コード例 #1
0
ファイル: trx0rec.c プロジェクト: isleon/Jaxer
/**************************************************************************
Reports in the undo log of an update or delete marking of a clustered index
record. */
static
ulint
trx_undo_page_report_modify(
/*========================*/
					/* out: byte offset of the inserted
					undo log entry on the page if succeed,
					0 if fail */
	page_t* 	undo_page,	/* in: undo log page */
	trx_t*		trx,		/* in: transaction */
	dict_index_t*	index,		/* in: clustered index where update or
					delete marking is done */
	rec_t*		rec,		/* in: clustered index record which
					has NOT yet been modified */
	const ulint*	offsets,	/* in: rec_get_offsets(rec, index) */
	upd_t*		update,		/* in: update vector which tells the
					columns to be updated; in the case of
					a delete, this should be set to NULL */
	ulint		cmpl_info,	/* in: compiler info on secondary
					index updates */
	mtr_t*		mtr)		/* in: mtr */
{
	dict_table_t*	table;
	upd_field_t*	upd_field;
	dict_col_t*	col;
	ulint		first_free;
	byte*		ptr;
	ulint		len;
	byte* 		field;
	ulint		flen;
	ulint		pos;
	dulint		roll_ptr;
	dulint		trx_id;
	ulint		bits;
	ulint		col_no;
	byte*		old_ptr;
	ulint		type_cmpl;
	byte*		type_cmpl_ptr;
	ulint		i;
	
	ut_a(index->type & DICT_CLUSTERED);
	ut_ad(rec_offs_validate(rec, index, offsets));
	ut_ad(mach_read_from_2(undo_page + TRX_UNDO_PAGE_HDR
				+ TRX_UNDO_PAGE_TYPE) == TRX_UNDO_UPDATE);
	table = index->table;
	
	first_free = mach_read_from_2(undo_page + TRX_UNDO_PAGE_HDR
						+ TRX_UNDO_PAGE_FREE);
	ptr = undo_page + first_free;
						
	ut_ad(first_free <= UNIV_PAGE_SIZE);

	if (trx_undo_left(undo_page, ptr) < 50) {

		/* NOTE: the value 50 must be big enough so that the general
		fields written below fit on the undo log page */

		return(0);
	}

	/* Reserve 2 bytes for the pointer to the next undo log record */
	ptr += 2;

	/* Store first some general parameters to the undo log */

	if (update) {
		if (rec_get_deleted_flag(rec, table->comp)) {
			type_cmpl = TRX_UNDO_UPD_DEL_REC;
		} else {
			type_cmpl = TRX_UNDO_UPD_EXIST_REC;
		}
	} else {
		type_cmpl = TRX_UNDO_DEL_MARK_REC;
	}

	type_cmpl = type_cmpl | (cmpl_info * TRX_UNDO_CMPL_INFO_MULT);

	mach_write_to_1(ptr, type_cmpl);
	
	type_cmpl_ptr = ptr;

	ptr++;
	len = mach_dulint_write_much_compressed(ptr, trx->undo_no);
	ptr += len;

	len = mach_dulint_write_much_compressed(ptr, table->id);
	ptr += len;

	/*----------------------------------------*/
	/* Store the state of the info bits */

	bits = rec_get_info_bits(rec, table->comp);
	mach_write_to_1(ptr, bits);
	ptr += 1;

	/* Store the values of the system columns */
	field = rec_get_nth_field(rec, offsets,
		dict_index_get_sys_col_pos(index, DATA_TRX_ID), &len);
	ut_ad(len == DATA_TRX_ID_LEN);
	trx_id = trx_read_trx_id(field);
	field = rec_get_nth_field(rec, offsets,
		dict_index_get_sys_col_pos(index, DATA_ROLL_PTR), &len);
	ut_ad(len == DATA_ROLL_PTR_LEN);
	roll_ptr = trx_read_roll_ptr(field);

	len = mach_dulint_write_compressed(ptr, trx_id);
	ptr += len;

	len = mach_dulint_write_compressed(ptr, roll_ptr);
	ptr += len;

	/*----------------------------------------*/
	/* Store then the fields required to uniquely determine the
	record which will be modified in the clustered index */

	for (i = 0; i < dict_index_get_n_unique(index); i++) {

		field = rec_get_nth_field(rec, offsets, i, &flen);

		if (trx_undo_left(undo_page, ptr) < 4) {

			return(0);
		}

		len = mach_write_compressed(ptr, flen); 
		ptr += len;

		if (flen != UNIV_SQL_NULL) {
			if (trx_undo_left(undo_page, ptr) < flen) {

				return(0);
			}

			ut_memcpy(ptr, field, flen);
			ptr += flen;
		}
	}

	/*----------------------------------------*/
	/* Save to the undo log the old values of the columns to be updated. */

	if (update) {
	    if (trx_undo_left(undo_page, ptr) < 5) {

		return(0);
	    }

	    len = mach_write_compressed(ptr, upd_get_n_fields(update));
	    ptr += len;

	    for (i = 0; i < upd_get_n_fields(update); i++) {

		upd_field = upd_get_nth_field(update, i);
		pos = upd_field->field_no;

		/* Write field number to undo log */
		if (trx_undo_left(undo_page, ptr) < 5) {

			return(0);
		}

		len = mach_write_compressed(ptr, pos);
		ptr += len;

		/* Save the old value of field */
		field = rec_get_nth_field(rec, offsets, pos, &flen);

		if (trx_undo_left(undo_page, ptr) < 5) {

			return(0);
		}

		if (rec_offs_nth_extern(offsets, pos)) {
			/* If a field has external storage, we add to
			flen the flag */

			len = mach_write_compressed(ptr,
					UNIV_EXTERN_STORAGE_FIELD + flen);

			/* Notify purge that it eventually has to free the old
			externally stored field */
			
			trx->update_undo->del_marks = TRUE;

			*type_cmpl_ptr = *type_cmpl_ptr | TRX_UNDO_UPD_EXTERN;
		} else {
			len = mach_write_compressed(ptr, flen);
		}

		ptr += len;

		if (flen != UNIV_SQL_NULL) {
			if (trx_undo_left(undo_page, ptr) < flen) {

				return(0);
			}

			ut_memcpy(ptr, field, flen);
			ptr += flen;
		}
	    }
	}		

	/*----------------------------------------*/
	/* In the case of a delete marking, and also in the case of an update
	where any ordering field of any index changes, store the values of all
	columns which occur as ordering fields in any index. This info is used
	in the purge of old versions where we use it to build and search the
	delete marked index records, to look if we can remove them from the
	index tree. Note that starting from 4.0.14 also externally stored
	fields can be ordering in some index. But we always store at least
	384 first bytes locally to the clustered index record, which means
	we can construct the column prefix fields in the index from the
	stored data. */

	if (!update || !(cmpl_info & UPD_NODE_NO_ORD_CHANGE)) {	    

	    trx->update_undo->del_marks = TRUE;

	    if (trx_undo_left(undo_page, ptr) < 5) {

		return(0);
	    }
	    
	    old_ptr = ptr;

	    /* Reserve 2 bytes to write the number of bytes the stored fields
	    take in this undo record */

	    ptr += 2;

	    for (col_no = 0; col_no < dict_table_get_n_cols(table); col_no++) {

	    	col = dict_table_get_nth_col(table, col_no);

	    	if (col->ord_part > 0) {
	    		
			pos = dict_index_get_nth_col_pos(index, col_no);

			/* Write field number to undo log */
			if (trx_undo_left(undo_page, ptr) < 5) {
	
				return(0);
			}

			len = mach_write_compressed(ptr, pos);
			ptr += len;
	
			/* Save the old value of field */
			field = rec_get_nth_field(rec, offsets, pos, &flen);
	
			if (trx_undo_left(undo_page, ptr) < 5) {
	
				return(0);
			}

			len = mach_write_compressed(ptr, flen);
			ptr += len;

			if (flen != UNIV_SQL_NULL) {
				if (trx_undo_left(undo_page, ptr) < flen) {
	
					return(0);
				}

				ut_memcpy(ptr, field, flen);
				ptr += flen;
			}
		}
	    }

	    mach_write_to_2(old_ptr, ptr - old_ptr);	    
	}		

	/*----------------------------------------*/
	/* Write pointers to the previous and the next undo log records */
	if (trx_undo_left(undo_page, ptr) < 2) {

		return(0);
	}
	
	mach_write_to_2(ptr, first_free);
	ptr += 2;
	mach_write_to_2(undo_page + first_free, ptr - undo_page);

	mach_write_to_2(undo_page + TRX_UNDO_PAGE_HDR + TRX_UNDO_PAGE_FREE,
							ptr - undo_page);

	/* Write to the REDO log about this change in the UNDO log */

	trx_undof_page_add_undo_rec_log(undo_page, first_free,
							ptr - undo_page, mtr);
	return(first_free);	
}
コード例 #2
0
ファイル: row0vers.c プロジェクト: H0bby/GCS-SQL
/*****************************************************************//**
Constructs the last committed version of a clustered index record,
which should be seen by a semi-consistent read.
@return	DB_SUCCESS or DB_MISSING_HISTORY */
UNIV_INTERN
ulint
row_vers_build_for_semi_consistent_read(
/*====================================*/
	const rec_t*	rec,	/*!< in: record in a clustered index; the
				caller must have a latch on the page; this
				latch locks the top of the stack of versions
				of this records */
	mtr_t*		mtr,	/*!< in: mtr holding the latch on rec */
	dict_index_t*	index,	/*!< in: the clustered index */
	ulint**		offsets,/*!< in/out: offsets returned by
				rec_get_offsets(rec, index) */
	mem_heap_t**	offset_heap,/*!< in/out: memory heap from which
				the offsets are allocated */
	mem_heap_t*	in_heap,/*!< in: memory heap from which the memory for
				*old_vers is allocated; memory for possible
				intermediate versions is allocated and freed
				locally within the function */
	const rec_t**	old_vers)/*!< out: rec, old version, or NULL if the
				record does not exist in the view, that is,
				it was freshly inserted afterwards */
{
	const rec_t*	version;
	mem_heap_t*	heap		= NULL;
	byte*		buf;
	ulint		err;
	trx_id_t	rec_trx_id	= 0;

	ut_ad(dict_index_is_clust(index));
	ut_ad(mtr_memo_contains_page(mtr, rec, MTR_MEMO_PAGE_X_FIX)
	      || mtr_memo_contains_page(mtr, rec, MTR_MEMO_PAGE_S_FIX));
#ifdef UNIV_SYNC_DEBUG
	ut_ad(!rw_lock_own(&(purge_sys->latch), RW_LOCK_SHARED));
#endif /* UNIV_SYNC_DEBUG */

	ut_ad(rec_offs_validate(rec, index, *offsets));

	rw_lock_s_lock(&(purge_sys->latch));
	/* The S-latch on purge_sys prevents the purge view from
	changing.  Thus, if we have an uncommitted transaction at
	this point, then purge cannot remove its undo log even if
	the transaction could commit now. */

	version = rec;

	for (;;) {
		trx_t*		version_trx;
		mem_heap_t*	heap2;
		rec_t*		prev_version;
		trx_id_t	version_trx_id;

		version_trx_id = row_get_rec_trx_id(version, index, *offsets);
		if (rec == version) {
			rec_trx_id = version_trx_id;
		}

		mutex_enter(&kernel_mutex);
		version_trx = trx_get_on_id(version_trx_id);
		if (version_trx
		    && (version_trx->conc_state == TRX_COMMITTED_IN_MEMORY
			|| version_trx->conc_state == TRX_NOT_STARTED)) {

			version_trx = NULL;
		}
		mutex_exit(&kernel_mutex);

		if (!version_trx) {

			/* We found a version that belongs to a
			committed transaction: return it. */

#ifdef UNIV_BLOB_NULL_DEBUG
			ut_a(!rec_offs_any_null_extern(version, *offsets));
#endif /* UNIV_BLOB_NULL_DEBUG */

			if (rec == version) {
				*old_vers = rec;
				err = DB_SUCCESS;
				break;
			}

			/* We assume that a rolled-back transaction stays in
			TRX_ACTIVE state until all the changes have been
			rolled back and the transaction is removed from
			the global list of transactions. */

			if (rec_trx_id == version_trx_id) {
				/* The transaction was committed while
				we searched for earlier versions.
				Return the current version as a
				semi-consistent read. */

				version = rec;
				*offsets = rec_get_offsets(version,
							   index, *offsets,
							   ULINT_UNDEFINED,
							   offset_heap);
			}

			buf = mem_heap_alloc(in_heap, rec_offs_size(*offsets));
			*old_vers = rec_copy(buf, version, *offsets);
			rec_offs_make_valid(*old_vers, index, *offsets);
			err = DB_SUCCESS;

			break;
		}

		heap2 = heap;
		heap = mem_heap_create(1024);

		err = trx_undo_prev_version_build(rec, mtr, version, index,
						  *offsets, heap,
						  &prev_version);
		if (heap2) {
			mem_heap_free(heap2); /* free version */
		}

		if (UNIV_UNLIKELY(err != DB_SUCCESS)) {
			break;
		}

		if (prev_version == NULL) {
			/* It was a freshly inserted version */
			*old_vers = NULL;
			err = DB_SUCCESS;

			break;
		}

		version = prev_version;
		*offsets = rec_get_offsets(version, index, *offsets,
					   ULINT_UNDEFINED, offset_heap);
#ifdef UNIV_BLOB_NULL_DEBUG
		ut_a(!rec_offs_any_null_extern(version, *offsets));
#endif /* UNIV_BLOB_NULL_DEBUG */
	}/* for (;;) */

	if (heap) {
		mem_heap_free(heap);
	}
	rw_lock_s_unlock(&(purge_sys->latch));

	return(err);
}
コード例 #3
0
ファイル: trx0rec.c プロジェクト: AllenWeb/mariadb
/**********************************************************************//**
Reports in the undo log of an update or delete marking of a clustered index
record.
@return byte offset of the inserted undo log entry on the page if
succeed, 0 if fail */
static
ulint
trx_undo_page_report_modify(
/*========================*/
	page_t*		undo_page,	/*!< in: undo log page */
	trx_t*		trx,		/*!< in: transaction */
	dict_index_t*	index,		/*!< in: clustered index where update or
					delete marking is done */
	const rec_t*	rec,		/*!< in: clustered index record which
					has NOT yet been modified */
	const ulint*	offsets,	/*!< in: rec_get_offsets(rec, index) */
	const upd_t*	update,		/*!< in: update vector which tells the
					columns to be updated; in the case of
					a delete, this should be set to NULL */
	ulint		cmpl_info,	/*!< in: compiler info on secondary
					index updates */
	mtr_t*		mtr)		/*!< in: mtr */
{
	dict_table_t*	table;
	ulint		first_free;
	byte*		ptr;
	const byte*	field;
	ulint		flen;
	ulint		col_no;
	ulint		type_cmpl;
	byte*		type_cmpl_ptr;
	ulint		i;
	trx_id_t	trx_id;
	ibool		ignore_prefix = FALSE;
	byte		ext_buf[REC_VERSION_56_MAX_INDEX_COL_LEN
				+ BTR_EXTERN_FIELD_REF_SIZE];

	ut_a(dict_index_is_clust(index));
	ut_ad(rec_offs_validate(rec, index, offsets));
	ut_ad(mach_read_from_2(undo_page + TRX_UNDO_PAGE_HDR
			       + TRX_UNDO_PAGE_TYPE) == TRX_UNDO_UPDATE);
	table = index->table;

	first_free = mach_read_from_2(undo_page + TRX_UNDO_PAGE_HDR
				      + TRX_UNDO_PAGE_FREE);
	ptr = undo_page + first_free;

	ut_ad(first_free <= UNIV_PAGE_SIZE);

	if (trx_undo_left(undo_page, ptr) < 50) {

		/* NOTE: the value 50 must be big enough so that the general
		fields written below fit on the undo log page */

		return(0);
	}

	/* Reserve 2 bytes for the pointer to the next undo log record */
	ptr += 2;

	/* Store first some general parameters to the undo log */

	if (!update) {
		type_cmpl = TRX_UNDO_DEL_MARK_REC;
	} else if (rec_get_deleted_flag(rec, dict_table_is_comp(table))) {
		type_cmpl = TRX_UNDO_UPD_DEL_REC;
		/* We are about to update a delete marked record.
		We don't typically need the prefix in this case unless
		the delete marking is done by the same transaction
		(which we check below). */
		ignore_prefix = TRUE;
	} else {
		type_cmpl = TRX_UNDO_UPD_EXIST_REC;
	}

	type_cmpl |= cmpl_info * TRX_UNDO_CMPL_INFO_MULT;
	type_cmpl_ptr = ptr;

	*ptr++ = (byte) type_cmpl;
	ptr += mach_ull_write_much_compressed(ptr, trx->undo_no);

	ptr += mach_ull_write_much_compressed(ptr, table->id);

	/*----------------------------------------*/
	/* Store the state of the info bits */

	*ptr++ = (byte) rec_get_info_bits(rec, dict_table_is_comp(table));

	/* Store the values of the system columns */
	field = rec_get_nth_field(rec, offsets,
				  dict_index_get_sys_col_pos(
					  index, DATA_TRX_ID), &flen);
	ut_ad(flen == DATA_TRX_ID_LEN);

	trx_id = trx_read_trx_id(field);

	/* If it is an update of a delete marked record, then we are
	allowed to ignore blob prefixes if the delete marking was done
	by some other trx as it must have committed by now for us to
	allow an over-write. */
	if (ignore_prefix) {
		ignore_prefix = (trx_id != trx->id);
	}
	ptr += mach_ull_write_compressed(ptr, trx_id);

	field = rec_get_nth_field(rec, offsets,
				  dict_index_get_sys_col_pos(
					  index, DATA_ROLL_PTR), &flen);
	ut_ad(flen == DATA_ROLL_PTR_LEN);

	ptr += mach_ull_write_compressed(ptr, trx_read_roll_ptr(field));

	/*----------------------------------------*/
	/* Store then the fields required to uniquely determine the
	record which will be modified in the clustered index */

	for (i = 0; i < dict_index_get_n_unique(index); i++) {

		field = rec_get_nth_field(rec, offsets, i, &flen);

		/* The ordering columns must not be stored externally. */
		ut_ad(!rec_offs_nth_extern(offsets, i));
		ut_ad(dict_index_get_nth_col(index, i)->ord_part);

		if (trx_undo_left(undo_page, ptr) < 5) {

			return(0);
		}

		ptr += mach_write_compressed(ptr, flen);

		if (flen != UNIV_SQL_NULL) {
			if (trx_undo_left(undo_page, ptr) < flen) {

				return(0);
			}

			ut_memcpy(ptr, field, flen);
			ptr += flen;
		}
	}

	/*----------------------------------------*/
	/* Save to the undo log the old values of the columns to be updated. */

	if (update) {
		ulint	extended = 0;

		if (trx_undo_left(undo_page, ptr) < 5) {

			return(0);
		}

		if (srv_use_sys_stats_table
		    && index == UT_LIST_GET_FIRST(dict_sys->sys_stats->indexes)) {
			for (i = 0; i < upd_get_n_fields(update); i++) {
				ulint	pos = upd_get_nth_field(update, i)->field_no;

				if (pos >= rec_offs_n_fields(offsets)) {
					extended++;
				}
			}
		}

		ptr += mach_write_compressed(ptr, upd_get_n_fields(update) - extended);

		for (i = 0; i < upd_get_n_fields(update) - extended; i++) {

			ulint	pos = upd_get_nth_field(update, i)->field_no;

			/* Write field number to undo log */
			if (trx_undo_left(undo_page, ptr) < 5) {

				return(0);
			}

			ptr += mach_write_compressed(ptr, pos);

			/* Save the old value of field */
			field = rec_get_nth_field(rec, offsets, pos, &flen);

			if (trx_undo_left(undo_page, ptr) < 15) {

				return(0);
			}

			if (rec_offs_nth_extern(offsets, pos)) {
				const dict_col_t*	col
					= dict_index_get_nth_col(index, pos);
				ulint			prefix_len
					= dict_max_field_len_store_undo(
						table, col);

				ut_ad(prefix_len + BTR_EXTERN_FIELD_REF_SIZE
				      <= sizeof ext_buf);

				ptr = trx_undo_page_report_modify_ext(
					ptr,
					col->ord_part
					&& !ignore_prefix
					&& flen < REC_ANTELOPE_MAX_INDEX_COL_LEN
					? ext_buf : NULL, prefix_len,
					dict_table_zip_size(table),
					&field, &flen);

				/* Notify purge that it eventually has to
				free the old externally stored field */

				trx->update_undo->del_marks = TRUE;

				*type_cmpl_ptr |= TRX_UNDO_UPD_EXTERN;
			} else {
				ptr += mach_write_compressed(ptr, flen);
			}

			if (flen != UNIV_SQL_NULL) {
				if (trx_undo_left(undo_page, ptr) < flen) {

					return(0);
				}

				ut_memcpy(ptr, field, flen);
				ptr += flen;
			}
		}
	}

	/*----------------------------------------*/
	/* In the case of a delete marking, and also in the case of an update
	where any ordering field of any index changes, store the values of all
	columns which occur as ordering fields in any index. This info is used
	in the purge of old versions where we use it to build and search the
	delete marked index records, to look if we can remove them from the
	index tree. Note that starting from 4.0.14 also externally stored
	fields can be ordering in some index. Starting from 5.2, we no longer
	store REC_MAX_INDEX_COL_LEN first bytes to the undo log record,
	but we can construct the column prefix fields in the index by
	fetching the first page of the BLOB that is pointed to by the
	clustered index. This works also in crash recovery, because all pages
	(including BLOBs) are recovered before anything is rolled back. */

	if (!update || !(cmpl_info & UPD_NODE_NO_ORD_CHANGE)) {
		byte*	old_ptr = ptr;

		trx->update_undo->del_marks = TRUE;

		if (trx_undo_left(undo_page, ptr) < 5) {

			return(0);
		}

		/* Reserve 2 bytes to write the number of bytes the stored
		fields take in this undo record */

		ptr += 2;

		for (col_no = 0; col_no < dict_table_get_n_cols(table);
		     col_no++) {

			const dict_col_t*	col
				= dict_table_get_nth_col(table, col_no);

			if (col->ord_part) {
				ulint	pos;

				/* Write field number to undo log */
				if (trx_undo_left(undo_page, ptr) < 5 + 15) {

					return(0);
				}

				pos = dict_index_get_nth_col_pos(index,
								 col_no);
				ptr += mach_write_compressed(ptr, pos);

				/* Save the old value of field */
				field = rec_get_nth_field(rec, offsets, pos,
							  &flen);

				if (rec_offs_nth_extern(offsets, pos)) {
					const dict_col_t*	col =
						dict_index_get_nth_col(
							index, pos);
					ulint			prefix_len =
						dict_max_field_len_store_undo(
							table, col);

					ut_a(prefix_len < sizeof ext_buf);

					ptr = trx_undo_page_report_modify_ext(
						ptr,
						flen < REC_ANTELOPE_MAX_INDEX_COL_LEN
						&& !ignore_prefix
						? ext_buf : NULL, prefix_len,
						dict_table_zip_size(table),
						&field, &flen);
				} else {
					ptr += mach_write_compressed(
						ptr, flen);
				}

				if (flen != UNIV_SQL_NULL) {
					if (trx_undo_left(undo_page, ptr)
					    < flen) {

						return(0);
					}

					ut_memcpy(ptr, field, flen);
					ptr += flen;
				}
			}
		}

		mach_write_to_2(old_ptr, ptr - old_ptr);
	}

	/*----------------------------------------*/
	/* Write pointers to the previous and the next undo log records */
	if (trx_undo_left(undo_page, ptr) < 2) {

		return(0);
	}

	mach_write_to_2(ptr, first_free);
	ptr += 2;
	mach_write_to_2(undo_page + first_free, ptr - undo_page);

	mach_write_to_2(undo_page + TRX_UNDO_PAGE_HDR + TRX_UNDO_PAGE_FREE,
			ptr - undo_page);

	/* Write to the REDO log about this change in the UNDO log */

	trx_undof_page_add_undo_rec_log(undo_page, first_free,
					ptr - undo_page, mtr);
	return(first_free);
}
コード例 #4
0
ファイル: row0vers.c プロジェクト: H0bby/GCS-SQL
/*****************************************************************//**
Constructs the version of a clustered index record which a consistent
read should see. We assume that the trx id stored in rec is such that
the consistent read should not see rec in its present version.
@return	DB_SUCCESS or DB_MISSING_HISTORY */
UNIV_INTERN
ulint
row_vers_build_for_consistent_read(
/*===============================*/
	const rec_t*	rec,	/*!< in: record in a clustered index; the
				caller must have a latch on the page; this
				latch locks the top of the stack of versions
				of this records */
	mtr_t*		mtr,	/*!< in: mtr holding the latch on rec */
	dict_index_t*	index,	/*!< in: the clustered index */
	ulint**		offsets,/*!< in/out: offsets returned by
				rec_get_offsets(rec, index) */
	read_view_t*	view,	/*!< in: the consistent read view */
	mem_heap_t**	offset_heap,/*!< in/out: memory heap from which
				the offsets are allocated */
	mem_heap_t*	in_heap,/*!< in: memory heap from which the memory for
				*old_vers is allocated; memory for possible
				intermediate versions is allocated and freed
				locally within the function */
	rec_t**		old_vers)/*!< out, own: old version, or NULL if the
				record does not exist in the view, that is,
				it was freshly inserted afterwards */
{
	const rec_t*	version;
	rec_t*		prev_version;
	trx_id_t	trx_id;
	mem_heap_t*	heap		= NULL;
	byte*		buf;
	ulint		err;

	ut_ad(dict_index_is_clust(index));
	ut_ad(mtr_memo_contains_page(mtr, rec, MTR_MEMO_PAGE_X_FIX)
	      || mtr_memo_contains_page(mtr, rec, MTR_MEMO_PAGE_S_FIX));
#ifdef UNIV_SYNC_DEBUG
	ut_ad(!rw_lock_own(&(purge_sys->latch), RW_LOCK_SHARED));
#endif /* UNIV_SYNC_DEBUG */

	ut_ad(rec_offs_validate(rec, index, *offsets));

	trx_id = row_get_rec_trx_id(rec, index, *offsets);

	ut_ad(!read_view_sees_trx_id(view, trx_id));

	rw_lock_s_lock(&(purge_sys->latch));
	version = rec;

	for (;;) {
		mem_heap_t*	heap2	= heap;
		trx_undo_rec_t* undo_rec;
		roll_ptr_t	roll_ptr;
		undo_no_t	undo_no;
		heap = mem_heap_create(1024);

		/* If we have high-granularity consistent read view and
		creating transaction of the view is the same as trx_id in
		the record we see this record only in the case when
		undo_no of the record is < undo_no in the view. */

		if (view->type == VIEW_HIGH_GRANULARITY
		    && view->creator_trx_id == trx_id) {

			roll_ptr = row_get_rec_roll_ptr(version, index,
							*offsets);
			undo_rec = trx_undo_get_undo_rec_low(roll_ptr, heap);
			undo_no = trx_undo_rec_get_undo_no(undo_rec);
			mem_heap_empty(heap);

			if (view->undo_no > undo_no) {
				/* The view already sees this version: we can
				copy it to in_heap and return */

#ifdef UNIV_BLOB_NULL_DEBUG
				ut_a(!rec_offs_any_null_extern(
					     version, *offsets));
#endif /* UNIV_BLOB_NULL_DEBUG */

				buf = mem_heap_alloc(in_heap,
						     rec_offs_size(*offsets));
				*old_vers = rec_copy(buf, version, *offsets);
				rec_offs_make_valid(*old_vers, index,
						    *offsets);
				err = DB_SUCCESS;

				break;
			}
		}

		err = trx_undo_prev_version_build(rec, mtr, version, index,
						  *offsets, heap,
						  &prev_version);
		if (heap2) {
			mem_heap_free(heap2); /* free version */
		}

		if (err != DB_SUCCESS) {
			break;
		}

		if (prev_version == NULL) {
			/* It was a freshly inserted version */
			*old_vers = NULL;
			err = DB_SUCCESS;

			break;
		}

		*offsets = rec_get_offsets(prev_version, index, *offsets,
					   ULINT_UNDEFINED, offset_heap);

#ifdef UNIV_BLOB_NULL_DEBUG
		ut_a(!rec_offs_any_null_extern(prev_version, *offsets));
#endif /* UNIV_BLOB_NULL_DEBUG */

		trx_id = row_get_rec_trx_id(prev_version, index, *offsets);

		if (read_view_sees_trx_id(view, trx_id)) {

			/* The view already sees this version: we can copy
			it to in_heap and return */

			buf = mem_heap_alloc(in_heap, rec_offs_size(*offsets));
			*old_vers = rec_copy(buf, prev_version, *offsets);
			rec_offs_make_valid(*old_vers, index, *offsets);
			err = DB_SUCCESS;

			break;
		}

		version = prev_version;
	}/* for (;;) */

	mem_heap_free(heap);
	rw_lock_s_unlock(&(purge_sys->latch));

	return(err);
}