1
/*****************************************************************************
3
Copyright (c) 1997, 2009, Innobase Oy. All Rights Reserved.
5
This program is free software; you can redistribute it and/or modify it under
6
the terms of the GNU General Public License as published by the Free Software
7
Foundation; version 2 of the License.
9
This program is distributed in the hope that it will be useful, but WITHOUT
10
ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
11
FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
13
You should have received a copy of the GNU General Public License along with
14
this program; if not, write to the Free Software Foundation, Inc., 51 Franklin
15
St, Fifth Floor, Boston, MA 02110-1301 USA
17
*****************************************************************************/
19
/**************************************************//**
23
Created 1/8/1997 Heikki Tuuri
24
*******************************************************/
29
#include "row0undo.ic"
33
#include "mach0data.h"
38
#include "trx0purge.h"
45
#include "row0mysql.h"
48
/* How to undo row operations?
49
(1) For an insert, we have stored a prefix of the clustered index record
50
in the undo log. Using it, we look for the clustered record, and using
51
that we look for the records in the secondary indexes. The insert operation
52
may have been left incomplete, if the database crashed, for example.
53
We may have look at the trx id and roll ptr to make sure the record in the
54
clustered index is really the one for which the undo log record was
55
written. We can use the framework we get from the original insert op.
56
(2) Delete marking: We can use the framework we get from the original
57
delete mark op. We only have to check the trx id.
58
(3) Update: This may be the most complicated. We have to use the framework
59
we get from the original update op.
61
What if the same trx repeatedly deletes and inserts an identical row.
62
Then the row id changes and also roll ptr. What if the row id was not
63
part of the ordering fields in the clustered index? Maybe we have to write
64
it to undo log. Well, maybe not, because if we order the row id and trx id
65
in descending order, then the only undeleted copy is the first in the
66
index. Our searches in row operations always position the cursor before
67
the first record in the result set. But, if there is no key defined for
68
a table, then it would be desirable that row id is in ascending order.
69
So, lets store row id in descending order only if it is not an ordering
70
field in the clustered index.
72
NOTE: Deletes and inserts may lead to situation where there are identical
73
records in a secondary index. Is that a problem in the B-tree? Yes.
74
Also updates can lead to this, unless trx id and roll ptr are included in
76
(1) Fix in clustered indexes: include row id, trx id, and roll ptr
77
in node pointers of B-tree.
78
(2) Fix in secondary indexes: include all fields in node pointers, and
79
if an entry is inserted, check if it is equal to the right neighbor,
80
in which case update the right neighbor: the neighbor must be delete
81
marked, set it unmarked and write the trx id of the current transaction.
83
What if the same trx repeatedly updates the same row, updating a secondary
84
index field or not? Updating a clustered index ordering field?
86
(1) If it does not update the secondary index and not the clustered index
87
ord field. Then the secondary index record stays unchanged, but the
88
trx id in the secondary index record may be smaller than in the clustered
89
index record. This is no problem?
90
(2) If it updates secondary index ord field but not clustered: then in
91
secondary index there are delete marked records, which differ in an
92
ord field. No problem.
93
(3) Updates clustered ord field but not secondary, and secondary index
94
is unique. Then the record in secondary index is just updated at the
98
Problem with duplicate records:
99
Fix 1: Add a trx op no field to all indexes. A problem: if a trx with a
100
bigger trx id has inserted and delete marked a similar row, our trx inserts
101
again a similar row, and a trx with an even bigger id delete marks it. Then
102
the position of the row should change in the index if the trx id affects
103
the alphabetical ordering.
105
Fix 2: If an insert encounters a similar row marked deleted, we turn the
106
insert into an 'update' of the row marked deleted. Then we must write undo
107
info on the update. A problem: what if a purge operation tries to remove
108
the delete marked row?
110
We can think of the database row versions as a linked list which starts
111
from the record in the clustered index, and is linked by roll ptrs
112
through undo logs. The secondary index records are references which tell
113
what kinds of records can be found in this linked list for a record
114
in the clustered index.
116
How to do the purge? A record can be removed from the clustered index
117
if its linked list becomes empty, i.e., the row has been marked deleted
118
and its roll ptr points to the record in the undo log we are going through,
119
doing the purge. Similarly, during a rollback, a record can be removed
120
if the stored roll ptr in the undo log points to a trx already (being) purged,
121
or if the roll ptr is NULL, i.e., it was a fresh insert. */
123
/********************************************************************//**
124
Creates a row undo node to a query graph.
125
@return own: undo node */
128
row_undo_node_create(
129
/*=================*/
130
trx_t* trx, /*!< in: transaction */
131
que_thr_t* parent, /*!< in: parent node, i.e., a thr node */
132
mem_heap_t* heap) /*!< in: memory heap where created */
136
ut_ad(trx && parent && heap);
138
undo = mem_heap_alloc(heap, sizeof(undo_node_t));
140
undo->common.type = QUE_NODE_UNDO;
141
undo->common.parent = parent;
143
undo->state = UNDO_NODE_FETCH_NEXT;
146
btr_pcur_init(&(undo->pcur));
148
undo->heap = mem_heap_create(256);
153
/***********************************************************//**
154
Looks for the clustered index record when node has the row reference.
155
The pcur in node is used in the search. If found, stores the row to node,
156
and stores the position of pcur, and detaches it. The pcur must be closed
157
by the caller in any case.
158
@return TRUE if found; NOTE the node->pcur must be closed by the
159
caller, regardless of the return value */
162
row_undo_search_clust_to_pcur(
163
/*==========================*/
164
undo_node_t* node) /*!< in: row undo node */
166
dict_index_t* clust_index;
171
mem_heap_t* heap = NULL;
172
ulint offsets_[REC_OFFS_NORMAL_SIZE];
173
ulint* offsets = offsets_;
174
rec_offs_init(offsets_);
178
clust_index = dict_table_get_first_index(node->table);
180
found = row_search_on_row_ref(&(node->pcur), BTR_MODIFY_LEAF,
181
node->table, node->ref, &mtr);
183
rec = btr_pcur_get_rec(&(node->pcur));
185
offsets = rec_get_offsets(rec, clust_index, offsets,
186
ULINT_UNDEFINED, &heap);
188
if (!found || 0 != ut_dulint_cmp(node->roll_ptr,
189
row_get_rec_roll_ptr(rec, clust_index,
192
/* We must remove the reservation on the undo log record
193
BEFORE releasing the latch on the clustered index page: this
194
is to make sure that some thread will eventually undo the
195
modification corresponding to node->roll_ptr. */
197
/* fputs("--------------------undoing a previous version\n",
202
node->row = row_build(ROW_COPY_DATA, clust_index, rec,
203
offsets, NULL, &node->ext, node->heap);
205
node->undo_row = dtuple_copy(node->row, node->heap);
206
row_upd_replace(node->undo_row, &node->undo_ext,
207
clust_index, node->update, node->heap);
209
node->undo_row = NULL;
210
node->undo_ext = NULL;
213
btr_pcur_store_position(&(node->pcur), &mtr);
218
btr_pcur_commit_specify_mtr(&(node->pcur), &mtr);
220
if (UNIV_LIKELY_NULL(heap)) {
226
/***********************************************************//**
227
Fetches an undo log record and does the undo for the recorded operation.
228
If none left, or a partial rollback completed, returns control to the
229
parent node, which is always a query thread node.
230
@return DB_SUCCESS if operation successfully completed, else error code */
235
undo_node_t* node, /*!< in: row undo node */
236
que_thr_t* thr) /*!< in: query thread */
241
ibool locked_data_dict;
247
if (node->state == UNDO_NODE_FETCH_NEXT) {
249
node->undo_rec = trx_roll_pop_top_rec_of_trx(trx,
253
if (!node->undo_rec) {
254
/* Rollback completed for this query thread */
256
thr->run_node = que_node_get_parent(node);
261
node->roll_ptr = roll_ptr;
262
node->undo_no = trx_undo_rec_get_undo_no(node->undo_rec);
264
if (trx_undo_roll_ptr_is_insert(roll_ptr)) {
266
node->state = UNDO_NODE_INSERT;
268
node->state = UNDO_NODE_MODIFY;
271
} else if (node->state == UNDO_NODE_PREV_VERS) {
273
/* Undo should be done to the same clustered index record
274
again in this same rollback, restoring the previous version */
276
roll_ptr = node->new_roll_ptr;
278
node->undo_rec = trx_undo_get_undo_rec_low(roll_ptr,
280
node->roll_ptr = roll_ptr;
281
node->undo_no = trx_undo_rec_get_undo_no(node->undo_rec);
283
if (trx_undo_roll_ptr_is_insert(roll_ptr)) {
285
node->state = UNDO_NODE_INSERT;
287
node->state = UNDO_NODE_MODIFY;
291
/* Prevent DROP TABLE etc. while we are rolling back this row.
292
If we are doing a TABLE CREATE or some other dictionary operation,
293
then we already have dict_operation_lock locked in x-mode. Do not
294
try to lock again, because that would cause a hang. */
296
locked_data_dict = (trx->dict_operation_lock_mode == 0);
298
if (locked_data_dict) {
300
row_mysql_lock_data_dictionary(trx);
303
if (node->state == UNDO_NODE_INSERT) {
305
err = row_undo_ins(node);
307
node->state = UNDO_NODE_FETCH_NEXT;
309
ut_ad(node->state == UNDO_NODE_MODIFY);
310
err = row_undo_mod(node, thr);
313
if (locked_data_dict) {
315
row_mysql_unlock_data_dictionary(trx);
318
/* Do some cleanup */
319
btr_pcur_close(&(node->pcur));
321
mem_heap_empty(node->heap);
323
thr->run_node = node;
328
/***********************************************************//**
329
Undoes a row operation in a table. This is a high-level function used
330
in SQL execution graphs.
331
@return query thread to run next or NULL */
336
que_thr_t* thr) /*!< in: query thread */
344
srv_inc_activity_count();
346
trx = thr_get_trx(thr);
348
node = thr->run_node;
350
ut_ad(que_node_get_type(node) == QUE_NODE_UNDO);
352
err = row_undo(node, thr);
354
trx->error_state = err;
356
if (err != DB_SUCCESS) {
357
/* SQL error detected */
359
fprintf(stderr, "InnoDB: Fatal error %lu in rollback.\n",
362
if (err == DB_OUT_OF_FILE_SPACE) {
364
"InnoDB: Error 13 means out of tablespace.\n"
365
"InnoDB: Consider increasing"
366
" your tablespace.\n");