« back to all changes in this revision
Viewing changes to storage/innobase/row/row0sel.c.orig
- browse files at revision 563
- compare with another revision
- compare with revision 1757.1.2
- download diff
- download tarball
- view history from revision 563
- Committer: Brian Aker
- Date: 2008-10-28 08:36:02 UTC
- mfrom: (520.4.13 merge-innodb-plugin)
- Revision ID: brian@tangent.org-20081028083602-0p3zzlhlxr5q2sqo
Merging Monty's work
- files added:
- drizzled/query_id.cc
- storage/innobase/handler
- storage/innobase/mysql-test
- storage/innobase/scripts
- files renamed:
- storage/innobase/handler/ => storage/innobase/drizzle/
- files modified:
- configure.ac
added
removed
storage/innobase/row/row0sel.c.orig
1
/*******************************************************
2
Select
3
4
(c) 1997 Innobase Oy
5
6
Created 12/19/1997 Heikki Tuuri
7
*******************************************************/
8
9
#include "row0sel.h"
10
11
#ifdef UNIV_NONINL
12
#include "row0sel.ic"
13
#endif
14
15
#include "dict0dict.h"
16
#include "dict0boot.h"
17
#include "trx0undo.h"
18
#include "trx0trx.h"
19
#include "btr0btr.h"
20
#include "btr0cur.h"
21
#include "btr0sea.h"
22
#include "mach0data.h"
23
#include "que0que.h"
24
#include "row0upd.h"
25
#include "row0row.h"
26
#include "row0vers.h"
27
#include "rem0cmp.h"
28
#include "lock0lock.h"
29
#include "eval0eval.h"
30
#include "pars0sym.h"
31
#include "pars0pars.h"
32
#include "row0mysql.h"
33
#include "read0read.h"
34
#include "buf0lru.h"
35
36
/* Maximum number of rows to prefetch; MySQL interface has another parameter */
37
#define SEL_MAX_N_PREFETCH 16
38
39
/* Number of rows fetched, after which to start prefetching; MySQL interface
40
has another parameter */
41
#define SEL_PREFETCH_LIMIT 1
42
43
/* When a select has accessed about this many pages, it returns control back
44
to que_run_threads: this is to allow canceling runaway queries */
45
46
#define SEL_COST_LIMIT 100
47
48
/* Flags for search shortcut */
49
#define SEL_FOUND 0
50
#define SEL_EXHAUSTED 1
51
#define SEL_RETRY 2
52
53
/************************************************************************
54
Returns TRUE if the user-defined column in a secondary index record
55
is alphabetically the same as the corresponding BLOB column in the clustered
56
index record.
57
NOTE: the comparison is NOT done as a binary comparison, but character
58
fields are compared with collation! */
59
static
60
ibool
61
row_sel_sec_rec_is_for_blob(
62
/*========================*/
63
/* out: TRUE if the columns
64
are equal */
65
ulint mtype, /* in: main type */
66
ulint prtype, /* in: precise type */
67
ulint mbminlen, /* in: minimum length of a
68
multi-byte character */
69
ulint mbmaxlen, /* in: maximum length of a
70
multi-byte character */
71
const byte* clust_field, /* in: the locally stored part of
72
the clustered index column, including
73
the BLOB pointer; the clustered
74
index record must be covered by
75
a lock or a page latch to protect it
76
against deletion (rollback or purge) */
77
ulint clust_len, /* in: length of clust_field */
78
const byte* sec_field, /* in: column in secondary index */
79
ulint sec_len, /* in: length of sec_field */
80
ulint zip_size) /* in: compressed page size, or 0 */
81
{
82
ulint len;
83
byte buf[DICT_MAX_INDEX_COL_LEN];
84
85
len = btr_copy_externally_stored_field_prefix(buf, sizeof buf,
86
zip_size,
87
clust_field, clust_len);
88
len = dtype_get_at_most_n_mbchars(prtype, mbminlen, mbmaxlen,
89
sec_len, len, (const char*) buf);
90
91
return(!cmp_data_data(mtype, prtype, buf, len, sec_field, sec_len));
92
}
93
94
/************************************************************************
95
Returns TRUE if the user-defined column values in a secondary index record
96
are alphabetically the same as the corresponding columns in the clustered
97
index record.
98
NOTE: the comparison is NOT done as a binary comparison, but character
99
fields are compared with collation! */
100
static
101
ibool
102
row_sel_sec_rec_is_for_clust_rec(
103
/*=============================*/
104
/* out: TRUE if the secondary
105
record is equal to the corresponding
106
fields in the clustered record,
107
when compared with collation */
108
const rec_t* sec_rec, /* in: secondary index record */
109
dict_index_t* sec_index, /* in: secondary index */
110
const rec_t* clust_rec, /* in: clustered index record;
111
must be protected by a lock or
112
a page latch against deletion
113
in rollback or purge */
114
dict_index_t* clust_index) /* in: clustered index */
115
{
116
const byte* sec_field;
117
ulint sec_len;
118
const byte* clust_field;
119
ulint n;
120
ulint i;
121
mem_heap_t* heap = NULL;
122
ulint clust_offsets_[REC_OFFS_NORMAL_SIZE];
123
ulint sec_offsets_[REC_OFFS_SMALL_SIZE];
124
ulint* clust_offs = clust_offsets_;
125
ulint* sec_offs = sec_offsets_;
126
ibool is_equal = TRUE;
127
128
rec_offs_init(clust_offsets_);
129
rec_offs_init(sec_offsets_);
130
131
if (rec_get_deleted_flag(clust_rec,
132
dict_table_is_comp(clust_index->table))) {
133
134
/* The clustered index record is delete-marked;
135
it is not visible in the read view. Besides,
136
if there are any externally stored columns,
137
some of them may have already been purged. */
138
return(FALSE);
139
}
140
141
clust_offs = rec_get_offsets(clust_rec, clust_index, clust_offs,
142
ULINT_UNDEFINED, &heap);
143
sec_offs = rec_get_offsets(sec_rec, sec_index, sec_offs,
144
ULINT_UNDEFINED, &heap);
145
146
n = dict_index_get_n_ordering_defined_by_user(sec_index);
147
148
for (i = 0; i < n; i++) {
149
const dict_field_t* ifield;
150
const dict_col_t* col;
151
ulint clust_pos;
152
ulint clust_len;
153
ulint len;
154
155
ifield = dict_index_get_nth_field(sec_index, i);
156
col = dict_field_get_col(ifield);
157
clust_pos = dict_col_get_clust_pos(col, clust_index);
158
159
clust_field = rec_get_nth_field(
160
clust_rec, clust_offs, clust_pos, &clust_len);
161
sec_field = rec_get_nth_field(sec_rec, sec_offs, i, &sec_len);
162
163
len = clust_len;
164
165
if (ifield->prefix_len > 0 && len != UNIV_SQL_NULL) {
166
167
if (rec_offs_nth_extern(clust_offs, clust_pos)) {
168
len -= BTR_EXTERN_FIELD_REF_SIZE;
169
}
170
171
len = dtype_get_at_most_n_mbchars(
172
col->prtype, col->mbminlen, col->mbmaxlen,
173
ifield->prefix_len, len, (char*) clust_field);
174
175
if (rec_offs_nth_extern(clust_offs, clust_pos)
176
&& len < sec_len) {
177
if (!row_sel_sec_rec_is_for_blob(
178
col->mtype, col->prtype,
179
col->mbminlen, col->mbmaxlen,
180
clust_field, clust_len,
181
sec_field, sec_len,
182
dict_table_zip_size(
183
clust_index->table))) {
184
goto inequal;
185
}
186
187
continue;
188
}
189
}
190
191
if (0 != cmp_data_data(col->mtype, col->prtype,
192
clust_field, len,
193
sec_field, sec_len)) {
194
inequal:
195
is_equal = FALSE;
196
goto func_exit;
197
}
198
}
199
200
func_exit:
201
if (UNIV_LIKELY_NULL(heap)) {
202
mem_heap_free(heap);
203
}
204
return(is_equal);
205
}
206
207
/*************************************************************************
208
Creates a select node struct. */
209
UNIV_INTERN
210
sel_node_t*
211
sel_node_create(
212
/*============*/
213
/* out, own: select node struct */
214
mem_heap_t* heap) /* in: memory heap where created */
215
{
216
sel_node_t* node;
217
218
node = mem_heap_alloc(heap, sizeof(sel_node_t));
219
node->common.type = QUE_NODE_SELECT;
220
node->state = SEL_NODE_OPEN;
221
222
node->select_will_do_update = FALSE;
223
node->latch_mode = BTR_SEARCH_LEAF;
224
225
node->plans = NULL;
226
227
return(node);
228
}
229
230
/*************************************************************************
231
Frees the memory private to a select node when a query graph is freed,
232
does not free the heap where the node was originally created. */
233
UNIV_INTERN
234
void
235
sel_node_free_private(
236
/*==================*/
237
sel_node_t* node) /* in: select node struct */
238
{
239
ulint i;
240
plan_t* plan;
241
242
if (node->plans != NULL) {
243
for (i = 0; i < node->n_tables; i++) {
244
plan = sel_node_get_nth_plan(node, i);
245
246
btr_pcur_close(&(plan->pcur));
247
btr_pcur_close(&(plan->clust_pcur));
248
249
if (plan->old_vers_heap) {
250
mem_heap_free(plan->old_vers_heap);
251
}
252
}
253
}
254
}
255
256
/*************************************************************************
257
Evaluates the values in a select list. If there are aggregate functions,
258
their argument value is added to the aggregate total. */
259
UNIV_INLINE
260
void
261
sel_eval_select_list(
262
/*=================*/
263
sel_node_t* node) /* in: select node */
264
{
265
que_node_t* exp;
266
267
exp = node->select_list;
268
269
while (exp) {
270
eval_exp(exp);
271
272
exp = que_node_get_next(exp);
273
}
274
}
275
276
/*************************************************************************
277
Assigns the values in the select list to the possible into-variables in
278
SELECT ... INTO ... */
279
UNIV_INLINE
280
void
281
sel_assign_into_var_values(
282
/*=======================*/
283
sym_node_t* var, /* in: first variable in a list of variables */
284
sel_node_t* node) /* in: select node */
285
{
286
que_node_t* exp;
287
288
if (var == NULL) {
289
290
return;
291
}
292
293
exp = node->select_list;
294
295
while (var) {
296
ut_ad(exp);
297
298
eval_node_copy_val(var->alias, exp);
299
300
exp = que_node_get_next(exp);
301
var = que_node_get_next(var);
302
}
303
}
304
305
/*************************************************************************
306
Resets the aggregate value totals in the select list of an aggregate type
307
query. */
308
UNIV_INLINE
309
void
310
sel_reset_aggregate_vals(
311
/*=====================*/
312
sel_node_t* node) /* in: select node */
313
{
314
func_node_t* func_node;
315
316
ut_ad(node->is_aggregate);
317
318
func_node = node->select_list;
319
320
while (func_node) {
321
eval_node_set_int_val(func_node, 0);
322
323
func_node = que_node_get_next(func_node);
324
}
325
326
node->aggregate_already_fetched = FALSE;
327
}
328
329
/*************************************************************************
330
Copies the input variable values when an explicit cursor is opened. */
331
UNIV_INLINE
332
void
333
row_sel_copy_input_variable_vals(
334
/*=============================*/
335
sel_node_t* node) /* in: select node */
336
{
337
sym_node_t* var;
338
339
var = UT_LIST_GET_FIRST(node->copy_variables);
340
341
while (var) {
342
eval_node_copy_val(var, var->alias);
343
344
var->indirection = NULL;
345
346
var = UT_LIST_GET_NEXT(col_var_list, var);
347
}
348
}
349
350
/*************************************************************************
351
Fetches the column values from a record. */
352
static
353
void
354
row_sel_fetch_columns(
355
/*==================*/
356
dict_index_t* index, /* in: record index */
357
const rec_t* rec, /* in: record in a clustered or non-clustered
358
index; must be protected by a page latch */
359
const ulint* offsets,/* in: rec_get_offsets(rec, index) */
360
sym_node_t* column) /* in: first column in a column list, or
361
NULL */
362
{
363
dfield_t* val;
364
ulint index_type;
365
ulint field_no;
366
const byte* data;
367
ulint len;
368
369
ut_ad(rec_offs_validate(rec, index, offsets));
370
371
if (dict_index_is_clust(index)) {
372
index_type = SYM_CLUST_FIELD_NO;
373
} else {
374
index_type = SYM_SEC_FIELD_NO;
375
}
376
377
while (column) {
378
mem_heap_t* heap = NULL;
379
ibool needs_copy;
380
381
field_no = column->field_nos[index_type];
382
383
if (field_no != ULINT_UNDEFINED) {
384
385
if (UNIV_UNLIKELY(rec_offs_nth_extern(offsets,
386
field_no))) {
387
388
/* Copy an externally stored field to the
389
temporary heap */
390
391
heap = mem_heap_create(1);
392
393
data = btr_rec_copy_externally_stored_field(
394
rec, offsets,
395
dict_table_zip_size(index->table),
396
field_no, &len, heap);
397
398
ut_a(len != UNIV_SQL_NULL);
399
400
needs_copy = TRUE;
401
} else {
402
data = rec_get_nth_field(rec, offsets,
403
field_no, &len);
404
405
if (len == UNIV_SQL_NULL) {
406
len = UNIV_SQL_NULL;
407
}
408
409
needs_copy = column->copy_val;
410
}
411
412
if (needs_copy) {
413
eval_node_copy_and_alloc_val(column, data,
414
len);
415
} else {
416
val = que_node_get_val(column);
417
dfield_set_data(val, data, len);
418
}
419
420
if (UNIV_LIKELY_NULL(heap)) {
421
mem_heap_free(heap);
422
}
423
}
424
425
column = UT_LIST_GET_NEXT(col_var_list, column);
426
}
427
}
428
429
/*************************************************************************
430
Allocates a prefetch buffer for a column when prefetch is first time done. */
431
static
432
void
433
sel_col_prefetch_buf_alloc(
434
/*=======================*/
435
sym_node_t* column) /* in: symbol table node for a column */
436
{
437
sel_buf_t* sel_buf;
438
ulint i;
439
440
ut_ad(que_node_get_type(column) == QUE_NODE_SYMBOL);
441
442
column->prefetch_buf = mem_alloc(SEL_MAX_N_PREFETCH
443
* sizeof(sel_buf_t));
444
for (i = 0; i < SEL_MAX_N_PREFETCH; i++) {
445
sel_buf = column->prefetch_buf + i;
446
447
sel_buf->data = NULL;
448
449
sel_buf->val_buf_size = 0;
450
}
451
}
452
453
/*************************************************************************
454
Frees a prefetch buffer for a column, including the dynamically allocated
455
memory for data stored there. */
456
UNIV_INTERN
457
void
458
sel_col_prefetch_buf_free(
459
/*======================*/
460
sel_buf_t* prefetch_buf) /* in, own: prefetch buffer */
461
{
462
sel_buf_t* sel_buf;
463
ulint i;
464
465
for (i = 0; i < SEL_MAX_N_PREFETCH; i++) {
466
sel_buf = prefetch_buf + i;
467
468
if (sel_buf->val_buf_size > 0) {
469
470
mem_free(sel_buf->data);
471
}
472
}
473
}
474
475
/*************************************************************************
476
Pops the column values for a prefetched, cached row from the column prefetch
477
buffers and places them to the val fields in the column nodes. */
478
static
479
void
480
sel_pop_prefetched_row(
481
/*===================*/
482
plan_t* plan) /* in: plan node for a table */
483
{
484
sym_node_t* column;
485
sel_buf_t* sel_buf;
486
dfield_t* val;
487
byte* data;
488
ulint len;
489
ulint val_buf_size;
490
491
ut_ad(plan->n_rows_prefetched > 0);
492
493
column = UT_LIST_GET_FIRST(plan->columns);
494
495
while (column) {
496
val = que_node_get_val(column);
497
498
if (!column->copy_val) {
499
/* We did not really push any value for the
500
column */
501
502
ut_ad(!column->prefetch_buf);
503
ut_ad(que_node_get_val_buf_size(column) == 0);
504
ut_d(dfield_set_null(val));
505
506
goto next_col;
507
}
508
509
ut_ad(column->prefetch_buf);
510
ut_ad(!dfield_is_ext(val));
511
512
sel_buf = column->prefetch_buf + plan->first_prefetched;
513
514
data = sel_buf->data;
515
len = sel_buf->len;
516
val_buf_size = sel_buf->val_buf_size;
517
518
/* We must keep track of the allocated memory for
519
column values to be able to free it later: therefore
520
we swap the values for sel_buf and val */
521
522
sel_buf->data = dfield_get_data(val);
523
sel_buf->len = dfield_get_len(val);
524
sel_buf->val_buf_size = que_node_get_val_buf_size(column);
525
526
dfield_set_data(val, data, len);
527
que_node_set_val_buf_size(column, val_buf_size);
528
next_col:
529
column = UT_LIST_GET_NEXT(col_var_list, column);
530
}
531
532
plan->n_rows_prefetched--;
533
534
plan->first_prefetched++;
535
}
536
537
/*************************************************************************
538
Pushes the column values for a prefetched, cached row to the column prefetch
539
buffers from the val fields in the column nodes. */
540
UNIV_INLINE
541
void
542
sel_push_prefetched_row(
543
/*====================*/
544
plan_t* plan) /* in: plan node for a table */
545
{
546
sym_node_t* column;
547
sel_buf_t* sel_buf;
548
dfield_t* val;
549
byte* data;
550
ulint len;
551
ulint pos;
552
ulint val_buf_size;
553
554
if (plan->n_rows_prefetched == 0) {
555
pos = 0;
556
plan->first_prefetched = 0;
557
} else {
558
pos = plan->n_rows_prefetched;
559
560
/* We have the convention that pushing new rows starts only
561
after the prefetch stack has been emptied: */
562
563
ut_ad(plan->first_prefetched == 0);
564
}
565
566
plan->n_rows_prefetched++;
567
568
ut_ad(pos < SEL_MAX_N_PREFETCH);
569
570
column = UT_LIST_GET_FIRST(plan->columns);
571
572
while (column) {
573
if (!column->copy_val) {
574
/* There is no sense to push pointers to database
575
page fields when we do not keep latch on the page! */
576
577
goto next_col;
578
}
579
580
if (!column->prefetch_buf) {
581
/* Allocate a new prefetch buffer */
582
583
sel_col_prefetch_buf_alloc(column);
584
}
585
586
sel_buf = column->prefetch_buf + pos;
587
588
val = que_node_get_val(column);
589
590
data = dfield_get_data(val);
591
len = dfield_get_len(val);
592
val_buf_size = que_node_get_val_buf_size(column);
593
594
/* We must keep track of the allocated memory for
595
column values to be able to free it later: therefore
596
we swap the values for sel_buf and val */
597
598
dfield_set_data(val, sel_buf->data, sel_buf->len);
599
que_node_set_val_buf_size(column, sel_buf->val_buf_size);
600
601
sel_buf->data = data;
602
sel_buf->len = len;
603
sel_buf->val_buf_size = val_buf_size;
604
next_col:
605
column = UT_LIST_GET_NEXT(col_var_list, column);
606
}
607
}
608
609
/*************************************************************************
610
Builds a previous version of a clustered index record for a consistent read */
611
static
612
ulint
613
row_sel_build_prev_vers(
614
/*====================*/
615
/* out: DB_SUCCESS or error code */
616
read_view_t* read_view, /* in: read view */
617
dict_index_t* index, /* in: plan node for table */
618
rec_t* rec, /* in: record in a clustered index */
619
ulint** offsets, /* in/out: offsets returned by
620
rec_get_offsets(rec, plan->index) */
621
mem_heap_t** offset_heap, /* in/out: memory heap from which
622
the offsets are allocated */
623
mem_heap_t** old_vers_heap, /* out: old version heap to use */
624
rec_t** old_vers, /* out: old version, or NULL if the
625
record does not exist in the view:
626
i.e., it was freshly inserted
627
afterwards */
628
mtr_t* mtr) /* in: mtr */
629
{
630
ulint err;
631
632
if (*old_vers_heap) {
633
mem_heap_empty(*old_vers_heap);
634
} else {
635
*old_vers_heap = mem_heap_create(512);
636
}
637
638
err = row_vers_build_for_consistent_read(
639
rec, mtr, index, offsets, read_view, offset_heap,
640
*old_vers_heap, old_vers);
641
return(err);
642
}
643
644
/*************************************************************************
645
Builds the last committed version of a clustered index record for a
646
semi-consistent read. */
647
static
648
ulint
649
row_sel_build_committed_vers_for_mysql(
650
/*===================================*/
651
/* out: DB_SUCCESS or error code */
652
dict_index_t* clust_index, /* in: clustered index */
653
row_prebuilt_t* prebuilt, /* in: prebuilt struct */
654
const rec_t* rec, /* in: record in a clustered index */
655
ulint** offsets, /* in/out: offsets returned by
656
rec_get_offsets(rec, clust_index) */
657
mem_heap_t** offset_heap, /* in/out: memory heap from which
658
the offsets are allocated */
659
const rec_t** old_vers, /* out: old version, or NULL if the
660
record does not exist in the view:
661
i.e., it was freshly inserted
662
afterwards */
663
mtr_t* mtr) /* in: mtr */
664
{
665
ulint err;
666
667
if (prebuilt->old_vers_heap) {
668
mem_heap_empty(prebuilt->old_vers_heap);
669
} else {
670
prebuilt->old_vers_heap = mem_heap_create(200);
671
}
672
673
err = row_vers_build_for_semi_consistent_read(
674
rec, mtr, clust_index, offsets, offset_heap,
675
prebuilt->old_vers_heap, old_vers);
676
return(err);
677
}
678
679
/*************************************************************************
680
Tests the conditions which determine when the index segment we are searching
681
through has been exhausted. */
682
UNIV_INLINE
683
ibool
684
row_sel_test_end_conds(
685
/*===================*/
686
/* out: TRUE if row passed the tests */
687
plan_t* plan) /* in: plan for the table; the column values must
688
already have been retrieved and the right sides of
689
comparisons evaluated */
690
{
691
func_node_t* cond;
692
693
/* All conditions in end_conds are comparisons of a column to an
694
expression */
695
696
cond = UT_LIST_GET_FIRST(plan->end_conds);
697
698
while (cond) {
699
/* Evaluate the left side of the comparison, i.e., get the
700
column value if there is an indirection */
701
702
eval_sym(cond->args);
703
704
/* Do the comparison */
705
706
if (!eval_cmp(cond)) {
707
708
return(FALSE);
709
}
710
711
cond = UT_LIST_GET_NEXT(cond_list, cond);
712
}
713
714
return(TRUE);
715
}
716
717
/*************************************************************************
718
Tests the other conditions. */
719
UNIV_INLINE
720
ibool
721
row_sel_test_other_conds(
722
/*=====================*/
723
/* out: TRUE if row passed the tests */
724
plan_t* plan) /* in: plan for the table; the column values must
725
already have been retrieved */
726
{
727
func_node_t* cond;
728
729
cond = UT_LIST_GET_FIRST(plan->other_conds);
730
731
while (cond) {
732
eval_exp(cond);
733
734
if (!eval_node_get_ibool_val(cond)) {
735
736
return(FALSE);
737
}
738
739
cond = UT_LIST_GET_NEXT(cond_list, cond);
740
}
741
742
return(TRUE);
743
}
744
745
/*************************************************************************
746
Retrieves the clustered index record corresponding to a record in a
747
non-clustered index. Does the necessary locking. */
748
static
749
ulint
750
row_sel_get_clust_rec(
751
/*==================*/
752
/* out: DB_SUCCESS or error code */
753
sel_node_t* node, /* in: select_node */
754
plan_t* plan, /* in: plan node for table */
755
rec_t* rec, /* in: record in a non-clustered index */
756
que_thr_t* thr, /* in: query thread */
757
rec_t** out_rec,/* out: clustered record or an old version of
758
it, NULL if the old version did not exist
759
in the read view, i.e., it was a fresh
760
inserted version */
761
mtr_t* mtr) /* in: mtr used to get access to the
762
non-clustered record; the same mtr is used to
763
access the clustered index */
764
{
765
dict_index_t* index;
766
rec_t* clust_rec;
767
rec_t* old_vers;
768
ulint err;
769
mem_heap_t* heap = NULL;
770
ulint offsets_[REC_OFFS_NORMAL_SIZE];
771
ulint* offsets = offsets_;
772
rec_offs_init(offsets_);
773
774
*out_rec = NULL;
775
776
offsets = rec_get_offsets(rec,
777
btr_pcur_get_btr_cur(&plan->pcur)->index,
778
offsets, ULINT_UNDEFINED, &heap);
779
780
row_build_row_ref_fast(plan->clust_ref, plan->clust_map, rec, offsets);
781
782
index = dict_table_get_first_index(plan->table);
783
784
btr_pcur_open_with_no_init(index, plan->clust_ref, PAGE_CUR_LE,
785
node->latch_mode, &(plan->clust_pcur),
786
0, mtr);
787
788
clust_rec = btr_pcur_get_rec(&(plan->clust_pcur));
789
790
/* Note: only if the search ends up on a non-infimum record is the
791
low_match value the real match to the search tuple */
792
793
if (!page_rec_is_user_rec(clust_rec)
794
|| btr_pcur_get_low_match(&(plan->clust_pcur))
795
< dict_index_get_n_unique(index)) {
796
797
ut_a(rec_get_deleted_flag(rec,
798
dict_table_is_comp(plan->table)));
799
ut_a(node->read_view);
800
801
/* In a rare case it is possible that no clust rec is found
802
for a delete-marked secondary index record: if in row0umod.c
803
in row_undo_mod_remove_clust_low() we have already removed
804
the clust rec, while purge is still cleaning and removing
805
secondary index records associated with earlier versions of
806
the clustered index record. In that case we know that the
807
clustered index record did not exist in the read view of
808
trx. */
809
810
goto func_exit;
811
}
812
813
offsets = rec_get_offsets(clust_rec, index, offsets,
814
ULINT_UNDEFINED, &heap);
815
816
if (!node->read_view) {
817
/* Try to place a lock on the index record */
818
819
/* If innodb_locks_unsafe_for_binlog option is used
820
or this session is using READ COMMITTED isolation level
821
we lock only the record, i.e., next-key locking is
822
not used. */
823
ulint lock_type;
824
trx_t* trx;
825
826
trx = thr_get_trx(thr);
827
828
if (srv_locks_unsafe_for_binlog
829
|| trx->isolation_level == TRX_ISO_READ_COMMITTED) {
830
lock_type = LOCK_REC_NOT_GAP;
831
} else {
832
lock_type = LOCK_ORDINARY;
833
}
834
835
err = lock_clust_rec_read_check_and_lock(
836
0, btr_pcur_get_block(&plan->clust_pcur),
837
clust_rec, index, offsets,
838
node->row_lock_mode, lock_type, thr);
839
840
if (err != DB_SUCCESS) {
841
842
goto err_exit;
843
}
844
} else {
845
/* This is a non-locking consistent read: if necessary, fetch
846
a previous version of the record */
847
848
old_vers = NULL;
849
850
if (!lock_clust_rec_cons_read_sees(clust_rec, index, offsets,
851
node->read_view)) {
852
853
err = row_sel_build_prev_vers(
854
node->read_view, index, clust_rec,
855
&offsets, &heap, &plan->old_vers_heap,
856
&old_vers, mtr);
857
858
if (err != DB_SUCCESS) {
859
860
goto err_exit;
861
}
862
863
clust_rec = old_vers;
864
865
if (clust_rec == NULL) {
866
goto func_exit;
867
}
868
}
869
870
/* If we had to go to an earlier version of row or the
871
secondary index record is delete marked, then it may be that
872
the secondary index record corresponding to clust_rec
873
(or old_vers) is not rec; in that case we must ignore
874
such row because in our snapshot rec would not have existed.
875
Remember that from rec we cannot see directly which transaction
876
id corresponds to it: we have to go to the clustered index
877
record. A query where we want to fetch all rows where
878
the secondary index value is in some interval would return
879
a wrong result if we would not drop rows which we come to
880
visit through secondary index records that would not really
881
exist in our snapshot. */
882
883
if ((old_vers
884
|| rec_get_deleted_flag(rec, dict_table_is_comp(
885
plan->table)))
886
&& !row_sel_sec_rec_is_for_clust_rec(rec, plan->index,
887
clust_rec, index)) {
888
goto func_exit;
889
}
890
}
891
892
/* Fetch the columns needed in test conditions. The clustered
893
index record is protected by a page latch that was acquired
894
when plan->clust_pcur was positioned. The latch will not be
895
released until mtr_commit(mtr). */
896
897
row_sel_fetch_columns(index, clust_rec, offsets,
898
UT_LIST_GET_FIRST(plan->columns));
899
*out_rec = clust_rec;
900
func_exit:
901
err = DB_SUCCESS;
902
err_exit:
903
if (UNIV_LIKELY_NULL(heap)) {
904
mem_heap_free(heap);
905
}
906
return(err);
907
}
908
909
/*************************************************************************
910
Sets a lock on a record. */
911
UNIV_INLINE
912
ulint
913
sel_set_rec_lock(
914
/*=============*/
915
/* out: DB_SUCCESS or error code */
916
const buf_block_t* block, /* in: buffer block of rec */
917
const rec_t* rec, /* in: record */
918
dict_index_t* index, /* in: index */
919
const ulint* offsets,/* in: rec_get_offsets(rec, index) */
920
ulint mode, /* in: lock mode */
921
ulint type, /* in: LOCK_ORDINARY, LOCK_GAP, or
922
LOC_REC_NOT_GAP */
923
que_thr_t* thr) /* in: query thread */
924
{
925
trx_t* trx;
926
ulint err;
927
928
trx = thr_get_trx(thr);
929
930
if (UT_LIST_GET_LEN(trx->trx_locks) > 10000) {
931
if (buf_LRU_buf_pool_running_out()) {
932
933
return(DB_LOCK_TABLE_FULL);
934
}
935
}
936
937
if (dict_index_is_clust(index)) {
938
err = lock_clust_rec_read_check_and_lock(
939
0, block, rec, index, offsets, mode, type, thr);
940
} else {
941
err = lock_sec_rec_read_check_and_lock(
942
0, block, rec, index, offsets, mode, type, thr);
943
}
944
945
return(err);
946
}
947
948
/*************************************************************************
949
Opens a pcur to a table index. */
950
static
951
void
952
row_sel_open_pcur(
953
/*==============*/
954
sel_node_t* node, /* in: select node */
955
plan_t* plan, /* in: table plan */
956
ibool search_latch_locked,
957
/* in: TRUE if the thread currently
958
has the search latch locked in
959
s-mode */
960
mtr_t* mtr) /* in: mtr */
961
{
962
dict_index_t* index;
963
func_node_t* cond;
964
que_node_t* exp;
965
ulint n_fields;
966
ulint has_search_latch = 0; /* RW_S_LATCH or 0 */
967
ulint i;
968
969
if (search_latch_locked) {
970
has_search_latch = RW_S_LATCH;
971
}
972
973
index = plan->index;
974
975
/* Calculate the value of the search tuple: the exact match columns
976
get their expressions evaluated when we evaluate the right sides of
977
end_conds */
978
979
cond = UT_LIST_GET_FIRST(plan->end_conds);
980
981
while (cond) {
982
eval_exp(que_node_get_next(cond->args));
983
984
cond = UT_LIST_GET_NEXT(cond_list, cond);
985
}
986
987
if (plan->tuple) {
988
n_fields = dtuple_get_n_fields(plan->tuple);
989
990
if (plan->n_exact_match < n_fields) {
991
/* There is a non-exact match field which must be
992
evaluated separately */
993
994
eval_exp(plan->tuple_exps[n_fields - 1]);
995
}
996
997
for (i = 0; i < n_fields; i++) {
998
exp = plan->tuple_exps[i];
999
1000
dfield_copy_data(dtuple_get_nth_field(plan->tuple, i),
1001
que_node_get_val(exp));
1002
}
1003
1004
/* Open pcur to the index */
1005
1006
btr_pcur_open_with_no_init(index, plan->tuple, plan->mode,
1007
node->latch_mode, &(plan->pcur),
1008
has_search_latch, mtr);
1009
} else {
1010
/* Open the cursor to the start or the end of the index
1011
(FALSE: no init) */
1012
1013
btr_pcur_open_at_index_side(plan->asc, index, node->latch_mode,
1014
&(plan->pcur), FALSE, mtr);
1015
}
1016
1017
ut_ad(plan->n_rows_prefetched == 0);
1018
ut_ad(plan->n_rows_fetched == 0);
1019
ut_ad(plan->cursor_at_end == FALSE);
1020
1021
plan->pcur_is_open = TRUE;
1022
}
1023
1024
/*************************************************************************
1025
Restores a stored pcur position to a table index. */
1026
static
1027
ibool
1028
row_sel_restore_pcur_pos(
1029
/*=====================*/
1030
/* out: TRUE if the cursor should be moved to
1031
the next record after we return from this
1032
function (moved to the previous, in the case
1033
of a descending cursor) without processing
1034
again the current cursor record */
1035
sel_node_t* node, /* in: select node */
1036
plan_t* plan, /* in: table plan */
1037
mtr_t* mtr) /* in: mtr */
1038
{
1039
ibool equal_position;
1040
ulint relative_position;
1041
1042
ut_ad(!plan->cursor_at_end);
1043
1044
relative_position = btr_pcur_get_rel_pos(&(plan->pcur));
1045
1046
equal_position = btr_pcur_restore_position(node->latch_mode,
1047
&(plan->pcur), mtr);
1048
1049
/* If the cursor is traveling upwards, and relative_position is
1050
1051
(1) BTR_PCUR_BEFORE: this is not allowed, as we did not have a lock
1052
yet on the successor of the page infimum;
1053
(2) BTR_PCUR_AFTER: btr_pcur_restore_position placed the cursor on the
1054
first record GREATER than the predecessor of a page supremum; we have
1055
not yet processed the cursor record: no need to move the cursor to the
1056
next record;
1057
(3) BTR_PCUR_ON: btr_pcur_restore_position placed the cursor on the
1058
last record LESS or EQUAL to the old stored user record; (a) if
1059
equal_position is FALSE, this means that the cursor is now on a record
1060
less than the old user record, and we must move to the next record;
1061
(b) if equal_position is TRUE, then if
1062
plan->stored_cursor_rec_processed is TRUE, we must move to the next
1063
record, else there is no need to move the cursor. */
1064
1065
if (plan->asc) {
1066
if (relative_position == BTR_PCUR_ON) {
1067
1068
if (equal_position) {
1069
1070
return(plan->stored_cursor_rec_processed);
1071
}
1072
1073
return(TRUE);
1074
}
1075
1076
ut_ad(relative_position == BTR_PCUR_AFTER
1077
|| relative_position == BTR_PCUR_AFTER_LAST_IN_TREE);
1078
1079
return(FALSE);
1080
}
1081
1082
/* If the cursor is traveling downwards, and relative_position is
1083
1084
(1) BTR_PCUR_BEFORE: btr_pcur_restore_position placed the cursor on
1085
the last record LESS than the successor of a page infimum; we have not
1086
processed the cursor record: no need to move the cursor;
1087
(2) BTR_PCUR_AFTER: btr_pcur_restore_position placed the cursor on the
1088
first record GREATER than the predecessor of a page supremum; we have
1089
processed the cursor record: we should move the cursor to the previous
1090
record;
1091
(3) BTR_PCUR_ON: btr_pcur_restore_position placed the cursor on the
1092
last record LESS or EQUAL to the old stored user record; (a) if
1093
equal_position is FALSE, this means that the cursor is now on a record
1094
less than the old user record, and we need not move to the previous
1095
record; (b) if equal_position is TRUE, then if
1096
plan->stored_cursor_rec_processed is TRUE, we must move to the previous
1097
record, else there is no need to move the cursor. */
1098
1099
if (relative_position == BTR_PCUR_BEFORE
1100
|| relative_position == BTR_PCUR_BEFORE_FIRST_IN_TREE) {
1101
1102
return(FALSE);
1103
}
1104
1105
if (relative_position == BTR_PCUR_ON) {
1106
1107
if (equal_position) {
1108
1109
return(plan->stored_cursor_rec_processed);
1110
}
1111
1112
return(FALSE);
1113
}
1114
1115
ut_ad(relative_position == BTR_PCUR_AFTER
1116
|| relative_position == BTR_PCUR_AFTER_LAST_IN_TREE);
1117
1118
return(TRUE);
1119
}
1120
1121
/*************************************************************************
1122
Resets a plan cursor to a closed state. */
1123
UNIV_INLINE
1124
void
1125
plan_reset_cursor(
1126
/*==============*/
1127
plan_t* plan) /* in: plan */
1128
{
1129
plan->pcur_is_open = FALSE;
1130
plan->cursor_at_end = FALSE;
1131
plan->n_rows_fetched = 0;
1132
plan->n_rows_prefetched = 0;
1133
}
1134
1135
/*************************************************************************
1136
Tries to do a shortcut to fetch a clustered index record with a unique key,
1137
using the hash index if possible (not always). */
1138
static
1139
ulint
1140
row_sel_try_search_shortcut(
1141
/*========================*/
1142
/* out: SEL_FOUND, SEL_EXHAUSTED, SEL_RETRY */
1143
sel_node_t* node, /* in: select node for a consistent read */
1144
plan_t* plan, /* in: plan for a unique search in clustered
1145
index */
1146
mtr_t* mtr) /* in: mtr */
1147
{
1148
dict_index_t* index;
1149
rec_t* rec;
1150
mem_heap_t* heap = NULL;
1151
ulint offsets_[REC_OFFS_NORMAL_SIZE];
1152
ulint* offsets = offsets_;
1153
ulint ret;
1154
rec_offs_init(offsets_);
1155
1156
index = plan->index;
1157
1158
ut_ad(node->read_view);
1159
ut_ad(plan->unique_search);
1160
ut_ad(!plan->must_get_clust);
1161
#ifdef UNIV_SYNC_DEBUG
1162
ut_ad(rw_lock_own(&btr_search_latch, RW_LOCK_SHARED));
1163
#endif /* UNIV_SYNC_DEBUG */
1164
1165
row_sel_open_pcur(node, plan, TRUE, mtr);
1166
1167
rec = btr_pcur_get_rec(&(plan->pcur));
1168
1169
if (!page_rec_is_user_rec(rec)) {
1170
1171
return(SEL_RETRY);
1172
}
1173
1174
ut_ad(plan->mode == PAGE_CUR_GE);
1175
1176
/* As the cursor is now placed on a user record after a search with
1177
the mode PAGE_CUR_GE, the up_match field in the cursor tells how many
1178
fields in the user record matched to the search tuple */
1179
1180
if (btr_pcur_get_up_match(&(plan->pcur)) < plan->n_exact_match) {
1181
1182
return(SEL_EXHAUSTED);
1183
}
1184
1185
/* This is a non-locking consistent read: if necessary, fetch
1186
a previous version of the record */
1187
1188
offsets = rec_get_offsets(rec, index, offsets, ULINT_UNDEFINED, &heap);
1189
1190
if (dict_index_is_clust(index)) {
1191
if (!lock_clust_rec_cons_read_sees(rec, index, offsets,
1192
node->read_view)) {
1193
ret = SEL_RETRY;
1194
goto func_exit;
1195
}
1196
} else if (!lock_sec_rec_cons_read_sees(rec, node->read_view)) {
1197
1198
ret = SEL_RETRY;
1199
goto func_exit;
1200
}
1201
1202
/* Test the deleted flag. */
1203
1204
if (rec_get_deleted_flag(rec, dict_table_is_comp(plan->table))) {
1205
1206
ret = SEL_EXHAUSTED;
1207
goto func_exit;
1208
}
1209
1210
/* Fetch the columns needed in test conditions. The index
1211
record is protected by a page latch that was acquired when
1212
plan->pcur was positioned. The latch will not be released
1213
until mtr_commit(mtr). */
1214
1215
row_sel_fetch_columns(index, rec, offsets,
1216
UT_LIST_GET_FIRST(plan->columns));
1217
1218
/* Test the rest of search conditions */
1219
1220
if (!row_sel_test_other_conds(plan)) {
1221
1222
ret = SEL_EXHAUSTED;
1223
goto func_exit;
1224
}
1225
1226
ut_ad(plan->pcur.latch_mode == node->latch_mode);
1227
1228
plan->n_rows_fetched++;
1229
ret = SEL_FOUND;
1230
func_exit:
1231
if (UNIV_LIKELY_NULL(heap)) {
1232
mem_heap_free(heap);
1233
}
1234
return(ret);
1235
}
1236
1237
/*************************************************************************
1238
Performs a select step. */
1239
static
1240
ulint
1241
row_sel(
1242
/*====*/
1243
/* out: DB_SUCCESS or error code */
1244
sel_node_t* node, /* in: select node */
1245
que_thr_t* thr) /* in: query thread */
1246
{
1247
dict_index_t* index;
1248
plan_t* plan;
1249
mtr_t mtr;
1250
ibool moved;
1251
rec_t* rec;
1252
rec_t* old_vers;
1253
rec_t* clust_rec;
1254
ibool search_latch_locked;
1255
ibool consistent_read;
1256
1257
/* The following flag becomes TRUE when we are doing a
1258
consistent read from a non-clustered index and we must look
1259
at the clustered index to find out the previous delete mark
1260
state of the non-clustered record: */
1261
1262
ibool cons_read_requires_clust_rec = FALSE;
1263
ulint cost_counter = 0;
1264
ibool cursor_just_opened;
1265
ibool must_go_to_next;
1266
ibool leaf_contains_updates = FALSE;
1267
/* TRUE if select_will_do_update is
1268
TRUE and the current clustered index
1269
leaf page has been updated during
1270
the current mtr: mtr must be committed
1271
at the same time as the leaf x-latch
1272
is released */
1273
ibool mtr_has_extra_clust_latch = FALSE;
1274
/* TRUE if the search was made using
1275
a non-clustered index, and we had to
1276
access the clustered record: now &mtr
1277
contains a clustered index latch, and
1278
&mtr must be committed before we move
1279
to the next non-clustered record */
1280
ulint found_flag;
1281
ulint err;
1282
mem_heap_t* heap = NULL;
1283
ulint offsets_[REC_OFFS_NORMAL_SIZE];
1284
ulint* offsets = offsets_;
1285
rec_offs_init(offsets_);
1286
1287
ut_ad(thr->run_node == node);
1288
1289
search_latch_locked = FALSE;
1290
1291
if (node->read_view) {
1292
/* In consistent reads, we try to do with the hash index and
1293
not to use the buffer page get. This is to reduce memory bus
1294
load resulting from semaphore operations. The search latch
1295
will be s-locked when we access an index with a unique search
1296
condition, but not locked when we access an index with a
1297
less selective search condition. */
1298
1299
consistent_read = TRUE;
1300
} else {
1301
consistent_read = FALSE;
1302
}
1303
1304
table_loop:
1305
/* TABLE LOOP
1306
----------
1307
This is the outer major loop in calculating a join. We come here when
1308
node->fetch_table changes, and after adding a row to aggregate totals
1309
and, of course, when this function is called. */
1310
1311
ut_ad(leaf_contains_updates == FALSE);
1312
ut_ad(mtr_has_extra_clust_latch == FALSE);
1313
1314
plan = sel_node_get_nth_plan(node, node->fetch_table);
1315
index = plan->index;
1316
1317
if (plan->n_rows_prefetched > 0) {
1318
sel_pop_prefetched_row(plan);
1319
1320
goto next_table_no_mtr;
1321
}
1322
1323
if (plan->cursor_at_end) {
1324
/* The cursor has already reached the result set end: no more
1325
rows to process for this table cursor, as also the prefetch
1326
stack was empty */
1327
1328
ut_ad(plan->pcur_is_open);
1329
1330
goto table_exhausted_no_mtr;
1331
}
1332
1333
/* Open a cursor to index, or restore an open cursor position */
1334
1335
mtr_start(&mtr);
1336
1337
if (consistent_read && plan->unique_search && !plan->pcur_is_open
1338
&& !plan->must_get_clust
1339
&& !plan->table->big_rows) {
1340
if (!search_latch_locked) {
1341
rw_lock_s_lock(&btr_search_latch);
1342
1343
search_latch_locked = TRUE;
1344
} else if (btr_search_latch.writer_is_wait_ex) {
1345
1346
/* There is an x-latch request waiting: release the
1347
s-latch for a moment; as an s-latch here is often
1348
kept for some 10 searches before being released,
1349
a waiting x-latch request would block other threads
1350
from acquiring an s-latch for a long time, lowering
1351
performance significantly in multiprocessors. */
1352
1353
rw_lock_s_unlock(&btr_search_latch);
1354
rw_lock_s_lock(&btr_search_latch);
1355
}
1356
1357
found_flag = row_sel_try_search_shortcut(node, plan, &mtr);
1358
1359
if (found_flag == SEL_FOUND) {
1360
1361
goto next_table;
1362
1363
} else if (found_flag == SEL_EXHAUSTED) {
1364
1365
goto table_exhausted;
1366
}
1367
1368
ut_ad(found_flag == SEL_RETRY);
1369
1370
plan_reset_cursor(plan);
1371
1372
mtr_commit(&mtr);
1373
mtr_start(&mtr);
1374
}
1375
1376
if (search_latch_locked) {
1377
rw_lock_s_unlock(&btr_search_latch);
1378
1379
search_latch_locked = FALSE;
1380
}
1381
1382
if (!plan->pcur_is_open) {
1383
/* Evaluate the expressions to build the search tuple and
1384
open the cursor */
1385
1386
row_sel_open_pcur(node, plan, search_latch_locked, &mtr);
1387
1388
cursor_just_opened = TRUE;
1389
1390
/* A new search was made: increment the cost counter */
1391
cost_counter++;
1392
} else {
1393
/* Restore pcur position to the index */
1394
1395
must_go_to_next = row_sel_restore_pcur_pos(node, plan, &mtr);
1396
1397
cursor_just_opened = FALSE;
1398
1399
if (must_go_to_next) {
1400
/* We have already processed the cursor record: move
1401
to the next */
1402
1403
goto next_rec;
1404
}
1405
}
1406
1407
rec_loop:
1408
/* RECORD LOOP
1409
-----------
1410
In this loop we use pcur and try to fetch a qualifying row, and
1411
also fill the prefetch buffer for this table if n_rows_fetched has
1412
exceeded a threshold. While we are inside this loop, the following
1413
holds:
1414
(1) &mtr is started,
1415
(2) pcur is positioned and open.
1416
1417
NOTE that if cursor_just_opened is TRUE here, it means that we came
1418
to this point right after row_sel_open_pcur. */
1419
1420
ut_ad(mtr_has_extra_clust_latch == FALSE);
1421
1422
rec = btr_pcur_get_rec(&(plan->pcur));
1423
1424
/* PHASE 1: Set a lock if specified */
1425
1426
if (!node->asc && cursor_just_opened
1427
&& !page_rec_is_supremum(rec)) {
1428
1429
/* When we open a cursor for a descending search, we must set
1430
a next-key lock on the successor record: otherwise it would
1431
be possible to insert new records next to the cursor position,
1432
and it might be that these new records should appear in the
1433
search result set, resulting in the phantom problem. */
1434
1435
if (!consistent_read) {
1436
1437
/* If innodb_locks_unsafe_for_binlog option is used
1438
or this session is using READ COMMITTED isolation
1439
level, we lock only the record, i.e., next-key
1440
locking is not used. */
1441
1442
rec_t* next_rec = page_rec_get_next(rec);
1443
ulint lock_type;
1444
trx_t* trx;
1445
1446
trx = thr_get_trx(thr);
1447
1448
offsets = rec_get_offsets(next_rec, index, offsets,
1449
ULINT_UNDEFINED, &heap);
1450
1451
if (srv_locks_unsafe_for_binlog
1452
|| trx->isolation_level
1453
== TRX_ISO_READ_COMMITTED) {
1454
1455
if (page_rec_is_supremum(next_rec)) {
1456
1457
goto skip_lock;
1458
}
1459
1460
lock_type = LOCK_REC_NOT_GAP;
1461
} else {
1462
lock_type = LOCK_ORDINARY;
1463
}
1464
1465
err = sel_set_rec_lock(btr_pcur_get_block(&plan->pcur),
1466
next_rec, index, offsets,
1467
node->row_lock_mode,
1468
lock_type, thr);
1469
1470
if (err != DB_SUCCESS) {
1471
/* Note that in this case we will store in pcur
1472
the PREDECESSOR of the record we are waiting
1473
the lock for */
1474
1475
goto lock_wait_or_error;
1476
}
1477
}
1478
}
1479
1480
skip_lock:
1481
if (page_rec_is_infimum(rec)) {
1482
1483
/* The infimum record on a page cannot be in the result set,
1484
and neither can a record lock be placed on it: we skip such
1485
a record. We also increment the cost counter as we may have
1486
processed yet another page of index. */
1487
1488
cost_counter++;
1489
1490
goto next_rec;
1491
}
1492
1493
if (!consistent_read) {
1494
/* Try to place a lock on the index record */
1495
1496
/* If innodb_locks_unsafe_for_binlog option is used
1497
or this session is using READ COMMITTED isolation level,
1498
we lock only the record, i.e., next-key locking is
1499
not used. */
1500
1501
ulint lock_type;
1502
trx_t* trx;
1503
1504
offsets = rec_get_offsets(rec, index, offsets,
1505
ULINT_UNDEFINED, &heap);
1506
1507
trx = thr_get_trx(thr);
1508
1509
if (srv_locks_unsafe_for_binlog
1510
|| trx->isolation_level == TRX_ISO_READ_COMMITTED) {
1511
1512
if (page_rec_is_supremum(rec)) {
1513
1514
goto next_rec;
1515
}
1516
1517
lock_type = LOCK_REC_NOT_GAP;
1518
} else {
1519
lock_type = LOCK_ORDINARY;
1520
}
1521
1522
err = sel_set_rec_lock(btr_pcur_get_block(&plan->pcur),
1523
rec, index, offsets,
1524
node->row_lock_mode, lock_type, thr);
1525
1526
if (err != DB_SUCCESS) {
1527
1528
goto lock_wait_or_error;
1529
}
1530
}
1531
1532
if (page_rec_is_supremum(rec)) {
1533
1534
/* A page supremum record cannot be in the result set: skip
1535
it now when we have placed a possible lock on it */
1536
1537
goto next_rec;
1538
}
1539
1540
ut_ad(page_rec_is_user_rec(rec));
1541
1542
if (cost_counter > SEL_COST_LIMIT) {
1543
1544
/* Now that we have placed the necessary locks, we can stop
1545
for a while and store the cursor position; NOTE that if we
1546
would store the cursor position BEFORE placing a record lock,
1547
it might happen that the cursor would jump over some records
1548
that another transaction could meanwhile insert adjacent to
1549
the cursor: this would result in the phantom problem. */
1550
1551
goto stop_for_a_while;
1552
}
1553
1554
/* PHASE 2: Check a mixed index mix id if needed */
1555
1556
if (plan->unique_search && cursor_just_opened) {
1557
1558
ut_ad(plan->mode == PAGE_CUR_GE);
1559
1560
/* As the cursor is now placed on a user record after a search
1561
with the mode PAGE_CUR_GE, the up_match field in the cursor
1562
tells how many fields in the user record matched to the search
1563
tuple */
1564
1565
if (btr_pcur_get_up_match(&(plan->pcur))
1566
< plan->n_exact_match) {
1567
goto table_exhausted;
1568
}
1569
1570
/* Ok, no need to test end_conds or mix id */
1571
1572
}
1573
1574
/* We are ready to look at a possible new index entry in the result
1575
set: the cursor is now placed on a user record */
1576
1577
/* PHASE 3: Get previous version in a consistent read */
1578
1579
cons_read_requires_clust_rec = FALSE;
1580
offsets = rec_get_offsets(rec, index, offsets, ULINT_UNDEFINED, &heap);
1581
1582
if (consistent_read) {
1583
/* This is a non-locking consistent read: if necessary, fetch
1584
a previous version of the record */
1585
1586
if (dict_index_is_clust(index)) {
1587
1588
if (!lock_clust_rec_cons_read_sees(rec, index, offsets,
1589
node->read_view)) {
1590
1591
err = row_sel_build_prev_vers(
1592
node->read_view, index, rec,
1593
&offsets, &heap, &plan->old_vers_heap,
1594
&old_vers, &mtr);
1595
1596
if (err != DB_SUCCESS) {
1597
1598
goto lock_wait_or_error;
1599
}
1600
1601
if (old_vers == NULL) {
1602
offsets = rec_get_offsets(
1603
rec, index, offsets,
1604
ULINT_UNDEFINED, &heap);
1605
1606
/* Fetch the columns needed in
1607
test conditions. The clustered
1608
index record is protected by a
1609
page latch that was acquired
1610
by row_sel_open_pcur() or
1611
row_sel_restore_pcur_pos().
1612
The latch will not be released
1613
until mtr_commit(mtr). */
1614
1615
row_sel_fetch_columns(
1616
index, rec, offsets,
1617
UT_LIST_GET_FIRST(
1618
plan->columns));
1619
1620
if (!row_sel_test_end_conds(plan)) {
1621
1622
goto table_exhausted;
1623
}
1624
1625
goto next_rec;
1626
}
1627
1628
rec = old_vers;
1629
}
1630
} else if (!lock_sec_rec_cons_read_sees(rec,
1631
node->read_view)) {
1632
cons_read_requires_clust_rec = TRUE;
1633
}
1634
}
1635
1636
/* PHASE 4: Test search end conditions and deleted flag */
1637
1638
/* Fetch the columns needed in test conditions. The record is
1639
protected by a page latch that was acquired by
1640
row_sel_open_pcur() or row_sel_restore_pcur_pos(). The latch
1641
will not be released until mtr_commit(mtr). */
1642
1643
row_sel_fetch_columns(index, rec, offsets,
1644
UT_LIST_GET_FIRST(plan->columns));
1645
1646
/* Test the selection end conditions: these can only contain columns
1647
which already are found in the index, even though the index might be
1648
non-clustered */
1649
1650
if (plan->unique_search && cursor_just_opened) {
1651
1652
/* No test necessary: the test was already made above */
1653
1654
} else if (!row_sel_test_end_conds(plan)) {
1655
1656
goto table_exhausted;
1657
}
1658
1659
if (rec_get_deleted_flag(rec, dict_table_is_comp(plan->table))
1660
&& !cons_read_requires_clust_rec) {
1661
1662
/* The record is delete marked: we can skip it if this is
1663
not a consistent read which might see an earlier version
1664
of a non-clustered index record */
1665
1666
if (plan->unique_search) {
1667
1668
goto table_exhausted;
1669
}
1670
1671
goto next_rec;
1672
}
1673
1674
1675
/* PHASE 5: Get the clustered index record, if needed and if we did
1676
not do the search using the clustered index */
1677
1678
if (plan->must_get_clust || cons_read_requires_clust_rec) {
1679
1680
/* It was a non-clustered index and we must fetch also the
1681
clustered index record */
1682
1683
err = row_sel_get_clust_rec(node, plan, rec, thr, &clust_rec,
1684
&mtr);
1685
mtr_has_extra_clust_latch = TRUE;
1686
1687
if (err != DB_SUCCESS) {
1688
1689
goto lock_wait_or_error;
1690
}
1691
1692
/* Retrieving the clustered record required a search:
1693
increment the cost counter */
1694
1695
cost_counter++;
1696
1697
if (clust_rec == NULL) {
1698
/* The record did not exist in the read view */
1699
ut_ad(consistent_read);
1700
1701
goto next_rec;
1702
}
1703
1704
if (rec_get_deleted_flag(clust_rec,
1705
dict_table_is_comp(plan->table))) {
1706
1707
/* The record is delete marked: we can skip it */
1708
1709
goto next_rec;
1710
}
1711
1712
if (node->can_get_updated) {
1713
1714
btr_pcur_store_position(&(plan->clust_pcur), &mtr);
1715
}
1716
}
1717
1718
/* PHASE 6: Test the rest of search conditions */
1719
1720
if (!row_sel_test_other_conds(plan)) {
1721
1722
if (plan->unique_search) {
1723
1724
goto table_exhausted;
1725
}
1726
1727
goto next_rec;
1728
}
1729
1730
/* PHASE 7: We found a new qualifying row for the current table; push
1731
the row if prefetch is on, or move to the next table in the join */
1732
1733
plan->n_rows_fetched++;
1734
1735
ut_ad(plan->pcur.latch_mode == node->latch_mode);
1736
1737
if (node->select_will_do_update) {
1738
/* This is a searched update and we can do the update in-place,
1739
saving CPU time */
1740
1741
row_upd_in_place_in_select(node, thr, &mtr);
1742
1743
leaf_contains_updates = TRUE;
1744
1745
/* When the database is in the online backup mode, the number
1746
of log records for a single mtr should be small: increment the
1747
cost counter to ensure it */
1748
1749
cost_counter += 1 + (SEL_COST_LIMIT / 8);
1750
1751
if (plan->unique_search) {
1752
1753
goto table_exhausted;
1754
}
1755
1756
goto next_rec;
1757
}
1758
1759
if ((plan->n_rows_fetched <= SEL_PREFETCH_LIMIT)
1760
|| plan->unique_search || plan->no_prefetch
1761
|| plan->table->big_rows) {
1762
1763
/* No prefetch in operation: go to the next table */
1764
1765
goto next_table;
1766
}
1767
1768
sel_push_prefetched_row(plan);
1769
1770
if (plan->n_rows_prefetched == SEL_MAX_N_PREFETCH) {
1771
1772
/* The prefetch buffer is now full */
1773
1774
sel_pop_prefetched_row(plan);
1775
1776
goto next_table;
1777
}
1778
1779
next_rec:
1780
ut_ad(!search_latch_locked);
1781
1782
if (mtr_has_extra_clust_latch) {
1783
1784
/* We must commit &mtr if we are moving to the next
1785
non-clustered index record, because we could break the
1786
latching order if we would access a different clustered
1787
index page right away without releasing the previous. */
1788
1789
goto commit_mtr_for_a_while;
1790
}
1791
1792
if (leaf_contains_updates
1793
&& btr_pcur_is_after_last_on_page(&plan->pcur)) {
1794
1795
/* We must commit &mtr if we are moving to a different page,
1796
because we have done updates to the x-latched leaf page, and
1797
the latch would be released in btr_pcur_move_to_next, without
1798
&mtr getting committed there */
1799
1800
ut_ad(node->asc);
1801
1802
goto commit_mtr_for_a_while;
1803
}
1804
1805
if (node->asc) {
1806
moved = btr_pcur_move_to_next(&(plan->pcur), &mtr);
1807
} else {
1808
moved = btr_pcur_move_to_prev(&(plan->pcur), &mtr);
1809
}
1810
1811
if (!moved) {
1812
1813
goto table_exhausted;
1814
}
1815
1816
cursor_just_opened = FALSE;
1817
1818
/* END OF RECORD LOOP
1819
------------------ */
1820
goto rec_loop;
1821
1822
next_table:
1823
/* We found a record which satisfies the conditions: we can move to
1824
the next table or return a row in the result set */
1825
1826
ut_ad(btr_pcur_is_on_user_rec(&plan->pcur));
1827
1828
if (plan->unique_search && !node->can_get_updated) {
1829
1830
plan->cursor_at_end = TRUE;
1831
} else {
1832
ut_ad(!search_latch_locked);
1833
1834
plan->stored_cursor_rec_processed = TRUE;
1835
1836
btr_pcur_store_position(&(plan->pcur), &mtr);
1837
}
1838
1839
mtr_commit(&mtr);
1840
1841
leaf_contains_updates = FALSE;
1842
mtr_has_extra_clust_latch = FALSE;
1843
1844
next_table_no_mtr:
1845
/* If we use 'goto' to this label, it means that the row was popped
1846
from the prefetched rows stack, and &mtr is already committed */
1847
1848
if (node->fetch_table + 1 == node->n_tables) {
1849
1850
sel_eval_select_list(node);
1851
1852
if (node->is_aggregate) {
1853
1854
goto table_loop;
1855
}
1856
1857
sel_assign_into_var_values(node->into_list, node);
1858
1859
thr->run_node = que_node_get_parent(node);
1860
1861
err = DB_SUCCESS;
1862
goto func_exit;
1863
}
1864
1865
node->fetch_table++;
1866
1867
/* When we move to the next table, we first reset the plan cursor:
1868
we do not care about resetting it when we backtrack from a table */
1869
1870
plan_reset_cursor(sel_node_get_nth_plan(node, node->fetch_table));
1871
1872
goto table_loop;
1873
1874
table_exhausted:
1875
/* The table cursor pcur reached the result set end: backtrack to the
1876
previous table in the join if we do not have cached prefetched rows */
1877
1878
plan->cursor_at_end = TRUE;
1879
1880
mtr_commit(&mtr);
1881
1882
leaf_contains_updates = FALSE;
1883
mtr_has_extra_clust_latch = FALSE;
1884
1885
if (plan->n_rows_prefetched > 0) {
1886
/* The table became exhausted during a prefetch */
1887
1888
sel_pop_prefetched_row(plan);
1889
1890
goto next_table_no_mtr;
1891
}
1892
1893
table_exhausted_no_mtr:
1894
if (node->fetch_table == 0) {
1895
err = DB_SUCCESS;
1896
1897
if (node->is_aggregate && !node->aggregate_already_fetched) {
1898
1899
node->aggregate_already_fetched = TRUE;
1900
1901
sel_assign_into_var_values(node->into_list, node);
1902
1903
thr->run_node = que_node_get_parent(node);
1904
} else {
1905
node->state = SEL_NODE_NO_MORE_ROWS;
1906
1907
thr->run_node = que_node_get_parent(node);
1908
}
1909
1910
goto func_exit;
1911
}
1912
1913
node->fetch_table--;
1914
1915
goto table_loop;
1916
1917
stop_for_a_while:
1918
/* Return control for a while to que_run_threads, so that runaway
1919
queries can be canceled. NOTE that when we come here, we must, in a
1920
locking read, have placed the necessary (possibly waiting request)
1921
record lock on the cursor record or its successor: when we reposition
1922
the cursor, this record lock guarantees that nobody can meanwhile have
1923
inserted new records which should have appeared in the result set,
1924
which would result in the phantom problem. */
1925
1926
ut_ad(!search_latch_locked);
1927
1928
plan->stored_cursor_rec_processed = FALSE;
1929
btr_pcur_store_position(&(plan->pcur), &mtr);
1930
1931
mtr_commit(&mtr);
1932
1933
#ifdef UNIV_SYNC_DEBUG
1934
ut_ad(sync_thread_levels_empty_gen(TRUE));
1935
#endif /* UNIV_SYNC_DEBUG */
1936
err = DB_SUCCESS;
1937
goto func_exit;
1938
1939
commit_mtr_for_a_while:
1940
/* Stores the cursor position and commits &mtr; this is used if
1941
&mtr may contain latches which would break the latching order if
1942
&mtr would not be committed and the latches released. */
1943
1944
plan->stored_cursor_rec_processed = TRUE;
1945
1946
ut_ad(!search_latch_locked);
1947
btr_pcur_store_position(&(plan->pcur), &mtr);
1948
1949
mtr_commit(&mtr);
1950
1951
leaf_contains_updates = FALSE;
1952
mtr_has_extra_clust_latch = FALSE;
1953
1954
#ifdef UNIV_SYNC_DEBUG
1955
ut_ad(sync_thread_levels_empty_gen(TRUE));
1956
#endif /* UNIV_SYNC_DEBUG */
1957
1958
goto table_loop;
1959
1960
lock_wait_or_error:
1961
/* See the note at stop_for_a_while: the same holds for this case */
1962
1963
ut_ad(!btr_pcur_is_before_first_on_page(&plan->pcur) || !node->asc);
1964
ut_ad(!search_latch_locked);
1965
1966
plan->stored_cursor_rec_processed = FALSE;
1967
btr_pcur_store_position(&(plan->pcur), &mtr);
1968
1969
mtr_commit(&mtr);
1970
1971
#ifdef UNIV_SYNC_DEBUG
1972
ut_ad(sync_thread_levels_empty_gen(TRUE));
1973
#endif /* UNIV_SYNC_DEBUG */
1974
1975
func_exit:
1976
if (search_latch_locked) {
1977
rw_lock_s_unlock(&btr_search_latch);
1978
}
1979
if (UNIV_LIKELY_NULL(heap)) {
1980
mem_heap_free(heap);
1981
}
1982
return(err);
1983
}
1984
1985
/**************************************************************************
1986
Performs a select step. This is a high-level function used in SQL execution
1987
graphs. */
1988
UNIV_INTERN
1989
que_thr_t*
1990
row_sel_step(
1991
/*=========*/
1992
/* out: query thread to run next or NULL */
1993
que_thr_t* thr) /* in: query thread */
1994
{
1995
ulint i_lock_mode;
1996
sym_node_t* table_node;
1997
sel_node_t* node;
1998
ulint err;
1999
2000
ut_ad(thr);
2001
2002
node = thr->run_node;
2003
2004
ut_ad(que_node_get_type(node) == QUE_NODE_SELECT);
2005
2006
/* If this is a new time this node is executed (or when execution
2007
resumes after wait for a table intention lock), set intention locks
2008
on the tables, or assign a read view */
2009
2010
if (node->into_list && (thr->prev_node == que_node_get_parent(node))) {
2011
2012
node->state = SEL_NODE_OPEN;
2013
}
2014
2015
if (node->state == SEL_NODE_OPEN) {
2016
2017
/* It may be that the current session has not yet started
2018
its transaction, or it has been committed: */
2019
2020
trx_start_if_not_started(thr_get_trx(thr));
2021
2022
plan_reset_cursor(sel_node_get_nth_plan(node, 0));
2023
2024
if (node->consistent_read) {
2025
/* Assign a read view for the query */
2026
node->read_view = trx_assign_read_view(
2027
thr_get_trx(thr));
2028
} else {
2029
if (node->set_x_locks) {
2030
i_lock_mode = LOCK_IX;
2031
} else {
2032
i_lock_mode = LOCK_IS;
2033
}
2034
2035
table_node = node->table_list;
2036
2037
while (table_node) {
2038
err = lock_table(0, table_node->table,
2039
i_lock_mode, thr);
2040
if (err != DB_SUCCESS) {
2041
thr_get_trx(thr)->error_state = err;
2042
2043
return(NULL);
2044
}
2045
2046
table_node = que_node_get_next(table_node);
2047
}
2048
}
2049
2050
/* If this is an explicit cursor, copy stored procedure
2051
variable values, so that the values cannot change between
2052
fetches (currently, we copy them also for non-explicit
2053
cursors) */
2054
2055
if (node->explicit_cursor
2056
&& UT_LIST_GET_FIRST(node->copy_variables)) {
2057
2058
row_sel_copy_input_variable_vals(node);
2059
}
2060
2061
node->state = SEL_NODE_FETCH;
2062
node->fetch_table = 0;
2063
2064
if (node->is_aggregate) {
2065
/* Reset the aggregate total values */
2066
sel_reset_aggregate_vals(node);
2067
}
2068
}
2069
2070
err = row_sel(node, thr);
2071
2072
/* NOTE! if queries are parallelized, the following assignment may
2073
have problems; the assignment should be made only if thr is the
2074
only top-level thr in the graph: */
2075
2076
thr->graph->last_sel_node = node;
2077
2078
if (err != DB_SUCCESS) {
2079
thr_get_trx(thr)->error_state = err;
2080
2081
return(NULL);
2082
}
2083
2084
return(thr);
2085
}
2086
2087
/**************************************************************************
2088
Performs a fetch for a cursor. */
2089
UNIV_INTERN
2090
que_thr_t*
2091
fetch_step(
2092
/*=======*/
2093
/* out: query thread to run next or NULL */
2094
que_thr_t* thr) /* in: query thread */
2095
{
2096
sel_node_t* sel_node;
2097
fetch_node_t* node;
2098
2099
ut_ad(thr);
2100
2101
node = thr->run_node;
2102
sel_node = node->cursor_def;
2103
2104
ut_ad(que_node_get_type(node) == QUE_NODE_FETCH);
2105
2106
if (thr->prev_node != que_node_get_parent(node)) {
2107
2108
if (sel_node->state != SEL_NODE_NO_MORE_ROWS) {
2109
2110
if (node->into_list) {
2111
sel_assign_into_var_values(node->into_list,
2112
sel_node);
2113
} else {
2114
void* ret = (*node->func->func)(
2115
sel_node, node->func->arg);
2116
2117
if (!ret) {
2118
sel_node->state
2119
= SEL_NODE_NO_MORE_ROWS;
2120
}
2121
}
2122
}
2123
2124
thr->run_node = que_node_get_parent(node);
2125
2126
return(thr);
2127
}
2128
2129
/* Make the fetch node the parent of the cursor definition for
2130
the time of the fetch, so that execution knows to return to this
2131
fetch node after a row has been selected or we know that there is
2132
no row left */
2133
2134
sel_node->common.parent = node;
2135
2136
if (sel_node->state == SEL_NODE_CLOSED) {
2137
fprintf(stderr,
2138
"InnoDB: Error: fetch called on a closed cursor\n");
2139
2140
thr_get_trx(thr)->error_state = DB_ERROR;
2141
2142
return(NULL);
2143
}
2144
2145
thr->run_node = sel_node;
2146
2147
return(thr);
2148
}
2149
2150
/********************************************************************
2151
Sample callback function for fetch that prints each row.*/
2152
UNIV_INTERN
2153
void*
2154
row_fetch_print(
2155
/*============*/
2156
/* out: always returns non-NULL */
2157
void* row, /* in: sel_node_t* */
2158
void* user_arg) /* in: not used */
2159
{
2160
sel_node_t* node = row;
2161
que_node_t* exp;
2162
ulint i = 0;
2163
2164
UT_NOT_USED(user_arg);
2165
2166
fprintf(stderr, "row_fetch_print: row %p\n", row);
2167
2168
exp = node->select_list;
2169
2170
while (exp) {
2171
dfield_t* dfield = que_node_get_val(exp);
2172
const dtype_t* type = dfield_get_type(dfield);
2173
2174
fprintf(stderr, " column %lu:\n", (ulong)i);
2175
2176
dtype_print(type);
2177
fprintf(stderr, "\n");
2178
2179
if (dfield_get_len(dfield) != UNIV_SQL_NULL) {
2180
ut_print_buf(stderr, dfield_get_data(dfield),
2181
dfield_get_len(dfield));
2182
} else {
2183
fprintf(stderr, " <NULL>;");
2184
}
2185
2186
fprintf(stderr, "\n");
2187
2188
exp = que_node_get_next(exp);
2189
i++;
2190
}
2191
2192
return((void*)42);
2193
}
2194
2195
/********************************************************************
2196
Callback function for fetch that stores an unsigned 4 byte integer to the
2197
location pointed. The column's type must be DATA_INT, DATA_UNSIGNED, length
2198
= 4. */
2199
UNIV_INTERN
2200
void*
2201
row_fetch_store_uint4(
2202
/*==================*/
2203
/* out: always returns NULL */
2204
void* row, /* in: sel_node_t* */
2205
void* user_arg) /* in: data pointer */
2206
{
2207
sel_node_t* node = row;
2208
ib_uint32_t* val = user_arg;
2209
ulint tmp;
2210
2211
dfield_t* dfield = que_node_get_val(node->select_list);
2212
const dtype_t* type = dfield_get_type(dfield);
2213
ulint len = dfield_get_len(dfield);
2214
2215
ut_a(dtype_get_mtype(type) == DATA_INT);
2216
ut_a(dtype_get_prtype(type) & DATA_UNSIGNED);
2217
ut_a(len == 4);
2218
2219
tmp = mach_read_from_4(dfield_get_data(dfield));
2220
*val = (ib_uint32_t) tmp;
2221
2222
return(NULL);
2223
}
2224
2225
/***************************************************************
2226
Prints a row in a select result. */
2227
UNIV_INTERN
2228
que_thr_t*
2229
row_printf_step(
2230
/*============*/
2231
/* out: query thread to run next or NULL */
2232
que_thr_t* thr) /* in: query thread */
2233
{
2234
row_printf_node_t* node;
2235
sel_node_t* sel_node;
2236
que_node_t* arg;
2237
2238
ut_ad(thr);
2239
2240
node = thr->run_node;
2241
2242
sel_node = node->sel_node;
2243
2244
ut_ad(que_node_get_type(node) == QUE_NODE_ROW_PRINTF);
2245
2246
if (thr->prev_node == que_node_get_parent(node)) {
2247
2248
/* Reset the cursor */
2249
sel_node->state = SEL_NODE_OPEN;
2250
2251
/* Fetch next row to print */
2252
2253
thr->run_node = sel_node;
2254
2255
return(thr);
2256
}
2257
2258
if (sel_node->state != SEL_NODE_FETCH) {
2259
2260
ut_ad(sel_node->state == SEL_NODE_NO_MORE_ROWS);
2261
2262
/* No more rows to print */
2263
2264
thr->run_node = que_node_get_parent(node);
2265
2266
return(thr);
2267
}
2268
2269
arg = sel_node->select_list;
2270
2271
while (arg) {
2272
dfield_print_also_hex(que_node_get_val(arg));
2273
2274
fputs(" ::: ", stderr);
2275
2276
arg = que_node_get_next(arg);
2277
}
2278
2279
putc('\n', stderr);
2280
2281
/* Fetch next row to print */
2282
2283
thr->run_node = sel_node;
2284
2285
return(thr);
2286
}
2287
2288
/********************************************************************
2289
Converts a key value stored in MySQL format to an Innobase dtuple. The last
2290
field of the key value may be just a prefix of a fixed length field: hence
2291
the parameter key_len. But currently we do not allow search keys where the
2292
last field is only a prefix of the full key field len and print a warning if
2293
such appears. A counterpart of this function is
2294
ha_innobase::store_key_val_for_row() in ha_innodb.cc. */
2295
UNIV_INTERN
2296
void
2297
row_sel_convert_mysql_key_to_innobase(
2298
/*==================================*/
2299
dtuple_t* tuple, /* in/out: tuple where to build;
2300
NOTE: we assume that the type info
2301
in the tuple is already according
2302
to index! */
2303
byte* buf, /* in: buffer to use in field
2304
conversions */
2305
ulint buf_len, /* in: buffer length */
2306
dict_index_t* index, /* in: index of the key value */
2307
const byte* key_ptr, /* in: MySQL key value */
2308
ulint key_len, /* in: MySQL key value length */
2309
trx_t* trx) /* in: transaction */
2310
{
2311
byte* original_buf = buf;
2312
const byte* original_key_ptr = key_ptr;
2313
dict_field_t* field;
2314
dfield_t* dfield;
2315
ulint data_offset;
2316
ulint data_len;
2317
ulint data_field_len;
2318
ibool is_null;
2319
const byte* key_end;
2320
ulint n_fields = 0;
2321
2322
/* For documentation of the key value storage format in MySQL, see
2323
ha_innobase::store_key_val_for_row() in ha_innodb.cc. */
2324
2325
key_end = key_ptr + key_len;
2326
2327
/* Permit us to access any field in the tuple (ULINT_MAX): */
2328
2329
dtuple_set_n_fields(tuple, ULINT_MAX);
2330
2331
dfield = dtuple_get_nth_field(tuple, 0);
2332
field = dict_index_get_nth_field(index, 0);
2333
2334
if (UNIV_UNLIKELY(dfield_get_type(dfield)->mtype == DATA_SYS)) {
2335
/* A special case: we are looking for a position in the
2336
generated clustered index which InnoDB automatically added
2337
to a table with no primary key: the first and the only
2338
ordering column is ROW_ID which InnoDB stored to the key_ptr
2339
buffer. */
2340
2341
ut_a(key_len == DATA_ROW_ID_LEN);
2342
2343
dfield_set_data(dfield, key_ptr, DATA_ROW_ID_LEN);
2344
2345
dtuple_set_n_fields(tuple, 1);
2346
2347
return;
2348
}
2349
2350
while (key_ptr < key_end) {
2351
2352
ulint type = dfield_get_type(dfield)->mtype;
2353
ut_a(field->col->mtype == type);
2354
2355
data_offset = 0;
2356
is_null = FALSE;
2357
2358
if (!(dfield_get_type(dfield)->prtype & DATA_NOT_NULL)) {
2359
/* The first byte in the field tells if this is
2360
an SQL NULL value */
2361
2362
data_offset = 1;
2363
2364
if (*key_ptr != 0) {
2365
dfield_set_null(dfield);
2366
2367
is_null = TRUE;
2368
}
2369
}
2370
2371
/* Calculate data length and data field total length */
2372
2373
if (type == DATA_BLOB) {
2374
/* The key field is a column prefix of a BLOB or
2375
TEXT */
2376
2377
ut_a(field->prefix_len > 0);
2378
2379
/* MySQL stores the actual data length to the first 2
2380
bytes after the optional SQL NULL marker byte. The
2381
storage format is little-endian, that is, the most
2382
significant byte at a higher address. In UTF-8, MySQL
2383
seems to reserve field->prefix_len bytes for
2384
storing this field in the key value buffer, even
2385
though the actual value only takes data_len bytes
2386
from the start. */
2387
2388
data_len = key_ptr[data_offset]
2389
+ 256 * key_ptr[data_offset + 1];
2390
data_field_len = data_offset + 2 + field->prefix_len;
2391
2392
data_offset += 2;
2393
2394
/* Now that we know the length, we store the column
2395
value like it would be a fixed char field */
2396
2397
} else if (field->prefix_len > 0) {
2398
/* Looks like MySQL pads unused end bytes in the
2399
prefix with space. Therefore, also in UTF-8, it is ok
2400
to compare with a prefix containing full prefix_len
2401
bytes, and no need to take at most prefix_len / 3
2402
UTF-8 characters from the start.
2403
If the prefix is used as the upper end of a LIKE
2404
'abc%' query, then MySQL pads the end with chars
2405
0xff. TODO: in that case does it any harm to compare
2406
with the full prefix_len bytes. How do characters
2407
0xff in UTF-8 behave? */
2408
2409
data_len = field->prefix_len;
2410
data_field_len = data_offset + data_len;
2411
} else {
2412
data_len = dfield_get_type(dfield)->len;
2413
data_field_len = data_offset + data_len;
2414
}
2415
2416
if (UNIV_UNLIKELY
2417
(dtype_get_mysql_type(dfield_get_type(dfield))
2418
== DATA_MYSQL_TRUE_VARCHAR)
2419
&& UNIV_LIKELY(type != DATA_INT)) {
2420
/* In a MySQL key value format, a true VARCHAR is
2421
always preceded by 2 bytes of a length field.
2422
dfield_get_type(dfield)->len returns the maximum
2423
'payload' len in bytes. That does not include the
2424
2 bytes that tell the actual data length.
2425
2426
We added the check != DATA_INT to make sure we do
2427
not treat MySQL ENUM or SET as a true VARCHAR! */
2428
2429
data_len += 2;
2430
data_field_len += 2;
2431
}
2432
2433
/* Storing may use at most data_len bytes of buf */
2434
2435
if (UNIV_LIKELY(!is_null)) {
2436
row_mysql_store_col_in_innobase_format(
2437
dfield, buf,
2438
FALSE, /* MySQL key value format col */
2439
key_ptr + data_offset, data_len,
2440
dict_table_is_comp(index->table));
2441
buf += data_len;
2442
}
2443
2444
key_ptr += data_field_len;
2445
2446
if (UNIV_UNLIKELY(key_ptr > key_end)) {
2447
/* The last field in key was not a complete key field
2448
but a prefix of it.
2449
2450
Print a warning about this! HA_READ_PREFIX_LAST does
2451
not currently work in InnoDB with partial-field key
2452
value prefixes. Since MySQL currently uses a padding
2453
trick to calculate LIKE 'abc%' type queries there
2454
should never be partial-field prefixes in searches. */
2455
2456
ut_print_timestamp(stderr);
2457
2458
fputs(" InnoDB: Warning: using a partial-field"
2459
" key prefix in search.\n"
2460
"InnoDB: ", stderr);
2461
dict_index_name_print(stderr, trx, index);
2462
fprintf(stderr, ". Last data field length %lu bytes,\n"
2463
"InnoDB: key ptr now exceeds"
2464
" key end by %lu bytes.\n"
2465
"InnoDB: Key value in the MySQL format:\n",
2466
(ulong) data_field_len,
2467
(ulong) (key_ptr - key_end));
2468
fflush(stderr);
2469
ut_print_buf(stderr, original_key_ptr, key_len);
2470
fprintf(stderr, "\n");
2471
2472
if (!is_null) {
2473
ulint len = dfield_get_len(dfield);
2474
dfield_set_len(dfield, len
2475
- (ulint) (key_ptr - key_end));
2476
}
2477
}
2478
2479
n_fields++;
2480
field++;
2481
dfield++;
2482
}
2483
2484
ut_a(buf <= original_buf + buf_len);
2485
2486
/* We set the length of tuple to n_fields: we assume that the memory
2487
area allocated for it is big enough (usually bigger than n_fields). */
2488
2489
dtuple_set_n_fields(tuple, n_fields);
2490
}
2491
2492
/******************************************************************
2493
Stores the row id to the prebuilt struct. */
2494
static
2495
void
2496
row_sel_store_row_id_to_prebuilt(
2497
/*=============================*/
2498
row_prebuilt_t* prebuilt, /* in/out: prebuilt */
2499
const rec_t* index_rec, /* in: record */
2500
const dict_index_t* index, /* in: index of the record */
2501
const ulint* offsets) /* in: rec_get_offsets
2502
(index_rec, index) */
2503
{
2504
const byte* data;
2505
ulint len;
2506
2507
ut_ad(rec_offs_validate(index_rec, index, offsets));
2508
2509
data = rec_get_nth_field(
2510
index_rec, offsets,
2511
dict_index_get_sys_col_pos(index, DATA_ROW_ID), &len);
2512
2513
if (UNIV_UNLIKELY(len != DATA_ROW_ID_LEN)) {
2514
fprintf(stderr,
2515
"InnoDB: Error: Row id field is"
2516
" wrong length %lu in ", (ulong) len);
2517
dict_index_name_print(stderr, prebuilt->trx, index);
2518
fprintf(stderr, "\n"
2519
"InnoDB: Field number %lu, record:\n",
2520
(ulong) dict_index_get_sys_col_pos(index,
2521
DATA_ROW_ID));
2522
rec_print_new(stderr, index_rec, offsets);
2523
putc('\n', stderr);
2524
ut_error;
2525
}
2526
2527
ut_memcpy(prebuilt->row_id, data, len);
2528
}
2529
2530
/******************************************************************
2531
Stores a non-SQL-NULL field in the MySQL format. The counterpart of this
2532
function is row_mysql_store_col_in_innobase_format() in row0mysql.c. */
2533
static
2534
void
2535
row_sel_field_store_in_mysql_format(
2536
/*================================*/
2537
byte* dest, /* in/out: buffer where to store; NOTE
2538
that BLOBs are not in themselves
2539
stored here: the caller must allocate
2540
and copy the BLOB into buffer before,
2541
and pass the pointer to the BLOB in
2542
'data' */
2543
const mysql_row_templ_t* templ,
2544
/* in: MySQL column template.
2545
Its following fields are referenced:
2546
type, is_unsigned, mysql_col_len,
2547
mbminlen, mbmaxlen */
2548
const byte* data, /* in: data to store */
2549
ulint len) /* in: length of the data */
2550
{
2551
byte* ptr;
2552
byte* field_end;
2553
byte* pad_ptr;
2554
2555
ut_ad(len != UNIV_SQL_NULL);
2556
2557
switch (templ->type) {
2558
case DATA_INT:
2559
/* Convert integer data from Innobase to a little-endian
2560
format, sign bit restored to normal */
2561
2562
ptr = dest + len;
2563
2564
for (;;) {
2565
ptr--;
2566
*ptr = *data;
2567
if (ptr == dest) {
2568
break;
2569
}
2570
data++;
2571
}
2572
2573
if (!templ->is_unsigned) {
2574
dest[len - 1] = (byte) (dest[len - 1] ^ 128);
2575
}
2576
2577
ut_ad(templ->mysql_col_len == len);
2578
break;
2579
2580
case DATA_VARCHAR:
2581
case DATA_VARMYSQL:
2582
case DATA_BINARY:
2583
field_end = dest + templ->mysql_col_len;
2584
2585
if (templ->mysql_type == DATA_MYSQL_TRUE_VARCHAR) {
2586
/* This is a >= 5.0.3 type true VARCHAR. Store the
2587
length of the data to the first byte or the first
2588
two bytes of dest. */
2589
2590
dest = row_mysql_store_true_var_len(
2591
dest, len, templ->mysql_length_bytes);
2592
}
2593
2594
/* Copy the actual data */
2595
ut_memcpy(dest, data, len);
2596
2597
/* Pad with trailing spaces. We pad with spaces also the
2598
unused end of a >= 5.0.3 true VARCHAR column, just in case
2599
MySQL expects its contents to be deterministic. */
2600
2601
pad_ptr = dest + len;
2602
2603
ut_ad(templ->mbminlen <= templ->mbmaxlen);
2604
2605
/* We handle UCS2 charset strings differently. */
2606
if (templ->mbminlen == 2) {
2607
/* A space char is two bytes, 0x0020 in UCS2 */
2608
2609
if (len & 1) {
2610
/* A 0x20 has been stripped from the column.
2611
Pad it back. */
2612
2613
if (pad_ptr < field_end) {
2614
*pad_ptr = 0x20;
2615
pad_ptr++;
2616
}
2617
}
2618
2619
/* Pad the rest of the string with 0x0020 */
2620
2621
while (pad_ptr < field_end) {
2622
*pad_ptr = 0x00;
2623
pad_ptr++;
2624
*pad_ptr = 0x20;
2625
pad_ptr++;
2626
}
2627
} else {
2628
ut_ad(templ->mbminlen == 1);
2629
/* space=0x20 */
2630
2631
memset(pad_ptr, 0x20, field_end - pad_ptr);
2632
}
2633
break;
2634
2635
case DATA_BLOB:
2636
/* Store a pointer to the BLOB buffer to dest: the BLOB was
2637
already copied to the buffer in row_sel_store_mysql_rec */
2638
2639
row_mysql_store_blob_ref(dest, templ->mysql_col_len, data,
2640
len);
2641
break;
2642
2643
case DATA_MYSQL:
2644
memcpy(dest, data, len);
2645
2646
ut_ad(templ->mysql_col_len >= len);
2647
ut_ad(templ->mbmaxlen >= templ->mbminlen);
2648
2649
ut_ad(templ->mbmaxlen > templ->mbminlen
2650
|| templ->mysql_col_len == len);
2651
/* The following assertion would fail for old tables
2652
containing UTF-8 ENUM columns due to Bug #9526. */
2653
ut_ad(!templ->mbmaxlen
2654
|| !(templ->mysql_col_len % templ->mbmaxlen));
2655
ut_ad(len * templ->mbmaxlen >= templ->mysql_col_len);
2656
2657
if (templ->mbminlen != templ->mbmaxlen) {
2658
/* Pad with spaces. This undoes the stripping
2659
done in row0mysql.ic, function
2660
row_mysql_store_col_in_innobase_format(). */
2661
2662
memset(dest + len, 0x20, templ->mysql_col_len - len);
2663
}
2664
break;
2665
2666
default:
2667
#ifdef UNIV_DEBUG
2668
case DATA_SYS_CHILD:
2669
case DATA_SYS:
2670
/* These column types should never be shipped to MySQL. */
2671
ut_ad(0);
2672
2673
case DATA_CHAR:
2674
case DATA_FIXBINARY:
2675
case DATA_FLOAT:
2676
case DATA_DOUBLE:
2677
case DATA_DECIMAL:
2678
/* Above are the valid column types for MySQL data. */
2679
#endif /* UNIV_DEBUG */
2680
ut_ad(templ->mysql_col_len == len);
2681
memcpy(dest, data, len);
2682
}
2683
}
2684
2685
/******************************************************************
2686
Convert a row in the Innobase format to a row in the MySQL format.
2687
Note that the template in prebuilt may advise us to copy only a few
2688
columns to mysql_rec, other columns are left blank. All columns may not
2689
be needed in the query. */
2690
static
2691
ibool
2692
row_sel_store_mysql_rec(
2693
/*====================*/
2694
/* out: TRUE if success, FALSE if
2695
could not allocate memory for a BLOB
2696
(though we may also assert in that
2697
case) */
2698
byte* mysql_rec, /* out: row in the MySQL format */
2699
row_prebuilt_t* prebuilt, /* in: prebuilt struct */
2700
const rec_t* rec, /* in: Innobase record in the index
2701
which was described in prebuilt's
2702
template; must be protected by
2703
a page latch */
2704
const ulint* offsets, /* in: array returned by
2705
rec_get_offsets() */
2706
ulint start_field_no,
2707
ulint end_field_no)
2708
{
2709
mysql_row_templ_t* templ;
2710
mem_heap_t* extern_field_heap = NULL;
2711
mem_heap_t* heap;
2712
const byte* data;
2713
ulint len;
2714
ulint i;
2715
2716
ut_ad(prebuilt->mysql_template);
2717
ut_ad(rec_offs_validate(rec, NULL, offsets));
2718
2719
if (UNIV_LIKELY_NULL(prebuilt->blob_heap)) {
2720
mem_heap_free(prebuilt->blob_heap);
2721
prebuilt->blob_heap = NULL;
2722
}
2723
2724
for (i = start_field_no; i < end_field_no /* prebuilt->n_template */; i++) {
2725
2726
templ = prebuilt->mysql_template + i;
2727
2728
if (UNIV_UNLIKELY(rec_offs_nth_extern(offsets,
2729
templ->rec_field_no))) {
2730
2731
/* Copy an externally stored field to the temporary
2732
heap */
2733
2734
ut_a(!prebuilt->trx->has_search_latch);
2735
2736
if (UNIV_UNLIKELY(templ->type == DATA_BLOB)) {
2737
if (prebuilt->blob_heap == NULL) {
2738
prebuilt->blob_heap = mem_heap_create(
2739
UNIV_PAGE_SIZE);
2740
}
2741
2742
heap = prebuilt->blob_heap;
2743
} else {
2744
extern_field_heap
2745
= mem_heap_create(UNIV_PAGE_SIZE);
2746
2747
heap = extern_field_heap;
2748
}
2749
2750
/* NOTE: if we are retrieving a big BLOB, we may
2751
already run out of memory in the next call, which
2752
causes an assert */
2753
2754
data = btr_rec_copy_externally_stored_field(
2755
rec, offsets,
2756
dict_table_zip_size(prebuilt->table),
2757
templ->rec_field_no, &len, heap);
2758
2759
ut_a(len != UNIV_SQL_NULL);
2760
} else {
2761
/* Field is stored in the row. */
2762
2763
data = rec_get_nth_field(rec, offsets,
2764
templ->rec_field_no, &len);
2765
2766
if (UNIV_UNLIKELY(templ->type == DATA_BLOB)
2767
&& len != UNIV_SQL_NULL) {
2768
2769
/* It is a BLOB field locally stored in the
2770
InnoDB record: we MUST copy its contents to
2771
prebuilt->blob_heap here because later code
2772
assumes all BLOB values have been copied to a
2773
safe place. */
2774
2775
if (prebuilt->blob_heap == NULL) {
2776
prebuilt->blob_heap = mem_heap_create(
2777
UNIV_PAGE_SIZE);
2778
}
2779
2780
data = memcpy(mem_heap_alloc(
2781
prebuilt->blob_heap, len),
2782
data, len);
2783
}
2784
}
2785
2786
if (len != UNIV_SQL_NULL) {
2787
row_sel_field_store_in_mysql_format(
2788
mysql_rec + templ->mysql_col_offset,
2789
templ, data, len);
2790
2791
/* Cleanup */
2792
if (extern_field_heap) {
2793
mem_heap_free(extern_field_heap);
2794
extern_field_heap = NULL;
2795
}
2796
2797
if (templ->mysql_null_bit_mask) {
2798
/* It is a nullable column with a non-NULL
2799
value */
2800
mysql_rec[templ->mysql_null_byte_offset]
2801
&= ~(byte) templ->mysql_null_bit_mask;
2802
}
2803
} else {
2804
/* MySQL seems to assume the field for an SQL NULL
2805
value is set to zero or space. Not taking this into
2806
account caused seg faults with NULL BLOB fields, and
2807
bug number 154 in the MySQL bug database: GROUP BY
2808
and DISTINCT could treat NULL values inequal. */
2809
int pad_char;
2810
2811
mysql_rec[templ->mysql_null_byte_offset]
2812
|= (byte) templ->mysql_null_bit_mask;
2813
switch (templ->type) {
2814
case DATA_VARCHAR:
2815
case DATA_BINARY:
2816
case DATA_VARMYSQL:
2817
if (templ->mysql_type
2818
== DATA_MYSQL_TRUE_VARCHAR) {
2819
/* This is a >= 5.0.3 type
2820
true VARCHAR. Zero the field. */
2821
pad_char = 0x00;
2822
break;
2823
}
2824
/* Fall through */
2825
case DATA_CHAR:
2826
case DATA_FIXBINARY:
2827
case DATA_MYSQL:
2828
/* MySQL pads all string types (except
2829
BLOB, TEXT and true VARCHAR) with space. */
2830
if (UNIV_UNLIKELY(templ->mbminlen == 2)) {
2831
/* Treat UCS2 as a special case. */
2832
byte* d = mysql_rec
2833
+ templ->mysql_col_offset;
2834
len = templ->mysql_col_len;
2835
/* There are two UCS2 bytes per char,
2836
so the length has to be even. */
2837
ut_a(!(len & 1));
2838
/* Pad with 0x0020. */
2839
while (len) {
2840
*d++ = 0x00;
2841
*d++ = 0x20;
2842
len -= 2;
2843
}
2844
continue;
2845
}
2846
pad_char = 0x20;
2847
break;
2848
default:
2849
pad_char = 0x00;
2850
break;
2851
}
2852
2853
ut_ad(!pad_char || templ->mbminlen == 1);
2854
memset(mysql_rec + templ->mysql_col_offset,
2855
pad_char, templ->mysql_col_len);
2856
}
2857
}
2858
2859
return(TRUE);
2860
}
2861
2862
/*************************************************************************
2863
Builds a previous version of a clustered index record for a consistent read */
2864
static
2865
ulint
2866
row_sel_build_prev_vers_for_mysql(
2867
/*==============================*/
2868
/* out: DB_SUCCESS or error code */
2869
read_view_t* read_view, /* in: read view */
2870
dict_index_t* clust_index, /* in: clustered index */
2871
row_prebuilt_t* prebuilt, /* in: prebuilt struct */
2872
const rec_t* rec, /* in: record in a clustered index */
2873
ulint** offsets, /* in/out: offsets returned by
2874
rec_get_offsets(rec, clust_index) */
2875
mem_heap_t** offset_heap, /* in/out: memory heap from which
2876
the offsets are allocated */
2877
rec_t** old_vers, /* out: old version, or NULL if the
2878
record does not exist in the view:
2879
i.e., it was freshly inserted
2880
afterwards */
2881
mtr_t* mtr) /* in: mtr */
2882
{
2883
ulint err;
2884
2885
if (prebuilt->old_vers_heap) {
2886
mem_heap_empty(prebuilt->old_vers_heap);
2887
} else {
2888
prebuilt->old_vers_heap = mem_heap_create(200);
2889
}
2890
2891
err = row_vers_build_for_consistent_read(
2892
rec, mtr, clust_index, offsets, read_view, offset_heap,
2893
prebuilt->old_vers_heap, old_vers);
2894
return(err);
2895
}
2896
2897
/*************************************************************************
2898
Retrieves the clustered index record corresponding to a record in a
2899
non-clustered index. Does the necessary locking. Used in the MySQL
2900
interface. */
2901
static
2902
ulint
2903
row_sel_get_clust_rec_for_mysql(
2904
/*============================*/
2905
/* out: DB_SUCCESS or error code */
2906
row_prebuilt_t* prebuilt,/* in: prebuilt struct in the handle */
2907
dict_index_t* sec_index,/* in: secondary index where rec resides */
2908
const rec_t* rec, /* in: record in a non-clustered index; if
2909
this is a locking read, then rec is not
2910
allowed to be delete-marked, and that would
2911
not make sense either */
2912
que_thr_t* thr, /* in: query thread */
2913
const rec_t** out_rec,/* out: clustered record or an old version of
2914
it, NULL if the old version did not exist
2915
in the read view, i.e., it was a fresh
2916
inserted version */
2917
ulint** offsets,/* in: offsets returned by
2918
rec_get_offsets(rec, sec_index);
2919
out: offsets returned by
2920
rec_get_offsets(out_rec, clust_index) */
2921
mem_heap_t** offset_heap,/* in/out: memory heap from which
2922
the offsets are allocated */
2923
mtr_t* mtr) /* in: mtr used to get access to the
2924
non-clustered record; the same mtr is used to
2925
access the clustered index */
2926
{
2927
dict_index_t* clust_index;
2928
const rec_t* clust_rec;
2929
rec_t* old_vers;
2930
ulint err;
2931
trx_t* trx;
2932
2933
*out_rec = NULL;
2934
trx = thr_get_trx(thr);
2935
2936
row_build_row_ref_in_tuple(prebuilt->clust_ref, rec,
2937
sec_index, *offsets, trx);
2938
2939
clust_index = dict_table_get_first_index(sec_index->table);
2940
2941
btr_pcur_open_with_no_init(clust_index, prebuilt->clust_ref,
2942
PAGE_CUR_LE, BTR_SEARCH_LEAF,
2943
prebuilt->clust_pcur, 0, mtr);
2944
2945
clust_rec = btr_pcur_get_rec(prebuilt->clust_pcur);
2946
2947
prebuilt->clust_pcur->trx_if_known = trx;
2948
2949
/* Note: only if the search ends up on a non-infimum record is the
2950
low_match value the real match to the search tuple */
2951
2952
if (!page_rec_is_user_rec(clust_rec)
2953
|| btr_pcur_get_low_match(prebuilt->clust_pcur)
2954
< dict_index_get_n_unique(clust_index)) {
2955
2956
/* In a rare case it is possible that no clust rec is found
2957
for a delete-marked secondary index record: if in row0umod.c
2958
in row_undo_mod_remove_clust_low() we have already removed
2959
the clust rec, while purge is still cleaning and removing
2960
secondary index records associated with earlier versions of
2961
the clustered index record. In that case we know that the
2962
clustered index record did not exist in the read view of
2963
trx. */
2964
2965
if (!rec_get_deleted_flag(rec,
2966
dict_table_is_comp(sec_index->table))
2967
|| prebuilt->select_lock_type != LOCK_NONE) {
2968
ut_print_timestamp(stderr);
2969
fputs(" InnoDB: error clustered record"
2970
" for sec rec not found\n"
2971
"InnoDB: ", stderr);
2972
dict_index_name_print(stderr, trx, sec_index);
2973
fputs("\n"
2974
"InnoDB: sec index record ", stderr);
2975
rec_print(stderr, rec, sec_index);
2976
fputs("\n"
2977
"InnoDB: clust index record ", stderr);
2978
rec_print(stderr, clust_rec, clust_index);
2979
putc('\n', stderr);
2980
trx_print(stderr, trx, 600);
2981
2982
fputs("\n"
2983
"InnoDB: Submit a detailed bug report"
2984
" to http://bugs.mysql.com\n", stderr);
2985
}
2986
2987
clust_rec = NULL;
2988
2989
goto func_exit;
2990
}
2991
2992
*offsets = rec_get_offsets(clust_rec, clust_index, *offsets,
2993
ULINT_UNDEFINED, offset_heap);
2994
2995
if (prebuilt->select_lock_type != LOCK_NONE) {
2996
/* Try to place a lock on the index record; we are searching
2997
the clust rec with a unique condition, hence
2998
we set a LOCK_REC_NOT_GAP type lock */
2999
3000
err = lock_clust_rec_read_check_and_lock(
3001
0, btr_pcur_get_block(prebuilt->clust_pcur),
3002
clust_rec, clust_index, *offsets,
3003
prebuilt->select_lock_type, LOCK_REC_NOT_GAP, thr);
3004
if (err != DB_SUCCESS) {
3005
3006
goto err_exit;
3007
}
3008
} else {
3009
/* This is a non-locking consistent read: if necessary, fetch
3010
a previous version of the record */
3011
3012
old_vers = NULL;
3013
3014
/* If the isolation level allows reading of uncommitted data,
3015
then we never look for an earlier version */
3016
3017
if (trx->isolation_level > TRX_ISO_READ_UNCOMMITTED
3018
&& !lock_clust_rec_cons_read_sees(
3019
clust_rec, clust_index, *offsets,
3020
trx->read_view)) {
3021
3022
/* The following call returns 'offsets' associated with
3023
'old_vers' */
3024
err = row_sel_build_prev_vers_for_mysql(
3025
trx->read_view, clust_index, prebuilt,
3026
clust_rec, offsets, offset_heap, &old_vers,
3027
mtr);
3028
3029
if (err != DB_SUCCESS || old_vers == NULL) {
3030
3031
goto err_exit;
3032
}
3033
3034
clust_rec = old_vers;
3035
}
3036
3037
/* If we had to go to an earlier version of row or the
3038
secondary index record is delete marked, then it may be that
3039
the secondary index record corresponding to clust_rec
3040
(or old_vers) is not rec; in that case we must ignore
3041
such row because in our snapshot rec would not have existed.
3042
Remember that from rec we cannot see directly which transaction
3043
id corresponds to it: we have to go to the clustered index
3044
record. A query where we want to fetch all rows where
3045
the secondary index value is in some interval would return
3046
a wrong result if we would not drop rows which we come to
3047
visit through secondary index records that would not really
3048
exist in our snapshot. */
3049
3050
if (clust_rec
3051
&& (old_vers
3052
|| rec_get_deleted_flag(rec, dict_table_is_comp(
3053
sec_index->table)))
3054
&& !row_sel_sec_rec_is_for_clust_rec(
3055
rec, sec_index, clust_rec, clust_index)) {
3056
clust_rec = NULL;
3057
#ifdef UNIV_SEARCH_DEBUG
3058
} else {
3059
ut_a(clust_rec == NULL
3060
|| row_sel_sec_rec_is_for_clust_rec(
3061
rec, sec_index, clust_rec, clust_index));
3062
#endif
3063
}
3064
}
3065
3066
func_exit:
3067
*out_rec = clust_rec;
3068
3069
if (prebuilt->select_lock_type == LOCK_X) {
3070
/* We may use the cursor in update: store its position */
3071
3072
btr_pcur_store_position(prebuilt->clust_pcur, mtr);
3073
}
3074
3075
err = DB_SUCCESS;
3076
err_exit:
3077
return(err);
3078
}
3079
3080
/************************************************************************
3081
Restores cursor position after it has been stored. We have to take into
3082
account that the record cursor was positioned on may have been deleted.
3083
Then we may have to move the cursor one step up or down. */
3084
static
3085
ibool
3086
sel_restore_position_for_mysql(
3087
/*===========================*/
3088
/* out: TRUE if we may need to
3089
process the record the cursor is
3090
now positioned on (i.e. we should
3091
not go to the next record yet) */
3092
ibool* same_user_rec, /* out: TRUE if we were able to restore
3093
the cursor on a user record with the
3094
same ordering prefix in in the
3095
B-tree index */
3096
ulint latch_mode, /* in: latch mode wished in
3097
restoration */
3098
btr_pcur_t* pcur, /* in: cursor whose position
3099
has been stored */
3100
ibool moves_up, /* in: TRUE if the cursor moves up
3101
in the index */
3102
mtr_t* mtr) /* in: mtr; CAUTION: may commit
3103
mtr temporarily! */
3104
{
3105
ibool success;
3106
ulint relative_position;
3107
3108
relative_position = pcur->rel_pos;
3109
3110
success = btr_pcur_restore_position(latch_mode, pcur, mtr);
3111
3112
*same_user_rec = success;
3113
3114
if (relative_position == BTR_PCUR_ON) {
3115
if (success) {
3116
return(FALSE);
3117
}
3118
3119
if (moves_up) {
3120
btr_pcur_move_to_next(pcur, mtr);
3121
}
3122
3123
return(TRUE);
3124
}
3125
3126
if (relative_position == BTR_PCUR_AFTER
3127
|| relative_position == BTR_PCUR_AFTER_LAST_IN_TREE) {
3128
3129
if (moves_up) {
3130
return(TRUE);
3131
}
3132
3133
if (btr_pcur_is_on_user_rec(pcur)) {
3134
btr_pcur_move_to_prev(pcur, mtr);
3135
}
3136
3137
return(TRUE);
3138
}
3139
3140
ut_ad(relative_position == BTR_PCUR_BEFORE
3141
|| relative_position == BTR_PCUR_BEFORE_FIRST_IN_TREE);
3142
3143
if (moves_up && btr_pcur_is_on_user_rec(pcur)) {
3144
btr_pcur_move_to_next(pcur, mtr);
3145
}
3146
3147
return(TRUE);
3148
}
3149
3150
/************************************************************************
3151
Pops a cached row for MySQL from the fetch cache. */
3152
UNIV_INLINE
3153
void
3154
row_sel_pop_cached_row_for_mysql(
3155
/*=============================*/
3156
byte* buf, /* in/out: buffer where to copy the
3157
row */
3158
row_prebuilt_t* prebuilt) /* in: prebuilt struct */
3159
{
3160
ulint i;
3161
mysql_row_templ_t* templ;
3162
byte* cached_rec;
3163
ut_ad(prebuilt->n_fetch_cached > 0);
3164
ut_ad(prebuilt->mysql_prefix_len <= prebuilt->mysql_row_len);
3165
3166
if (UNIV_UNLIKELY(prebuilt->keep_other_fields_on_keyread)) {
3167
/* Copy cache record field by field, don't touch fields that
3168
are not covered by current key */
3169
cached_rec = prebuilt->fetch_cache[
3170
prebuilt->fetch_cache_first];
3171
3172
for (i = 0; i < prebuilt->n_template; i++) {
3173
templ = prebuilt->mysql_template + i;
3174
ut_memcpy(buf + templ->mysql_col_offset,
3175
cached_rec + templ->mysql_col_offset,
3176
templ->mysql_col_len);
3177
/* Copy NULL bit of the current field from cached_rec
3178
to buf */
3179
if (templ->mysql_null_bit_mask) {
3180
buf[templ->mysql_null_byte_offset]
3181
^= (buf[templ->mysql_null_byte_offset]
3182
^ cached_rec[templ->mysql_null_byte_offset])
3183
& (byte)templ->mysql_null_bit_mask;
3184
}
3185
}
3186
}
3187
else {
3188
ut_memcpy(buf,
3189
prebuilt->fetch_cache[prebuilt->fetch_cache_first],
3190
prebuilt->mysql_prefix_len);
3191
}
3192
prebuilt->n_fetch_cached--;
3193
prebuilt->fetch_cache_first++;
3194
3195
if (prebuilt->n_fetch_cached == 0) {
3196
prebuilt->fetch_cache_first = 0;
3197
}
3198
}
3199
3200
/************************************************************************
3201
Pushes a row for MySQL to the fetch cache. */
3202
UNIV_INLINE
3203
void
3204
row_sel_push_cache_row_for_mysql(
3205
/*=============================*/
3206
row_prebuilt_t* prebuilt, /* in: prebuilt struct */
3207
const rec_t* rec, /* in: record to push; must
3208
be protected by a page latch */
3209
const ulint* offsets, /* in: rec_get_offsets() */
3210
ulint start_field_no, /* psergy: start from this field */
3211
byte* remainder_buf) /* if above !=0 -> where to take
3212
prev fields */
3213
{
3214
byte* buf;
3215
ulint i;
3216
3217
ut_ad(prebuilt->n_fetch_cached < MYSQL_FETCH_CACHE_SIZE);
3218
ut_ad(rec_offs_validate(rec, NULL, offsets));
3219
ut_a(!prebuilt->templ_contains_blob);
3220
3221
if (prebuilt->fetch_cache[0] == NULL) {
3222
/* Allocate memory for the fetch cache */
3223
3224
for (i = 0; i < MYSQL_FETCH_CACHE_SIZE; i++) {
3225
3226
/* A user has reported memory corruption in these
3227
buffers in Linux. Put magic numbers there to help
3228
to track a possible bug. */
3229
3230
buf = mem_alloc(prebuilt->mysql_row_len + 8);
3231
3232
prebuilt->fetch_cache[i] = buf + 4;
3233
3234
mach_write_to_4(buf, ROW_PREBUILT_FETCH_MAGIC_N);
3235
mach_write_to_4(buf + 4 + prebuilt->mysql_row_len,
3236
ROW_PREBUILT_FETCH_MAGIC_N);
3237
}
3238
}
3239
3240
ut_ad(prebuilt->fetch_cache_first == 0);
3241
3242
if (UNIV_UNLIKELY(!row_sel_store_mysql_rec(
3243
prebuilt->fetch_cache[
3244
prebuilt->n_fetch_cached],
3245
prebuilt, rec, offsets, start_field_no,
3246
prebuilt->n_template))) {
3247
ut_error;
3248
}
3249
3250
if (start_field_no) {
3251
for (i=0; i < start_field_no; i++) {
3252
register ulint offs;
3253
mysql_row_templ_t* templ;
3254
templ = prebuilt->mysql_template + i;
3255
3256
if (templ->mysql_null_bit_mask) {
3257
offs= templ->mysql_null_byte_offset;
3258
*(prebuilt->fetch_cache[prebuilt->n_fetch_cached] + offs) ^=
3259
(*(remainder_buf + offs) & templ->mysql_null_bit_mask);
3260
}
3261
offs= templ->mysql_col_offset;
3262
memcpy(prebuilt->fetch_cache[prebuilt->n_fetch_cached] + offs,
3263
remainder_buf + offs,
3264
templ->mysql_col_len);
3265
}
3266
}
3267
3268
3269
prebuilt->n_fetch_cached++;
3270
}
3271
3272
/*************************************************************************
3273
Tries to do a shortcut to fetch a clustered index record with a unique key,
3274
using the hash index if possible (not always). We assume that the search
3275
mode is PAGE_CUR_GE, it is a consistent read, there is a read view in trx,
3276
btr search latch has been locked in S-mode. */
3277
static
3278
ulint
3279
row_sel_try_search_shortcut_for_mysql(
3280
/*==================================*/
3281
/* out: SEL_FOUND, SEL_EXHAUSTED, SEL_RETRY */
3282
const rec_t** out_rec,/* out: record if found */
3283
row_prebuilt_t* prebuilt,/* in: prebuilt struct */
3284
ulint** offsets,/* in/out: for rec_get_offsets(*out_rec) */
3285
mem_heap_t** heap, /* in/out: heap for rec_get_offsets() */
3286
mtr_t* mtr) /* in: started mtr */
3287
{
3288
dict_index_t* index = prebuilt->index;
3289
const dtuple_t* search_tuple = prebuilt->search_tuple;
3290
btr_pcur_t* pcur = prebuilt->pcur;
3291
trx_t* trx = prebuilt->trx;
3292
const rec_t* rec;
3293
3294
ut_ad(dict_index_is_clust(index));
3295
ut_ad(!prebuilt->templ_contains_blob);
3296
3297
btr_pcur_open_with_no_init(index, search_tuple, PAGE_CUR_GE,
3298
BTR_SEARCH_LEAF, pcur,
3299
#ifndef UNIV_SEARCH_DEBUG
3300
RW_S_LATCH,
3301
#else
3302
0,
3303
#endif
3304
mtr);
3305
rec = btr_pcur_get_rec(pcur);
3306
3307
if (!page_rec_is_user_rec(rec)) {
3308
3309
return(SEL_RETRY);
3310
}
3311
3312
/* As the cursor is now placed on a user record after a search with
3313
the mode PAGE_CUR_GE, the up_match field in the cursor tells how many
3314
fields in the user record matched to the search tuple */
3315
3316
if (btr_pcur_get_up_match(pcur) < dtuple_get_n_fields(search_tuple)) {
3317
3318
return(SEL_EXHAUSTED);
3319
}
3320
3321
/* This is a non-locking consistent read: if necessary, fetch
3322
a previous version of the record */
3323
3324
*offsets = rec_get_offsets(rec, index, *offsets,
3325
ULINT_UNDEFINED, heap);
3326
3327
if (!lock_clust_rec_cons_read_sees(rec, index,
3328
*offsets, trx->read_view)) {
3329
3330
return(SEL_RETRY);
3331
}
3332
3333
if (rec_get_deleted_flag(rec, dict_table_is_comp(index->table))) {
3334
3335
return(SEL_EXHAUSTED);
3336
}
3337
3338
*out_rec = rec;
3339
3340
return(SEL_FOUND);
3341
}
3342
3343
/************************************************************************
3344
Searches for rows in the database. This is used in the interface to
3345
MySQL. This function opens a cursor, and also implements fetch next
3346
and fetch prev. NOTE that if we do a search with a full key value
3347
from a unique index (ROW_SEL_EXACT), then we will not store the cursor
3348
position and fetch next or fetch prev must not be tried to the cursor! */
3349
UNIV_INTERN
3350
ulint
3351
row_search_for_mysql(
3352
/*=================*/
3353
/* out: DB_SUCCESS,
3354
DB_RECORD_NOT_FOUND,
3355
DB_END_OF_INDEX, DB_DEADLOCK,
3356
DB_LOCK_TABLE_FULL, DB_CORRUPTION,
3357
or DB_TOO_BIG_RECORD */
3358
byte* buf, /* in/out: buffer for the fetched
3359
row in the MySQL format */
3360
ulint mode, /* in: search mode PAGE_CUR_L, ... */
3361
row_prebuilt_t* prebuilt, /* in: prebuilt struct for the
3362
table handle; this contains the info
3363
of search_tuple, index; if search
3364
tuple contains 0 fields then we
3365
position the cursor at the start or
3366
the end of the index, depending on
3367
'mode' */
3368
ulint match_mode, /* in: 0 or ROW_SEL_EXACT or
3369
ROW_SEL_EXACT_PREFIX */
3370
ulint direction) /* in: 0 or ROW_SEL_NEXT or
3371
ROW_SEL_PREV; NOTE: if this is != 0,
3372
then prebuilt must have a pcur
3373
with stored position! In opening of a
3374
cursor 'direction' should be 0. */
3375
{
3376
dict_index_t* index = prebuilt->index;
3377
ibool comp = dict_table_is_comp(index->table);
3378
const dtuple_t* search_tuple = prebuilt->search_tuple;
3379
btr_pcur_t* pcur = prebuilt->pcur;
3380
trx_t* trx = prebuilt->trx;
3381
dict_index_t* clust_index;
3382
que_thr_t* thr;
3383
const rec_t* rec;
3384
const rec_t* result_rec;
3385
const rec_t* clust_rec;
3386
ulint err = DB_SUCCESS;
3387
ibool unique_search = FALSE;
3388
ibool unique_search_from_clust_index = FALSE;
3389
ibool mtr_has_extra_clust_latch = FALSE;
3390
ibool moves_up = FALSE;
3391
ibool set_also_gap_locks = TRUE;
3392
/* if the query is a plain locking SELECT, and the isolation level
3393
is <= TRX_ISO_READ_COMMITTED, then this is set to FALSE */
3394
ibool did_semi_consistent_read = FALSE;
3395
/* if the returned record was locked and we did a semi-consistent
3396
read (fetch the newest committed version), then this is set to
3397
TRUE */
3398
#ifdef UNIV_SEARCH_DEBUG
3399
ulint cnt = 0;
3400
#endif /* UNIV_SEARCH_DEBUG */
3401
ulint next_offs;
3402
ibool same_user_rec;
3403
mtr_t mtr;
3404
mem_heap_t* heap = NULL;
3405
ulint offsets_[REC_OFFS_NORMAL_SIZE];
3406
ulint* offsets = offsets_;
3407
ibool some_fields_in_buffer;
3408
ibool get_clust_rec= 0;
3409
3410
rec_offs_init(offsets_);
3411
3412
ut_ad(index && pcur && search_tuple);
3413
ut_ad(trx->mysql_thread_id == os_thread_get_curr_id());
3414
3415
if (UNIV_UNLIKELY(prebuilt->table->ibd_file_missing)) {
3416
ut_print_timestamp(stderr);
3417
fprintf(stderr, " InnoDB: Error:\n"
3418
"InnoDB: MySQL is trying to use a table handle"
3419
" but the .ibd file for\n"
3420
"InnoDB: table %s does not exist.\n"
3421
"InnoDB: Have you deleted the .ibd file"
3422
" from the database directory under\n"
3423
"InnoDB: the MySQL datadir, or have you used"
3424
" DISCARD TABLESPACE?\n"
3425
"InnoDB: Look from\n"
3426
"InnoDB: http://dev.mysql.com/doc/refman/5.1/en/"
3427
"innodb-troubleshooting.html\n"
3428
"InnoDB: how you can resolve the problem.\n",
3429
prebuilt->table->name);
3430
3431
return(DB_ERROR);
3432
}
3433
3434
if (UNIV_UNLIKELY(prebuilt->magic_n != ROW_PREBUILT_ALLOCATED)) {
3435
fprintf(stderr,
3436
"InnoDB: Error: trying to free a corrupt\n"
3437
"InnoDB: table handle. Magic n %lu, table name ",
3438
(ulong) prebuilt->magic_n);
3439
ut_print_name(stderr, trx, TRUE, prebuilt->table->name);
3440
putc('\n', stderr);
3441
3442
mem_analyze_corruption(prebuilt);
3443
3444
ut_error;
3445
}
3446
3447
#if 0
3448
/* August 19, 2005 by Heikki: temporarily disable this error
3449
print until the cursor lock count is done correctly.
3450
See bugs #12263 and #12456!*/
3451
3452
if (trx->n_mysql_tables_in_use == 0
3453
&& UNIV_UNLIKELY(prebuilt->select_lock_type == LOCK_NONE)) {
3454
/* Note that if MySQL uses an InnoDB temp table that it
3455
created inside LOCK TABLES, then n_mysql_tables_in_use can
3456
be zero; in that case select_lock_type is set to LOCK_X in
3457
::start_stmt. */
3458
3459
fputs("InnoDB: Error: MySQL is trying to perform a SELECT\n"
3460
"InnoDB: but it has not locked"
3461
" any tables in ::external_lock()!\n",
3462
stderr);
3463
trx_print(stderr, trx, 600);
3464
fputc('\n', stderr);
3465
}
3466
#endif
3467
3468
#if 0
3469
fprintf(stderr, "Match mode %lu\n search tuple ",
3470
(ulong) match_mode);
3471
dtuple_print(search_tuple);
3472
fprintf(stderr, "N tables locked %lu\n",
3473
(ulong) trx->mysql_n_tables_locked);
3474
#endif
3475
/*-------------------------------------------------------------*/
3476
/* PHASE 0: Release a possible s-latch we are holding on the
3477
adaptive hash index latch if there is someone waiting behind */
3478
3479
if (UNIV_UNLIKELY(btr_search_latch.writer != RW_LOCK_NOT_LOCKED)
3480
&& trx->has_search_latch) {
3481
3482
/* There is an x-latch request on the adaptive hash index:
3483
release the s-latch to reduce starvation and wait for
3484
BTR_SEA_TIMEOUT rounds before trying to keep it again over
3485
calls from MySQL */
3486
3487
rw_lock_s_unlock(&btr_search_latch);
3488
trx->has_search_latch = FALSE;
3489
3490
trx->search_latch_timeout = BTR_SEA_TIMEOUT;
3491
}
3492
3493
/* Reset the new record lock info if srv_locks_unsafe_for_binlog
3494
is set or session is using a READ COMMITED isolation level. Then
3495
we are able to remove the record locks set here on an individual
3496
row. */
3497
3498
if ((srv_locks_unsafe_for_binlog
3499
|| trx->isolation_level == TRX_ISO_READ_COMMITTED)
3500
&& prebuilt->select_lock_type != LOCK_NONE) {
3501
3502
trx_reset_new_rec_lock_info(trx);
3503
}
3504
3505
/*-------------------------------------------------------------*/
3506
/* PHASE 1: Try to pop the row from the prefetch cache */
3507
3508
if (UNIV_UNLIKELY(direction == 0)) {
3509
trx->op_info = "starting index read";
3510
3511
prebuilt->n_rows_fetched = 0;
3512
prebuilt->n_fetch_cached = 0;
3513
prebuilt->fetch_cache_first = 0;
3514
3515
if (prebuilt->sel_graph == NULL) {
3516
/* Build a dummy select query graph */
3517
row_prebuild_sel_graph(prebuilt);
3518
}
3519
} else {
3520
trx->op_info = "fetching rows";
3521
3522
if (prebuilt->n_rows_fetched == 0) {
3523
prebuilt->fetch_direction = direction;
3524
}
3525
3526
if (UNIV_UNLIKELY(direction != prebuilt->fetch_direction)) {
3527
if (UNIV_UNLIKELY(prebuilt->n_fetch_cached > 0)) {
3528
ut_error;
3529
/* TODO: scrollable cursor: restore cursor to
3530
the place of the latest returned row,
3531
or better: prevent caching for a scroll
3532
cursor! */
3533
}
3534
3535
prebuilt->n_rows_fetched = 0;
3536
prebuilt->n_fetch_cached = 0;
3537
prebuilt->fetch_cache_first = 0;
3538
3539
} else if (UNIV_LIKELY(prebuilt->n_fetch_cached > 0)) {
3540
row_sel_pop_cached_row_for_mysql(buf, prebuilt);
3541
3542
prebuilt->n_rows_fetched++;
3543
3544
srv_n_rows_read++;
3545
err = DB_SUCCESS;
3546
goto func_exit;
3547
}
3548
3549
if (prebuilt->fetch_cache_first > 0
3550
&& prebuilt->fetch_cache_first < MYSQL_FETCH_CACHE_SIZE) {
3551
3552
/* The previous returned row was popped from the fetch
3553
cache, but the cache was not full at the time of the
3554
popping: no more rows can exist in the result set */
3555
3556
err = DB_RECORD_NOT_FOUND;
3557
goto func_exit;
3558
}
3559
3560
prebuilt->n_rows_fetched++;
3561
3562
if (prebuilt->n_rows_fetched > 1000000000) {
3563
/* Prevent wrap-over */
3564
prebuilt->n_rows_fetched = 500000000;
3565
}
3566
3567
mode = pcur->search_mode;
3568
}
3569
3570
/* In a search where at most one record in the index may match, we
3571
can use a LOCK_REC_NOT_GAP type record lock when locking a
3572
non-delete-marked matching record.
3573
3574
Note that in a unique secondary index there may be different
3575
delete-marked versions of a record where only the primary key
3576
values differ: thus in a secondary index we must use next-key
3577
locks when locking delete-marked records. */
3578
3579
if (match_mode == ROW_SEL_EXACT
3580
&& dict_index_is_unique(index)
3581
&& dtuple_get_n_fields(search_tuple)
3582
== dict_index_get_n_unique(index)
3583
&& (dict_index_is_clust(index)
3584
|| !dtuple_contains_null(search_tuple))) {
3585
3586
/* Note above that a UNIQUE secondary index can contain many
3587
rows with the same key value if one of the columns is the SQL
3588
null. A clustered index under MySQL can never contain null
3589
columns because we demand that all the columns in primary key
3590
are non-null. */
3591
3592
unique_search = TRUE;
3593
3594
/* Even if the condition is unique, MySQL seems to try to
3595
retrieve also a second row if a primary key contains more than
3596
1 column. Return immediately if this is not a HANDLER
3597
command. */
3598
3599
if (UNIV_UNLIKELY(direction != 0
3600
&& !prebuilt->used_in_HANDLER)) {
3601
3602
err = DB_RECORD_NOT_FOUND;
3603
goto func_exit;
3604
}
3605
}
3606
3607
mtr_start(&mtr);
3608
3609
/*-------------------------------------------------------------*/
3610
/* PHASE 2: Try fast adaptive hash index search if possible */
3611
3612
/* Next test if this is the special case where we can use the fast
3613
adaptive hash index to try the search. Since we must release the
3614
search system latch when we retrieve an externally stored field, we
3615
cannot use the adaptive hash index in a search in the case the row
3616
may be long and there may be externally stored fields */
3617
3618
if (UNIV_UNLIKELY(direction == 0)
3619
&& unique_search
3620
&& dict_index_is_clust(index)
3621
&& !prebuilt->templ_contains_blob
3622
&& !prebuilt->used_in_HANDLER
3623
&& (prebuilt->mysql_row_len < UNIV_PAGE_SIZE / 8)) {
3624
3625
mode = PAGE_CUR_GE;
3626
3627
unique_search_from_clust_index = TRUE;
3628
3629
if (trx->mysql_n_tables_locked == 0
3630
&& prebuilt->select_lock_type == LOCK_NONE
3631
&& trx->isolation_level > TRX_ISO_READ_UNCOMMITTED
3632
&& trx->read_view) {
3633
3634
/* This is a SELECT query done as a consistent read,
3635
and the read view has already been allocated:
3636
let us try a search shortcut through the hash
3637
index.
3638
NOTE that we must also test that
3639
mysql_n_tables_locked == 0, because this might
3640
also be INSERT INTO ... SELECT ... or
3641
CREATE TABLE ... SELECT ... . Our algorithm is
3642
NOT prepared to inserts interleaved with the SELECT,
3643
and if we try that, we can deadlock on the adaptive
3644
hash index semaphore! */
3645
3646
#ifndef UNIV_SEARCH_DEBUG
3647
if (!trx->has_search_latch) {
3648
rw_lock_s_lock(&btr_search_latch);
3649
trx->has_search_latch = TRUE;
3650
}
3651
#endif
3652
switch (row_sel_try_search_shortcut_for_mysql(
3653
&rec, prebuilt, &offsets, &heap,
3654
&mtr)) {
3655
case SEL_FOUND:
3656
#ifdef UNIV_SEARCH_DEBUG
3657
ut_a(0 == cmp_dtuple_rec(search_tuple,
3658
rec, offsets));
3659
#endif
3660
/* At this point, rec is protected by
3661
a page latch that was acquired by
3662
row_sel_try_search_shortcut_for_mysql().
3663
The latch will not be released until
3664
mtr_commit(&mtr). */
3665
3666
if (!row_sel_store_mysql_rec(buf, prebuilt,
3667
rec, offsets, 0,
3668
prebuilt->n_template)) {
3669
err = DB_TOO_BIG_RECORD;
3670
3671
/* We let the main loop to do the
3672
error handling */
3673
goto shortcut_fails_too_big_rec;
3674
}
3675
3676
mtr_commit(&mtr);
3677
3678
/* ut_print_name(stderr, index->name);
3679
fputs(" shortcut\n", stderr); */
3680
3681
srv_n_rows_read++;
3682
3683
err = DB_SUCCESS;
3684
goto release_search_latch_if_needed;
3685
3686
case SEL_EXHAUSTED:
3687
mtr_commit(&mtr);
3688
3689
/* ut_print_name(stderr, index->name);
3690
fputs(" record not found 2\n", stderr); */
3691
3692
err = DB_RECORD_NOT_FOUND;
3693
release_search_latch_if_needed:
3694
if (trx->search_latch_timeout > 0
3695
&& trx->has_search_latch) {
3696
3697
trx->search_latch_timeout--;
3698
3699
rw_lock_s_unlock(&btr_search_latch);
3700
trx->has_search_latch = FALSE;
3701
}
3702
3703
/* NOTE that we do NOT store the cursor
3704
position */
3705
goto func_exit;
3706
3707
case SEL_RETRY:
3708
break;
3709
3710
default:
3711
ut_ad(0);
3712
}
3713
shortcut_fails_too_big_rec:
3714
mtr_commit(&mtr);
3715
mtr_start(&mtr);
3716
}
3717
}
3718
3719
/*-------------------------------------------------------------*/
3720
/* PHASE 3: Open or restore index cursor position */
3721
3722
if (trx->has_search_latch) {
3723
rw_lock_s_unlock(&btr_search_latch);
3724
trx->has_search_latch = FALSE;
3725
}
3726
3727
trx_start_if_not_started(trx);
3728
3729
if (trx->isolation_level <= TRX_ISO_READ_COMMITTED
3730
&& prebuilt->select_lock_type != LOCK_NONE
3731
&& trx->mysql_thd != NULL
3732
&& trx->mysql_query_str != NULL
3733
&& *trx->mysql_query_str != NULL) {
3734
3735
/* Scan the MySQL query string; check if SELECT is the first
3736
word there */
3737
3738
if (dict_str_starts_with_keyword(
3739
trx->mysql_thd, *trx->mysql_query_str, "SELECT")) {
3740
/* It is a plain locking SELECT and the isolation
3741
level is low: do not lock gaps */
3742
3743
set_also_gap_locks = FALSE;
3744
}
3745
}
3746
3747
/* Note that if the search mode was GE or G, then the cursor
3748
naturally moves upward (in fetch next) in alphabetical order,
3749
otherwise downward */
3750
3751
if (UNIV_UNLIKELY(direction == 0)) {
3752
if (mode == PAGE_CUR_GE || mode == PAGE_CUR_G) {
3753
moves_up = TRUE;
3754
}
3755
} else if (direction == ROW_SEL_NEXT) {
3756
moves_up = TRUE;
3757
}
3758
3759
thr = que_fork_get_first_thr(prebuilt->sel_graph);
3760
3761
que_thr_move_to_run_state_for_mysql(thr, trx);
3762
3763
clust_index = dict_table_get_first_index(index->table);
3764
3765
if (UNIV_LIKELY(direction != 0)) {
3766
ibool need_to_process = sel_restore_position_for_mysql(
3767
&same_user_rec, BTR_SEARCH_LEAF,
3768
pcur, moves_up, &mtr);
3769
3770
if (UNIV_UNLIKELY(need_to_process)) {
3771
if (UNIV_UNLIKELY(prebuilt->row_read_type
3772
== ROW_READ_DID_SEMI_CONSISTENT)) {
3773
/* We did a semi-consistent read,
3774
but the record was removed in
3775
the meantime. */
3776
prebuilt->row_read_type
3777
= ROW_READ_TRY_SEMI_CONSISTENT;
3778
}
3779
} else if (UNIV_LIKELY(prebuilt->row_read_type
3780
!= ROW_READ_DID_SEMI_CONSISTENT)) {
3781
3782
/* The cursor was positioned on the record
3783
that we returned previously. If we need
3784
to repeat a semi-consistent read as a
3785
pessimistic locking read, the record
3786
cannot be skipped. */
3787
3788
goto next_rec;
3789
}
3790
3791
} else if (dtuple_get_n_fields(search_tuple) > 0) {
3792
3793
btr_pcur_open_with_no_init(index, search_tuple, mode,
3794
BTR_SEARCH_LEAF,
3795
pcur, 0, &mtr);
3796
3797
pcur->trx_if_known = trx;
3798
3799
rec = btr_pcur_get_rec(pcur);
3800
3801
if (!moves_up
3802
&& !page_rec_is_supremum(rec)
3803
&& set_also_gap_locks
3804
&& !(srv_locks_unsafe_for_binlog
3805
|| trx->isolation_level == TRX_ISO_READ_COMMITTED)
3806
&& prebuilt->select_lock_type != LOCK_NONE) {
3807
3808
/* Try to place a gap lock on the next index record
3809
to prevent phantoms in ORDER BY ... DESC queries */
3810
const rec_t* next = page_rec_get_next_const(rec);
3811
3812
offsets = rec_get_offsets(next, index, offsets,
3813
ULINT_UNDEFINED, &heap);
3814
err = sel_set_rec_lock(btr_pcur_get_block(pcur),
3815
next, index, offsets,
3816
prebuilt->select_lock_type,
3817
LOCK_GAP, thr);
3818
3819
if (err != DB_SUCCESS) {
3820
3821
goto lock_wait_or_error;
3822
}
3823
}
3824
} else {
3825
if (mode == PAGE_CUR_G) {
3826
btr_pcur_open_at_index_side(
3827
TRUE, index, BTR_SEARCH_LEAF, pcur, FALSE,
3828
&mtr);
3829
} else if (mode == PAGE_CUR_L) {
3830
btr_pcur_open_at_index_side(
3831
FALSE, index, BTR_SEARCH_LEAF, pcur, FALSE,
3832
&mtr);
3833
}
3834
}
3835
3836
if (!prebuilt->sql_stat_start) {
3837
/* No need to set an intention lock or assign a read view */
3838
3839
if (trx->read_view == NULL
3840
&& prebuilt->select_lock_type == LOCK_NONE) {
3841
3842
fputs("InnoDB: Error: MySQL is trying to"
3843
" perform a consistent read\n"
3844
"InnoDB: but the read view is not assigned!\n",
3845
stderr);
3846
trx_print(stderr, trx, 600);
3847
fputc('\n', stderr);
3848
ut_a(0);
3849
}
3850
} else if (prebuilt->select_lock_type == LOCK_NONE) {
3851
/* This is a consistent read */
3852
/* Assign a read view for the query */
3853
3854
trx_assign_read_view(trx);
3855
prebuilt->sql_stat_start = FALSE;
3856
} else {
3857
ulint lock_mode;
3858
if (prebuilt->select_lock_type == LOCK_S) {
3859
lock_mode = LOCK_IS;
3860
} else {
3861
lock_mode = LOCK_IX;
3862
}
3863
err = lock_table(0, index->table, lock_mode, thr);
3864
3865
if (err != DB_SUCCESS) {
3866
3867
goto lock_wait_or_error;
3868
}
3869
prebuilt->sql_stat_start = FALSE;
3870
}
3871
3872
rec_loop:
3873
/*-------------------------------------------------------------*/
3874
/* PHASE 4: Look for matching records in a loop */
3875
3876
rec = btr_pcur_get_rec(pcur);
3877
ut_ad(!!page_rec_is_comp(rec) == comp);
3878
#ifdef UNIV_SEARCH_DEBUG
3879
/*
3880
fputs("Using ", stderr);
3881
dict_index_name_print(stderr, index);
3882
fprintf(stderr, " cnt %lu ; Page no %lu\n", cnt,
3883
page_get_page_no(page_align(rec)));
3884
rec_print(rec);
3885
*/
3886
#endif /* UNIV_SEARCH_DEBUG */
3887
3888
if (page_rec_is_infimum(rec)) {
3889
3890
/* The infimum record on a page cannot be in the result set,
3891
and neither can a record lock be placed on it: we skip such
3892
a record. */
3893
3894
goto next_rec;
3895
}
3896
3897
if (page_rec_is_supremum(rec)) {
3898
3899
if (set_also_gap_locks
3900
&& !(srv_locks_unsafe_for_binlog
3901
|| trx->isolation_level == TRX_ISO_READ_COMMITTED)
3902
&& prebuilt->select_lock_type != LOCK_NONE) {
3903
3904
/* Try to place a lock on the index record */
3905
3906
/* If innodb_locks_unsafe_for_binlog option is used
3907
or this session is using a READ COMMITTED isolation
3908
level we do not lock gaps. Supremum record is really
3909
a gap and therefore we do not set locks there. */
3910
3911
offsets = rec_get_offsets(rec, index, offsets,
3912
ULINT_UNDEFINED, &heap);
3913
err = sel_set_rec_lock(btr_pcur_get_block(pcur),
3914
rec, index, offsets,
3915
prebuilt->select_lock_type,
3916
LOCK_ORDINARY, thr);
3917
3918
if (err != DB_SUCCESS) {
3919
3920
goto lock_wait_or_error;
3921
}
3922
}
3923
/* A page supremum record cannot be in the result set: skip
3924
it now that we have placed a possible lock on it */
3925
3926
goto next_rec;
3927
}
3928
3929
/*-------------------------------------------------------------*/
3930
/* Do sanity checks in case our cursor has bumped into page
3931
corruption */
3932
3933
if (comp) {
3934
next_offs = rec_get_next_offs(rec, TRUE);
3935
if (UNIV_UNLIKELY(next_offs < PAGE_NEW_SUPREMUM)) {
3936
3937
goto wrong_offs;
3938
}
3939
} else {
3940
next_offs = rec_get_next_offs(rec, FALSE);
3941
if (UNIV_UNLIKELY(next_offs < PAGE_OLD_SUPREMUM)) {
3942
3943
goto wrong_offs;
3944
}
3945
}
3946
3947
if (UNIV_UNLIKELY(next_offs >= UNIV_PAGE_SIZE - PAGE_DIR)) {
3948
3949
wrong_offs:
3950
if (srv_force_recovery == 0 || moves_up == FALSE) {
3951
ut_print_timestamp(stderr);
3952
buf_page_print(page_align(rec), 0);
3953
fprintf(stderr,
3954
"\nInnoDB: rec address %p,"
3955
" buf block fix count %lu\n",
3956
(void*) rec, (ulong)
3957
btr_cur_get_block(btr_pcur_get_btr_cur(pcur))
3958
->page.buf_fix_count);
3959
fprintf(stderr,
3960
"InnoDB: Index corruption: rec offs %lu"
3961
" next offs %lu, page no %lu,\n"
3962
"InnoDB: ",
3963
(ulong) page_offset(rec),
3964
(ulong) next_offs,
3965
(ulong) page_get_page_no(page_align(rec)));
3966
dict_index_name_print(stderr, trx, index);
3967
fputs(". Run CHECK TABLE. You may need to\n"
3968
"InnoDB: restore from a backup, or"
3969
" dump + drop + reimport the table.\n",
3970
stderr);
3971
3972
err = DB_CORRUPTION;
3973
3974
goto lock_wait_or_error;
3975
} else {
3976
/* The user may be dumping a corrupt table. Jump
3977
over the corruption to recover as much as possible. */
3978
3979
fprintf(stderr,
3980
"InnoDB: Index corruption: rec offs %lu"
3981
" next offs %lu, page no %lu,\n"
3982
"InnoDB: ",
3983
(ulong) page_offset(rec),
3984
(ulong) next_offs,
3985
(ulong) page_get_page_no(page_align(rec)));
3986
dict_index_name_print(stderr, trx, index);
3987
fputs(". We try to skip the rest of the page.\n",
3988
stderr);
3989
3990
btr_pcur_move_to_last_on_page(pcur, &mtr);
3991
3992
goto next_rec;
3993
}
3994
}
3995
/*-------------------------------------------------------------*/
3996
3997
/* Calculate the 'offsets' associated with 'rec' */
3998
3999
offsets = rec_get_offsets(rec, index, offsets, ULINT_UNDEFINED, &heap);
4000
4001
if (UNIV_UNLIKELY(srv_force_recovery > 0)) {
4002
if (!rec_validate(rec, offsets)
4003
|| !btr_index_rec_validate(rec, index, FALSE)) {
4004
fprintf(stderr,
4005
"InnoDB: Index corruption: rec offs %lu"
4006
" next offs %lu, page no %lu,\n"
4007
"InnoDB: ",
4008
(ulong) page_offset(rec),
4009
(ulong) next_offs,
4010
(ulong) page_get_page_no(page_align(rec)));
4011
dict_index_name_print(stderr, trx, index);
4012
fputs(". We try to skip the record.\n",
4013
stderr);
4014
4015
goto next_rec;
4016
}
4017
}
4018
4019
/* Note that we cannot trust the up_match value in the cursor at this
4020
place because we can arrive here after moving the cursor! Thus
4021
we have to recompare rec and search_tuple to determine if they
4022
match enough. */
4023
4024
if (match_mode == ROW_SEL_EXACT) {
4025
/* Test if the index record matches completely to search_tuple
4026
in prebuilt: if not, then we return with DB_RECORD_NOT_FOUND */
4027
4028
/* fputs("Comparing rec and search tuple\n", stderr); */
4029
4030
if (0 != cmp_dtuple_rec(search_tuple, rec, offsets)) {
4031
4032
if (set_also_gap_locks
4033
&& !(srv_locks_unsafe_for_binlog
4034
|| trx->isolation_level
4035
== TRX_ISO_READ_COMMITTED)
4036
&& prebuilt->select_lock_type != LOCK_NONE) {
4037
4038
/* Try to place a gap lock on the index
4039
record only if innodb_locks_unsafe_for_binlog
4040
option is not set or this session is not
4041
using a READ COMMITTED isolation level. */
4042
4043
err = sel_set_rec_lock(
4044
btr_pcur_get_block(pcur),
4045
rec, index, offsets,
4046
prebuilt->select_lock_type, LOCK_GAP,
4047
thr);
4048
4049
if (err != DB_SUCCESS) {
4050
4051
goto lock_wait_or_error;
4052
}
4053
}
4054
4055
btr_pcur_store_position(pcur, &mtr);
4056
4057
err = DB_RECORD_NOT_FOUND;
4058
/* ut_print_name(stderr, index->name);
4059
fputs(" record not found 3\n", stderr); */
4060
4061
goto normal_return;
4062
}
4063
4064
} else if (match_mode == ROW_SEL_EXACT_PREFIX) {
4065
4066
if (!cmp_dtuple_is_prefix_of_rec(search_tuple, rec, offsets)) {
4067
4068
if (set_also_gap_locks
4069
&& !(srv_locks_unsafe_for_binlog
4070
|| trx->isolation_level
4071
== TRX_ISO_READ_COMMITTED)
4072
&& prebuilt->select_lock_type != LOCK_NONE) {
4073
4074
/* Try to place a gap lock on the index
4075
record only if innodb_locks_unsafe_for_binlog
4076
option is not set or this session is not
4077
using a READ COMMITTED isolation level. */
4078
4079
err = sel_set_rec_lock(
4080
btr_pcur_get_block(pcur),
4081
rec, index, offsets,
4082
prebuilt->select_lock_type, LOCK_GAP,
4083
thr);
4084
4085
if (err != DB_SUCCESS) {
4086
4087
goto lock_wait_or_error;
4088
}
4089
}
4090
4091
btr_pcur_store_position(pcur, &mtr);
4092
4093
err = DB_RECORD_NOT_FOUND;
4094
/* ut_print_name(stderr, index->name);
4095
fputs(" record not found 4\n", stderr); */
4096
4097
goto normal_return;
4098
}
4099
}
4100
4101
/* We are ready to look at a possible new index entry in the result
4102
set: the cursor is now placed on a user record */
4103
4104
if (prebuilt->select_lock_type != LOCK_NONE) {
4105
/* Try to place a lock on the index record; note that delete
4106
marked records are a special case in a unique search. If there
4107
is a non-delete marked record, then it is enough to lock its
4108
existence with LOCK_REC_NOT_GAP. */
4109
4110
/* If innodb_locks_unsafe_for_binlog option is used
4111
or this session is using a READ COMMITED isolation
4112
level we lock only the record, i.e., next-key locking is
4113
not used. */
4114
4115
ulint lock_type;
4116
4117
if (!set_also_gap_locks
4118
|| srv_locks_unsafe_for_binlog
4119
|| trx->isolation_level == TRX_ISO_READ_COMMITTED
4120
|| (unique_search
4121
&& !UNIV_UNLIKELY(rec_get_deleted_flag(rec, comp)))) {
4122
4123
goto no_gap_lock;
4124
} else {
4125
lock_type = LOCK_ORDINARY;
4126
}
4127
4128
/* If we are doing a 'greater or equal than a primary key
4129
value' search from a clustered index, and we find a record
4130
that has that exact primary key value, then there is no need
4131
to lock the gap before the record, because no insert in the
4132
gap can be in our search range. That is, no phantom row can
4133
appear that way.
4134
4135
An example: if col1 is the primary key, the search is WHERE
4136
col1 >= 100, and we find a record where col1 = 100, then no
4137
need to lock the gap before that record. */
4138
4139
if (index == clust_index
4140
&& mode == PAGE_CUR_GE
4141
&& direction == 0
4142
&& dtuple_get_n_fields_cmp(search_tuple)
4143
== dict_index_get_n_unique(index)
4144
&& 0 == cmp_dtuple_rec(search_tuple, rec, offsets)) {
4145
no_gap_lock:
4146
lock_type = LOCK_REC_NOT_GAP;
4147
}
4148
4149
err = sel_set_rec_lock(btr_pcur_get_block(pcur),
4150
rec, index, offsets,
4151
prebuilt->select_lock_type,
4152
lock_type, thr);
4153
4154
switch (err) {
4155
const rec_t* old_vers;
4156
case DB_SUCCESS:
4157
break;
4158
case DB_LOCK_WAIT:
4159
if (UNIV_LIKELY(prebuilt->row_read_type
4160
!= ROW_READ_TRY_SEMI_CONSISTENT)
4161
|| index != clust_index) {
4162
4163
goto lock_wait_or_error;
4164
}
4165
4166
/* The following call returns 'offsets'
4167
associated with 'old_vers' */
4168
err = row_sel_build_committed_vers_for_mysql(
4169
clust_index, prebuilt, rec,
4170
&offsets, &heap, &old_vers, &mtr);
4171
4172
if (err != DB_SUCCESS) {
4173
4174
goto lock_wait_or_error;
4175
}
4176
4177
mutex_enter(&kernel_mutex);
4178
if (trx->was_chosen_as_deadlock_victim) {
4179
mutex_exit(&kernel_mutex);
4180
err = DB_DEADLOCK;
4181
4182
goto lock_wait_or_error;
4183
}
4184
if (UNIV_LIKELY(trx->wait_lock != NULL)) {
4185
lock_cancel_waiting_and_release(
4186
trx->wait_lock);
4187
trx_reset_new_rec_lock_info(trx);
4188
} else {
4189
mutex_exit(&kernel_mutex);
4190
4191
/* The lock was granted while we were
4192
searching for the last committed version.
4193
Do a normal locking read. */
4194
4195
offsets = rec_get_offsets(rec, index, offsets,
4196
ULINT_UNDEFINED,
4197
&heap);
4198
err = DB_SUCCESS;
4199
break;
4200
}
4201
mutex_exit(&kernel_mutex);
4202
4203
if (old_vers == NULL) {
4204
/* The row was not yet committed */
4205
4206
goto next_rec;
4207
}
4208
4209
did_semi_consistent_read = TRUE;
4210
rec = old_vers;
4211
break;
4212
default:
4213
4214
goto lock_wait_or_error;
4215
}
4216
} else {
4217
/* This is a non-locking consistent read: if necessary, fetch
4218
a previous version of the record */
4219
4220
if (trx->isolation_level == TRX_ISO_READ_UNCOMMITTED) {
4221
4222
/* Do nothing: we let a non-locking SELECT read the
4223
latest version of the record */
4224
4225
} else if (index == clust_index) {
4226
4227
/* Fetch a previous version of the row if the current
4228
one is not visible in the snapshot; if we have a very
4229
high force recovery level set, we try to avoid crashes
4230
by skipping this lookup */
4231
4232
if (UNIV_LIKELY(srv_force_recovery < 5)
4233
&& !lock_clust_rec_cons_read_sees(
4234
rec, index, offsets, trx->read_view)) {
4235
4236
rec_t* old_vers;
4237
/* The following call returns 'offsets'
4238
associated with 'old_vers' */
4239
err = row_sel_build_prev_vers_for_mysql(
4240
trx->read_view, clust_index,
4241
prebuilt, rec, &offsets, &heap,
4242
&old_vers, &mtr);
4243
4244
if (err != DB_SUCCESS) {
4245
4246
goto lock_wait_or_error;
4247
}
4248
4249
if (old_vers == NULL) {
4250
/* The row did not exist yet in
4251
the read view */
4252
4253
goto next_rec;
4254
}
4255
4256
rec = old_vers;
4257
}
4258
} else if (!lock_sec_rec_cons_read_sees(rec, trx->read_view)) {
4259
/* We are looking into a non-clustered index,
4260
and to get the right version of the record we
4261
have to look also into the clustered index: this
4262
is necessary, because we can only get the undo
4263
information via the clustered index record. */
4264
4265
ut_ad(index != clust_index);
4266
get_clust_rec= TRUE;
4267
goto idx_cond_check;
4268
}
4269
}
4270
4271
/* NOTE that at this point rec can be an old version of a clustered
4272
index record built for a consistent read. We cannot assume after this
4273
point that rec is on a buffer pool page. Functions like
4274
page_rec_is_comp() cannot be used! */
4275
4276
if (UNIV_UNLIKELY(rec_get_deleted_flag(rec, comp))) {
4277
4278
/* The record is delete-marked: we can skip it */
4279
4280
if ((srv_locks_unsafe_for_binlog
4281
|| trx->isolation_level == TRX_ISO_READ_COMMITTED)
4282
&& prebuilt->select_lock_type != LOCK_NONE
4283
&& !did_semi_consistent_read) {
4284
4285
/* No need to keep a lock on a delete-marked record
4286
if we do not want to use next-key locking. */
4287
4288
row_unlock_for_mysql(prebuilt, TRUE);
4289
}
4290
4291
/* This is an optimization to skip setting the next key lock
4292
on the record that follows this delete-marked record. This
4293
optimization works because of the unique search criteria
4294
which precludes the presence of a range lock between this
4295
delete marked record and the record following it.
4296
4297
For now this is applicable only to clustered indexes while
4298
doing a unique search. There is scope for further optimization
4299
applicable to unique secondary indexes. Current behaviour is
4300
to widen the scope of a lock on an already delete marked record
4301
if the same record is deleted twice by the same transaction */
4302
if (index == clust_index && unique_search) {
4303
err = DB_RECORD_NOT_FOUND;
4304
4305
goto normal_return;
4306
}
4307
4308
goto next_rec;
4309
}
4310
4311
idx_cond_check:
4312
if (prebuilt->idx_cond_func)
4313
{
4314
int res;
4315
ut_ad(prebuilt->template_type != ROW_DRIZZLE_DUMMY_TEMPLATE);
4316
offsets = rec_get_offsets(rec, index, offsets, ULINT_UNDEFINED, &heap);
4317
row_sel_store_mysql_rec(buf, prebuilt, rec,
4318
offsets, 0, prebuilt->n_index_fields);
4319
res= prebuilt->idx_cond_func(prebuilt->idx_cond_func_arg);
4320
if (res == 0)
4321
goto next_rec;
4322
if (res == 2)
4323
{
4324
err = DB_RECORD_NOT_FOUND;
4325
goto idx_cond_failed;
4326
}
4327
}
4328
4329
/* Get the clustered index record if needed, if we did not do the
4330
search using the clustered index. */
4331
4332
if (get_clust_rec || (index != clust_index &&
4333
prebuilt->need_to_access_clustered)) {
4334
4335
/* We use a 'goto' to the preceding label if a consistent
4336
read of a secondary index record requires us to look up old
4337
versions of the associated clustered index record. */
4338
4339
ut_ad(rec_offs_validate(rec, index, offsets));
4340
4341
/* It was a non-clustered index and we must fetch also the
4342
clustered index record */
4343
4344
mtr_has_extra_clust_latch = TRUE;
4345
4346
/* The following call returns 'offsets' associated with
4347
'clust_rec'. Note that 'clust_rec' can be an old version
4348
built for a consistent read. */
4349
4350
err = row_sel_get_clust_rec_for_mysql(prebuilt, index, rec,
4351
thr, &clust_rec,
4352
&offsets, &heap, &mtr);
4353
if (err != DB_SUCCESS) {
4354
4355
goto lock_wait_or_error;
4356
}
4357
4358
if (clust_rec == NULL) {
4359
/* The record did not exist in the read view */
4360
ut_ad(prebuilt->select_lock_type == LOCK_NONE);
4361
4362
goto next_rec;
4363
}
4364
4365
if (UNIV_UNLIKELY(rec_get_deleted_flag(clust_rec, comp))) {
4366
4367
/* The record is delete marked: we can skip it */
4368
4369
if ((srv_locks_unsafe_for_binlog
4370
|| trx->isolation_level == TRX_ISO_READ_COMMITTED)
4371
&& prebuilt->select_lock_type != LOCK_NONE) {
4372
4373
/* No need to keep a lock on a delete-marked
4374
record if we do not want to use next-key
4375
locking. */
4376
4377
row_unlock_for_mysql(prebuilt, TRUE);
4378
}
4379
4380
goto next_rec;
4381
}
4382
4383
if (prebuilt->need_to_access_clustered) {
4384
4385
result_rec = clust_rec;
4386
4387
ut_ad(rec_offs_validate(result_rec, clust_index,
4388
offsets));
4389
} else {
4390
/* We used 'offsets' for the clust rec, recalculate
4391
them for 'rec' */
4392
offsets = rec_get_offsets(rec, index, offsets,
4393
ULINT_UNDEFINED, &heap);
4394
result_rec = rec;
4395
}
4396
} else {
4397
result_rec = rec;
4398
}
4399
4400
/* We found a qualifying record 'result_rec'. At this point,
4401
'offsets' are associated with 'result_rec'. */
4402
4403
ut_ad(rec_offs_validate(result_rec,
4404
result_rec != rec ? clust_index : index,
4405
offsets));
4406
4407
/* At this point, the clustered index record is protected
4408
by a page latch that was acquired when pcur was positioned.
4409
The latch will not be released until mtr_commit(&mtr). */
4410
4411
if ((match_mode == ROW_SEL_EXACT
4412
|| prebuilt->n_rows_fetched >= MYSQL_FETCH_CACHE_THRESHOLD)
4413
&& prebuilt->select_lock_type == LOCK_NONE
4414
&& !prebuilt->templ_contains_blob
4415
&& !prebuilt->clust_index_was_generated
4416
&& !prebuilt->used_in_HANDLER
4417
&& prebuilt->template_type
4418
!= ROW_MYSQL_DUMMY_TEMPLATE) {
4419
4420
/* Inside an update, for example, we do not cache rows,
4421
since we may use the cursor position to do the actual
4422
update, that is why we require ...lock_type == LOCK_NONE.
4423
Since we keep space in prebuilt only for the BLOBs of
4424
a single row, we cannot cache rows in the case there
4425
are BLOBs in the fields to be fetched. In HANDLER we do
4426
not cache rows because there the cursor is a scrollable
4427
cursor. */
4428
some_fields_in_buffer= (index != clust_index &&
4429
prebuilt->idx_cond_func);
4430
4431
row_sel_push_cache_row_for_mysql(prebuilt, result_rec,
4432
offsets,
4433
some_fields_in_buffer?
4434
prebuilt->n_index_fields: 0,
4435
buf);
4436
if (prebuilt->n_fetch_cached == MYSQL_FETCH_CACHE_SIZE) {
4437
4438
goto got_row;
4439
}
4440
4441
goto next_rec;
4442
} else {
4443
if (prebuilt->template_type == ROW_MYSQL_DUMMY_TEMPLATE) {
4444
memcpy(buf + 4, result_rec
4445
- rec_offs_extra_size(offsets),
4446
rec_offs_size(offsets));
4447
mach_write_to_4(buf,
4448
rec_offs_extra_size(offsets) + 4);
4449
} else {
4450
if (!row_sel_store_mysql_rec(buf, prebuilt,
4451
result_rec, offsets,
4452
prebuilt->idx_cond_func?
4453
prebuilt->n_index_fields: 0,
4454
prebuilt->n_template)) {
4455
err = DB_TOO_BIG_RECORD;
4456
4457
goto lock_wait_or_error;
4458
}
4459
}
4460
4461
if (prebuilt->clust_index_was_generated) {
4462
if (result_rec != rec) {
4463
offsets = rec_get_offsets(
4464
rec, index, offsets, ULINT_UNDEFINED,
4465
&heap);
4466
}
4467
row_sel_store_row_id_to_prebuilt(prebuilt, rec,
4468
index, offsets);
4469
}
4470
}
4471
4472
/* From this point on, 'offsets' are invalid. */
4473
4474
got_row:
4475
/* We have an optimization to save CPU time: if this is a consistent
4476
read on a unique condition on the clustered index, then we do not
4477
store the pcur position, because any fetch next or prev will anyway
4478
return 'end of file'. Exceptions are locking reads and the MySQL
4479
HANDLER command where the user can move the cursor with PREV or NEXT
4480
even after a unique search. */
4481
4482
idx_cond_failed:
4483
if (!unique_search_from_clust_index
4484
|| prebuilt->select_lock_type != LOCK_NONE
4485
|| prebuilt->used_in_HANDLER) {
4486
4487
/* Inside an update always store the cursor position */
4488
4489
btr_pcur_store_position(pcur, &mtr);
4490
}
4491
4492
err = DB_SUCCESS;
4493
4494
goto normal_return;
4495
4496
next_rec:
4497
/* Reset the old and new "did semi-consistent read" flags. */
4498
if (UNIV_UNLIKELY(prebuilt->row_read_type
4499
== ROW_READ_DID_SEMI_CONSISTENT)) {
4500
prebuilt->row_read_type = ROW_READ_TRY_SEMI_CONSISTENT;
4501
}
4502
did_semi_consistent_read = FALSE;
4503
4504
if (UNIV_UNLIKELY(srv_locks_unsafe_for_binlog
4505
|| trx->isolation_level == TRX_ISO_READ_COMMITTED)
4506
&& prebuilt->select_lock_type != LOCK_NONE) {
4507
4508
trx_reset_new_rec_lock_info(trx);
4509
}
4510
4511
/*-------------------------------------------------------------*/
4512
/* PHASE 5: Move the cursor to the next index record */
4513
4514
if (UNIV_UNLIKELY(mtr_has_extra_clust_latch)) {
4515
/* We must commit mtr if we are moving to the next
4516
non-clustered index record, because we could break the
4517
latching order if we would access a different clustered
4518
index page right away without releasing the previous. */
4519
4520
btr_pcur_store_position(pcur, &mtr);
4521
4522
mtr_commit(&mtr);
4523
mtr_has_extra_clust_latch = FALSE;
4524
4525
mtr_start(&mtr);
4526
if (sel_restore_position_for_mysql(&same_user_rec,
4527
BTR_SEARCH_LEAF,
4528
pcur, moves_up, &mtr)) {
4529
#ifdef UNIV_SEARCH_DEBUG
4530
cnt++;
4531
#endif /* UNIV_SEARCH_DEBUG */
4532
4533
goto rec_loop;
4534
}
4535
}
4536
4537
if (moves_up) {
4538
if (UNIV_UNLIKELY(!btr_pcur_move_to_next(pcur, &mtr))) {
4539
not_moved:
4540
btr_pcur_store_position(pcur, &mtr);
4541
4542
if (match_mode != 0) {
4543
err = DB_RECORD_NOT_FOUND;
4544
} else {
4545
err = DB_END_OF_INDEX;
4546
}
4547
4548
goto normal_return;
4549
}
4550
} else {
4551
if (UNIV_UNLIKELY(!btr_pcur_move_to_prev(pcur, &mtr))) {
4552
goto not_moved;
4553
}
4554
}
4555
4556
#ifdef UNIV_SEARCH_DEBUG
4557
cnt++;
4558
#endif /* UNIV_SEARCH_DEBUG */
4559
4560
goto rec_loop;
4561
4562
lock_wait_or_error:
4563
/* Reset the old and new "did semi-consistent read" flags. */
4564
if (UNIV_UNLIKELY(prebuilt->row_read_type
4565
== ROW_READ_DID_SEMI_CONSISTENT)) {
4566
prebuilt->row_read_type = ROW_READ_TRY_SEMI_CONSISTENT;
4567
}
4568
did_semi_consistent_read = FALSE;
4569
4570
/*-------------------------------------------------------------*/
4571
4572
btr_pcur_store_position(pcur, &mtr);
4573
4574
mtr_commit(&mtr);
4575
mtr_has_extra_clust_latch = FALSE;
4576
4577
trx->error_state = err;
4578
4579
/* The following is a patch for MySQL */
4580
4581
que_thr_stop_for_mysql(thr);
4582
4583
thr->lock_state = QUE_THR_LOCK_ROW;
4584
4585
if (row_mysql_handle_errors(&err, trx, thr, NULL)) {
4586
/* It was a lock wait, and it ended */
4587
4588
thr->lock_state = QUE_THR_LOCK_NOLOCK;
4589
mtr_start(&mtr);
4590
4591
sel_restore_position_for_mysql(&same_user_rec,
4592
BTR_SEARCH_LEAF, pcur,
4593
moves_up, &mtr);
4594
4595
if ((srv_locks_unsafe_for_binlog
4596
|| trx->isolation_level == TRX_ISO_READ_COMMITTED)
4597
&& !same_user_rec) {
4598
4599
/* Since we were not able to restore the cursor
4600
on the same user record, we cannot use
4601
row_unlock_for_mysql() to unlock any records, and
4602
we must thus reset the new rec lock info. Since
4603
in lock0lock.c we have blocked the inheriting of gap
4604
X-locks, we actually do not have any new record locks
4605
set in this case.
4606
4607
Note that if we were able to restore on the 'same'
4608
user record, it is still possible that we were actually
4609
waiting on a delete-marked record, and meanwhile
4610
it was removed by purge and inserted again by some
4611
other user. But that is no problem, because in
4612
rec_loop we will again try to set a lock, and
4613
new_rec_lock_info in trx will be right at the end. */
4614
4615
trx_reset_new_rec_lock_info(trx);
4616
}
4617
4618
mode = pcur->search_mode;
4619
4620
goto rec_loop;
4621
}
4622
4623
thr->lock_state = QUE_THR_LOCK_NOLOCK;
4624
4625
#ifdef UNIV_SEARCH_DEBUG
4626
/* fputs("Using ", stderr);
4627
dict_index_name_print(stderr, index);
4628
fprintf(stderr, " cnt %lu ret value %lu err\n", cnt, err); */
4629
#endif /* UNIV_SEARCH_DEBUG */
4630
goto func_exit;
4631
4632
normal_return:
4633
/*-------------------------------------------------------------*/
4634
que_thr_stop_for_mysql_no_error(thr, trx);
4635
4636
mtr_commit(&mtr);
4637
4638
if (prebuilt->n_fetch_cached > 0) {
4639
row_sel_pop_cached_row_for_mysql(buf, prebuilt);
4640
4641
err = DB_SUCCESS;
4642
}
4643
4644
#ifdef UNIV_SEARCH_DEBUG
4645
/* fputs("Using ", stderr);
4646
dict_index_name_print(stderr, index);
4647
fprintf(stderr, " cnt %lu ret value %lu err\n", cnt, err); */
4648
#endif /* UNIV_SEARCH_DEBUG */
4649
if (err == DB_SUCCESS) {
4650
srv_n_rows_read++;
4651
}
4652
4653
func_exit:
4654
trx->op_info = "";
4655
if (UNIV_LIKELY_NULL(heap)) {
4656
mem_heap_free(heap);
4657
}
4658
4659
/* Set or reset the "did semi-consistent read" flag on return.
4660
The flag did_semi_consistent_read is set if and only if
4661
the record being returned was fetched with a semi-consistent read. */
4662
ut_ad(prebuilt->row_read_type != ROW_READ_WITH_LOCKS
4663
|| !did_semi_consistent_read);
4664
4665
if (UNIV_UNLIKELY(prebuilt->row_read_type != ROW_READ_WITH_LOCKS)) {
4666
if (UNIV_UNLIKELY(did_semi_consistent_read)) {
4667
prebuilt->row_read_type = ROW_READ_DID_SEMI_CONSISTENT;
4668
} else {
4669
prebuilt->row_read_type = ROW_READ_TRY_SEMI_CONSISTENT;
4670
}
4671
}
4672
return(err);
4673
}
4674
4675
/***********************************************************************
4676
Checks if MySQL at the moment is allowed for this table to retrieve a
4677
consistent read result, or store it to the query cache. */
4678
UNIV_INTERN
4679
ibool
4680
row_search_check_if_query_cache_permitted(
4681
/*======================================*/
4682
/* out: TRUE if storing or retrieving
4683
from the query cache is permitted */
4684
trx_t* trx, /* in: transaction object */
4685
const char* norm_name) /* in: concatenation of database name,
4686
'/' char, table name */
4687
{
4688
dict_table_t* table;
4689
ibool ret = FALSE;
4690
4691
table = dict_table_get(norm_name, FALSE);
4692
4693
if (table == NULL) {
4694
4695
return(FALSE);
4696
}
4697
4698
mutex_enter(&kernel_mutex);
4699
4700
/* Start the transaction if it is not started yet */
4701
4702
trx_start_if_not_started_low(trx);
4703
4704
/* If there are locks on the table or some trx has invalidated the
4705
cache up to our trx id, then ret = FALSE.
4706
We do not check what type locks there are on the table, though only
4707
IX type locks actually would require ret = FALSE. */
4708
4709
if (UT_LIST_GET_LEN(table->locks) == 0
4710
&& ut_dulint_cmp(trx->id,
4711
table->query_cache_inv_trx_id) >= 0) {
4712
4713
ret = TRUE;
4714
4715
/* If the isolation level is high, assign a read view for the
4716
transaction if it does not yet have one */
4717
4718
if (trx->isolation_level >= TRX_ISO_REPEATABLE_READ
4719
&& !trx->read_view) {
4720
4721
trx->read_view = read_view_open_now(
4722
trx->id, trx->global_read_view_heap);
4723
trx->global_read_view = trx->read_view;
4724
}
4725
}
4726
4727
mutex_exit(&kernel_mutex);
4728
4729
return(ret);
4730
}
4731
4732
/***********************************************************************
4733
Read the AUTOINC column from the current row. If the value is less than
4734
0 and the type is not unsigned then we reset the value to 0. */
4735
static
4736
ib_uint64_t
4737
row_search_autoinc_read_column(
4738
/*===========================*/
4739
/* out: value read from the column */
4740
dict_index_t* index, /* in: index to read from */
4741
const rec_t* rec, /* in: current rec */
4742
ulint col_no, /* in: column number */
4743
ibool unsigned_type) /* in: signed or unsigned flag */
4744
{
4745
ulint len;
4746
const byte* data;
4747
ib_uint64_t value;
4748
mem_heap_t* heap = NULL;
4749
ulint offsets_[REC_OFFS_NORMAL_SIZE];
4750
ulint* offsets = offsets_;
4751
4752
rec_offs_init(offsets_);
4753
4754
offsets = rec_get_offsets(rec, index, offsets, ULINT_UNDEFINED, &heap);
4755
4756
data = rec_get_nth_field(rec, offsets, col_no, &len);
4757
4758
ut_a(len != UNIV_SQL_NULL);
4759
ut_a(len <= sizeof value);
4760
4761
/* we assume AUTOINC value cannot be negative */
4762
value = mach_read_int_type(data, len, unsigned_type);
4763
4764
if (UNIV_LIKELY_NULL(heap)) {
4765
mem_heap_free(heap);
4766
}
4767
4768
if (!unsigned_type && (ib_int64_t) value < 0) {
4769
value = 0;
4770
}
4771
4772
return(value);
4773
}
4774
4775
/***********************************************************************
4776
Get the last row. */
4777
static
4778
const rec_t*
4779
row_search_autoinc_get_rec(
4780
/*=======================*/
4781
/* out: current rec or NULL */
4782
btr_pcur_t* pcur, /* in: the current cursor */
4783
mtr_t* mtr) /* in: mini transaction */
4784
{
4785
do {
4786
const rec_t* rec = btr_pcur_get_rec(pcur);
4787
4788
if (page_rec_is_user_rec(rec)) {
4789
return(rec);
4790
}
4791
} while (btr_pcur_move_to_prev(pcur, mtr));
4792
4793
return(NULL);
4794
}
4795
4796
/***********************************************************************
4797
Read the max AUTOINC value from an index. */
4798
UNIV_INTERN
4799
ulint
4800
row_search_max_autoinc(
4801
/*===================*/
4802
/* out: DB_SUCCESS if all OK else
4803
error code, DB_RECORD_NOT_FOUND if
4804
column name can't be found in index */
4805
dict_index_t* index, /* in: index to search */
4806
const char* col_name, /* in: name of autoinc column */
4807
ib_uint64_t* value) /* out: AUTOINC value read */
4808
{
4809
ulint i;
4810
ulint n_cols;
4811
dict_field_t* dfield = NULL;
4812
ulint error = DB_SUCCESS;
4813
4814
n_cols = dict_index_get_n_ordering_defined_by_user(index);
4815
4816
/* Search the index for the AUTOINC column name */
4817
for (i = 0; i < n_cols; ++i) {
4818
dfield = dict_index_get_nth_field(index, i);
4819
4820
if (strcmp(col_name, dfield->name) == 0) {
4821
break;
4822
}
4823
}
4824
4825
*value = 0;
4826
4827
/* Must find the AUTOINC column name */
4828
if (i < n_cols && dfield) {
4829
mtr_t mtr;
4830
btr_pcur_t pcur;
4831
4832
mtr_start(&mtr);
4833
4834
/* Open at the high/right end (FALSE), and INIT
4835
cursor (TRUE) */
4836
btr_pcur_open_at_index_side(
4837
FALSE, index, BTR_SEARCH_LEAF, &pcur, TRUE, &mtr);
4838
4839
if (page_get_n_recs(btr_pcur_get_page(&pcur)) > 0) {
4840
const rec_t* rec;
4841
4842
rec = row_search_autoinc_get_rec(&pcur, &mtr);
4843
4844
if (rec != NULL) {
4845
ibool unsigned_type = (
4846
dfield->col->prtype & DATA_UNSIGNED);
4847
4848
*value = row_search_autoinc_read_column(
4849
index, rec, i, unsigned_type);
4850
}
4851
}
4852
4853
btr_pcur_close(&pcur);
4854
4855
mtr_commit(&mtr);
4856
} else {
4857
error = DB_RECORD_NOT_FOUND;
4858
}
4859
4860
return(error);
4861
}
Loggerhead is a web-based interface for Breezy
Version: 2.0.1