branches/zip: In the rollback of incomplete transactions after crash

recovery, tolerate clustered index records whose externally stored
columns have not been written.  This should remove the assertion failures
that were reported as Mantis issue#58, issue#62, issue#64.

trx_is_recv(): New function: TRUE if this transaction is rolling back
an incomplete transaction in crash recovery.

enum trx_rbmode: Rollback modes: no rollback, normal rollback, crash recovery.

btr_cur_pessimistic_delete(), btr_free_externally_stored_field(),
btr_rec_free_externally_stored_fields():
Replace the ibool parameter with enum trx_rbmode.

btr_free_externally_stored_field(): If field_ref is zero, return
but assert ut_a(rbmode == RB_RECOVERY).  Unless InnoDB has crashed
while inserting a clustered index record, field_ref should not be zero.

btr_rec_free_updated_extern_fields(): Add the parameter enum trx_rbmode.

btr_cur_pessimistic_update(): Pass the rbmode parameter to
btr_rec_free_updated_extern_fields().

row_undo_ins(), row_undo_mod_upd_del_sec(): If row_build_index_entry()
fails, assert trx_is_recv() and skip this secondary index.

row_undo_mod_upd_del_sec(): Empty the heap at the end of each loop
iteration in order to conserve memory and to reduce the number of
low-level memory allocations.

btr/btr0btr.c

@@ -2202,7 +2202,7 @@ btr_node_ptr_delete(
 	/* Delete node pointer on father page */
 	btr_page_get_father(index, block, mtr, &cursor);
 
-	compressed = btr_cur_pessimistic_delete(&err, TRUE, &cursor, FALSE,
+	compressed = btr_cur_pessimistic_delete(&err, TRUE, &cursor, RB_NONE,
 						mtr);
 	ut_a(err == DB_SUCCESS);
 

btr/btr0cur.c

@@ -32,6 +32,7 @@ Created 10/16/1994 Heikki Tuuri
 #include "btr0sea.h"
 #include "row0upd.h"
 #include "trx0rec.h"
+#include "trx0roll.h" /* trx_is_recv() */
 #include "que0que.h"
 #include "row0row.h"
 #include "srv0srv.h"
@@ -112,6 +113,7 @@ btr_rec_free_updated_extern_fields(
 				part will be updated, or NULL */
 	const ulint*	offsets,/* in: rec_get_offsets(rec, index) */
 	const upd_t*	update,	/* in: update vector */
+	enum trx_rbmode	rbmode,	/* in: rollback mode */
 	mtr_t*		mtr);	/* in: mini-transaction handle which contains
 				an X-latch to record page and to the tree */
 /***************************************************************
@@ -126,9 +128,7 @@ btr_rec_free_externally_stored_fields(
 	const ulint*	offsets,/* in: rec_get_offsets(rec, index) */
 	page_zip_des_t*	page_zip,/* in: compressed page whose uncompressed
 				part will be updated, or NULL */
-	ibool		do_not_free_inherited,/* in: TRUE if called in a
-				rollback and we do not want to free
-				inherited fields */
+	enum trx_rbmode	rbmode,	/* in: rollback mode */
 	mtr_t*		mtr);	/* in: mini-transaction handle which contains
 				an X-latch to record page and to the index
 				tree */
@@ -2162,8 +2162,9 @@ btr_cur_pessimistic_update(
 
 		ut_ad(big_rec_vec == NULL);
 
-		btr_rec_free_updated_extern_fields(index, rec, page_zip,
-						   offsets, update, mtr);
+		btr_rec_free_updated_extern_fields(
+			index, rec, page_zip, offsets, update,
+			trx_is_recv(trx) ? RB_RECOVERY : RB_NORMAL, mtr);
 	}
 
 	/* We have to set appropriate extern storage bits in the new
@@ -2811,7 +2812,7 @@ btr_cur_pessimistic_delete(
 				if compression does not occur, the cursor
 				stays valid: it points to successor of
 				deleted record on function exit */
-	ibool		in_rollback,/* in: TRUE if called in rollback */
+	enum trx_rbmode	rbmode,	/* in: rollback mode */
 	mtr_t*		mtr)	/* in: mtr */
 {
 	buf_block_t*	block;
@@ -2865,7 +2866,7 @@ btr_cur_pessimistic_delete(
 	if (rec_offs_any_extern(offsets)) {
 		btr_rec_free_externally_stored_fields(index,
 						      rec, offsets, page_zip,
-						      in_rollback, mtr);
+						      rbmode, mtr);
 #ifdef UNIV_ZIP_DEBUG
 		ut_a(!page_zip || page_zip_validate(page_zip, page));
 #endif /* UNIV_ZIP_DEBUG */
@@ -4084,9 +4085,7 @@ btr_free_externally_stored_field(
 					to rec, or NULL if rec == NULL */
 	ulint		i,		/* in: field number of field_ref;
 					ignored if rec == NULL */
-	ibool		do_not_free_inherited,/* in: TRUE if called in a
-					rollback and we do not want to free
-					inherited fields */
+	enum trx_rbmode	rbmode,		/* in: rollback mode */
 	mtr_t*		local_mtr __attribute__((unused))) /* in: mtr
 					containing the latch to data an an
 					X-latch to the index tree */
@@ -4116,6 +4115,15 @@ btr_free_externally_stored_field(
 	}
 #endif /* UNIV_DEBUG */
 
+	if (UNIV_UNLIKELY(!memcmp(field_ref, field_ref_zero,
+				  BTR_EXTERN_FIELD_REF_SIZE))) {
+		/* In the rollback of uncommitted transactions, we may
+		encounter a clustered index record whose BLOBs have
+		not been written.  There is nothing to free then. */
+		ut_a(rbmode == RB_RECOVERY);
+		return;
+	}
+
 	space_id = mach_read_from_4(field_ref + BTR_EXTERN_SPACE_ID);
 
 	if (UNIV_UNLIKELY(space_id != dict_index_get_space(index))) {
@@ -4160,7 +4168,7 @@ btr_free_externally_stored_field(
 		    || (mach_read_from_1(field_ref + BTR_EXTERN_LEN)
 			& BTR_EXTERN_OWNER_FLAG)
 		    /* Rollback and inherited field */
-		    || (do_not_free_inherited
+		    || (rbmode != RB_NONE
 			&& (mach_read_from_1(field_ref + BTR_EXTERN_LEN)
 			    & BTR_EXTERN_INHERITED_FLAG))) {
 
@@ -4262,9 +4270,7 @@ btr_rec_free_externally_stored_fields(
 	const ulint*	offsets,/* in: rec_get_offsets(rec, index) */
 	page_zip_des_t*	page_zip,/* in: compressed page whose uncompressed
 				part will be updated, or NULL */
-	ibool		do_not_free_inherited,/* in: TRUE if called in a
-				rollback and we do not want to free
-				inherited fields */
+	enum trx_rbmode	rbmode,	/* in: rollback mode */
 	mtr_t*		mtr)	/* in: mini-transaction handle which contains
 				an X-latch to record page and to the index
 				tree */
@@ -4288,8 +4294,7 @@ btr_rec_free_externally_stored_fields(
 
 			btr_free_externally_stored_field(
 				index, data + len - BTR_EXTERN_FIELD_REF_SIZE,
-				rec, offsets, page_zip, i,
-				do_not_free_inherited, mtr);
+				rec, offsets, page_zip, i, rbmode, mtr);
 		}
 	}
 }
@@ -4308,6 +4313,7 @@ btr_rec_free_updated_extern_fields(
 				part will be updated, or NULL */
 	const ulint*	offsets,/* in: rec_get_offsets(rec, index) */
 	const upd_t*	update,	/* in: update vector */
+	enum trx_rbmode	rbmode,	/* in: rollback mode */
 	mtr_t*		mtr)	/* in: mini-transaction handle which contains
 				an X-latch to record page and to the tree */
 {
@@ -4333,7 +4339,7 @@ btr_rec_free_updated_extern_fields(
 			btr_free_externally_stored_field(
 				index, data + len - BTR_EXTERN_FIELD_REF_SIZE,
 				rec, offsets, page_zip,
-				ufield->field_no, TRUE, mtr);
+				ufield->field_no, rbmode, mtr);
 		}
 	}
 }

ibuf/ibuf0ibuf.c

@@ -3140,7 +3140,7 @@ ibuf_delete_rec(
 	root = ibuf_tree_root_get(ibuf_data, 0, mtr);
 
 	btr_cur_pessimistic_delete(&err, TRUE, btr_pcur_get_btr_cur(pcur),
-				   FALSE, mtr);
+				   RB_NONE, mtr);
 	ut_a(err == DB_SUCCESS);
 
 #ifdef UNIV_IBUF_COUNT_DEBUG

include/btr0cur.h

@@ -379,7 +379,7 @@ btr_cur_pessimistic_delete(
 				if compression does not occur, the cursor
 				stays valid: it points to successor of
 				deleted record on function exit */
-	ibool		in_rollback,/* in: TRUE if called in rollback */
+	enum trx_rbmode	rbmode,	/* in: rollback mode */
 	mtr_t*		mtr);	/* in: mtr */
 /***************************************************************
 Parses a redo log record of updating a record in-place. */
@@ -521,9 +521,7 @@ btr_free_externally_stored_field(
 					to rec, or NULL if rec == NULL */
 	ulint		i,		/* in: field number of field_ref;
 					ignored if rec == NULL */
-	ibool		do_not_free_inherited,/* in: TRUE if called in a
-					rollback and we do not want to free
-					inherited fields */
+	enum trx_rbmode	rbmode,		/* in: rollback mode */
 	mtr_t*		local_mtr);	/* in: mtr containing the latch to
 					data an an X-latch to the index
 					tree */

include/trx0roll.h

@@ -15,6 +15,17 @@ Created 3/26/1996 Heikki Tuuri
 #include "mtr0mtr.h"
 #include "trx0sys.h"
 
+/***********************************************************************
+Determines if this transaction is rolling back an incomplete transaction
+in crash recovery. */
+UNIV_INTERN
+ibool
+trx_is_recv(
+/*========*/
+				/* out: TRUE if trx is an incomplete
+				transaction that is being rolled back
+				in crash recovery */
+	const trx_t*	trx);	/* in: transaction */
 /***********************************************************************
 Returns a transaction savepoint taken at this point in time. */
 UNIV_INTERN

include/trx0types.h

@@ -36,6 +36,14 @@ typedef struct roll_node_struct	roll_node_t;
 typedef struct commit_node_struct commit_node_t;
 typedef struct trx_named_savept_struct trx_named_savept_t;
 
+/* Rollback modes */
+enum trx_rbmode {
+	RB_NONE = 0,	/* no rollback */
+	RB_NORMAL,	/* normal rollback */
+	RB_RECOVERY,	/* rolling back an incomplete transaction,
+			in crash recovery */
+};
+
 /* Transaction savepoint */
 typedef struct trx_savept_struct trx_savept_t;
 struct trx_savept_struct{

row/row0purge.c

@@ -144,7 +144,8 @@ row_purge_remove_clust_if_poss_low(
 		success = btr_cur_optimistic_delete(btr_cur, &mtr);
 	} else {
 		ut_ad(mode == BTR_MODIFY_TREE);
-		btr_cur_pessimistic_delete(&err, FALSE, btr_cur, FALSE, &mtr);
+		btr_cur_pessimistic_delete(&err, FALSE, btr_cur,
+					   RB_NONE, &mtr);
 
 		if (err == DB_SUCCESS) {
 			success = TRUE;
@@ -265,7 +266,7 @@ row_purge_remove_sec_if_poss_low(
 		} else {
 			ut_ad(mode == BTR_MODIFY_TREE);
 			btr_cur_pessimistic_delete(&err, FALSE, btr_cur,
-						   FALSE, &mtr);
+						   RB_NONE, &mtr);
 			success = err == DB_SUCCESS;
 			ut_a(success || err == DB_OUT_OF_FILE_SPACE);
 		}
@@ -458,7 +459,7 @@ skip_secondaries:
 				index,
 				data_field + dfield_get_len(&ufield->new_val)
 				- BTR_EXTERN_FIELD_REF_SIZE,
-				NULL, NULL, NULL, 0, FALSE, &mtr);
+				NULL, NULL, NULL, 0, RB_NONE, &mtr);
 			mtr_commit(&mtr);
 		}
 	}

row/row0uins.c

@@ -87,7 +87,10 @@ retry:
 					    &(node->pcur), &mtr);
 	ut_a(success);
 
-	btr_cur_pessimistic_delete(&err, FALSE, btr_cur, TRUE, &mtr);
+	btr_cur_pessimistic_delete(&err, FALSE, btr_cur,
+				   trx_is_recv(node->trx)
+				   ? RB_RECOVERY
+				   : RB_NORMAL, &mtr);
 
 	/* The delete operation may fail if we have little
 	file space left: TODO: easiest to crash the database
@@ -160,7 +163,14 @@ row_undo_ins_remove_sec_low(
 	} else {
 		ut_ad(mode == BTR_MODIFY_TREE);
 
-		btr_cur_pessimistic_delete(&err, FALSE, btr_cur, TRUE, &mtr);
+		/* No need to distinguish RB_RECOVERY here, because we
+		are deleting a secondary index record: the distinction
+		between RB_NORMAL and RB_RECOVERY only matters when
+		deleting a record that contains externally stored
+		columns. */
+		ut_ad(!dict_index_is_clust(index));
+		btr_cur_pessimistic_delete(&err, FALSE, btr_cur,
+					   RB_NORMAL, &mtr);
 	}
 
 	btr_pcur_close(&pcur);
@@ -284,13 +294,25 @@ row_undo_ins(
 
 		entry = row_build_index_entry(node->row, node->ext,
 					      node->index, node->heap);
-		ut_a(entry);
+		if (UNIV_UNLIKELY(!entry)) {
+			/* The database must have crashed after
+			inserting a clustered index record but before
+			writing all the externally stored columns of
+			that record.  Because secondary index entries
+			are inserted after the clustered index record,
+			we may assume that the secondary index record
+			does not exist.  However, this situation may
+			only occur during the rollback of incomplete
+			transactions. */
+			ut_a(trx_is_recv(node->trx));
+		} else {
 			err = row_undo_ins_remove_sec(node->index, entry);
 
 			if (err != DB_SUCCESS) {
 
 				return(err);
 			}
+		}
 
 		node->index = dict_table_get_next_index(node->index);
 	}

row/row0umod.c

@@ -178,9 +178,9 @@ row_undo_mod_remove_clust_low(
 
 		/* Note that since this operation is analogous to purge,
 		we can free also inherited externally stored fields:
-		hence the last FALSE in the call below */
+		hence the RB_NONE in the call below */
 
-		btr_cur_pessimistic_delete(&err, FALSE, btr_cur, FALSE, mtr);
+		btr_cur_pessimistic_delete(&err, FALSE, btr_cur, RB_NONE, mtr);
 
 		/* The delete operation may fail if we have little
 		file space left: TODO: easiest to crash the database
@@ -350,8 +350,14 @@ row_undo_mod_del_mark_or_remove_sec_low(
 		} else {
 			ut_ad(mode == BTR_MODIFY_TREE);
 
+			/* No need to distinguish RB_RECOVERY here, because we
+			are deleting a secondary index record: the distinction
+			between RB_NORMAL and RB_RECOVERY only matters when
+			deleting a record that contains externally stored
+			columns. */
+			ut_ad(!dict_index_is_clust(index));
 			btr_cur_pessimistic_delete(&err, FALSE, btr_cur,
-						   TRUE, &mtr);
+						   RB_NORMAL, &mtr);
 
 			/* The delete operation may fail if we have little
 			file space left: TODO: easiest to crash the database
@@ -506,7 +512,7 @@ row_undo_mod_upd_del_sec(
 	mem_heap_t*	heap;
 	dtuple_t*	entry;
 	dict_index_t*	index;
-	ulint		err;
+	ulint		err	= DB_SUCCESS;
 
 	heap = mem_heap_create(1024);
 
@@ -515,22 +521,35 @@ row_undo_mod_upd_del_sec(
 
 		entry = row_build_index_entry(node->row, node->ext,
 					      index, heap);
-		ut_a(entry);
-		err = row_undo_mod_del_mark_or_remove_sec(node, thr, index,
-							  entry);
-		if (err != DB_SUCCESS) {
+		if (UNIV_UNLIKELY(!entry)) {
+			/* The database must have crashed after
+			inserting a clustered index record but before
+			writing all the externally stored columns of
+			that record.  Because secondary index entries
+			are inserted after the clustered index record,
+			we may assume that the secondary index record
+			does not exist.  However, this situation may
+			only occur during the rollback of incomplete
+			transactions. */
+			ut_a(trx_is_recv(thr_get_trx(thr)));
+		} else {
+			err = row_undo_mod_del_mark_or_remove_sec(
+				node, thr, index, entry);
 
-			mem_heap_free(heap);
+			if (err != DB_SUCCESS) {
 
-			return(err);
+				break;
+			}
 		}
 
+		mem_heap_empty(heap);
+
 		node->index = dict_table_get_next_index(node->index);
 	}
 
 	mem_heap_free(heap);
 
-	return(DB_SUCCESS);
+	return(err);
 }
 
 /***************************************************************

trx/trx0roll.c

@@ -373,6 +373,21 @@ trx_release_savepoint_for_mysql(
 	return(DB_SUCCESS);
 }
 
+/***********************************************************************
+Determines if this transaction is rolling back an incomplete transaction
+in crash recovery. */
+UNIV_INTERN
+ibool
+trx_is_recv(
+/*========*/
+				/* out: TRUE if trx is an incomplete
+				transaction that is being rolled back
+				in crash recovery */
+	const trx_t*	trx)	/* in: transaction */
+{
+	return(trx == trx_roll_crash_recv_trx);
+}
+
 /***********************************************************************
 Returns a transaction savepoint taken at this point in time. */
 UNIV_INTERN