nbak.cpp

/*
 *	PROGRAM:	JRD Access Method
 *	MODULE:		nbak.cpp
 *	DESCRIPTION:	Incremental backup technology
 *
 *  The contents of this file are subject to the Initial
 *  Developer's 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.ibphoenix.com/main.nfs?a=ibphoenix&page=ibp_idpl.
 *
 *  Software distributed under the License is distributed AS IS,
 *  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 Nickolay Samofatov
 *  for the Firebird Open Source RDBMS project.
 *
 *  Copyright (c) 2004 Nickolay Samofatov <nickolay@broadviewsoftware.com>
 *  and all contributors signed below.
 *
 *  All Rights Reserved.
 *  Contributor(s):
 *
 *  Roman Simakov <roman-simakov@users.sourceforge.net>
 *  Khorsun Vladyslav <hvlad@users.sourceforge.net>
 *
 */

#include "firebird.h"
#include "jrd.h"
#include "nbak.h"
#include "ods.h"
#include "lck.h"
#include "cch.h"
#include "lck_proto.h"
#include "pag_proto.h"
#include "err_proto.h"
#include "cch_proto.h"
#include "../common/isc_proto.h"
#include "os/pio_proto.h"
#include "iberror.h"
#include "../yvalve/gds_proto.h"
#include "../common/os/guid.h"
#include "../common/os/isc_i_proto.h"
#include "../jrd/CryptoManager.h"
#include "../jrd/replication/Publisher.h"

#ifdef HAVE_SYS_TYPES_H
#include <sys/types.h>
#endif

#ifdef HAVE_SYS_STAT_H
#include <sys/stat.h>
#endif

#ifdef HAVE_UNISTD_H
#include <unistd.h>
#endif

#ifdef HAVE_ERRNO_H
#include <errno.h>
#endif

#ifdef NBAK_DEBUG
#include <stdarg.h>
IMPLEMENT_TRACE_ROUTINE(nbak_trace, "NBAK")
#endif


using namespace Jrd;
using namespace Firebird;


/******************************** NBackupStateLock ******************************/

NBackupStateLock::NBackupStateLock(thread_db* tdbb, MemoryPool& p, BackupManager* bakMan):
	GlobalRWLock(tdbb, p, LCK_backup_database), backup_manager(bakMan)
{
}

bool NBackupStateLock::fetch(thread_db* tdbb)
{
	backup_manager->endFlush();
	NBAK_TRACE( ("backup_manager->endFlush()") );
	if (!backup_manager->actualizeState(tdbb))
	{
		ERR_bugcheck_msg("Can't actualize backup state");
	}
	return true;
}

void NBackupStateLock::invalidate(thread_db* tdbb)
{
	GlobalRWLock::invalidate(tdbb);
	NBAK_TRACE( ("invalidate state stateLock(%p)", this) );
	backup_manager->setState(Ods::hdr_nbak_unknown);
	backup_manager->closeDelta(tdbb);
}

void NBackupStateLock::blockingAstHandler(thread_db* tdbb)
{
	// master instance should not try to acquire localStateLock or enter "flush" mode
	if (backup_manager->isMaster())
	{
		GlobalRWLock::blockingAstHandler(tdbb);
		return;
	}

	// Ensure we have no dirty pages in local cache before releasing
	// of state lock
	if (!backup_manager->databaseFlushInProgress())
	{
		backup_manager->beginFlush();
		NBAK_TRACE_AST( ("backup_manager->beginFlush()") );

		Firebird::MutexUnlockGuard counterGuard(counterMutex, FB_FUNCTION);
		CCH_flush_ast(tdbb);
		NBAK_TRACE_AST(("database FLUSHED"));
	}

	{	// scope
		Firebird::MutexUnlockGuard counterGuard(counterMutex, FB_FUNCTION);
		backup_manager->stateBlocking = !backup_manager->localStateLock.tryBeginWrite(FB_FUNCTION);
		if (backup_manager->stateBlocking)
			return;
	}

	GlobalRWLock::blockingAstHandler(tdbb);

	if (cachedLock->lck_physical == LCK_read)
		backup_manager->endFlush();

	backup_manager->localStateLock.endWrite();
}


/******************************** NBackupAllocLock ******************************/

NBackupAllocLock::NBackupAllocLock(thread_db* tdbb, MemoryPool& p, BackupManager* bakMan):
	GlobalRWLock(tdbb, p, LCK_backup_alloc), backup_manager(bakMan)
{
}

bool NBackupAllocLock::fetch(thread_db* tdbb)
{
	if (!backup_manager->actualizeAlloc(tdbb, true))
		ERR_bugcheck_msg("Can't actualize alloc table");
	return true;
}

void NBackupAllocLock::invalidate(thread_db* tdbb)
{
	backup_manager->invalidateAlloc(tdbb);
	GlobalRWLock::invalidate(tdbb);
	NBAK_TRACE(("invalidate alloc table allocLock(%p)", this));
}

/******************************** BackupManager::StateWriteGuard ******************************/

BackupManager::StateWriteGuard::StateWriteGuard(thread_db* tdbb, Jrd::WIN* window)
	: m_tdbb(tdbb), m_window(NULL), m_success(false)
{
	Database* const dbb = tdbb->getDatabase();
	Jrd::Attachment* const att = tdbb->getAttachment();

	dbb->dbb_backup_manager->beginFlush();
	CCH_flush(tdbb, FLUSH_ALL, 0); // Flush local cache to release all dirty pages

	CCH_FETCH(tdbb, window, LCK_write, pag_header);

	dbb->dbb_backup_manager->localStateLock.beginWrite(FB_FUNCTION);

	if (!dbb->dbb_backup_manager->lockStateWrite(tdbb, LCK_WAIT))
	{
		dbb->dbb_backup_manager->localStateLock.endWrite();
		ERR_bugcheck_msg("Can't lock state for write");
	}

	dbb->dbb_backup_manager->endFlush();

	NBAK_TRACE(("backup state locked for write"));
	m_window = window;
}

BackupManager::StateWriteGuard::~StateWriteGuard()
{
	Database* const dbb = m_tdbb->getDatabase();
	Jrd::Attachment* const att = m_tdbb->getAttachment();

	// It is important to set state into nbak_state_unknown *before* release of state lock,
	// otherwise someone could acquire state lock, fetch and modify some page before state will
	// be set into unknown. But dirty page can't be written when backup state is unknown
	// because write target (database or delta) is also unknown.

	if (!m_success)
	{
		NBAK_TRACE( ("invalidate state") );
		dbb->dbb_backup_manager->setState(Ods::hdr_nbak_unknown);
	}

	releaseHeader();
	dbb->dbb_backup_manager->unlockStateWrite(m_tdbb);
	dbb->dbb_backup_manager->localStateLock.endWrite();
}

void BackupManager::StateWriteGuard::releaseHeader()
{
	if (m_window)
	{
		CCH_RELEASE(m_tdbb, m_window);
		m_window = NULL;
	}
}

/********************************** CORE LOGIC ********************************/

void BackupManager::generateFilename()
{
	diff_name = database->dbb_filename + ".delta";
	explicit_diff_name = false;
}

void BackupManager::openDelta(thread_db* tdbb)
{
	fb_assert(!diff_file);
	diff_file = PIO_open(tdbb, diff_name, diff_name);

	if (database->dbb_flags & (DBB_force_write | DBB_no_fs_cache))
	{
		setForcedWrites(database->dbb_flags & DBB_force_write,
						database->dbb_flags & DBB_no_fs_cache);
	}
}

void BackupManager::closeDelta(thread_db* tdbb)
{
	if (diff_file)
	{
		PIO_flush(tdbb, diff_file);
		PIO_close(diff_file);
		diff_file = NULL;
	}
}

// Initialize and open difference file for writing
void BackupManager::beginBackup(thread_db* tdbb)
{
	NBAK_TRACE(("beginBackup"));

	SET_TDBB(tdbb);

	// Check for raw device
	if ((!explicit_diff_name) && database->onRawDevice()) {
		ERR_post(Arg::Gds(isc_need_difference));
	}

	MasterGuard masterGuard(*this);

	WIN window(HEADER_PAGE_NUMBER);

	StateWriteGuard stateGuard(tdbb, &window);
	Ods::header_page* header = (Ods::header_page*) window.win_buffer;

	// Check state
	if (backup_state != Ods::hdr_nbak_normal)
	{
		NBAK_TRACE(("begin backup - invalid state %d", backup_state));
		stateGuard.setSuccess();
		return;
	}

	// Check crypt state
	if (header->hdr_flags & Ods::hdr_crypt_process)
	{
		NBAK_TRACE(("begin backup - crypt thread runs"));
		stateGuard.setSuccess();
		ERR_post(Arg::Gds(isc_wish_list) << Arg::Gds(isc_random) <<
			"Cannot begin backup: please wait for crypt thread completion");
	}

	try
	{
		// Create file
		NBAK_TRACE(("Creating difference file %s", diff_name.c_str()));
		diff_file = PIO_create(tdbb, diff_name, true, false);
	}
	catch (const Firebird::Exception&)
	{
		// no reasons to set it to unknown if we just failed to create difference file
		stateGuard.setSuccess();
		backup_state = Ods::hdr_nbak_normal;
		throw;
	}

	{ // logical scope
		if (database->dbb_flags & (DBB_force_write | DBB_no_fs_cache))
		{
			setForcedWrites(database->dbb_flags & DBB_force_write,
							database->dbb_flags & DBB_no_fs_cache);
		}

#ifdef UNIX
		// adjust difference file access rights to make it match main DB ones
		if (diff_file && geteuid() == 0)
		{
			struct STAT st;
			PageSpace* pageSpace = database->dbb_page_manager.findPageSpace(DB_PAGE_SPACE);
			const char* func = NULL;

			{
				jrd_file* file = pageSpace->file;
				ReadLockGuard readGuard(file->fil_desc_lock, FB_FUNCTION);
				if (os_utils::fstat(file->fil_desc, &st) != 0)
				{
					func = "fstat";
				}
			}

			while (!func && fchown(diff_file->fil_desc, st.st_uid, st.st_gid) != 0)
			{
				if (errno != EINTR)
					func = "fchown";
			}

			while (!func && fchmod(diff_file->fil_desc, st.st_mode) != 0)
			{
				if (errno != EINTR)
					func = "fchmod";
			}

			if (func)
			{
				stateGuard.setSuccess();
				Firebird::system_call_failed::raise(func);
			}
		}
#endif

		// Zero out first page (empty allocation table)
		BufferDesc temp_bdb(database->dbb_bcb);
		temp_bdb.bdb_page = 0;
		temp_bdb.bdb_buffer = reinterpret_cast<Ods::pag*>(alloc_buffer);
		memset(alloc_buffer, 0, database->dbb_page_size);
		if (!PIO_write(tdbb, diff_file, &temp_bdb, temp_bdb.bdb_buffer, tdbb->tdbb_status_vector))
			ERR_punt();
		NBAK_TRACE(("Set backup state in header"));
		Guid guid;
		GenerateGuid(&guid);
		// Set state in database header page. All changes are written to main database file yet.
		CCH_MARK_MUST_WRITE(tdbb, &window);
		const int newState = Ods::hdr_nbak_stalled; // Should be USHORT?
		header->hdr_flags = (header->hdr_flags & ~Ods::hdr_backup_mask) | newState;
		const ULONG adjusted_scn = ++header->hdr_header.pag_scn; // Generate new SCN
		PAG_replace_entry_first(tdbb, header, Ods::HDR_backup_guid, sizeof(guid),
			reinterpret_cast<const UCHAR*>(&guid));

		REPL_journal_switch(tdbb);

		stateGuard.releaseHeader();
		stateGuard.setSuccess();

		backup_state = newState;
		current_scn = adjusted_scn;

		// All changes go to the difference file now
	}
}


// Determine actual DB size (raw devices support)
ULONG BackupManager::getPageCount(thread_db* tdbb)
{
	if (backup_state != Ods::hdr_nbak_stalled)
	{
		// calculate pages only when database is locked for backup:
		// other case such service is just dangerous
		return 0;
	}

	return PAG_page_count(tdbb);
}


bool BackupManager::extendDatabase(thread_db* tdbb)
{
	if (!alloc_table)
	{
		LocalAllocWriteGuard localAllocGuard(this);
		actualizeAlloc(tdbb, false);
	}

	ULONG maxPage = 0;
	{
		LocalAllocReadGuard localAllocGuard(this);
		AllocItemTree::Accessor all(alloc_table);

		if (all.getFirst())
		{
			do
			{
				const ULONG pg = all.current().db_page;
				if (maxPage < pg)
					maxPage = pg;
			} while (all.getNext());
		}
	}

	PageSpace *pgSpace = database->dbb_page_manager.findPageSpace(DB_PAGE_SPACE);
	ULONG maxAllocPage = pgSpace->maxAlloc();
	if (maxAllocPage >= maxPage)
		return true;

	if (!pgSpace->extend(tdbb, maxPage, true))
		return false;

	maxAllocPage = pgSpace->maxAlloc();
	while (maxAllocPage < maxPage)
	{
		const USHORT ret = PIO_init_data(tdbb, pgSpace->file, tdbb->tdbb_status_vector,
										 maxAllocPage, 256);

		if (ret != 256)
			return false;

		maxAllocPage += ret;
	}

	return true;
}


// Merge difference file to main files (if needed) and unlink() difference
// file then. If merge is already in progress method silently returns and
// does nothing (so it can be used for recovery on database startup).
void BackupManager::endBackup(thread_db* tdbb, bool recover)
{
	NBAK_TRACE(("end_backup, recover=%i", recover));

	// Check for recover

	GlobalRWLock endLock(tdbb, *database->dbb_permanent, LCK_backup_end, false);

	if (!endLock.lockWrite(tdbb, LCK_NO_WAIT))
	{
		// Someboby holds write lock on LCK_backup_end. We need not to do end_backup
		return;
	}

	MasterGuard masterGuard(*this);

	// STEP 1. Change state in header to "merge"
	WIN window(HEADER_PAGE_NUMBER);
	Ods::header_page* header;

#ifdef NBAK_DEBUG
	ULONG adjusted_scn; // We use this value to prevent race conditions.
						// They are possible because we release state lock
						// for some instants and anything is possible at
						// that times.
#endif

	try
	{
		// Check state under PR lock of backup state for speed
		{ // scope
			StateReadGuard stateGuard(tdbb);
			// Nobody is doing end_backup but database isn't in merge state.
			if ( (recover || backup_state != Ods::hdr_nbak_stalled) && (backup_state != Ods::hdr_nbak_merge ) )
			{
				NBAK_TRACE(("invalid state %d", backup_state));
				endLock.unlockWrite(tdbb);
				return;
			}

			if (backup_state == Ods::hdr_nbak_stalled && !extendDatabase(tdbb))
				status_exception::raise(tdbb->tdbb_status_vector);
		}

		// Here backup state can be changed. Need to check it again after lock
		StateWriteGuard stateGuard(tdbb, &window);

		if ( (recover || backup_state != Ods::hdr_nbak_stalled) && (backup_state != Ods::hdr_nbak_merge ) )
		{
			stateGuard.setSuccess();
			NBAK_TRACE(("invalid state %d", backup_state));
			endLock.unlockWrite(tdbb);
			return;
		}

		if (!extendDatabase(tdbb))
		{
			stateGuard.setSuccess();
			status_exception::raise(tdbb->tdbb_status_vector);
		}

		header = (Ods::header_page*) window.win_buffer;

		NBAK_TRACE(("difference file %s, current backup state is %d", diff_name.c_str(), backup_state));

		// Set state in database header
		backup_state = Ods::hdr_nbak_merge;
#ifdef NBAK_DEBUG
		adjusted_scn =
#endif
					   ++current_scn;
		NBAK_TRACE(("New state is getting to become %d", backup_state));
		CCH_MARK_MUST_WRITE(tdbb, &window);
		// Generate new SCN
		header->hdr_header.pag_scn = current_scn;
		NBAK_TRACE(("new SCN=%d is getting written to header", header->hdr_header.pag_scn));
		// Adjust state
		header->hdr_flags = (header->hdr_flags & ~Ods::hdr_backup_mask) | backup_state;
		NBAK_TRACE(("Setting state %d in header page is over", backup_state));
		stateGuard.setSuccess();
	}
	catch (const Firebird::Exception&)
	{
		endLock.unlockWrite(tdbb);
		throw;
	}

	// STEP 2. Merging database and delta
	// Here comes the dirty work. We need to reapply all changes from difference file to database
	// Release write state lock and get read lock.
	// Merge process should not inhibit normal operations.

	try
	{
		NBAK_TRACE(("database locked to merge"));

		StateReadGuard stateGuard(tdbb);

		NBAK_TRACE(("Merge. State=%d, current_scn=%d, adjusted_scn=%d",
			backup_state, current_scn, adjusted_scn));

		{ // scope
			LocalAllocWriteGuard localAllocGuard(this);

			// In merge mode allocation table can not be changed, so we don't
			// need to acquire global allocLock.
			actualizeAlloc(tdbb, true);
		}

		LocalAllocReadGuard localAllocGuard(this);

		NBAK_TRACE(("Merge. Alloc table is actualized."));
		AllocItemTree::Accessor all(alloc_table);

		if (all.getFirst())
		{
			int n = 0;
			do {
				JRD_reschedule(tdbb);

				WIN window2(DB_PAGE_SPACE, all.current().db_page);
				NBAK_TRACE(("Merge page %d, diff=%d", all.current().db_page, all.current().diff_page));
				Ods::pag* page = CCH_FETCH(tdbb, &window2, LCK_write, pag_undefined);
				NBAK_TRACE(("Merge: page %d is fetched", all.current().db_page));
				if (page->pag_scn != current_scn)
				{
					CCH_MARK_SYSTEM(tdbb, &window2);
					NBAK_TRACE(("Merge: page %d is marked", all.current().db_page));
				}
				CCH_RELEASE(tdbb, &window2);
				NBAK_TRACE(("Merge: page %d is released", all.current().db_page));

				if (++n == 512)
				{
					CCH_flush(tdbb, FLUSH_SYSTEM, 0);
					n = 0;
				}
			} while (all.getNext());
		}

		CCH_flush(tdbb, FLUSH_ALL, 0);
		NBAK_TRACE(("Merging is over. Database unlocked"));
	}
	catch (const Firebird::Exception&)
	{
		endLock.unlockWrite(tdbb);
		throw;
	}

	// STEP 3. Change state in header to "normal"
	// We finished. We need to reflect it in our database header page
	try {
		window.win_page = HEADER_PAGE;
		window.win_flags = 0;

		StateWriteGuard stateGuard(tdbb, &window);

		header = (Ods::header_page*) window.win_buffer;

		// Set state in database header
		backup_state = Ods::hdr_nbak_normal;
		CCH_MARK_MUST_WRITE(tdbb, &window);
		// Adjust state
		header->hdr_flags = (header->hdr_flags & ~Ods::hdr_backup_mask) | backup_state;
		NBAK_TRACE(("Set state %d in header page", backup_state));
		// Generate new SCN
		header->hdr_header.pag_scn = ++current_scn;
		NBAK_TRACE(("new SCN=%d is getting written to header", header->hdr_header.pag_scn));

		stateGuard.releaseHeader();
		stateGuard.setSuccess();

		// Page allocation table cache is no longer valid
		NBAK_TRACE(("Dropping alloc table"));
		{
			LocalAllocWriteGuard localAllocGuard(this);

			delete alloc_table;
			alloc_table = NULL;
			last_allocated_page = 0;
			if (!allocLock->tryReleaseLock(tdbb))
				ERR_bugcheck_msg("There are holders of alloc_lock after end_backup finish");
		}

		closeDelta(tdbb);
		unlink(diff_name.c_str());

		NBAK_TRACE(("backup is over"));
		endLock.unlockWrite(tdbb);
	}
	catch (const Firebird::Exception&)
	{
		endLock.unlockWrite(tdbb);
		throw;
	}

	return;
}

void BackupManager::initializeAlloc(thread_db* tdbb)
{
	StateReadGuard stateGuard(tdbb);
	if (getState() != Ods::hdr_nbak_normal)
		actualizeAlloc(tdbb, false);
}

bool BackupManager::actualizeAlloc(thread_db* tdbb, bool haveGlobalLock)
{
	// localAllocLock must be already locked for write here !

	// Difference file pointer pages have one ULONG as number of pages allocated on the page and
	// then go physical numbers of pages from main database file. Offsets of numbers correspond
	// to difference file pages.
	const size_t PAGES_PER_ALLOC_PAGE = database->dbb_page_size / sizeof(ULONG) - 1;

	FbStatusVector *status_vector = tdbb->tdbb_status_vector;
	try {
		NBAK_TRACE(("actualize_alloc last_allocated_page=%d alloc_table=%p",
			last_allocated_page, alloc_table));
		// For SuperServer this routine is really executed only at database startup when
		// it has exlock or when exclusive access to database is enabled
		if (!alloc_table)
			alloc_table = FB_NEW_POOL(*database->dbb_permanent) AllocItemTree(database->dbb_permanent);

		while (true)
		{
			BufferDesc temp_bdb(database->dbb_bcb);

			// Get offset of pointer page. We can do so because page sizes are powers of 2
			temp_bdb.bdb_page = last_allocated_page & ~PAGES_PER_ALLOC_PAGE;
			temp_bdb.bdb_buffer = reinterpret_cast<Ods::pag*>(alloc_buffer);

			if (!PIO_read(tdbb, diff_file, &temp_bdb, temp_bdb.bdb_buffer, status_vector)) {
				return false;
			}

			// If we have not acquired global allocLock, don't read last page of allocation
			// table as it could be modified concurrently. Lets read it later when global
			// alloc lock will be acquired.
			if (!haveGlobalLock && (alloc_buffer[0] != PAGES_PER_ALLOC_PAGE))
				break;

			for (ULONG i = last_allocated_page - temp_bdb.bdb_page.getPageNum(); i < alloc_buffer[0]; i++)
			{
				NBAK_TRACE(("alloc item page=%d, diff=%d", alloc_buffer[i + 1], temp_bdb.bdb_page.getPageNum() + i + 1));
				if (!alloc_table->add(AllocItem(alloc_buffer[i + 1], temp_bdb.bdb_page.getPageNum() + i + 1)))
				{
					database->dbb_flags |= DBB_bugcheck;
					ERR_build_status(status_vector,
						Arg::Gds(isc_bug_check) << Arg::Str("Duplicated item in allocation table detected"));
					return false;
				}
			}
			last_allocated_page = temp_bdb.bdb_page.getPageNum() + alloc_buffer[0];
			if (alloc_buffer[0] == PAGES_PER_ALLOC_PAGE)
				last_allocated_page++;	// if page is full adjust position for next pointer page
			else
				break;	// We finished reading allocation table
		}
	}
	catch (const Firebird::Exception& ex)
	{
		// Handle out of memory error, etc
		delete alloc_table;
		ex.stuffException(status_vector);
		alloc_table = NULL;
		last_allocated_page = 0;
		return false;
	}

	allocIsValid = haveGlobalLock;
	return true;
}

ULONG BackupManager::findPageIndex(thread_db* tdbb, ULONG db_page)
{
	// localAllocLock must be already locked here

	NBAK_TRACE(("find_page_index"));
	if (!alloc_table)
		return 0;

	AllocItemTree::Accessor a(alloc_table);
	ULONG diff_page = a.locate(db_page) ? a.current().diff_page : 0;

	return diff_page;
}

// Return page index in difference file that can be used in
// writeDifference call later.
ULONG BackupManager::getPageIndex(thread_db* tdbb, ULONG db_page)
{
	NBAK_TRACE(("get_page_index"));

	{
		LocalAllocReadGuard localAllocGuard(this);

		const ULONG diff_page = findPageIndex(tdbb, db_page);
		if (diff_page || backup_state == Ods::hdr_nbak_merge && allocIsValid)
			return diff_page;
	}

	LocalAllocWriteGuard localAllocGuard(this);
	GlobalAllocReadGuard globalAllocGuard(tdbb, this);
	return findPageIndex(tdbb, db_page);
}

// Mark next difference page as used by some database page
ULONG BackupManager::allocateDifferencePage(thread_db* tdbb, ULONG db_page)
{
	LocalAllocWriteGuard localAllocGuard(this);

	// This page may be allocated while we wait for a local lock above
	if (ULONG diff_page = findPageIndex(tdbb, db_page)) {
		return diff_page;
	}

	GlobalAllocWriteGuard globalAllocGuard(tdbb, this);

	// This page may be allocated by other process
	if (ULONG diff_page = findPageIndex(tdbb, db_page)) {
		return diff_page;
	}

	NBAK_TRACE(("allocate_difference_page"));
	fb_assert(last_allocated_page % (database->dbb_page_size / sizeof(ULONG)) == alloc_buffer[0]);

	FbStatusVector* status_vector = tdbb->tdbb_status_vector;
	// Grow file first. This is done in such order to keep difference
	// file consistent in case of write error. We should always be able
	// to read next alloc page when previous one is full.
	BufferDesc temp_bdb(database->dbb_bcb);
	temp_bdb.bdb_page = last_allocated_page + 1;
	temp_bdb.bdb_buffer = reinterpret_cast<Ods::pag*>(empty_buffer);
	if (!PIO_write(tdbb, diff_file, &temp_bdb, temp_bdb.bdb_buffer, status_vector))
		return 0;

	const bool alloc_page_full = alloc_buffer[0] == database->dbb_page_size / sizeof(ULONG) - 2;
	if (alloc_page_full)
	{
		// Pointer page is full. Its time to create new one.
		temp_bdb.bdb_page = last_allocated_page + 2;
		temp_bdb.bdb_buffer = reinterpret_cast<Ods::pag*>(empty_buffer);
		if (!PIO_write(tdbb, diff_file, &temp_bdb, temp_bdb.bdb_buffer, status_vector))
			return 0;
	}

	// Write new item to the allocation table
	temp_bdb.bdb_page = last_allocated_page & ~(database->dbb_page_size / sizeof(ULONG) - 1);
	temp_bdb.bdb_buffer = reinterpret_cast<Ods::pag*>(alloc_buffer);
	alloc_buffer[++alloc_buffer[0]] = db_page;
	if (!PIO_write(tdbb, diff_file, &temp_bdb, temp_bdb.bdb_buffer, status_vector))
		return 0;

	last_allocated_page++;

	// Register new page in the alloc table
	try
	{
		alloc_table->add(AllocItem(db_page, last_allocated_page));
	}
	catch (const Firebird::Exception& ex)
	{
		// Handle out of memory error
		delete alloc_table;
		alloc_table = NULL;
		last_allocated_page = 0;
		ex.stuffException(status_vector);
		return 0;
	}

	// Adjust buffer and counters if we allocated new alloc page earlier
	if (alloc_page_full)
	{
		last_allocated_page++;
		memset(alloc_buffer, 0, database->dbb_page_size);
		return last_allocated_page - 1;
	}

	return last_allocated_page;
}

bool BackupManager::writeDifference(thread_db* tdbb, FbStatusVector* status, ULONG diff_page, Ods::pag* page)
{
	if (!diff_page)
	{
		// We should never be here but if it happens let not overwrite first allocation
		// page with garbage.

		(Arg::Gds(isc_random) << "Can't allocate difference page").copyTo(status);
		return false;
	}

	NBAK_TRACE(("write_diff page=%d, diff=%d", page->pag_pageno, diff_page));
	BufferDesc temp_bdb(database->dbb_bcb);
	temp_bdb.bdb_page = diff_page;
	temp_bdb.bdb_buffer = page;

	// Check that diff page is not allocation page
	fb_assert(diff_page % (database->dbb_page_size / sizeof(ULONG)));

	class Pio : public CryptoManager::IOCallback
	{
	public:
		Pio(jrd_file* p_file, BufferDesc* p_bdb)
			: file(p_file), bdb(p_bdb)
		{ }

		bool callback(thread_db* tdbb, FbStatusVector* sv, Ods::pag* page)
		{
			return PIO_write(tdbb, file, bdb, page, sv);
		}

	private:
		jrd_file* file;
		BufferDesc* bdb;
	};

	Pio io(diff_file, &temp_bdb);

	if (!database->dbb_crypto_manager->write(tdbb, status, page, &io))
		return false;
	return true;
}

bool BackupManager::readDifference(thread_db* tdbb, ULONG diff_page, Ods::pag* page)
{
	BufferDesc temp_bdb(database->dbb_bcb);
	temp_bdb.bdb_page = diff_page;
	temp_bdb.bdb_buffer = page;

	class Pio : public CryptoManager::IOCallback
	{
	public:
		Pio(jrd_file* p_file, BufferDesc* p_bdb)
			: file(p_file), bdb(p_bdb)
		{ }

		bool callback(thread_db* tdbb, FbStatusVector* sv, Ods::pag* page)
		{
			return PIO_read(tdbb, file, bdb, page, sv);
		}

	private:
		jrd_file* file;
		BufferDesc* bdb;
	};

	Pio io(diff_file, &temp_bdb);

	if (!database->dbb_crypto_manager->read(tdbb, tdbb->tdbb_status_vector, page, &io))
		return false;
	NBAK_TRACE(("read_diff page=%d, diff=%d", page->pag_pageno, diff_page));
	return true;
}

void BackupManager::flushDifference(thread_db* tdbb)
{
	PIO_flush(tdbb, diff_file);
}

void BackupManager::setForcedWrites(const bool forceWrite, const bool notUseFSCache)
{
	if (diff_file)
		PIO_force_write(diff_file, forceWrite, notUseFSCache);
}

BackupManager::BackupManager(thread_db* tdbb, Database* _database, int ini_state) :
	dbCreating(false), database(_database), diff_file(NULL), alloc_table(NULL),
	last_allocated_page(0), temp_buffers_space(*database->dbb_permanent),
	current_scn(0), diff_name(*database->dbb_permanent),
	explicit_diff_name(false), flushInProgress(false), shutDown(false), allocIsValid(false),
	master(false), stateBlocking(false),
	stateLock(FB_NEW_POOL(*database->dbb_permanent) NBackupStateLock(tdbb, *database->dbb_permanent, this)),
	allocLock(FB_NEW_POOL(*database->dbb_permanent) NBackupAllocLock(tdbb, *database->dbb_permanent, this))
{
	// Allocate various database page buffers needed for operation
	// Align it at sector boundary for faster IO (also guarantees correct alignment for ULONG later)
	UCHAR* temp_buffers = reinterpret_cast<UCHAR*>
		(temp_buffers_space.getAlignedBuffer(database->dbb_page_size * 3, database->getIOBlockSize()));

	backup_state = ini_state;

	empty_buffer = reinterpret_cast<ULONG*>(temp_buffers);
	spare_buffer = reinterpret_cast<ULONG*>(temp_buffers + database->dbb_page_size);
	alloc_buffer = reinterpret_cast<ULONG*>(temp_buffers + database->dbb_page_size * 2);

	NBAK_TRACE(("Create BackupManager, database=%s", database->dbb_filename.c_str()));
}

BackupManager::~BackupManager()
{
	delete stateLock;
	delete allocLock;
	delete alloc_table;
}

void BackupManager::setDifference(thread_db* tdbb, const char* filename)
{
	SET_TDBB(tdbb);

	if (filename)
	{
		WIN window(HEADER_PAGE_NUMBER);
		Ods::header_page* header =
			(Ods::header_page*) CCH_FETCH(tdbb, &window, LCK_write, pag_header);
		CCH_MARK_MUST_WRITE(tdbb, &window);
		PAG_replace_entry_first(tdbb, header, Ods::HDR_difference_file,
			static_cast<USHORT>(strlen(filename)), reinterpret_cast<const UCHAR*>(filename));
		CCH_RELEASE(tdbb, &window);
		diff_name = filename;
		explicit_diff_name = true;
	}
	else
	{
		PAG_delete_clump_entry(tdbb, Ods::HDR_difference_file);
		generateFilename();
	}
}

bool BackupManager::actualizeState(thread_db* tdbb)
{
	// State is unknown. We need to read it from the disk.
	// We cannot use CCH for this because of likely recursion.
	NBAK_TRACE(("actualizeState: current_state=%i", backup_state));

	if (dbCreating)
	{
		backup_state = Ods::hdr_nbak_normal;
		return true;
	}

	SET_TDBB(tdbb);

	FbStatusVector* status = tdbb->tdbb_status_vector;

	// Read original page from database file or shadows.
	SSHORT retryCount = 0;
	Ods::header_page* header = reinterpret_cast<Ods::header_page*>(spare_buffer);
	BufferDesc temp_bdb(database->dbb_bcb);
	temp_bdb.bdb_page = HEADER_PAGE_NUMBER;
	temp_bdb.bdb_buffer = &header->hdr_header;
	PageSpace* pageSpace = database->dbb_page_manager.findPageSpace(DB_PAGE_SPACE);
	fb_assert(pageSpace);
	jrd_file* file = pageSpace->file;

	// It's header page, never encrypted
	while (!PIO_read(tdbb, file, &temp_bdb, temp_bdb.bdb_buffer, status))
	{
		if (!CCH_rollover_to_shadow(tdbb, database, file, false))
		{
			NBAK_TRACE(("Shadow change error"));
			return false;
		}
		if (file != pageSpace->file)
			file = pageSpace->file;
		else
		{
			if (retryCount++ == 3)
			{
				NBAK_TRACE(("IO error"));
				return false;
			}
		}
	}

	const int new_backup_state = header->hdr_flags & Ods::hdr_backup_mask;
	NBAK_TRACE(("backup state read from header is %d", new_backup_state));
	// Check is we missed lock/unlock cycle and need to invalidate
	// our allocation table and file handle
	const bool missed_cycle = (header->hdr_header.pag_scn - current_scn) > 1;
	current_scn = header->hdr_header.pag_scn;

	// Read difference file name from header clumplets
	explicit_diff_name = false;
	const UCHAR* p = header->hdr_data;
	const UCHAR* const end = reinterpret_cast<UCHAR*>(header) + header->hdr_page_size;
	while (p < end)
	{
		switch (*p)
		{
		case Ods::HDR_difference_file:
			explicit_diff_name = true;
			diff_name.assign(reinterpret_cast<const char*>(p + 2), p[1]);
			break;

		default:
			p += p[1] + 2;
			continue;
		}
		break;
	}
	if (!explicit_diff_name)
		generateFilename();

	if (new_backup_state == Ods::hdr_nbak_normal || missed_cycle)
	{
		// Page allocation table cache is no longer valid.
		LocalAllocWriteGuard localAllocGuard(this);

		if (alloc_table)
		{
			NBAK_TRACE(("Dropping alloc table"));
			delete alloc_table;
			alloc_table = NULL;
			last_allocated_page = 0;
			if (!allocLock->tryReleaseLock(tdbb))
				ERR_bugcheck_msg("There are holders of alloc_lock after end_backup finish");
		}

		closeDelta(tdbb);
	}

	if (new_backup_state != Ods::hdr_nbak_normal && !diff_file)
		openDelta(tdbb);
	// Adjust state at the very and to ensure proper error handling
	backup_state = new_backup_state;

	return true;
}

void BackupManager::shutdown(thread_db* tdbb)
{
	shutDown = true;

	closeDelta(tdbb);
	stateLock->shutdownLock(tdbb);
	allocLock->shutdownLock(tdbb);
}