trace)
		return;

	m_tdbb->setRequest(&m_request);
	m_start_clock = fb_utils::query_performance_counter();

	TraceConnectionImpl conn(att);
	trace_mgr->event_sweep(&conn, &m_sweep_info, process_state_started);

	m_relation_clock = fb_utils::query_performance_counter();
}


TraceSweepEvent::~TraceSweepEvent()
{
	m_tdbb->setRequest(NULL);
	report(process_state_failed);
}


void TraceSweepEvent::beginSweepRelation(jrd_rel* relation)
{
	if (!m_need_trace)
		return;

	if (relation && relation->rel_name.isEmpty())
	{
		// don't accumulate per-relation stats for metadata query below
		MET_lookup_relation_id(m_tdbb, relation->rel_id, false);
	}

	m_relation_clock = fb_utils::query_performance_counter();
	m_request.req_stats.reset();
}


void TraceSweepEvent::endSweepRelation(jrd_rel* relation)
{
	if (!m_need_trace)
		return;

	// don't report empty relation
	if (m_request.req_stats.getValue(RuntimeStatistics::RECORD_SEQ_READS) == 0 &&
		m_request.req_stats.getValue(RuntimeStatistics::RECORD_BACKOUTS) == 0 &&
		m_request.req_stats.getValue(RuntimeStatistics::RECORD_PURGES) == 0 &&
		m_request.req_stats.getValue(RuntimeStatistics::RECORD_EXPUNGES) == 0)
	{
		return;
	}

	// we need to compare stats against zero base 
	m_request.req_base_stats.reset();

	Database* dbb = m_tdbb->getDatabase();
	TraceRuntimeStats stats(dbb, &m_request.req_base_stats, &m_request.req_stats,
		fb_utils::query_performance_counter() - m_relation_clock, 
		0);

	m_sweep_info.setPerf(stats.getPerf());

	Attachment* att = m_tdbb->getAttachment();
	TraceConnectionImpl conn(att);
	TraceManager* trace_mgr = att->att_trace_manager;
	trace_mgr->event_sweep(&conn, &m_sweep_info, process_state_progress);
}


void TraceSweepEvent::report(ntrace_process_state_t state)
{
	Attachment* att = m_tdbb->getAttachment();

	if (state == process_state_finished) 
	{
		gds__log("Sweep is finished\n"
			"\tDatabase \"%s\" \n"
			"\tOIT %"SLONGFORMAT", OAT %"SLONGFORMAT", OST %"SLONGFORMAT", Next %"SLONGFORMAT,
			att->att_filename.c_str(),
			m_sweep_info.getOIT(),
			m_sweep_info.getOAT(),
			m_sweep_info.getOST(),
			m_sweep_info.getNext());
	}

	if (!m_need_trace)
		return;

	Database* dbb = m_tdbb->getDatabase();
	TraceManager* trace_mgr = att->att_trace_manager;

	TraceConnectionImpl conn(att);

	// we need to compare stats against zero base 
	m_request.req_base_stats.reset(); 

	TraceRuntimeStats stats(dbb, &m_request.req_base_stats, &att->att_stats,
		fb_utils::query_performance_counter() - m_start_clock, 
		0);

	m_sweep_info.setPerf(stats.getPerf());
	trace_mgr->event_sweep(&conn, &m_sweep_info, state);

	if (state == process_state_failed || state == process_state_finished)
		m_need_trace = false;
}