mariadb/row/row0purge.c

/******************************************************
Purge obsolete records

(c) 1997 Innobase Oy

Created 3/14/1997 Heikki Tuuri
*******************************************************/

#include "row0purge.h"

#ifdef UNIV_NONINL
#include "row0purge.ic"
#endif

#include "fsp0fsp.h"
#include "mach0data.h"
#include "trx0rseg.h"
#include "trx0trx.h"
#include "trx0roll.h"
#include "trx0undo.h"
#include "trx0purge.h"
#include "trx0rec.h"
#include "que0que.h"
#include "row0row.h"
#include "row0upd.h"
#include "row0vers.h"
#include "row0mysql.h"
#include "log0log.h"

/************************************************************************
Creates a purge node to a query graph. */
UNIV_INTERN
purge_node_t*
row_purge_node_create(
/*==================*/
				/* out, own: purge node */
	que_thr_t*	parent,	/* in: parent node, i.e., a thr node */
	mem_heap_t*	heap)	/* in: memory heap where created */
{
	purge_node_t*	node;

	ut_ad(parent && heap);

	node = mem_heap_alloc(heap, sizeof(purge_node_t));

	node->common.type = QUE_NODE_PURGE;
	node->common.parent = parent;

	node->heap = mem_heap_create(256);

	return(node);
}

/***************************************************************
Repositions the pcur in the purge node on the clustered index record,
if found. */
static
ibool
row_purge_reposition_pcur(
/*======================*/
				/* out: TRUE if the record was found */
	ulint		mode,	/* in: latching mode */
	purge_node_t*	node,	/* in: row purge node */
	mtr_t*		mtr)	/* in: mtr */
{
	ibool	found;

	if (node->found_clust) {
		found = btr_pcur_restore_position(mode, &(node->pcur), mtr);

		return(found);
	}

	found = row_search_on_row_ref(&(node->pcur), mode, node->table,
				      node->ref, mtr);
	node->found_clust = found;

	if (found) {
		btr_pcur_store_position(&(node->pcur), mtr);
	}

	return(found);
}

/***************************************************************
Removes a delete marked clustered index record if possible. */
static
ibool
row_purge_remove_clust_if_poss_low(
/*===============================*/
				/* out: TRUE if success, or if not found, or
				if modified after the delete marking */
	purge_node_t*	node,	/* in: row purge node */
	ulint		mode)	/* in: BTR_MODIFY_LEAF or BTR_MODIFY_TREE */
{
	dict_index_t*	index;
	btr_pcur_t*	pcur;
	btr_cur_t*	btr_cur;
	ibool		success;
	ulint		err;
	mtr_t		mtr;
	rec_t*		rec;
	mem_heap_t*	heap		= NULL;
	ulint		offsets_[REC_OFFS_NORMAL_SIZE];
	rec_offs_init(offsets_);

	index = dict_table_get_first_index(node->table);

	pcur = &(node->pcur);
	btr_cur = btr_pcur_get_btr_cur(pcur);

	mtr_start(&mtr);

	success = row_purge_reposition_pcur(mode, node, &mtr);

	if (!success) {
		/* The record is already removed */

		btr_pcur_commit_specify_mtr(pcur, &mtr);

		return(TRUE);
	}

	rec = btr_pcur_get_rec(pcur);

	if (0 != ut_dulint_cmp(node->roll_ptr, row_get_rec_roll_ptr(
				       rec, index, rec_get_offsets(
					       rec, index, offsets_,
					       ULINT_UNDEFINED, &heap)))) {
		if (UNIV_LIKELY_NULL(heap)) {
			mem_heap_free(heap);
		}
		/* Someone else has modified the record later: do not remove */
		btr_pcur_commit_specify_mtr(pcur, &mtr);

		return(TRUE);
	}

	if (UNIV_LIKELY_NULL(heap)) {
		mem_heap_free(heap);
	}

	if (mode == BTR_MODIFY_LEAF) {
		success = btr_cur_optimistic_delete(btr_cur, &mtr);
	} else {
		ut_ad(mode == BTR_MODIFY_TREE);
		btr_cur_pessimistic_delete(&err, FALSE, btr_cur,
					   RB_NONE, &mtr);

		if (err == DB_SUCCESS) {
			success = TRUE;
		} else if (err == DB_OUT_OF_FILE_SPACE) {
			success = FALSE;
		} else {
			ut_error;
		}
	}

	btr_pcur_commit_specify_mtr(pcur, &mtr);

	return(success);
}

/***************************************************************
Removes a clustered index record if it has not been modified after the delete
marking. */
static
void
row_purge_remove_clust_if_poss(
/*===========================*/
	purge_node_t*	node)	/* in: row purge node */
{
	ibool	success;
	ulint	n_tries	= 0;

	/*	fputs("Purge: Removing clustered record\n", stderr); */

	success = row_purge_remove_clust_if_poss_low(node, BTR_MODIFY_LEAF);
	if (success) {

		return;
	}
retry:
	success = row_purge_remove_clust_if_poss_low(node, BTR_MODIFY_TREE);
	/* The delete operation may fail if we have little
	file space left: TODO: easiest to crash the database
	and restart with more file space */

	if (!success && n_tries < BTR_CUR_RETRY_DELETE_N_TIMES) {
		n_tries++;

		os_thread_sleep(BTR_CUR_RETRY_SLEEP_TIME);

		goto retry;
	}

	ut_a(success);
}

/***************************************************************
Determines if it is possible to remove a secondary index entry.
Removal is possible if the secondary index entry does not refer to any
not delete marked version of a clustered index record where DB_TRX_ID
is newer than the purge view.

NOTE: This function should only be called by the purge thread, only
while holding a latch on the leaf page of the secondary index entry
(or keeping the buffer pool watch on the page).  It is possible that
this function first returns TRUE and then FALSE, if a user transaction
inserts a record that the secondary index entry would refer to.
However, in that case, the user transaction would also re-insert the
secondary index entry after purge has removed it and released the leaf
page latch. */
UNIV_INTERN
ibool
row_purge_poss_sec(
/*===============*/
				/* out: TRUE if the secondary index
				record can be purged */
	purge_node_t*	node,	/* in/out: row purge node */
	dict_index_t*	index,	/* in: secondary index */
	const dtuple_t*	entry)	/* in: secondary index entry */
{
	ibool	can_delete;
	mtr_t	mtr;

	ut_ad(!dict_index_is_clust(index));
	mtr_start(&mtr);

	can_delete = !row_purge_reposition_pcur(BTR_SEARCH_LEAF, node, &mtr)
		|| !row_vers_old_has_index_entry(TRUE,
						 btr_pcur_get_rec(&node->pcur),
						 &mtr, index, entry);

	btr_pcur_commit_specify_mtr(&node->pcur, &mtr);

	return(can_delete);
}

/***************************************************************
Removes a secondary index entry if possible, by modifying the
index tree.  Does not try to buffer the delete. */
static
ibool
row_purge_remove_sec_if_poss_tree(
/*==============================*/
				/* out: TRUE if success or if not found */
	purge_node_t*	node,	/* in: row purge node */
	dict_index_t*	index,	/* in: index */
	const dtuple_t*	entry)	/* in: index entry */
{
	btr_pcur_t		pcur;
	btr_cur_t*		btr_cur;
	ibool			success	= TRUE;
	ulint			err;
	mtr_t			mtr;
	enum row_search_result	search_result;

	log_free_check();
	mtr_start(&mtr);

	search_result = row_search_index_entry(index, entry, BTR_MODIFY_TREE,
					       &pcur, &mtr);

	switch (search_result) {
	case ROW_NOT_FOUND:
		/* Not found.  This is a legitimate condition.  In a
		rollback, InnoDB will remove secondary recs that would
		be purged anyway.  Then the actual purge will not find
		the secondary index record.  Also, the purge itself is
		eager: if it comes to consider a secondary index
		record, and notices it does not need to exist in the
		index, it will remove it.  Then if/when the purge
		comes to consider the secondary index record a second
		time, it will not exist any more in the index. */

		/* fputs("PURGE:........sec entry not found\n", stderr); */
		/* dtuple_print(stderr, entry); */
		goto func_exit;
	case ROW_FOUND:
		break;
	case ROW_BUFFERED:
	case ROW_NOT_DELETED_REF:
	case ROW_NOT_DELETED:
		/* These are invalid outcomes, because the mode passed
		to row_search_index_entry() did not include any of the
		flags BTR_INSERT, BTR_DELETE, or BTR_DELETE_MARK. */
		ut_error;
	}

	btr_cur = btr_pcur_get_btr_cur(&pcur);

	/* We should remove the index record if no later version of the row,
	which cannot be purged yet, requires its existence. If some requires,
	we should do nothing. */

	if (row_purge_poss_sec(node, index, entry)) {
		/* Remove the index record, which should have been
		marked for deletion. */
		ut_ad(REC_INFO_DELETED_FLAG
		      & rec_get_info_bits(btr_cur_get_rec(btr_cur),
					  dict_table_is_comp(index->table)));

		btr_cur_pessimistic_delete(&err, FALSE, btr_cur,
					   RB_NONE, &mtr);
		switch (UNIV_EXPECT(err, DB_SUCCESS)) {
		case DB_SUCCESS:
			break;
		case DB_OUT_OF_FILE_SPACE:
			success = FALSE;
			break;
		default:
			ut_error;
		}
	}

func_exit:
	btr_pcur_close(&pcur);
	mtr_commit(&mtr);

	return(success);
}

/***************************************************************
Removes a secondary index entry without modifying the index tree,
if possible. */
static
ibool
row_purge_remove_sec_if_poss_leaf(
/*==============================*/
				/* out: TRUE if success or if not found */
	purge_node_t*	node,	/* in: row purge node */
	dict_index_t*	index,	/* in: index */
	const dtuple_t*	entry)	/* in: index entry */
{
	mtr_t			mtr;
	btr_pcur_t		pcur;
	enum row_search_result	search_result;

	log_free_check();

	mtr_start(&mtr);

	/* Set the purge node for the call to row_purge_poss_sec(). */
	pcur.btr_cur.purge_node = node;
	/* Set the query thread, so that ibuf_insert_low() will be
	able to invoke thd_get_trx(). */
	pcur.btr_cur.thr = que_node_get_parent(node);

	search_result = row_search_index_entry(
		index, entry, BTR_MODIFY_LEAF | BTR_DELETE, &pcur, &mtr);

	btr_pcur_close(&pcur);
	mtr_commit(&mtr);

	switch (search_result) {
	case ROW_NOT_DELETED:
		/* The index entry could not be deleted. */
		return(FALSE);

	case ROW_NOT_DELETED_REF:
		/* The index entry is still needed. */
	case ROW_NOT_FOUND:
		/* The index entry does not exist, nothing to do. */
	case ROW_FOUND:
		/* The index entry existed in the buffer pool
		and was deleted because of the BTR_DELETE. */
	case ROW_BUFFERED:
		/* The deletion was buffered. */
		return(TRUE);
	}

	ut_error;
	return(FALSE);
}

/***************************************************************
Removes a secondary index entry if possible. */
UNIV_INLINE
void
row_purge_remove_sec_if_poss(
/*=========================*/
	purge_node_t*	node,	/* in: row purge node */
	dict_index_t*	index,	/* in: index */
	dtuple_t*	entry)	/* in: index entry */
{
	ibool	success;
	ulint	n_tries		= 0;

	/*	fputs("Purge: Removing secondary record\n", stderr); */

	if (row_purge_remove_sec_if_poss_leaf(node, index, entry)) {

		return;
	}
retry:
	success = row_purge_remove_sec_if_poss_tree(node, index, entry);
	/* The delete operation may fail if we have little
	file space left: TODO: easiest to crash the database
	and restart with more file space */

	if (!success && n_tries < BTR_CUR_RETRY_DELETE_N_TIMES) {

		n_tries++;

		os_thread_sleep(BTR_CUR_RETRY_SLEEP_TIME);

		goto retry;
	}

	ut_a(success);
}

/***************************************************************
Purges a delete marking of a record. */
static
void
row_purge_del_mark(
/*===============*/
	purge_node_t*	node)	/* in: row purge node */
{
	mem_heap_t*	heap;
	dtuple_t*	entry;
	dict_index_t*	index;

	ut_ad(node);

	heap = mem_heap_create(1024);

	while (node->index != NULL) {
		index = node->index;

		/* Build the index entry */
		entry = row_build_index_entry(node->row, NULL, index, heap);
		ut_a(entry);
		row_purge_remove_sec_if_poss(node, index, entry);

		node->index = dict_table_get_next_index(node->index);
	}

	mem_heap_free(heap);

	row_purge_remove_clust_if_poss(node);
}

/***************************************************************
Purges an update of an existing record. Also purges an update of a delete
marked record if that record contained an externally stored field. */
static
void
row_purge_upd_exist_or_extern(
/*==========================*/
	purge_node_t*	node)	/* in: row purge node */
{
	mem_heap_t*	heap;
	dtuple_t*	entry;
	dict_index_t*	index;
	ibool		is_insert;
	ulint		rseg_id;
	ulint		page_no;
	ulint		offset;
	ulint		i;
	mtr_t		mtr;

	ut_ad(node);

	if (node->rec_type == TRX_UNDO_UPD_DEL_REC) {

		goto skip_secondaries;
	}

	heap = mem_heap_create(1024);

	while (node->index != NULL) {
		index = node->index;

		if (row_upd_changes_ord_field_binary(NULL, node->index,
						     node->update)) {
			/* Build the older version of the index entry */
			entry = row_build_index_entry(node->row, NULL,
						      index, heap);
			ut_a(entry);
			row_purge_remove_sec_if_poss(node, index, entry);
		}

		node->index = dict_table_get_next_index(node->index);
	}

	mem_heap_free(heap);

skip_secondaries:
	/* Free possible externally stored fields */
	for (i = 0; i < upd_get_n_fields(node->update); i++) {

		const upd_field_t*	ufield
			= upd_get_nth_field(node->update, i);

		if (dfield_is_ext(&ufield->new_val)) {
			buf_block_t*	block;
			ulint		internal_offset;
			byte*		data_field;

			/* We use the fact that new_val points to
			node->undo_rec and get thus the offset of
			dfield data inside the undo record. Then we
			can calculate from node->roll_ptr the file
			address of the new_val data */

			internal_offset
				= ((const byte*)
				   dfield_get_data(&ufield->new_val))
				- node->undo_rec;

			ut_a(internal_offset < UNIV_PAGE_SIZE);

			trx_undo_decode_roll_ptr(node->roll_ptr,
						 &is_insert, &rseg_id,
						 &page_no, &offset);
			mtr_start(&mtr);

			/* We have to acquire an X-latch to the clustered
			index tree */

			index = dict_table_get_first_index(node->table);

			mtr_x_lock(dict_index_get_lock(index), &mtr);

			/* NOTE: we must also acquire an X-latch to the
			root page of the tree. We will need it when we
			free pages from the tree. If the tree is of height 1,
			the tree X-latch does NOT protect the root page,
			because it is also a leaf page. Since we will have a
			latch on an undo log page, we would break the
			latching order if we would only later latch the
			root page of such a tree! */

			btr_root_get(index, &mtr);

			/* We assume in purge of externally stored fields
			that the space id of the undo log record is 0! */

			block = buf_page_get(0, 0, page_no, RW_X_LATCH, &mtr);
			buf_block_dbg_add_level(block, SYNC_TRX_UNDO_PAGE);

			data_field = buf_block_get_frame(block)
				+ offset + internal_offset;

			ut_a(dfield_get_len(&ufield->new_val)
			     >= BTR_EXTERN_FIELD_REF_SIZE);
			btr_free_externally_stored_field(
				index,
				data_field + dfield_get_len(&ufield->new_val)
				- BTR_EXTERN_FIELD_REF_SIZE,
				NULL, NULL, NULL, 0, RB_NONE, &mtr);
			mtr_commit(&mtr);
		}
	}
}

/***************************************************************
Parses the row reference and other info in a modify undo log record. */
static
ibool
row_purge_parse_undo_rec(
/*=====================*/
				/* out: TRUE if purge operation required:
				NOTE that then the CALLER must unfreeze
				data dictionary! */
	purge_node_t*	node,	/* in: row undo node */
	ibool*		updated_extern,
				/* out: TRUE if an externally stored field
				was updated */
	que_thr_t*	thr)	/* in: query thread */
{
	dict_index_t*	clust_index;
	byte*		ptr;
	trx_t*		trx;
	dulint		undo_no;
	dulint		table_id;
	dulint		trx_id;
	dulint		roll_ptr;
	ulint		info_bits;
	ulint		type;
	ulint		cmpl_info;

	ut_ad(node && thr);

	trx = thr_get_trx(thr);

	ptr = trx_undo_rec_get_pars(node->undo_rec, &type, &cmpl_info,
				    updated_extern, &undo_no, &table_id);
	node->rec_type = type;

	if (type == TRX_UNDO_UPD_DEL_REC && !(*updated_extern)) {

		return(FALSE);
	}

	ptr = trx_undo_update_rec_get_sys_cols(ptr, &trx_id, &roll_ptr,
					       &info_bits);
	node->table = NULL;

	if (type == TRX_UNDO_UPD_EXIST_REC
	    && cmpl_info & UPD_NODE_NO_ORD_CHANGE && !(*updated_extern)) {

		/* Purge requires no changes to indexes: we may return */

		return(FALSE);
	}

	/* Prevent DROP TABLE etc. from running when we are doing the purge
	for this row */

	row_mysql_freeze_data_dictionary(trx);

	mutex_enter(&(dict_sys->mutex));

	node->table = dict_table_get_on_id_low(table_id);

	mutex_exit(&(dict_sys->mutex));

	if (node->table == NULL) {
		/* The table has been dropped: no need to do purge */
err_exit:
		row_mysql_unfreeze_data_dictionary(trx);
		return(FALSE);
	}

	if (node->table->ibd_file_missing) {
		/* We skip purge of missing .ibd files */

		node->table = NULL;

		goto err_exit;
	}

	clust_index = dict_table_get_first_index(node->table);

	if (clust_index == NULL) {
		/* The table was corrupt in the data dictionary */

		goto err_exit;
	}

	ptr = trx_undo_rec_get_row_ref(ptr, clust_index, &(node->ref),
				       node->heap);

	ptr = trx_undo_update_rec_get_update(ptr, clust_index, type, trx_id,
					     roll_ptr, info_bits, trx,
					     node->heap, &(node->update));

	/* Read to the partial row the fields that occur in indexes */

	if (!(cmpl_info & UPD_NODE_NO_ORD_CHANGE)) {
		ptr = trx_undo_rec_get_partial_row(
			ptr, clust_index, &node->row,
			type == TRX_UNDO_UPD_DEL_REC,
			node->heap);
	}

	return(TRUE);
}

/***************************************************************
Fetches an undo log record and does the purge for the recorded operation.
If none left, or the current purge completed, returns the control to the
parent node, which is always a query thread node. */
static
ulint
row_purge(
/*======*/
				/* out: DB_SUCCESS if operation successfully
				completed, else error code */
	purge_node_t*	node,	/* in: row purge node */
	que_thr_t*	thr)	/* in: query thread */
{
	dulint	roll_ptr;
	ibool	purge_needed;
	ibool	updated_extern;
	trx_t*	trx;

	ut_ad(node && thr);

	trx = thr_get_trx(thr);

	node->undo_rec = trx_purge_fetch_next_rec(&roll_ptr,
						  &(node->reservation),
						  node->heap);
	if (!node->undo_rec) {
		/* Purge completed for this query thread */

		thr->run_node = que_node_get_parent(node);

		return(DB_SUCCESS);
	}

	node->roll_ptr = roll_ptr;

	if (node->undo_rec == &trx_purge_dummy_rec) {
		purge_needed = FALSE;
	} else {
		purge_needed = row_purge_parse_undo_rec(node, &updated_extern,
							thr);
		/* If purge_needed == TRUE, we must also remember to unfreeze
		data dictionary! */
	}

	if (purge_needed) {
		node->found_clust = FALSE;

		node->index = dict_table_get_next_index(
			dict_table_get_first_index(node->table));

		if (node->rec_type == TRX_UNDO_DEL_MARK_REC) {
			row_purge_del_mark(node);

		} else if (updated_extern
			   || node->rec_type == TRX_UNDO_UPD_EXIST_REC) {

			row_purge_upd_exist_or_extern(node);
		}

		if (node->found_clust) {
			btr_pcur_close(&(node->pcur));
		}

		row_mysql_unfreeze_data_dictionary(trx);
	}

	/* Do some cleanup */
	trx_purge_rec_release(node->reservation);
	mem_heap_empty(node->heap);

	thr->run_node = node;

	return(DB_SUCCESS);
}

/***************************************************************
Does the purge operation for a single undo log record. This is a high-level
function used in an SQL execution graph. */
UNIV_INTERN
que_thr_t*
row_purge_step(
/*===========*/
				/* out: query thread to run next or NULL */
	que_thr_t*	thr)	/* in: query thread */
{
	purge_node_t*	node;
	ulint		err;

	ut_ad(thr);

	node = thr->run_node;

	ut_ad(que_node_get_type(node) == QUE_NODE_PURGE);

	err = row_purge(node, thr);

	ut_ad(err == DB_SUCCESS);

	return(thr);
}