row0sel.c 125 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
/*****************************************************************************

Copyright (c) 1997, 2009, Innobase Oy. All Rights Reserved.
Copyright (c) 2008, Google Inc.

Portions of this file contain modifications contributed and copyrighted by
Google, Inc. Those modifications are gratefully acknowledged and are described
briefly in the InnoDB documentation. The contributions by Google are
incorporated with their permission, and subject to the conditions contained in
the file COPYING.Google.

This program is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free Software
Foundation; version 2 of the License.

This program is distributed in the hope that it will be useful, but WITHOUT
ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.

You should have received a copy of the GNU General Public License along with
this program; if not, write to the Free Software Foundation, Inc., 59 Temple
Place, Suite 330, Boston, MA 02111-1307 USA

*****************************************************************************/

osku's avatar
osku committed
26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57
/*******************************************************
Select

Created 12/19/1997 Heikki Tuuri
*******************************************************/

#include "row0sel.h"

#ifdef UNIV_NONINL
#include "row0sel.ic"
#endif

#include "dict0dict.h"
#include "dict0boot.h"
#include "trx0undo.h"
#include "trx0trx.h"
#include "btr0btr.h"
#include "btr0cur.h"
#include "btr0sea.h"
#include "mach0data.h"
#include "que0que.h"
#include "row0upd.h"
#include "row0row.h"
#include "row0vers.h"
#include "rem0cmp.h"
#include "lock0lock.h"
#include "eval0eval.h"
#include "pars0sym.h"
#include "pars0pars.h"
#include "row0mysql.h"
#include "read0read.h"
#include "buf0lru.h"
58
#include "ha_prototypes.h"
osku's avatar
osku committed
59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76

/* Maximum number of rows to prefetch; MySQL interface has another parameter */
#define SEL_MAX_N_PREFETCH	16

/* Number of rows fetched, after which to start prefetching; MySQL interface
has another parameter */
#define SEL_PREFETCH_LIMIT	1

/* When a select has accessed about this many pages, it returns control back
to que_run_threads: this is to allow canceling runaway queries */

#define SEL_COST_LIMIT	100

/* Flags for search shortcut */
#define SEL_FOUND	0
#define	SEL_EXHAUSTED	1
#define SEL_RETRY	2

77 78 79 80 81
/************************************************************************
Returns TRUE if the user-defined column in a secondary index record
is alphabetically the same as the corresponding BLOB column in the clustered
index record.
NOTE: the comparison is NOT done as a binary comparison, but character
82 83
fields are compared with collation!
@return	TRUE if the columns are equal */
84 85 86 87
static
ibool
row_sel_sec_rec_is_for_blob(
/*========================*/
88 89 90
	ulint		mtype,		/*!< in: main type */
	ulint		prtype,		/*!< in: precise type */
	ulint		mbminlen,	/*!< in: minimum length of a
marko's avatar
marko committed
91
					multi-byte character */
92
	ulint		mbmaxlen,	/*!< in: maximum length of a
93
					multi-byte character */
94
	const byte*	clust_field,	/*!< in: the locally stored part of
95 96 97 98 99
					the clustered index column, including
					the BLOB pointer; the clustered
					index record must be covered by
					a lock or a page latch to protect it
					against deletion (rollback or purge) */
100 101 102 103
	ulint		clust_len,	/*!< in: length of clust_field */
	const byte*	sec_field,	/*!< in: column in secondary index */
	ulint		sec_len,	/*!< in: length of sec_field */
	ulint		zip_size)	/*!< in: compressed page size, or 0 */
104 105 106 107 108 109 110
{
	ulint	len;
	byte	buf[DICT_MAX_INDEX_COL_LEN];

	len = btr_copy_externally_stored_field_prefix(buf, sizeof buf,
						      zip_size,
						      clust_field, clust_len);
111 112 113 114 115 116 117 118 119 120

	if (UNIV_UNLIKELY(len == 0)) {
		/* The BLOB was being deleted as the server crashed.
		There should not be any secondary index records
		referring to this clustered index record, because
		btr_free_externally_stored_field() is called after all
		secondary index entries of the row have been purged. */
		return(FALSE);
	}

121
	len = dtype_get_at_most_n_mbchars(prtype, mbminlen, mbmaxlen,
122
					  sec_len, len, (const char*) buf);
123 124 125 126

	return(!cmp_data_data(mtype, prtype, buf, len, sec_field, sec_len));
}

osku's avatar
osku committed
127 128 129 130 131
/************************************************************************
Returns TRUE if the user-defined column values in a secondary index record
are alphabetically the same as the corresponding columns in the clustered
index record.
NOTE: the comparison is NOT done as a binary comparison, but character
132 133
fields are compared with collation!
@return	TRUE if the secondary record is equal to the corresponding fields in the clustered record, when compared with collation */
osku's avatar
osku committed
134 135 136 137
static
ibool
row_sel_sec_rec_is_for_clust_rec(
/*=============================*/
138 139 140
	const rec_t*	sec_rec,	/*!< in: secondary index record */
	dict_index_t*	sec_index,	/*!< in: secondary index */
	const rec_t*	clust_rec,	/*!< in: clustered index record;
141 142 143
					must be protected by a lock or
					a page latch against deletion
					in rollback or purge */
144
	dict_index_t*	clust_index)	/*!< in: clustered index */
osku's avatar
osku committed
145
{
marko's avatar
marko committed
146
	const byte*	sec_field;
147
	ulint		sec_len;
marko's avatar
marko committed
148
	const byte*	clust_field;
149 150
	ulint		n;
	ulint		i;
osku's avatar
osku committed
151 152 153 154 155 156 157
	mem_heap_t*	heap		= NULL;
	ulint		clust_offsets_[REC_OFFS_NORMAL_SIZE];
	ulint		sec_offsets_[REC_OFFS_SMALL_SIZE];
	ulint*		clust_offs	= clust_offsets_;
	ulint*		sec_offs	= sec_offsets_;
	ibool		is_equal	= TRUE;

158 159
	rec_offs_init(clust_offsets_);
	rec_offs_init(sec_offsets_);
osku's avatar
osku committed
160

161 162 163 164 165 166 167 168 169 170
	if (rec_get_deleted_flag(clust_rec,
				 dict_table_is_comp(clust_index->table))) {

		/* The clustered index record is delete-marked;
		it is not visible in the read view.  Besides,
		if there are any externally stored columns,
		some of them may have already been purged. */
		return(FALSE);
	}

osku's avatar
osku committed
171
	clust_offs = rec_get_offsets(clust_rec, clust_index, clust_offs,
172
				     ULINT_UNDEFINED, &heap);
osku's avatar
osku committed
173
	sec_offs = rec_get_offsets(sec_rec, sec_index, sec_offs,
174
				   ULINT_UNDEFINED, &heap);
osku's avatar
osku committed
175

176
	n = dict_index_get_n_ordering_defined_by_user(sec_index);
osku's avatar
osku committed
177

178
	for (i = 0; i < n; i++) {
179 180
		const dict_field_t*	ifield;
		const dict_col_t*	col;
181 182 183
		ulint			clust_pos;
		ulint			clust_len;
		ulint			len;
184

osku's avatar
osku committed
185
		ifield = dict_index_get_nth_field(sec_index, i);
186
		col = dict_field_get_col(ifield);
187
		clust_pos = dict_col_get_clust_pos(col, clust_index);
188

189
		clust_field = rec_get_nth_field(
190
			clust_rec, clust_offs, clust_pos, &clust_len);
osku's avatar
osku committed
191 192
		sec_field = rec_get_nth_field(sec_rec, sec_offs, i, &sec_len);

193 194 195 196 197 198 199
		len = clust_len;

		if (ifield->prefix_len > 0 && len != UNIV_SQL_NULL) {

			if (rec_offs_nth_extern(clust_offs, clust_pos)) {
				len -= BTR_EXTERN_FIELD_REF_SIZE;
			}
osku's avatar
osku committed
200

201
			len = dtype_get_at_most_n_mbchars(
202
				col->prtype, col->mbminlen, col->mbmaxlen,
203 204 205 206 207 208
				ifield->prefix_len, len, (char*) clust_field);

			if (rec_offs_nth_extern(clust_offs, clust_pos)
			    && len < sec_len) {
				if (!row_sel_sec_rec_is_for_blob(
					    col->mtype, col->prtype,
209
					    col->mbminlen, col->mbmaxlen,
210 211 212 213 214 215
					    clust_field, clust_len,
					    sec_field, sec_len,
					    dict_table_zip_size(
						    clust_index->table))) {
					goto inequal;
				}
216 217

				continue;
218
			}
osku's avatar
osku committed
219 220
		}

221
		if (0 != cmp_data_data(col->mtype, col->prtype,
222
				       clust_field, len,
223
				       sec_field, sec_len)) {
224
inequal:
osku's avatar
osku committed
225 226
			is_equal = FALSE;
			goto func_exit;
227 228
		}
	}
osku's avatar
osku committed
229 230 231 232 233 234 235 236 237

func_exit:
	if (UNIV_LIKELY_NULL(heap)) {
		mem_heap_free(heap);
	}
	return(is_equal);
}

/*************************************************************************
238 239
Creates a select node struct.
@return	own: select node struct */
240
UNIV_INTERN
osku's avatar
osku committed
241 242 243
sel_node_t*
sel_node_create(
/*============*/
244
	mem_heap_t*	heap)	/*!< in: memory heap where created */
osku's avatar
osku committed
245 246 247 248 249 250 251 252
{
	sel_node_t*	node;

	node = mem_heap_alloc(heap, sizeof(sel_node_t));
	node->common.type = QUE_NODE_SELECT;
	node->state = SEL_NODE_OPEN;

	node->plans = NULL;
253

osku's avatar
osku committed
254 255 256 257 258 259
	return(node);
}

/*************************************************************************
Frees the memory private to a select node when a query graph is freed,
does not free the heap where the node was originally created. */
260
UNIV_INTERN
osku's avatar
osku committed
261 262 263
void
sel_node_free_private(
/*==================*/
264
	sel_node_t*	node)	/*!< in: select node struct */
osku's avatar
osku committed
265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289
{
	ulint	i;
	plan_t*	plan;

	if (node->plans != NULL) {
		for (i = 0; i < node->n_tables; i++) {
			plan = sel_node_get_nth_plan(node, i);

			btr_pcur_close(&(plan->pcur));
			btr_pcur_close(&(plan->clust_pcur));

			if (plan->old_vers_heap) {
				mem_heap_free(plan->old_vers_heap);
			}
		}
	}
}

/*************************************************************************
Evaluates the values in a select list. If there are aggregate functions,
their argument value is added to the aggregate total. */
UNIV_INLINE
void
sel_eval_select_list(
/*=================*/
290
	sel_node_t*	node)	/*!< in: select node */
osku's avatar
osku committed
291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309
{
	que_node_t*	exp;

	exp = node->select_list;

	while (exp) {
		eval_exp(exp);

		exp = que_node_get_next(exp);
	}
}

/*************************************************************************
Assigns the values in the select list to the possible into-variables in
SELECT ... INTO ... */
UNIV_INLINE
void
sel_assign_into_var_values(
/*=======================*/
310 311
	sym_node_t*	var,	/*!< in: first variable in a list of variables */
	sel_node_t*	node)	/*!< in: select node */
osku's avatar
osku committed
312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338
{
	que_node_t*	exp;

	if (var == NULL) {

		return;
	}

	exp = node->select_list;

	while (var) {
		ut_ad(exp);

		eval_node_copy_val(var->alias, exp);

		exp = que_node_get_next(exp);
		var = que_node_get_next(var);
	}
}

/*************************************************************************
Resets the aggregate value totals in the select list of an aggregate type
query. */
UNIV_INLINE
void
sel_reset_aggregate_vals(
/*=====================*/
339
	sel_node_t*	node)	/*!< in: select node */
osku's avatar
osku committed
340 341 342 343 344 345 346 347 348 349 350
{
	func_node_t*	func_node;

	ut_ad(node->is_aggregate);

	func_node = node->select_list;

	while (func_node) {
		eval_node_set_int_val(func_node, 0);

		func_node = que_node_get_next(func_node);
351
	}
osku's avatar
osku committed
352 353 354 355 356 357 358 359 360 361

	node->aggregate_already_fetched = FALSE;
}

/*************************************************************************
Copies the input variable values when an explicit cursor is opened. */
UNIV_INLINE
void
row_sel_copy_input_variable_vals(
/*=============================*/
362
	sel_node_t*	node)	/*!< in: select node */
osku's avatar
osku committed
363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382
{
	sym_node_t*	var;

	var = UT_LIST_GET_FIRST(node->copy_variables);

	while (var) {
		eval_node_copy_val(var, var->alias);

		var->indirection = NULL;

		var = UT_LIST_GET_NEXT(col_var_list, var);
	}
}

/*************************************************************************
Fetches the column values from a record. */
static
void
row_sel_fetch_columns(
/*==================*/
383 384
	dict_index_t*	index,	/*!< in: record index */
	const rec_t*	rec,	/*!< in: record in a clustered or non-clustered
385
				index; must be protected by a page latch */
386 387
	const ulint*	offsets,/*!< in: rec_get_offsets(rec, index) */
	sym_node_t*	column)	/*!< in: first column in a column list, or
osku's avatar
osku committed
388 389 390 391 392
				NULL */
{
	dfield_t*	val;
	ulint		index_type;
	ulint		field_no;
393
	const byte*	data;
osku's avatar
osku committed
394
	ulint		len;
395

osku's avatar
osku committed
396 397
	ut_ad(rec_offs_validate(rec, index, offsets));

398
	if (dict_index_is_clust(index)) {
osku's avatar
osku committed
399 400 401 402 403 404
		index_type = SYM_CLUST_FIELD_NO;
	} else {
		index_type = SYM_SEC_FIELD_NO;
	}

	while (column) {
405 406 407
		mem_heap_t*	heap = NULL;
		ibool		needs_copy;

osku's avatar
osku committed
408 409 410
		field_no = column->field_nos[index_type];

		if (field_no != ULINT_UNDEFINED) {
411

412
			if (UNIV_UNLIKELY(rec_offs_nth_extern(offsets,
413
							      field_no))) {
414 415 416 417 418 419

				/* Copy an externally stored field to the
				temporary heap */

				heap = mem_heap_create(1);

420 421 422 423
				data = btr_rec_copy_externally_stored_field(
					rec, offsets,
					dict_table_zip_size(index->table),
					field_no, &len, heap);
424 425

				ut_a(len != UNIV_SQL_NULL);
426

427 428 429
				needs_copy = TRUE;
			} else {
				data = rec_get_nth_field(rec, offsets,
430
							 field_no, &len);
431

432 433 434 435
				if (len == UNIV_SQL_NULL) {
					len = UNIV_SQL_NULL;
				}

436 437 438 439
				needs_copy = column->copy_val;
			}

			if (needs_copy) {
osku's avatar
osku committed
440
				eval_node_copy_and_alloc_val(column, data,
441
							     len);
osku's avatar
osku committed
442 443 444 445
			} else {
				val = que_node_get_val(column);
				dfield_set_data(val, data, len);
			}
446 447 448 449

			if (UNIV_LIKELY_NULL(heap)) {
				mem_heap_free(heap);
			}
osku's avatar
osku committed
450 451 452 453 454 455 456 457 458 459 460 461
		}

		column = UT_LIST_GET_NEXT(col_var_list, column);
	}
}

/*************************************************************************
Allocates a prefetch buffer for a column when prefetch is first time done. */
static
void
sel_col_prefetch_buf_alloc(
/*=======================*/
462
	sym_node_t*	column)	/*!< in: symbol table node for a column */
osku's avatar
osku committed
463 464 465 466 467
{
	sel_buf_t*	sel_buf;
	ulint		i;

	ut_ad(que_node_get_type(column) == QUE_NODE_SYMBOL);
468

osku's avatar
osku committed
469
	column->prefetch_buf = mem_alloc(SEL_MAX_N_PREFETCH
470
					 * sizeof(sel_buf_t));
osku's avatar
osku committed
471 472 473 474 475 476 477 478 479 480 481 482
	for (i = 0; i < SEL_MAX_N_PREFETCH; i++) {
		sel_buf = column->prefetch_buf + i;

		sel_buf->data = NULL;

		sel_buf->val_buf_size = 0;
	}
}

/*************************************************************************
Frees a prefetch buffer for a column, including the dynamically allocated
memory for data stored there. */
483
UNIV_INTERN
osku's avatar
osku committed
484 485 486
void
sel_col_prefetch_buf_free(
/*======================*/
487
	sel_buf_t*	prefetch_buf)	/*!< in, own: prefetch buffer */
osku's avatar
osku committed
488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508
{
	sel_buf_t*	sel_buf;
	ulint		i;

	for (i = 0; i < SEL_MAX_N_PREFETCH; i++) {
		sel_buf = prefetch_buf + i;

		if (sel_buf->val_buf_size > 0) {

			mem_free(sel_buf->data);
		}
	}
}

/*************************************************************************
Pops the column values for a prefetched, cached row from the column prefetch
buffers and places them to the val fields in the column nodes. */
static
void
sel_pop_prefetched_row(
/*===================*/
509
	plan_t*	plan)	/*!< in: plan node for a table */
osku's avatar
osku committed
510 511 512 513 514 515 516
{
	sym_node_t*	column;
	sel_buf_t*	sel_buf;
	dfield_t*	val;
	byte*		data;
	ulint		len;
	ulint		val_buf_size;
517

osku's avatar
osku committed
518 519 520 521 522 523 524 525 526 527 528 529 530
	ut_ad(plan->n_rows_prefetched > 0);

	column = UT_LIST_GET_FIRST(plan->columns);

	while (column) {
		val = que_node_get_val(column);

		if (!column->copy_val) {
			/* We did not really push any value for the
			column */

			ut_ad(!column->prefetch_buf);
			ut_ad(que_node_get_val_buf_size(column) == 0);
531 532
			ut_d(dfield_set_null(val));

osku's avatar
osku committed
533 534 535 536
			goto next_col;
		}

		ut_ad(column->prefetch_buf);
537
		ut_ad(!dfield_is_ext(val));
osku's avatar
osku committed
538 539 540 541 542 543 544 545 546 547 548 549 550 551

		sel_buf = column->prefetch_buf + plan->first_prefetched;

		data = sel_buf->data;
		len = sel_buf->len;
		val_buf_size = sel_buf->val_buf_size;

		/* We must keep track of the allocated memory for
		column values to be able to free it later: therefore
		we swap the values for sel_buf and val */

		sel_buf->data = dfield_get_data(val);
		sel_buf->len = dfield_get_len(val);
		sel_buf->val_buf_size = que_node_get_val_buf_size(column);
552

osku's avatar
osku committed
553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570
		dfield_set_data(val, data, len);
		que_node_set_val_buf_size(column, val_buf_size);
next_col:
		column = UT_LIST_GET_NEXT(col_var_list, column);
	}

	plan->n_rows_prefetched--;

	plan->first_prefetched++;
}

/*************************************************************************
Pushes the column values for a prefetched, cached row to the column prefetch
buffers from the val fields in the column nodes. */
UNIV_INLINE
void
sel_push_prefetched_row(
/*====================*/
571
	plan_t*	plan)	/*!< in: plan node for a table */
osku's avatar
osku committed
572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588
{
	sym_node_t*	column;
	sel_buf_t*	sel_buf;
	dfield_t*	val;
	byte*		data;
	ulint		len;
	ulint		pos;
	ulint		val_buf_size;

	if (plan->n_rows_prefetched == 0) {
		pos = 0;
		plan->first_prefetched = 0;
	} else {
		pos = plan->n_rows_prefetched;

		/* We have the convention that pushing new rows starts only
		after the prefetch stack has been emptied: */
589

osku's avatar
osku committed
590 591 592 593
		ut_ad(plan->first_prefetched == 0);
	}

	plan->n_rows_prefetched++;
594

osku's avatar
osku committed
595
	ut_ad(pos < SEL_MAX_N_PREFETCH);
596

osku's avatar
osku committed
597 598 599 600 601 602 603 604 605
	column = UT_LIST_GET_FIRST(plan->columns);

	while (column) {
		if (!column->copy_val) {
			/* There is no sense to push pointers to database
			page fields when we do not keep latch on the page! */

			goto next_col;
		}
606

osku's avatar
osku committed
607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626
		if (!column->prefetch_buf) {
			/* Allocate a new prefetch buffer */

			sel_col_prefetch_buf_alloc(column);
		}

		sel_buf = column->prefetch_buf + pos;

		val = que_node_get_val(column);

		data = dfield_get_data(val);
		len = dfield_get_len(val);
		val_buf_size = que_node_get_val_buf_size(column);

		/* We must keep track of the allocated memory for
		column values to be able to free it later: therefore
		we swap the values for sel_buf and val */

		dfield_set_data(val, sel_buf->data, sel_buf->len);
		que_node_set_val_buf_size(column, sel_buf->val_buf_size);
627

osku's avatar
osku committed
628 629 630
		sel_buf->data = data;
		sel_buf->len = len;
		sel_buf->val_buf_size = val_buf_size;
631
next_col:
osku's avatar
osku committed
632 633 634 635 636
		column = UT_LIST_GET_NEXT(col_var_list, column);
	}
}

/*************************************************************************
637 638
Builds a previous version of a clustered index record for a consistent read
@return	DB_SUCCESS or error code */
osku's avatar
osku committed
639 640 641 642
static
ulint
row_sel_build_prev_vers(
/*====================*/
643 644 645 646
	read_view_t*	read_view,	/*!< in: read view */
	dict_index_t*	index,		/*!< in: plan node for table */
	rec_t*		rec,		/*!< in: record in a clustered index */
	ulint**		offsets,	/*!< in/out: offsets returned by
osku's avatar
osku committed
647
					rec_get_offsets(rec, plan->index) */
648
	mem_heap_t**	offset_heap,	/*!< in/out: memory heap from which
osku's avatar
osku committed
649
					the offsets are allocated */
650 651
	mem_heap_t**    old_vers_heap,  /*!< out: old version heap to use */
	rec_t**		old_vers,	/*!< out: old version, or NULL if the
osku's avatar
osku committed
652 653 654
					record does not exist in the view:
					i.e., it was freshly inserted
					afterwards */
655
	mtr_t*		mtr)		/*!< in: mtr */
osku's avatar
osku committed
656 657 658
{
	ulint	err;

659 660
	if (*old_vers_heap) {
		mem_heap_empty(*old_vers_heap);
osku's avatar
osku committed
661
	} else {
662
		*old_vers_heap = mem_heap_create(512);
osku's avatar
osku committed
663
	}
664

665
	err = row_vers_build_for_consistent_read(
666 667
		rec, mtr, index, offsets, read_view, offset_heap,
		*old_vers_heap, old_vers);
osku's avatar
osku committed
668 669 670
	return(err);
}

671 672
/*************************************************************************
Builds the last committed version of a clustered index record for a
673 674
semi-consistent read.
@return	DB_SUCCESS or error code */
675 676 677 678
static
ulint
row_sel_build_committed_vers_for_mysql(
/*===================================*/
679 680 681 682
	dict_index_t*	clust_index,	/*!< in: clustered index */
	row_prebuilt_t*	prebuilt,	/*!< in: prebuilt struct */
	const rec_t*	rec,		/*!< in: record in a clustered index */
	ulint**		offsets,	/*!< in/out: offsets returned by
683
					rec_get_offsets(rec, clust_index) */
684
	mem_heap_t**	offset_heap,	/*!< in/out: memory heap from which
685
					the offsets are allocated */
686
	const rec_t**	old_vers,	/*!< out: old version, or NULL if the
687 688 689
					record does not exist in the view:
					i.e., it was freshly inserted
					afterwards */
690
	mtr_t*		mtr)		/*!< in: mtr */
691 692 693 694 695 696 697 698
{
	ulint	err;

	if (prebuilt->old_vers_heap) {
		mem_heap_empty(prebuilt->old_vers_heap);
	} else {
		prebuilt->old_vers_heap = mem_heap_create(200);
	}
699

700 701 702
	err = row_vers_build_for_semi_consistent_read(
		rec, mtr, clust_index, offsets, offset_heap,
		prebuilt->old_vers_heap, old_vers);
703 704 705
	return(err);
}

osku's avatar
osku committed
706 707
/*************************************************************************
Tests the conditions which determine when the index segment we are searching
708 709
through has been exhausted.
@return	TRUE if row passed the tests */
osku's avatar
osku committed
710 711 712 713
UNIV_INLINE
ibool
row_sel_test_end_conds(
/*===================*/
714
	plan_t*	plan)	/*!< in: plan for the table; the column values must
osku's avatar
osku committed
715 716 717 718 719 720 721
			already have been retrieved and the right sides of
			comparisons evaluated */
{
	func_node_t*	cond;

	/* All conditions in end_conds are comparisons of a column to an
	expression */
722

osku's avatar
osku committed
723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744
	cond = UT_LIST_GET_FIRST(plan->end_conds);

	while (cond) {
		/* Evaluate the left side of the comparison, i.e., get the
		column value if there is an indirection */

		eval_sym(cond->args);

		/* Do the comparison */

		if (!eval_cmp(cond)) {

			return(FALSE);
		}

		cond = UT_LIST_GET_NEXT(cond_list, cond);
	}

	return(TRUE);
}

/*************************************************************************
745 746
Tests the other conditions.
@return	TRUE if row passed the tests */
osku's avatar
osku committed
747 748 749 750
UNIV_INLINE
ibool
row_sel_test_other_conds(
/*=====================*/
751
	plan_t*	plan)	/*!< in: plan for the table; the column values must
osku's avatar
osku committed
752 753 754
			already have been retrieved */
{
	func_node_t*	cond;
755

osku's avatar
osku committed
756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773
	cond = UT_LIST_GET_FIRST(plan->other_conds);

	while (cond) {
		eval_exp(cond);

		if (!eval_node_get_ibool_val(cond)) {

			return(FALSE);
		}

		cond = UT_LIST_GET_NEXT(cond_list, cond);
	}

	return(TRUE);
}

/*************************************************************************
Retrieves the clustered index record corresponding to a record in a
774 775
non-clustered index. Does the necessary locking.
@return	DB_SUCCESS or error code */
osku's avatar
osku committed
776 777 778 779
static
ulint
row_sel_get_clust_rec(
/*==================*/
780 781 782 783 784
	sel_node_t*	node,	/*!< in: select_node */
	plan_t*		plan,	/*!< in: plan node for table */
	rec_t*		rec,	/*!< in: record in a non-clustered index */
	que_thr_t*	thr,	/*!< in: query thread */
	rec_t**		out_rec,/*!< out: clustered record or an old version of
osku's avatar
osku committed
785 786 787
				it, NULL if the old version did not exist
				in the read view, i.e., it was a fresh
				inserted version */
788
	mtr_t*		mtr)	/*!< in: mtr used to get access to the
osku's avatar
osku committed
789 790 791 792 793 794 795 796 797 798
				non-clustered record; the same mtr is used to
				access the clustered index */
{
	dict_index_t*	index;
	rec_t*		clust_rec;
	rec_t*		old_vers;
	ulint		err;
	mem_heap_t*	heap		= NULL;
	ulint		offsets_[REC_OFFS_NORMAL_SIZE];
	ulint*		offsets		= offsets_;
799
	rec_offs_init(offsets_);
osku's avatar
osku committed
800 801 802 803

	*out_rec = NULL;

	offsets = rec_get_offsets(rec,
804 805
				  btr_pcur_get_btr_cur(&plan->pcur)->index,
				  offsets, ULINT_UNDEFINED, &heap);
806

osku's avatar
osku committed
807 808 809
	row_build_row_ref_fast(plan->clust_ref, plan->clust_map, rec, offsets);

	index = dict_table_get_first_index(plan->table);
810

osku's avatar
osku committed
811
	btr_pcur_open_with_no_init(index, plan->clust_ref, PAGE_CUR_LE,
812
				   BTR_SEARCH_LEAF, &plan->clust_pcur,
813
				   0, mtr);
osku's avatar
osku committed
814 815 816 817 818 819 820

	clust_rec = btr_pcur_get_rec(&(plan->clust_pcur));

	/* Note: only if the search ends up on a non-infimum record is the
	low_match value the real match to the search tuple */

	if (!page_rec_is_user_rec(clust_rec)
821 822
	    || btr_pcur_get_low_match(&(plan->clust_pcur))
	    < dict_index_get_n_unique(index)) {
823

824
		ut_a(rec_get_deleted_flag(rec,
825
					  dict_table_is_comp(plan->table)));
osku's avatar
osku committed
826 827 828 829 830 831 832 833 834 835 836 837 838 839 840
		ut_a(node->read_view);

		/* In a rare case it is possible that no clust rec is found
		for a delete-marked secondary index record: if in row0umod.c
		in row_undo_mod_remove_clust_low() we have already removed
		the clust rec, while purge is still cleaning and removing
		secondary index records associated with earlier versions of
		the clustered index record. In that case we know that the
		clustered index record did not exist in the read view of
		trx. */

		goto func_exit;
	}

	offsets = rec_get_offsets(clust_rec, index, offsets,
841
				  ULINT_UNDEFINED, &heap);
osku's avatar
osku committed
842 843 844

	if (!node->read_view) {
		/* Try to place a lock on the index record */
845

846 847
		/* If innodb_locks_unsafe_for_binlog option is used
		or this session is using READ COMMITTED isolation level
osku's avatar
osku committed
848 849 850
		we lock only the record, i.e., next-key locking is
		not used. */
		ulint	lock_type;
851
		trx_t*	trx;
osku's avatar
osku committed
852

853 854 855
		trx = thr_get_trx(thr);

		if (srv_locks_unsafe_for_binlog
856
		    || trx->isolation_level == TRX_ISO_READ_COMMITTED) {
osku's avatar
osku committed
857 858 859 860 861
			lock_type = LOCK_REC_NOT_GAP;
		} else {
			lock_type = LOCK_ORDINARY;
		}

862
		err = lock_clust_rec_read_check_and_lock(
863 864
			0, btr_pcur_get_block(&plan->clust_pcur),
			clust_rec, index, offsets,
865
			node->row_lock_mode, lock_type, thr);
osku's avatar
osku committed
866 867 868 869 870 871 872 873 874 875 876 877

		if (err != DB_SUCCESS) {

			goto err_exit;
		}
	} else {
		/* This is a non-locking consistent read: if necessary, fetch
		a previous version of the record */

		old_vers = NULL;

		if (!lock_clust_rec_cons_read_sees(clust_rec, index, offsets,
878
						   node->read_view)) {
osku's avatar
osku committed
879

880 881 882 883 884
			err = row_sel_build_prev_vers(
				node->read_view, index, clust_rec,
				&offsets, &heap, &plan->old_vers_heap,
				&old_vers, mtr);

osku's avatar
osku committed
885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908
			if (err != DB_SUCCESS) {

				goto err_exit;
			}

			clust_rec = old_vers;

			if (clust_rec == NULL) {
				goto func_exit;
			}
		}

		/* If we had to go to an earlier version of row or the
		secondary index record is delete marked, then it may be that
		the secondary index record corresponding to clust_rec
		(or old_vers) is not rec; in that case we must ignore
		such row because in our snapshot rec would not have existed.
		Remember that from rec we cannot see directly which transaction
		id corresponds to it: we have to go to the clustered index
		record. A query where we want to fetch all rows where
		the secondary index value is in some interval would return
		a wrong result if we would not drop rows which we come to
		visit through secondary index records that would not really
		exist in our snapshot. */
909

910
		if ((old_vers
911 912
		     || rec_get_deleted_flag(rec, dict_table_is_comp(
						     plan->table)))
913 914
		    && !row_sel_sec_rec_is_for_clust_rec(rec, plan->index,
							 clust_rec, index)) {
osku's avatar
osku committed
915 916 917 918
			goto func_exit;
		}
	}

919 920 921 922
	/* Fetch the columns needed in test conditions.  The clustered
	index record is protected by a page latch that was acquired
	when plan->clust_pcur was positioned.  The latch will not be
	released until mtr_commit(mtr). */
osku's avatar
osku committed
923 924

	row_sel_fetch_columns(index, clust_rec, offsets,
925
			      UT_LIST_GET_FIRST(plan->columns));
osku's avatar
osku committed
926 927 928 929 930 931 932 933 934 935 936
	*out_rec = clust_rec;
func_exit:
	err = DB_SUCCESS;
err_exit:
	if (UNIV_LIKELY_NULL(heap)) {
		mem_heap_free(heap);
	}
	return(err);
}

/*************************************************************************
937 938
Sets a lock on a record.
@return	DB_SUCCESS or error code */
osku's avatar
osku committed
939 940 941 942
UNIV_INLINE
ulint
sel_set_rec_lock(
/*=============*/
943 944 945 946 947 948
	const buf_block_t*	block,	/*!< in: buffer block of rec */
	const rec_t*		rec,	/*!< in: record */
	dict_index_t*		index,	/*!< in: index */
	const ulint*		offsets,/*!< in: rec_get_offsets(rec, index) */
	ulint			mode,	/*!< in: lock mode */
	ulint			type,	/*!< in: LOCK_ORDINARY, LOCK_GAP, or
949
					LOC_REC_NOT_GAP */
950
	que_thr_t*		thr)	/*!< in: query thread */
osku's avatar
osku committed
951 952 953 954
{
	trx_t*	trx;
	ulint	err;

955
	trx = thr_get_trx(thr);
osku's avatar
osku committed
956 957 958

	if (UT_LIST_GET_LEN(trx->trx_locks) > 10000) {
		if (buf_LRU_buf_pool_running_out()) {
959

osku's avatar
osku committed
960 961 962 963
			return(DB_LOCK_TABLE_FULL);
		}
	}

964
	if (dict_index_is_clust(index)) {
965
		err = lock_clust_rec_read_check_and_lock(
966
			0, block, rec, index, offsets, mode, type, thr);
osku's avatar
osku committed
967
	} else {
968
		err = lock_sec_rec_read_check_and_lock(
969
			0, block, rec, index, offsets, mode, type, thr);
osku's avatar
osku committed
970 971 972 973 974 975 976 977 978 979 980
	}

	return(err);
}

/*************************************************************************
Opens a pcur to a table index. */
static
void
row_sel_open_pcur(
/*==============*/
981
	plan_t*		plan,		/*!< in: table plan */
osku's avatar
osku committed
982
	ibool		search_latch_locked,
983
					/*!< in: TRUE if the thread currently
osku's avatar
osku committed
984 985
					has the search latch locked in
					s-mode */
986
	mtr_t*		mtr)		/*!< in: mtr */
osku's avatar
osku committed
987 988 989 990 991
{
	dict_index_t*	index;
	func_node_t*	cond;
	que_node_t*	exp;
	ulint		n_fields;
992
	ulint		has_search_latch = 0;	/* RW_S_LATCH or 0 */
osku's avatar
osku committed
993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008
	ulint		i;

	if (search_latch_locked) {
		has_search_latch = RW_S_LATCH;
	}

	index = plan->index;

	/* Calculate the value of the search tuple: the exact match columns
	get their expressions evaluated when we evaluate the right sides of
	end_conds */

	cond = UT_LIST_GET_FIRST(plan->end_conds);

	while (cond) {
		eval_exp(que_node_get_next(cond->args));
1009

osku's avatar
osku committed
1010 1011
		cond = UT_LIST_GET_NEXT(cond_list, cond);
	}
1012

osku's avatar
osku committed
1013 1014
	if (plan->tuple) {
		n_fields = dtuple_get_n_fields(plan->tuple);
1015

osku's avatar
osku committed
1016 1017 1018
		if (plan->n_exact_match < n_fields) {
			/* There is a non-exact match field which must be
			evaluated separately */
1019

osku's avatar
osku committed
1020 1021
			eval_exp(plan->tuple_exps[n_fields - 1]);
		}
1022

osku's avatar
osku committed
1023 1024
		for (i = 0; i < n_fields; i++) {
			exp = plan->tuple_exps[i];
1025

1026
			dfield_copy_data(dtuple_get_nth_field(plan->tuple, i),
1027
					 que_node_get_val(exp));
osku's avatar
osku committed
1028
		}
1029

osku's avatar
osku committed
1030
		/* Open pcur to the index */
1031

osku's avatar
osku committed
1032
		btr_pcur_open_with_no_init(index, plan->tuple, plan->mode,
1033
					   BTR_SEARCH_LEAF, &plan->pcur,
1034
					   has_search_latch, mtr);
osku's avatar
osku committed
1035 1036 1037 1038
	} else {
		/* Open the cursor to the start or the end of the index
		(FALSE: no init) */

1039
		btr_pcur_open_at_index_side(plan->asc, index, BTR_SEARCH_LEAF,
1040
					    &(plan->pcur), FALSE, mtr);
osku's avatar
osku committed
1041 1042 1043 1044 1045
	}

	ut_ad(plan->n_rows_prefetched == 0);
	ut_ad(plan->n_rows_fetched == 0);
	ut_ad(plan->cursor_at_end == FALSE);
1046

osku's avatar
osku committed
1047 1048 1049 1050
	plan->pcur_is_open = TRUE;
}

/*************************************************************************
1051 1052
Restores a stored pcur position to a table index.
@return	TRUE if the cursor should be moved to the next record after we return from this function (moved to the previous, in the case of a descending cursor) without processing again the current cursor record */
osku's avatar
osku committed
1053 1054 1055 1056
static
ibool
row_sel_restore_pcur_pos(
/*=====================*/
1057 1058
	plan_t*		plan,	/*!< in: table plan */
	mtr_t*		mtr)	/*!< in: mtr */
osku's avatar
osku committed
1059 1060 1061 1062 1063
{
	ibool	equal_position;
	ulint	relative_position;

	ut_ad(!plan->cursor_at_end);
1064

osku's avatar
osku committed
1065 1066
	relative_position = btr_pcur_get_rel_pos(&(plan->pcur));

1067
	equal_position = btr_pcur_restore_position(BTR_SEARCH_LEAF,
1068
						   &(plan->pcur), mtr);
osku's avatar
osku committed
1069 1070

	/* If the cursor is traveling upwards, and relative_position is
1071

osku's avatar
osku committed
1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097
	(1) BTR_PCUR_BEFORE: this is not allowed, as we did not have a lock
	yet on the successor of the page infimum;
	(2) BTR_PCUR_AFTER: btr_pcur_restore_position placed the cursor on the
	first record GREATER than the predecessor of a page supremum; we have
	not yet processed the cursor record: no need to move the cursor to the
	next record;
	(3) BTR_PCUR_ON: btr_pcur_restore_position placed the cursor on the
	last record LESS or EQUAL to the old stored user record; (a) if
	equal_position is FALSE, this means that the cursor is now on a record
	less than the old user record, and we must move to the next record;
	(b) if equal_position is TRUE, then if
	plan->stored_cursor_rec_processed is TRUE, we must move to the next
	record, else there is no need to move the cursor. */

	if (plan->asc) {
		if (relative_position == BTR_PCUR_ON) {

			if (equal_position) {

				return(plan->stored_cursor_rec_processed);
			}

			return(TRUE);
		}

		ut_ad(relative_position == BTR_PCUR_AFTER
1098
		      || relative_position == BTR_PCUR_AFTER_LAST_IN_TREE);
osku's avatar
osku committed
1099 1100 1101 1102 1103

		return(FALSE);
	}

	/* If the cursor is traveling downwards, and relative_position is
1104

osku's avatar
osku committed
1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120
	(1) BTR_PCUR_BEFORE: btr_pcur_restore_position placed the cursor on
	the last record LESS than the successor of a page infimum; we have not
	processed the cursor record: no need to move the cursor;
	(2) BTR_PCUR_AFTER: btr_pcur_restore_position placed the cursor on the
	first record GREATER than the predecessor of a page supremum; we have
	processed the cursor record: we should move the cursor to the previous
	record;
	(3) BTR_PCUR_ON: btr_pcur_restore_position placed the cursor on the
	last record LESS or EQUAL to the old stored user record; (a) if
	equal_position is FALSE, this means that the cursor is now on a record
	less than the old user record, and we need not move to the previous
	record; (b) if equal_position is TRUE, then if
	plan->stored_cursor_rec_processed is TRUE, we must move to the previous
	record, else there is no need to move the cursor. */

	if (relative_position == BTR_PCUR_BEFORE
1121
	    || relative_position == BTR_PCUR_BEFORE_FIRST_IN_TREE) {
osku's avatar
osku committed
1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136

		return(FALSE);
	}

	if (relative_position == BTR_PCUR_ON) {

		if (equal_position) {

			return(plan->stored_cursor_rec_processed);
		}

		return(FALSE);
	}

	ut_ad(relative_position == BTR_PCUR_AFTER
1137
	      || relative_position == BTR_PCUR_AFTER_LAST_IN_TREE);
osku's avatar
osku committed
1138 1139 1140 1141 1142 1143 1144 1145 1146 1147

	return(TRUE);
}

/*************************************************************************
Resets a plan cursor to a closed state. */
UNIV_INLINE
void
plan_reset_cursor(
/*==============*/
1148
	plan_t*	plan)	/*!< in: plan */
1149
{
osku's avatar
osku committed
1150
	plan->pcur_is_open = FALSE;
1151
	plan->cursor_at_end = FALSE;
osku's avatar
osku committed
1152 1153 1154
	plan->n_rows_fetched = 0;
	plan->n_rows_prefetched = 0;
}
1155

osku's avatar
osku committed
1156 1157
/*************************************************************************
Tries to do a shortcut to fetch a clustered index record with a unique key,
1158 1159
using the hash index if possible (not always).
@return	SEL_FOUND, SEL_EXHAUSTED, SEL_RETRY */
osku's avatar
osku committed
1160 1161 1162 1163
static
ulint
row_sel_try_search_shortcut(
/*========================*/
1164 1165
	sel_node_t*	node,	/*!< in: select node for a consistent read */
	plan_t*		plan,	/*!< in: plan for a unique search in clustered
osku's avatar
osku committed
1166
				index */
1167
	mtr_t*		mtr)	/*!< in: mtr */
osku's avatar
osku committed
1168 1169 1170 1171 1172 1173 1174
{
	dict_index_t*	index;
	rec_t*		rec;
	mem_heap_t*	heap		= NULL;
	ulint		offsets_[REC_OFFS_NORMAL_SIZE];
	ulint*		offsets		= offsets_;
	ulint		ret;
1175
	rec_offs_init(offsets_);
osku's avatar
osku committed
1176 1177 1178 1179 1180 1181 1182 1183 1184

	index = plan->index;

	ut_ad(node->read_view);
	ut_ad(plan->unique_search);
	ut_ad(!plan->must_get_clust);
#ifdef UNIV_SYNC_DEBUG
	ut_ad(rw_lock_own(&btr_search_latch, RW_LOCK_SHARED));
#endif /* UNIV_SYNC_DEBUG */
1185

1186
	row_sel_open_pcur(plan, TRUE, mtr);
osku's avatar
osku committed
1187 1188

	rec = btr_pcur_get_rec(&(plan->pcur));
1189

osku's avatar
osku committed
1190 1191 1192 1193 1194 1195
	if (!page_rec_is_user_rec(rec)) {

		return(SEL_RETRY);
	}

	ut_ad(plan->mode == PAGE_CUR_GE);
1196

osku's avatar
osku committed
1197 1198
	/* As the cursor is now placed on a user record after a search with
	the mode PAGE_CUR_GE, the up_match field in the cursor tells how many
1199
	fields in the user record matched to the search tuple */
osku's avatar
osku committed
1200 1201 1202 1203 1204 1205 1206 1207

	if (btr_pcur_get_up_match(&(plan->pcur)) < plan->n_exact_match) {

		return(SEL_EXHAUSTED);
	}

	/* This is a non-locking consistent read: if necessary, fetch
	a previous version of the record */
1208

osku's avatar
osku committed
1209 1210
	offsets = rec_get_offsets(rec, index, offsets, ULINT_UNDEFINED, &heap);

1211
	if (dict_index_is_clust(index)) {
osku's avatar
osku committed
1212
		if (!lock_clust_rec_cons_read_sees(rec, index, offsets,
1213
						   node->read_view)) {
osku's avatar
osku committed
1214 1215 1216
			ret = SEL_RETRY;
			goto func_exit;
		}
1217
	} else if (!lock_sec_rec_cons_read_sees(rec, node->read_view)) {
osku's avatar
osku committed
1218 1219 1220 1221 1222

		ret = SEL_RETRY;
		goto func_exit;
	}

1223
	/* Test the deleted flag. */
osku's avatar
osku committed
1224

1225
	if (rec_get_deleted_flag(rec, dict_table_is_comp(plan->table))) {
osku's avatar
osku committed
1226 1227 1228 1229 1230

		ret = SEL_EXHAUSTED;
		goto func_exit;
	}

1231 1232 1233 1234 1235 1236 1237 1238
	/* Fetch the columns needed in test conditions.  The index
	record is protected by a page latch that was acquired when
	plan->pcur was positioned.  The latch will not be released
	until mtr_commit(mtr). */

	row_sel_fetch_columns(index, rec, offsets,
			      UT_LIST_GET_FIRST(plan->columns));

osku's avatar
osku committed
1239
	/* Test the rest of search conditions */
1240

osku's avatar
osku committed
1241 1242 1243 1244 1245 1246
	if (!row_sel_test_other_conds(plan)) {

		ret = SEL_EXHAUSTED;
		goto func_exit;
	}

1247
	ut_ad(plan->pcur.latch_mode == BTR_SEARCH_LEAF);
osku's avatar
osku committed
1248 1249

	plan->n_rows_fetched++;
1250
	ret = SEL_FOUND;
osku's avatar
osku committed
1251 1252 1253 1254
func_exit:
	if (UNIV_LIKELY_NULL(heap)) {
		mem_heap_free(heap);
	}
1255
	return(ret);
osku's avatar
osku committed
1256 1257 1258
}

/*************************************************************************
1259 1260
Performs a select step.
@return	DB_SUCCESS or error code */
osku's avatar
osku committed
1261 1262 1263 1264
static
ulint
row_sel(
/*====*/
1265 1266
	sel_node_t*	node,	/*!< in: select node */
	que_thr_t*	thr)	/*!< in: query thread */
osku's avatar
osku committed
1267 1268 1269 1270 1271 1272 1273 1274 1275 1276
{
	dict_index_t*	index;
	plan_t*		plan;
	mtr_t		mtr;
	ibool		moved;
	rec_t*		rec;
	rec_t*		old_vers;
	rec_t*		clust_rec;
	ibool		search_latch_locked;
	ibool		consistent_read;
1277

1278 1279 1280 1281
	/* The following flag becomes TRUE when we are doing a
	consistent read from a non-clustered index and we must look
	at the clustered index to find out the previous delete mark
	state of the non-clustered record: */
osku's avatar
osku committed
1282 1283 1284 1285 1286

	ibool		cons_read_requires_clust_rec	= FALSE;
	ulint		cost_counter			= 0;
	ibool		cursor_just_opened;
	ibool		must_go_to_next;
1287
	ibool		mtr_has_extra_clust_latch	= FALSE;
1288 1289 1290 1291 1292 1293
	/* TRUE if the search was made using
	a non-clustered index, and we had to
	access the clustered record: now &mtr
	contains a clustered index latch, and
	&mtr must be committed before we move
	to the next non-clustered record */
osku's avatar
osku committed
1294 1295 1296 1297 1298
	ulint		found_flag;
	ulint		err;
	mem_heap_t*	heap				= NULL;
	ulint		offsets_[REC_OFFS_NORMAL_SIZE];
	ulint*		offsets				= offsets_;
1299
	rec_offs_init(offsets_);
osku's avatar
osku committed
1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319

	ut_ad(thr->run_node == node);

	search_latch_locked = FALSE;

	if (node->read_view) {
		/* In consistent reads, we try to do with the hash index and
		not to use the buffer page get. This is to reduce memory bus
		load resulting from semaphore operations. The search latch
		will be s-locked when we access an index with a unique search
		condition, but not locked when we access an index with a
		less selective search condition. */

		consistent_read = TRUE;
	} else {
		consistent_read = FALSE;
	}

table_loop:
	/* TABLE LOOP
1320
	----------
osku's avatar
osku committed
1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346
	This is the outer major loop in calculating a join. We come here when
	node->fetch_table changes, and after adding a row to aggregate totals
	and, of course, when this function is called. */

	ut_ad(mtr_has_extra_clust_latch == FALSE);

	plan = sel_node_get_nth_plan(node, node->fetch_table);
	index = plan->index;

	if (plan->n_rows_prefetched > 0) {
		sel_pop_prefetched_row(plan);

		goto next_table_no_mtr;
	}

	if (plan->cursor_at_end) {
		/* The cursor has already reached the result set end: no more
		rows to process for this table cursor, as also the prefetch
		stack was empty */

		ut_ad(plan->pcur_is_open);

		goto table_exhausted_no_mtr;
	}

	/* Open a cursor to index, or restore an open cursor position */
1347

osku's avatar
osku committed
1348 1349 1350
	mtr_start(&mtr);

	if (consistent_read && plan->unique_search && !plan->pcur_is_open
1351
	    && !plan->must_get_clust
1352
	    && !plan->table->big_rows) {
osku's avatar
osku committed
1353 1354 1355 1356
		if (!search_latch_locked) {
			rw_lock_s_lock(&btr_search_latch);

			search_latch_locked = TRUE;
1357
		} else if (rw_lock_get_writer(&btr_search_latch) == RW_LOCK_WAIT_EX) {
osku's avatar
osku committed
1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379

			/* There is an x-latch request waiting: release the
			s-latch for a moment; as an s-latch here is often
			kept for some 10 searches before being released,
			a waiting x-latch request would block other threads
			from acquiring an s-latch for a long time, lowering
			performance significantly in multiprocessors. */

			rw_lock_s_unlock(&btr_search_latch);
			rw_lock_s_lock(&btr_search_latch);
		}

		found_flag = row_sel_try_search_shortcut(node, plan, &mtr);

		if (found_flag == SEL_FOUND) {

			goto next_table;

		} else if (found_flag == SEL_EXHAUSTED) {

			goto table_exhausted;
		}
1380

osku's avatar
osku committed
1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398
		ut_ad(found_flag == SEL_RETRY);

		plan_reset_cursor(plan);

		mtr_commit(&mtr);
		mtr_start(&mtr);
	}

	if (search_latch_locked) {
		rw_lock_s_unlock(&btr_search_latch);

		search_latch_locked = FALSE;
	}

	if (!plan->pcur_is_open) {
		/* Evaluate the expressions to build the search tuple and
		open the cursor */

1399
		row_sel_open_pcur(plan, search_latch_locked, &mtr);
osku's avatar
osku committed
1400 1401 1402 1403 1404 1405 1406 1407

		cursor_just_opened = TRUE;

		/* A new search was made: increment the cost counter */
		cost_counter++;
	} else {
		/* Restore pcur position to the index */

1408
		must_go_to_next = row_sel_restore_pcur_pos(plan, &mtr);
osku's avatar
osku committed
1409 1410 1411 1412 1413 1414

		cursor_just_opened = FALSE;

		if (must_go_to_next) {
			/* We have already processed the cursor record: move
			to the next */
1415

osku's avatar
osku committed
1416 1417 1418
			goto next_rec;
		}
	}
1419

osku's avatar
osku committed
1420 1421
rec_loop:
	/* RECORD LOOP
1422
	-----------
osku's avatar
osku committed
1423 1424 1425 1426 1427 1428 1429 1430 1431
	In this loop we use pcur and try to fetch a qualifying row, and
	also fill the prefetch buffer for this table if n_rows_fetched has
	exceeded a threshold. While we are inside this loop, the following
	holds:
	(1) &mtr is started,
	(2) pcur is positioned and open.

	NOTE that if cursor_just_opened is TRUE here, it means that we came
	to this point right after row_sel_open_pcur. */
1432

osku's avatar
osku committed
1433 1434 1435
	ut_ad(mtr_has_extra_clust_latch == FALSE);

	rec = btr_pcur_get_rec(&(plan->pcur));
1436

osku's avatar
osku committed
1437 1438 1439
	/* PHASE 1: Set a lock if specified */

	if (!node->asc && cursor_just_opened
1440
	    && !page_rec_is_supremum(rec)) {
osku's avatar
osku committed
1441 1442 1443 1444 1445 1446

		/* When we open a cursor for a descending search, we must set
		a next-key lock on the successor record: otherwise it would
		be possible to insert new records next to the cursor position,
		and it might be that these new records should appear in the
		search result set, resulting in the phantom problem. */
1447

osku's avatar
osku committed
1448 1449
		if (!consistent_read) {

1450 1451 1452 1453
			/* If innodb_locks_unsafe_for_binlog option is used
			or this session is using READ COMMITTED isolation
			level, we lock only the record, i.e., next-key
			locking is not used. */
osku's avatar
osku committed
1454 1455 1456

			rec_t*	next_rec = page_rec_get_next(rec);
			ulint	lock_type;
1457 1458 1459 1460
			trx_t*	trx;

			trx = thr_get_trx(thr);

osku's avatar
osku committed
1461
			offsets = rec_get_offsets(next_rec, index, offsets,
1462
						  ULINT_UNDEFINED, &heap);
osku's avatar
osku committed
1463

1464
			if (srv_locks_unsafe_for_binlog
1465 1466
			    || trx->isolation_level
			    == TRX_ISO_READ_COMMITTED) {
1467 1468 1469 1470 1471 1472

				if (page_rec_is_supremum(next_rec)) {

					goto skip_lock;
				}

osku's avatar
osku committed
1473 1474 1475 1476 1477
				lock_type = LOCK_REC_NOT_GAP;
			} else {
				lock_type = LOCK_ORDINARY;
			}

1478 1479
			err = sel_set_rec_lock(btr_pcur_get_block(&plan->pcur),
					       next_rec, index, offsets,
1480 1481
					       node->row_lock_mode,
					       lock_type, thr);
osku's avatar
osku committed
1482 1483 1484 1485 1486

			if (err != DB_SUCCESS) {
				/* Note that in this case we will store in pcur
				the PREDECESSOR of the record we are waiting
				the lock for */
1487

osku's avatar
osku committed
1488 1489 1490 1491 1492
				goto lock_wait_or_error;
			}
		}
	}

1493
skip_lock:
osku's avatar
osku committed
1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506
	if (page_rec_is_infimum(rec)) {

		/* The infimum record on a page cannot be in the result set,
		and neither can a record lock be placed on it: we skip such
		a record. We also increment the cost counter as we may have
		processed yet another page of index. */

		cost_counter++;

		goto next_rec;
	}

	if (!consistent_read) {
1507
		/* Try to place a lock on the index record */
osku's avatar
osku committed
1508

1509 1510
		/* If innodb_locks_unsafe_for_binlog option is used
		or this session is using READ COMMITTED isolation level,
osku's avatar
osku committed
1511 1512 1513 1514
		we lock only the record, i.e., next-key locking is
		not used. */

		ulint	lock_type;
1515 1516
		trx_t*	trx;

osku's avatar
osku committed
1517
		offsets = rec_get_offsets(rec, index, offsets,
1518
					  ULINT_UNDEFINED, &heap);
osku's avatar
osku committed
1519

1520 1521 1522
		trx = thr_get_trx(thr);

		if (srv_locks_unsafe_for_binlog
1523
		    || trx->isolation_level == TRX_ISO_READ_COMMITTED) {
1524 1525 1526 1527 1528 1529

			if (page_rec_is_supremum(rec)) {

				goto next_rec;
			}

osku's avatar
osku committed
1530 1531 1532 1533 1534
			lock_type = LOCK_REC_NOT_GAP;
		} else {
			lock_type = LOCK_ORDINARY;
		}

1535 1536
		err = sel_set_rec_lock(btr_pcur_get_block(&plan->pcur),
				       rec, index, offsets,
1537
				       node->row_lock_mode, lock_type, thr);
osku's avatar
osku committed
1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555

		if (err != DB_SUCCESS) {

			goto lock_wait_or_error;
		}
	}

	if (page_rec_is_supremum(rec)) {

		/* A page supremum record cannot be in the result set: skip
		it now when we have placed a possible lock on it */

		goto next_rec;
	}

	ut_ad(page_rec_is_user_rec(rec));

	if (cost_counter > SEL_COST_LIMIT) {
1556

osku's avatar
osku committed
1557 1558 1559 1560 1561 1562 1563 1564 1565
		/* Now that we have placed the necessary locks, we can stop
		for a while and store the cursor position; NOTE that if we
		would store the cursor position BEFORE placing a record lock,
		it might happen that the cursor would jump over some records
		that another transaction could meanwhile insert adjacent to
		the cursor: this would result in the phantom problem. */

		goto stop_for_a_while;
	}
1566

osku's avatar
osku committed
1567 1568 1569 1570 1571
	/* PHASE 2: Check a mixed index mix id if needed */

	if (plan->unique_search && cursor_just_opened) {

		ut_ad(plan->mode == PAGE_CUR_GE);
1572

osku's avatar
osku committed
1573 1574 1575
		/* As the cursor is now placed on a user record after a search
		with the mode PAGE_CUR_GE, the up_match field in the cursor
		tells how many fields in the user record matched to the search
1576
		tuple */
osku's avatar
osku committed
1577 1578

		if (btr_pcur_get_up_match(&(plan->pcur))
1579
		    < plan->n_exact_match) {
osku's avatar
osku committed
1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598
			goto table_exhausted;
		}

		/* Ok, no need to test end_conds or mix id */

	}

	/* We are ready to look at a possible new index entry in the result
	set: the cursor is now placed on a user record */

	/* PHASE 3: Get previous version in a consistent read */

	cons_read_requires_clust_rec = FALSE;
	offsets = rec_get_offsets(rec, index, offsets, ULINT_UNDEFINED, &heap);

	if (consistent_read) {
		/* This is a non-locking consistent read: if necessary, fetch
		a previous version of the record */

1599
		if (dict_index_is_clust(index)) {
1600

osku's avatar
osku committed
1601
			if (!lock_clust_rec_cons_read_sees(rec, index, offsets,
1602
							   node->read_view)) {
osku's avatar
osku committed
1603

1604 1605 1606 1607 1608
				err = row_sel_build_prev_vers(
					node->read_view, index, rec,
					&offsets, &heap, &plan->old_vers_heap,
					&old_vers, &mtr);

osku's avatar
osku committed
1609 1610 1611 1612 1613 1614
				if (err != DB_SUCCESS) {

					goto lock_wait_or_error;
				}

				if (old_vers == NULL) {
1615 1616 1617
					offsets = rec_get_offsets(
						rec, index, offsets,
						ULINT_UNDEFINED, &heap);
1618 1619 1620 1621 1622 1623 1624 1625 1626 1627

					/* Fetch the columns needed in
					test conditions. The clustered
					index record is protected by a
					page latch that was acquired
					by row_sel_open_pcur() or
					row_sel_restore_pcur_pos().
					The latch will not be released
					until mtr_commit(mtr). */

1628 1629 1630 1631
					row_sel_fetch_columns(
						index, rec, offsets,
						UT_LIST_GET_FIRST(
							plan->columns));
osku's avatar
osku committed
1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642

					if (!row_sel_test_end_conds(plan)) {

						goto table_exhausted;
					}

					goto next_rec;
				}

				rec = old_vers;
			}
1643
		} else if (!lock_sec_rec_cons_read_sees(rec,
osku's avatar
osku committed
1644 1645 1646 1647 1648 1649 1650
							node->read_view)) {
			cons_read_requires_clust_rec = TRUE;
		}
	}

	/* PHASE 4: Test search end conditions and deleted flag */

1651 1652 1653 1654
	/* Fetch the columns needed in test conditions.  The record is
	protected by a page latch that was acquired by
	row_sel_open_pcur() or row_sel_restore_pcur_pos().  The latch
	will not be released until mtr_commit(mtr). */
1655

osku's avatar
osku committed
1656
	row_sel_fetch_columns(index, rec, offsets,
1657
			      UT_LIST_GET_FIRST(plan->columns));
osku's avatar
osku committed
1658 1659 1660 1661 1662 1663 1664 1665 1666 1667 1668 1669 1670 1671

	/* Test the selection end conditions: these can only contain columns
	which already are found in the index, even though the index might be
	non-clustered */

	if (plan->unique_search && cursor_just_opened) {

		/* No test necessary: the test was already made above */

	} else if (!row_sel_test_end_conds(plan)) {

		goto table_exhausted;
	}

1672
	if (rec_get_deleted_flag(rec, dict_table_is_comp(plan->table))
1673
	    && !cons_read_requires_clust_rec) {
osku's avatar
osku committed
1674 1675 1676 1677 1678 1679

		/* The record is delete marked: we can skip it if this is
		not a consistent read which might see an earlier version
		of a non-clustered index record */

		if (plan->unique_search) {
1680

osku's avatar
osku committed
1681 1682
			goto table_exhausted;
		}
1683

osku's avatar
osku committed
1684 1685 1686 1687 1688 1689 1690 1691 1692 1693 1694 1695
		goto next_rec;
	}

	/* PHASE 5: Get the clustered index record, if needed and if we did
	not do the search using the clustered index */

	if (plan->must_get_clust || cons_read_requires_clust_rec) {

		/* It was a non-clustered index and we must fetch also the
		clustered index record */

		err = row_sel_get_clust_rec(node, plan, rec, thr, &clust_rec,
1696
					    &mtr);
osku's avatar
osku committed
1697
		mtr_has_extra_clust_latch = TRUE;
1698

osku's avatar
osku committed
1699 1700 1701 1702 1703 1704 1705 1706 1707 1708 1709 1710 1711 1712 1713 1714 1715
		if (err != DB_SUCCESS) {

			goto lock_wait_or_error;
		}

		/* Retrieving the clustered record required a search:
		increment the cost counter */

		cost_counter++;

		if (clust_rec == NULL) {
			/* The record did not exist in the read view */
			ut_ad(consistent_read);

			goto next_rec;
		}

1716
		if (rec_get_deleted_flag(clust_rec,
1717
					 dict_table_is_comp(plan->table))) {
osku's avatar
osku committed
1718 1719 1720 1721 1722 1723 1724 1725 1726 1727

			/* The record is delete marked: we can skip it */

			goto next_rec;
		}

		if (node->can_get_updated) {

			btr_pcur_store_position(&(plan->clust_pcur), &mtr);
		}
1728
	}
osku's avatar
osku committed
1729 1730

	/* PHASE 6: Test the rest of search conditions */
1731

osku's avatar
osku committed
1732 1733 1734
	if (!row_sel_test_other_conds(plan)) {

		if (plan->unique_search) {
1735

osku's avatar
osku committed
1736 1737 1738 1739 1740 1741 1742 1743
			goto table_exhausted;
		}

		goto next_rec;
	}

	/* PHASE 7: We found a new qualifying row for the current table; push
	the row if prefetch is on, or move to the next table in the join */
1744

osku's avatar
osku committed
1745 1746
	plan->n_rows_fetched++;

1747
	ut_ad(plan->pcur.latch_mode == BTR_SEARCH_LEAF);
osku's avatar
osku committed
1748 1749

	if ((plan->n_rows_fetched <= SEL_PREFETCH_LIMIT)
1750
	    || plan->unique_search || plan->no_prefetch
1751
	    || plan->table->big_rows) {
osku's avatar
osku committed
1752 1753

		/* No prefetch in operation: go to the next table */
1754

osku's avatar
osku committed
1755 1756 1757 1758 1759 1760 1761 1762
		goto next_table;
	}

	sel_push_prefetched_row(plan);

	if (plan->n_rows_prefetched == SEL_MAX_N_PREFETCH) {

		/* The prefetch buffer is now full */
1763

osku's avatar
osku committed
1764 1765 1766 1767 1768
		sel_pop_prefetched_row(plan);

		goto next_table;
	}

1769
next_rec:
osku's avatar
osku committed
1770 1771 1772 1773 1774 1775 1776 1777 1778 1779 1780
	ut_ad(!search_latch_locked);

	if (mtr_has_extra_clust_latch) {

		/* We must commit &mtr if we are moving to the next
		non-clustered index record, because we could break the
		latching order if we would access a different clustered
		index page right away without releasing the previous. */

		goto commit_mtr_for_a_while;
	}
1781

osku's avatar
osku committed
1782 1783 1784 1785 1786 1787 1788
	if (node->asc) {
		moved = btr_pcur_move_to_next(&(plan->pcur), &mtr);
	} else {
		moved = btr_pcur_move_to_prev(&(plan->pcur), &mtr);
	}

	if (!moved) {
1789

osku's avatar
osku committed
1790 1791 1792 1793 1794 1795
		goto table_exhausted;
	}

	cursor_just_opened = FALSE;

	/* END OF RECORD LOOP
1796
	------------------ */
osku's avatar
osku committed
1797 1798 1799 1800 1801 1802
	goto rec_loop;

next_table:
	/* We found a record which satisfies the conditions: we can move to
	the next table or return a row in the result set */

1803
	ut_ad(btr_pcur_is_on_user_rec(&plan->pcur));
1804

osku's avatar
osku committed
1805 1806 1807 1808 1809 1810 1811 1812 1813 1814 1815 1816 1817 1818 1819 1820 1821 1822
	if (plan->unique_search && !node->can_get_updated) {

		plan->cursor_at_end = TRUE;
	} else {
		ut_ad(!search_latch_locked);

		plan->stored_cursor_rec_processed = TRUE;

		btr_pcur_store_position(&(plan->pcur), &mtr);
	}

	mtr_commit(&mtr);

	mtr_has_extra_clust_latch = FALSE;

next_table_no_mtr:
	/* If we use 'goto' to this label, it means that the row was popped
	from the prefetched rows stack, and &mtr is already committed */
1823

osku's avatar
osku committed
1824 1825 1826 1827 1828 1829
	if (node->fetch_table + 1 == node->n_tables) {

		sel_eval_select_list(node);

		if (node->is_aggregate) {

1830
			goto table_loop;
osku's avatar
osku committed
1831 1832 1833
		}

		sel_assign_into_var_values(node->into_list, node);
1834

osku's avatar
osku committed
1835 1836 1837 1838 1839 1840 1841 1842 1843 1844
		thr->run_node = que_node_get_parent(node);

		err = DB_SUCCESS;
		goto func_exit;
	}

	node->fetch_table++;

	/* When we move to the next table, we first reset the plan cursor:
	we do not care about resetting it when we backtrack from a table */
1845

osku's avatar
osku committed
1846
	plan_reset_cursor(sel_node_get_nth_plan(node, node->fetch_table));
1847

osku's avatar
osku committed
1848 1849 1850 1851
	goto table_loop;

table_exhausted:
	/* The table cursor pcur reached the result set end: backtrack to the
1852
	previous table in the join if we do not have cached prefetched rows */
osku's avatar
osku committed
1853 1854 1855 1856 1857 1858

	plan->cursor_at_end = TRUE;

	mtr_commit(&mtr);

	mtr_has_extra_clust_latch = FALSE;
1859

osku's avatar
osku committed
1860 1861
	if (plan->n_rows_prefetched > 0) {
		/* The table became exhausted during a prefetch */
1862

osku's avatar
osku committed
1863 1864 1865 1866 1867 1868 1869 1870 1871 1872 1873 1874 1875 1876 1877 1878
		sel_pop_prefetched_row(plan);

		goto next_table_no_mtr;
	}

table_exhausted_no_mtr:
	if (node->fetch_table == 0) {
		err = DB_SUCCESS;

		if (node->is_aggregate && !node->aggregate_already_fetched) {

			node->aggregate_already_fetched = TRUE;

			sel_assign_into_var_values(node->into_list, node);

			thr->run_node = que_node_get_parent(node);
1879 1880
		} else {
			node->state = SEL_NODE_NO_MORE_ROWS;
osku's avatar
osku committed
1881

1882
			thr->run_node = que_node_get_parent(node);
osku's avatar
osku committed
1883
		}
1884

osku's avatar
osku committed
1885 1886 1887 1888 1889 1890 1891 1892 1893 1894 1895 1896 1897 1898
		goto func_exit;
	}

	node->fetch_table--;

	goto table_loop;

stop_for_a_while:
	/* Return control for a while to que_run_threads, so that runaway
	queries can be canceled. NOTE that when we come here, we must, in a
	locking read, have placed the necessary (possibly waiting request)
	record lock on the cursor record or its successor: when we reposition
	the cursor, this record lock guarantees that nobody can meanwhile have
	inserted new records which should have appeared in the result set,
1899
	which would result in the phantom problem. */
osku's avatar
osku committed
1900 1901 1902 1903 1904 1905 1906

	ut_ad(!search_latch_locked);

	plan->stored_cursor_rec_processed = FALSE;
	btr_pcur_store_position(&(plan->pcur), &mtr);

	mtr_commit(&mtr);
1907

1908
#ifdef UNIV_SYNC_DEBUG
osku's avatar
osku committed
1909
	ut_ad(sync_thread_levels_empty_gen(TRUE));
1910
#endif /* UNIV_SYNC_DEBUG */
osku's avatar
osku committed
1911 1912 1913 1914 1915 1916
	err = DB_SUCCESS;
	goto func_exit;

commit_mtr_for_a_while:
	/* Stores the cursor position and commits &mtr; this is used if
	&mtr may contain latches which would break the latching order if
1917
	&mtr would not be committed and the latches released. */
osku's avatar
osku committed
1918 1919 1920 1921 1922 1923 1924 1925 1926

	plan->stored_cursor_rec_processed = TRUE;

	ut_ad(!search_latch_locked);
	btr_pcur_store_position(&(plan->pcur), &mtr);

	mtr_commit(&mtr);

	mtr_has_extra_clust_latch = FALSE;
1927

1928
#ifdef UNIV_SYNC_DEBUG
osku's avatar
osku committed
1929
	ut_ad(sync_thread_levels_empty_gen(TRUE));
1930
#endif /* UNIV_SYNC_DEBUG */
osku's avatar
osku committed
1931 1932 1933 1934 1935 1936

	goto table_loop;

lock_wait_or_error:
	/* See the note at stop_for_a_while: the same holds for this case */

1937
	ut_ad(!btr_pcur_is_before_first_on_page(&plan->pcur) || !node->asc);
osku's avatar
osku committed
1938 1939 1940 1941
	ut_ad(!search_latch_locked);

	plan->stored_cursor_rec_processed = FALSE;
	btr_pcur_store_position(&(plan->pcur), &mtr);
1942

osku's avatar
osku committed
1943
	mtr_commit(&mtr);
1944

1945
#ifdef UNIV_SYNC_DEBUG
osku's avatar
osku committed
1946
	ut_ad(sync_thread_levels_empty_gen(TRUE));
1947
#endif /* UNIV_SYNC_DEBUG */
osku's avatar
osku committed
1948 1949

func_exit:
1950 1951 1952
	if (search_latch_locked) {
		rw_lock_s_unlock(&btr_search_latch);
	}
osku's avatar
osku committed
1953 1954 1955 1956 1957 1958 1959 1960
	if (UNIV_LIKELY_NULL(heap)) {
		mem_heap_free(heap);
	}
	return(err);
}

/**************************************************************************
Performs a select step. This is a high-level function used in SQL execution
1961 1962
graphs.
@return	query thread to run next or NULL */
1963
UNIV_INTERN
osku's avatar
osku committed
1964 1965 1966
que_thr_t*
row_sel_step(
/*=========*/
1967
	que_thr_t*	thr)	/*!< in: query thread */
osku's avatar
osku committed
1968 1969 1970 1971 1972 1973 1974
{
	ulint		i_lock_mode;
	sym_node_t*	table_node;
	sel_node_t*	node;
	ulint		err;

	ut_ad(thr);
1975

osku's avatar
osku committed
1976 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 1996 1997 1998 1999
	node = thr->run_node;

	ut_ad(que_node_get_type(node) == QUE_NODE_SELECT);

	/* If this is a new time this node is executed (or when execution
	resumes after wait for a table intention lock), set intention locks
	on the tables, or assign a read view */

	if (node->into_list && (thr->prev_node == que_node_get_parent(node))) {

		node->state = SEL_NODE_OPEN;
	}

	if (node->state == SEL_NODE_OPEN) {

		/* It may be that the current session has not yet started
		its transaction, or it has been committed: */

		trx_start_if_not_started(thr_get_trx(thr));

		plan_reset_cursor(sel_node_get_nth_plan(node, 0));

		if (node->consistent_read) {
			/* Assign a read view for the query */
2000 2001
			node->read_view = trx_assign_read_view(
				thr_get_trx(thr));
osku's avatar
osku committed
2002 2003 2004 2005 2006 2007
		} else {
			if (node->set_x_locks) {
				i_lock_mode = LOCK_IX;
			} else {
				i_lock_mode = LOCK_IS;
			}
2008

osku's avatar
osku committed
2009
			table_node = node->table_list;
2010

osku's avatar
osku committed
2011 2012
			while (table_node) {
				err = lock_table(0, table_node->table,
2013
						 i_lock_mode, thr);
osku's avatar
osku committed
2014
				if (err != DB_SUCCESS) {
2015
					thr_get_trx(thr)->error_state = err;
2016

osku's avatar
osku committed
2017 2018
					return(NULL);
				}
2019

osku's avatar
osku committed
2020 2021 2022
				table_node = que_node_get_next(table_node);
			}
		}
2023

osku's avatar
osku committed
2024 2025 2026 2027 2028
		/* If this is an explicit cursor, copy stored procedure
		variable values, so that the values cannot change between
		fetches (currently, we copy them also for non-explicit
		cursors) */

2029 2030
		if (node->explicit_cursor
		    && UT_LIST_GET_FIRST(node->copy_variables)) {
osku's avatar
osku committed
2031 2032 2033

			row_sel_copy_input_variable_vals(node);
		}
2034

osku's avatar
osku committed
2035 2036 2037 2038 2039 2040 2041 2042 2043 2044 2045 2046 2047 2048
		node->state = SEL_NODE_FETCH;
		node->fetch_table = 0;

		if (node->is_aggregate) {
			/* Reset the aggregate total values */
			sel_reset_aggregate_vals(node);
		}
	}

	err = row_sel(node, thr);

	/* NOTE! if queries are parallelized, the following assignment may
	have problems; the assignment should be made only if thr is the
	only top-level thr in the graph: */
2049

osku's avatar
osku committed
2050 2051
	thr->graph->last_sel_node = node;

2052 2053
	if (err != DB_SUCCESS) {
		thr_get_trx(thr)->error_state = err;
osku's avatar
osku committed
2054 2055 2056 2057 2058

		return(NULL);
	}

	return(thr);
2059
}
osku's avatar
osku committed
2060 2061

/**************************************************************************
2062 2063
Performs a fetch for a cursor.
@return	query thread to run next or NULL */
2064
UNIV_INTERN
osku's avatar
osku committed
2065 2066 2067
que_thr_t*
fetch_step(
/*=======*/
2068
	que_thr_t*	thr)	/*!< in: query thread */
osku's avatar
osku committed
2069 2070 2071 2072 2073
{
	sel_node_t*	sel_node;
	fetch_node_t*	node;

	ut_ad(thr);
2074

osku's avatar
osku committed
2075 2076
	node = thr->run_node;
	sel_node = node->cursor_def;
2077

osku's avatar
osku committed
2078 2079 2080 2081 2082
	ut_ad(que_node_get_type(node) == QUE_NODE_FETCH);

	if (thr->prev_node != que_node_get_parent(node)) {

		if (sel_node->state != SEL_NODE_NO_MORE_ROWS) {
2083

2084 2085
			if (node->into_list) {
				sel_assign_into_var_values(node->into_list,
2086
							   sel_node);
2087
			} else {
2088 2089
				void* ret = (*node->func->func)(
					sel_node, node->func->arg);
2090 2091

				if (!ret) {
2092 2093
					sel_node->state
						= SEL_NODE_NO_MORE_ROWS;
2094 2095
				}
			}
osku's avatar
osku committed
2096 2097 2098 2099 2100 2101 2102 2103 2104 2105 2106
		}

		thr->run_node = que_node_get_parent(node);

		return(thr);
	}

	/* Make the fetch node the parent of the cursor definition for
	the time of the fetch, so that execution knows to return to this
	fetch node after a row has been selected or we know that there is
	no row left */
2107

osku's avatar
osku committed
2108
	sel_node->common.parent = node;
2109

osku's avatar
osku committed
2110
	if (sel_node->state == SEL_NODE_CLOSED) {
2111 2112
		fprintf(stderr,
			"InnoDB: Error: fetch called on a closed cursor\n");
osku's avatar
osku committed
2113

2114
		thr_get_trx(thr)->error_state = DB_ERROR;
osku's avatar
osku committed
2115 2116 2117 2118 2119 2120 2121

		return(NULL);
	}

	thr->run_node = sel_node;

	return(thr);
2122
}
osku's avatar
osku committed
2123

2124
/********************************************************************
2125 2126
Sample callback function for fetch that prints each row.
@return	always returns non-NULL */
2127
UNIV_INTERN
2128 2129 2130
void*
row_fetch_print(
/*============*/
2131 2132
	void*	row,		/*!< in:  sel_node_t* */
	void*	user_arg)	/*!< in:  not used */
2133 2134 2135 2136 2137 2138 2139 2140 2141 2142 2143 2144 2145
{
	sel_node_t*	node = row;
	que_node_t*	exp;
	ulint		i = 0;

	UT_NOT_USED(user_arg);

	fprintf(stderr, "row_fetch_print: row %p\n", row);

	exp = node->select_list;

	while (exp) {
		dfield_t*	dfield = que_node_get_val(exp);
2146
		const dtype_t*	type = dfield_get_type(dfield);
2147 2148 2149 2150

		fprintf(stderr, " column %lu:\n", (ulong)i);

		dtype_print(type);
2151
		putc('\n', stderr);
2152

2153 2154
		if (dfield_get_len(dfield) != UNIV_SQL_NULL) {
			ut_print_buf(stderr, dfield_get_data(dfield),
2155
				     dfield_get_len(dfield));
2156
			putc('\n', stderr);
2157
		} else {
2158
			fputs(" <NULL>;\n", stderr);
2159 2160
		}

2161 2162 2163 2164 2165 2166 2167 2168 2169 2170
		exp = que_node_get_next(exp);
		i++;
	}

	return((void*)42);
}

/********************************************************************
Callback function for fetch that stores an unsigned 4 byte integer to the
location pointed. The column's type must be DATA_INT, DATA_UNSIGNED, length
2171 2172
= 4.
@return	always returns NULL */
2173
UNIV_INTERN
2174 2175 2176
void*
row_fetch_store_uint4(
/*==================*/
2177 2178
	void*	row,		/*!< in:  sel_node_t* */
	void*	user_arg)	/*!< in:  data pointer */
2179 2180 2181 2182 2183 2184
{
	sel_node_t*	node = row;
	ib_uint32_t*	val = user_arg;
	ulint		tmp;

	dfield_t*	dfield = que_node_get_val(node->select_list);
2185
	const dtype_t*	type = dfield_get_type(dfield);
2186 2187 2188 2189 2190 2191 2192
	ulint		len = dfield_get_len(dfield);

	ut_a(dtype_get_mtype(type) == DATA_INT);
	ut_a(dtype_get_prtype(type) & DATA_UNSIGNED);
	ut_a(len == 4);

	tmp = mach_read_from_4(dfield_get_data(dfield));
2193
	*val = (ib_uint32_t) tmp;
2194 2195 2196 2197

	return(NULL);
}

osku's avatar
osku committed
2198
/***************************************************************
2199 2200
Prints a row in a select result.
@return	query thread to run next or NULL */
2201
UNIV_INTERN
osku's avatar
osku committed
2202 2203 2204
que_thr_t*
row_printf_step(
/*============*/
2205
	que_thr_t*	thr)	/*!< in: query thread */
osku's avatar
osku committed
2206 2207 2208 2209 2210 2211
{
	row_printf_node_t*	node;
	sel_node_t*		sel_node;
	que_node_t*		arg;

	ut_ad(thr);
2212

osku's avatar
osku committed
2213
	node = thr->run_node;
2214

osku's avatar
osku committed
2215 2216 2217 2218 2219
	sel_node = node->sel_node;

	ut_ad(que_node_get_type(node) == QUE_NODE_ROW_PRINTF);

	if (thr->prev_node == que_node_get_parent(node)) {
2220

osku's avatar
osku committed
2221 2222 2223 2224 2225 2226
		/* Reset the cursor */
		sel_node->state = SEL_NODE_OPEN;

		/* Fetch next row to print */

		thr->run_node = sel_node;
2227

osku's avatar
osku committed
2228 2229 2230 2231 2232 2233 2234 2235 2236 2237
		return(thr);
	}

	if (sel_node->state != SEL_NODE_FETCH) {

		ut_ad(sel_node->state == SEL_NODE_NO_MORE_ROWS);

		/* No more rows to print */

		thr->run_node = que_node_get_parent(node);
2238

osku's avatar
osku committed
2239 2240 2241 2242 2243 2244 2245 2246 2247 2248 2249 2250 2251 2252 2253 2254 2255 2256 2257 2258
		return(thr);
	}

	arg = sel_node->select_list;

	while (arg) {
		dfield_print_also_hex(que_node_get_val(arg));

		fputs(" ::: ", stderr);

		arg = que_node_get_next(arg);
	}

	putc('\n', stderr);

	/* Fetch next row to print */

	thr->run_node = sel_node;

	return(thr);
2259
}
osku's avatar
osku committed
2260 2261 2262 2263 2264 2265 2266 2267

/********************************************************************
Converts a key value stored in MySQL format to an Innobase dtuple. The last
field of the key value may be just a prefix of a fixed length field: hence
the parameter key_len. But currently we do not allow search keys where the
last field is only a prefix of the full key field len and print a warning if
such appears. A counterpart of this function is
ha_innobase::store_key_val_for_row() in ha_innodb.cc. */
2268
UNIV_INTERN
osku's avatar
osku committed
2269 2270 2271
void
row_sel_convert_mysql_key_to_innobase(
/*==================================*/
2272
	dtuple_t*	tuple,		/*!< in/out: tuple where to build;
osku's avatar
osku committed
2273 2274 2275
					NOTE: we assume that the type info
					in the tuple is already according
					to index! */
2276
	byte*		buf,		/*!< in: buffer to use in field
osku's avatar
osku committed
2277
					conversions */
2278 2279 2280 2281 2282
	ulint		buf_len,	/*!< in: buffer length */
	dict_index_t*	index,		/*!< in: index of the key value */
	const byte*	key_ptr,	/*!< in: MySQL key value */
	ulint		key_len,	/*!< in: MySQL key value length */
	trx_t*		trx)		/*!< in: transaction */
osku's avatar
osku committed
2283 2284
{
	byte*		original_buf	= buf;
2285
	const byte*	original_key_ptr = key_ptr;
osku's avatar
osku committed
2286 2287 2288 2289 2290 2291
	dict_field_t*	field;
	dfield_t*	dfield;
	ulint		data_offset;
	ulint		data_len;
	ulint		data_field_len;
	ibool		is_null;
2292
	const byte*	key_end;
osku's avatar
osku committed
2293
	ulint		n_fields = 0;
2294

osku's avatar
osku committed
2295 2296 2297 2298 2299 2300
	/* For documentation of the key value storage format in MySQL, see
	ha_innobase::store_key_val_for_row() in ha_innodb.cc. */

	key_end = key_ptr + key_len;

	/* Permit us to access any field in the tuple (ULINT_MAX): */
2301

osku's avatar
osku committed
2302 2303
	dtuple_set_n_fields(tuple, ULINT_MAX);

2304
	dfield = dtuple_get_nth_field(tuple, 0);
osku's avatar
osku committed
2305 2306
	field = dict_index_get_nth_field(index, 0);

marko's avatar
marko committed
2307
	if (UNIV_UNLIKELY(dfield_get_type(dfield)->mtype == DATA_SYS)) {
osku's avatar
osku committed
2308 2309 2310 2311 2312 2313 2314 2315 2316
		/* A special case: we are looking for a position in the
		generated clustered index which InnoDB automatically added
		to a table with no primary key: the first and the only
		ordering column is ROW_ID which InnoDB stored to the key_ptr
		buffer. */

		ut_a(key_len == DATA_ROW_ID_LEN);

		dfield_set_data(dfield, key_ptr, DATA_ROW_ID_LEN);
2317

osku's avatar
osku committed
2318 2319 2320 2321 2322 2323 2324
		dtuple_set_n_fields(tuple, 1);

		return;
	}

	while (key_ptr < key_end) {

2325 2326
		ulint	type = dfield_get_type(dfield)->mtype;
		ut_a(field->col->mtype == type);
osku's avatar
osku committed
2327 2328 2329 2330

		data_offset = 0;
		is_null = FALSE;

2331 2332 2333 2334
		if (!(dfield_get_type(dfield)->prtype & DATA_NOT_NULL)) {
			/* The first byte in the field tells if this is
			an SQL NULL value */

osku's avatar
osku committed
2335 2336
			data_offset = 1;

2337
			if (*key_ptr != 0) {
2338
				dfield_set_null(dfield);
osku's avatar
osku committed
2339 2340

				is_null = TRUE;
2341 2342
			}
		}
osku's avatar
osku committed
2343 2344

		/* Calculate data length and data field total length */
2345

osku's avatar
osku committed
2346 2347 2348 2349 2350 2351 2352 2353 2354 2355 2356 2357 2358 2359 2360 2361
		if (type == DATA_BLOB) {
			/* The key field is a column prefix of a BLOB or
			TEXT */

			ut_a(field->prefix_len > 0);

			/* MySQL stores the actual data length to the first 2
			bytes after the optional SQL NULL marker byte. The
			storage format is little-endian, that is, the most
			significant byte at a higher address. In UTF-8, MySQL
			seems to reserve field->prefix_len bytes for
			storing this field in the key value buffer, even
			though the actual value only takes data_len bytes
			from the start. */

			data_len = key_ptr[data_offset]
2362
				+ 256 * key_ptr[data_offset + 1];
osku's avatar
osku committed
2363 2364 2365 2366 2367 2368 2369 2370 2371 2372 2373 2374 2375 2376 2377 2378 2379 2380 2381
			data_field_len = data_offset + 2 + field->prefix_len;

			data_offset += 2;

			/* Now that we know the length, we store the column
			value like it would be a fixed char field */

		} else if (field->prefix_len > 0) {
			/* Looks like MySQL pads unused end bytes in the
			prefix with space. Therefore, also in UTF-8, it is ok
			to compare with a prefix containing full prefix_len
			bytes, and no need to take at most prefix_len / 3
			UTF-8 characters from the start.
			If the prefix is used as the upper end of a LIKE
			'abc%' query, then MySQL pads the end with chars
			0xff. TODO: in that case does it any harm to compare
			with the full prefix_len bytes. How do characters
			0xff in UTF-8 behave? */

2382
			data_len = field->prefix_len;
osku's avatar
osku committed
2383 2384 2385 2386 2387 2388
			data_field_len = data_offset + data_len;
		} else {
			data_len = dfield_get_type(dfield)->len;
			data_field_len = data_offset + data_len;
		}

2389 2390 2391 2392
		if (UNIV_UNLIKELY
		    (dtype_get_mysql_type(dfield_get_type(dfield))
		     == DATA_MYSQL_TRUE_VARCHAR)
		    && UNIV_LIKELY(type != DATA_INT)) {
osku's avatar
osku committed
2393 2394 2395 2396 2397 2398 2399 2400 2401 2402 2403 2404 2405 2406
			/* In a MySQL key value format, a true VARCHAR is
			always preceded by 2 bytes of a length field.
			dfield_get_type(dfield)->len returns the maximum
			'payload' len in bytes. That does not include the
			2 bytes that tell the actual data length.

			We added the check != DATA_INT to make sure we do
			not treat MySQL ENUM or SET as a true VARCHAR! */

			data_len += 2;
			data_field_len += 2;
		}

		/* Storing may use at most data_len bytes of buf */
2407

marko's avatar
marko committed
2408
		if (UNIV_LIKELY(!is_null)) {
2409 2410 2411 2412 2413
			row_mysql_store_col_in_innobase_format(
				dfield, buf,
				FALSE, /* MySQL key value format col */
				key_ptr + data_offset, data_len,
				dict_table_is_comp(index->table));
osku's avatar
osku committed
2414 2415 2416
			buf += data_len;
		}

2417
		key_ptr += data_field_len;
osku's avatar
osku committed
2418

marko's avatar
marko committed
2419
		if (UNIV_UNLIKELY(key_ptr > key_end)) {
osku's avatar
osku committed
2420 2421 2422
			/* The last field in key was not a complete key field
			but a prefix of it.

2423
			Print a warning about this! HA_READ_PREFIX_LAST does
osku's avatar
osku committed
2424 2425 2426 2427 2428
			not currently work in InnoDB with partial-field key
			value prefixes. Since MySQL currently uses a padding
			trick to calculate LIKE 'abc%' type queries there
			should never be partial-field prefixes in searches. */

2429 2430
			ut_print_timestamp(stderr);

2431 2432 2433
			fputs("  InnoDB: Warning: using a partial-field"
			      " key prefix in search.\n"
			      "InnoDB: ", stderr);
osku's avatar
osku committed
2434 2435
			dict_index_name_print(stderr, trx, index);
			fprintf(stderr, ". Last data field length %lu bytes,\n"
2436 2437 2438 2439 2440
				"InnoDB: key ptr now exceeds"
				" key end by %lu bytes.\n"
				"InnoDB: Key value in the MySQL format:\n",
				(ulong) data_field_len,
				(ulong) (key_ptr - key_end));
osku's avatar
osku committed
2441 2442
			fflush(stderr);
			ut_print_buf(stderr, original_key_ptr, key_len);
2443
			putc('\n', stderr);
osku's avatar
osku committed
2444 2445

			if (!is_null) {
2446 2447 2448
				ulint	len = dfield_get_len(dfield);
				dfield_set_len(dfield, len
					       - (ulint) (key_ptr - key_end));
osku's avatar
osku committed
2449 2450 2451
			}
		}

2452 2453
		n_fields++;
		field++;
osku's avatar
osku committed
2454
		dfield++;
2455
	}
osku's avatar
osku committed
2456 2457 2458

	ut_a(buf <= original_buf + buf_len);

2459
	/* We set the length of tuple to n_fields: we assume that the memory
osku's avatar
osku committed
2460
	area allocated for it is big enough (usually bigger than n_fields). */
2461 2462

	dtuple_set_n_fields(tuple, n_fields);
osku's avatar
osku committed
2463 2464 2465 2466 2467 2468 2469 2470
}

/******************************************************************
Stores the row id to the prebuilt struct. */
static
void
row_sel_store_row_id_to_prebuilt(
/*=============================*/
2471 2472 2473 2474
	row_prebuilt_t*		prebuilt,	/*!< in/out: prebuilt */
	const rec_t*		index_rec,	/*!< in: record */
	const dict_index_t*	index,		/*!< in: index of the record */
	const ulint*		offsets)	/*!< in: rec_get_offsets
2475
						(index_rec, index) */
osku's avatar
osku committed
2476
{
2477 2478
	const byte*	data;
	ulint		len;
osku's avatar
osku committed
2479 2480 2481

	ut_ad(rec_offs_validate(index_rec, index, offsets));

2482 2483 2484
	data = rec_get_nth_field(
		index_rec, offsets,
		dict_index_get_sys_col_pos(index, DATA_ROW_ID), &len);
osku's avatar
osku committed
2485

2486
	if (UNIV_UNLIKELY(len != DATA_ROW_ID_LEN)) {
2487
		fprintf(stderr,
2488 2489
			"InnoDB: Error: Row id field is"
			" wrong length %lu in ", (ulong) len);
osku's avatar
osku committed
2490 2491
		dict_index_name_print(stderr, prebuilt->trx, index);
		fprintf(stderr, "\n"
2492
			"InnoDB: Field number %lu, record:\n",
2493
			(ulong) dict_index_get_sys_col_pos(index,
2494
							   DATA_ROW_ID));
osku's avatar
osku committed
2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509
		rec_print_new(stderr, index_rec, offsets);
		putc('\n', stderr);
		ut_error;
	}

	ut_memcpy(prebuilt->row_id, data, len);
}

/******************************************************************
Stores a non-SQL-NULL field in the MySQL format. The counterpart of this
function is row_mysql_store_col_in_innobase_format() in row0mysql.c. */
static
void
row_sel_field_store_in_mysql_format(
/*================================*/
2510
	byte*		dest,	/*!< in/out: buffer where to store; NOTE
2511 2512 2513 2514 2515 2516
				that BLOBs are not in themselves
				stored here: the caller must allocate
				and copy the BLOB into buffer before,
				and pass the pointer to the BLOB in
				'data' */
	const mysql_row_templ_t* templ,
2517
				/*!< in: MySQL column template.
2518 2519 2520
				Its following fields are referenced:
				type, is_unsigned, mysql_col_len,
				mbminlen, mbmaxlen */
2521 2522
	const byte*	data,	/*!< in: data to store */
	ulint		len)	/*!< in: length of the data */
osku's avatar
osku committed
2523 2524 2525 2526 2527 2528 2529
{
	byte*	ptr;
	byte*	field_end;
	byte*	pad_ptr;

	ut_ad(len != UNIV_SQL_NULL);

2530 2531
	switch (templ->type) {
	case DATA_INT:
osku's avatar
osku committed
2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550
		/* Convert integer data from Innobase to a little-endian
		format, sign bit restored to normal */

		ptr = dest + len;

		for (;;) {
			ptr--;
			*ptr = *data;
			if (ptr == dest) {
				break;
			}
			data++;
		}

		if (!templ->is_unsigned) {
			dest[len - 1] = (byte) (dest[len - 1] ^ 128);
		}

		ut_ad(templ->mysql_col_len == len);
2551
		break;
osku's avatar
osku committed
2552

2553 2554 2555
	case DATA_VARCHAR:
	case DATA_VARMYSQL:
	case DATA_BINARY:
osku's avatar
osku committed
2556 2557 2558 2559 2560 2561
		field_end = dest + templ->mysql_col_len;

		if (templ->mysql_type == DATA_MYSQL_TRUE_VARCHAR) {
			/* This is a >= 5.0.3 type true VARCHAR. Store the
			length of the data to the first byte or the first
			two bytes of dest. */
2562

2563 2564
			dest = row_mysql_store_true_var_len(
				dest, len, templ->mysql_length_bytes);
osku's avatar
osku committed
2565 2566 2567 2568
		}

		/* Copy the actual data */
		ut_memcpy(dest, data, len);
2569

osku's avatar
osku committed
2570 2571 2572
		/* Pad with trailing spaces. We pad with spaces also the
		unused end of a >= 5.0.3 true VARCHAR column, just in case
		MySQL expects its contents to be deterministic. */
2573

osku's avatar
osku committed
2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584
		pad_ptr = dest + len;

		ut_ad(templ->mbminlen <= templ->mbmaxlen);

		/* We handle UCS2 charset strings differently. */
		if (templ->mbminlen == 2) {
			/* A space char is two bytes, 0x0020 in UCS2 */

			if (len & 1) {
				/* A 0x20 has been stripped from the column.
				Pad it back. */
2585

osku's avatar
osku committed
2586 2587 2588 2589 2590
				if (pad_ptr < field_end) {
					*pad_ptr = 0x20;
					pad_ptr++;
				}
			}
2591

osku's avatar
osku committed
2592 2593 2594 2595 2596 2597 2598 2599 2600 2601 2602 2603 2604 2605
			/* Pad the rest of the string with 0x0020 */

			while (pad_ptr < field_end) {
				*pad_ptr = 0x00;
				pad_ptr++;
				*pad_ptr = 0x20;
				pad_ptr++;
			}
		} else {
			ut_ad(templ->mbminlen == 1);
			/* space=0x20 */

			memset(pad_ptr, 0x20, field_end - pad_ptr);
		}
2606 2607 2608
		break;

	case DATA_BLOB:
osku's avatar
osku committed
2609 2610 2611 2612
		/* Store a pointer to the BLOB buffer to dest: the BLOB was
		already copied to the buffer in row_sel_store_mysql_rec */

		row_mysql_store_blob_ref(dest, templ->mysql_col_len, data,
2613
					 len);
2614 2615 2616
		break;

	case DATA_MYSQL:
osku's avatar
osku committed
2617 2618
		memcpy(dest, data, len);

2619 2620
		ut_ad(templ->mysql_col_len >= len);
		ut_ad(templ->mbmaxlen >= templ->mbminlen);
osku's avatar
osku committed
2621

2622
		ut_ad(templ->mbmaxlen > templ->mbminlen
2623
		      || templ->mysql_col_len == len);
osku's avatar
osku committed
2624 2625 2626
		/* The following assertion would fail for old tables
		containing UTF-8 ENUM columns due to Bug #9526. */
		ut_ad(!templ->mbmaxlen
2627
		      || !(templ->mysql_col_len % templ->mbmaxlen));
2628
		ut_ad(len * templ->mbmaxlen >= templ->mysql_col_len);
osku's avatar
osku committed
2629 2630 2631 2632 2633 2634 2635 2636

		if (templ->mbminlen != templ->mbmaxlen) {
			/* Pad with spaces. This undoes the stripping
			done in row0mysql.ic, function
			row_mysql_store_col_in_innobase_format(). */

			memset(dest + len, 0x20, templ->mysql_col_len - len);
		}
2637
		break;
osku's avatar
osku committed
2638

2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653
	default:
#ifdef UNIV_DEBUG
	case DATA_SYS_CHILD:
	case DATA_SYS:
		/* These column types should never be shipped to MySQL. */
		ut_ad(0);

	case DATA_CHAR:
	case DATA_FIXBINARY:
	case DATA_FLOAT:
	case DATA_DOUBLE:
	case DATA_DECIMAL:
		/* Above are the valid column types for MySQL data. */
#endif /* UNIV_DEBUG */
		ut_ad(templ->mysql_col_len == len);
osku's avatar
osku committed
2654 2655 2656 2657 2658 2659 2660 2661
		memcpy(dest, data, len);
	}
}

/******************************************************************
Convert a row in the Innobase format to a row in the MySQL format.
Note that the template in prebuilt may advise us to copy only a few
columns to mysql_rec, other columns are left blank. All columns may not
2662 2663
be needed in the query.
@return	TRUE if success, FALSE if could not allocate memory for a BLOB (though we may also assert in that case) */
osku's avatar
osku committed
2664 2665 2666 2667
static
ibool
row_sel_store_mysql_rec(
/*====================*/
2668 2669 2670
	byte*		mysql_rec,	/*!< out: row in the MySQL format */
	row_prebuilt_t*	prebuilt,	/*!< in: prebuilt struct */
	const rec_t*	rec,		/*!< in: Innobase record in the index
osku's avatar
osku committed
2671
					which was described in prebuilt's
2672 2673
					template; must be protected by
					a page latch */
2674
	const ulint*	offsets)	/*!< in: array returned by
osku's avatar
osku committed
2675 2676 2677 2678
					rec_get_offsets() */
{
	mysql_row_templ_t*	templ;
	mem_heap_t*		extern_field_heap	= NULL;
2679
	mem_heap_t*		heap;
2680
	const byte*		data;
osku's avatar
osku committed
2681 2682
	ulint			len;
	ulint			i;
2683

osku's avatar
osku committed
2684
	ut_ad(prebuilt->mysql_template);
2685
	ut_ad(prebuilt->default_rec);
osku's avatar
osku committed
2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697
	ut_ad(rec_offs_validate(rec, NULL, offsets));

	if (UNIV_LIKELY_NULL(prebuilt->blob_heap)) {
		mem_heap_free(prebuilt->blob_heap);
		prebuilt->blob_heap = NULL;
	}

	for (i = 0; i < prebuilt->n_template; i++) {

		templ = prebuilt->mysql_template + i;

		if (UNIV_UNLIKELY(rec_offs_nth_extern(offsets,
2698
						      templ->rec_field_no))) {
osku's avatar
osku committed
2699 2700 2701 2702 2703 2704

			/* Copy an externally stored field to the temporary
			heap */

			ut_a(!prebuilt->trx->has_search_latch);

2705 2706
			if (UNIV_UNLIKELY(templ->type == DATA_BLOB)) {
				if (prebuilt->blob_heap == NULL) {
2707 2708
					prebuilt->blob_heap = mem_heap_create(
						UNIV_PAGE_SIZE);
2709 2710 2711 2712
				}

				heap = prebuilt->blob_heap;
			} else {
2713 2714
				extern_field_heap
					= mem_heap_create(UNIV_PAGE_SIZE);
2715 2716 2717

				heap = extern_field_heap;
			}
osku's avatar
osku committed
2718 2719 2720 2721 2722

			/* NOTE: if we are retrieving a big BLOB, we may
			already run out of memory in the next call, which
			causes an assert */

2723 2724 2725 2726
			data = btr_rec_copy_externally_stored_field(
				rec, offsets,
				dict_table_zip_size(prebuilt->table),
				templ->rec_field_no, &len, heap);
osku's avatar
osku committed
2727 2728

			ut_a(len != UNIV_SQL_NULL);
2729 2730 2731 2732
		} else {
			/* Field is stored in the row. */

			data = rec_get_nth_field(rec, offsets,
2733
						 templ->rec_field_no, &len);
2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752

			if (UNIV_UNLIKELY(templ->type == DATA_BLOB)
			    && len != UNIV_SQL_NULL) {

				/* It is a BLOB field locally stored in the
				InnoDB record: we MUST copy its contents to
				prebuilt->blob_heap here because later code
				assumes all BLOB values have been copied to a
				safe place. */

				if (prebuilt->blob_heap == NULL) {
					prebuilt->blob_heap = mem_heap_create(
						UNIV_PAGE_SIZE);
				}

				data = memcpy(mem_heap_alloc(
						prebuilt->blob_heap, len),
						data, len);
			}
osku's avatar
osku committed
2753 2754 2755
		}

		if (len != UNIV_SQL_NULL) {
2756 2757 2758
			row_sel_field_store_in_mysql_format(
				mysql_rec + templ->mysql_col_offset,
				templ, data, len);
osku's avatar
osku committed
2759 2760 2761

			/* Cleanup */
			if (extern_field_heap) {
2762
				mem_heap_free(extern_field_heap);
osku's avatar
osku committed
2763
				extern_field_heap = NULL;
2764 2765
			}

osku's avatar
osku committed
2766 2767 2768
			if (templ->mysql_null_bit_mask) {
				/* It is a nullable column with a non-NULL
				value */
2769 2770
				mysql_rec[templ->mysql_null_byte_offset]
					&= ~(byte) templ->mysql_null_bit_mask;
osku's avatar
osku committed
2771 2772
			}
		} else {
2773 2774
			/* MySQL assumes that the field for an SQL
			NULL value is set to the default value. */
osku's avatar
osku committed
2775

2776 2777
			mysql_rec[templ->mysql_null_byte_offset]
				|= (byte) templ->mysql_null_bit_mask;
2778 2779 2780
			memcpy(mysql_rec + templ->mysql_col_offset,
			       prebuilt->default_rec + templ->mysql_col_offset,
			       templ->mysql_col_len);
osku's avatar
osku committed
2781
		}
2782
	}
osku's avatar
osku committed
2783 2784 2785 2786 2787

	return(TRUE);
}

/*************************************************************************
2788 2789
Builds a previous version of a clustered index record for a consistent read
@return	DB_SUCCESS or error code */
osku's avatar
osku committed
2790 2791 2792 2793
static
ulint
row_sel_build_prev_vers_for_mysql(
/*==============================*/
2794 2795 2796 2797 2798
	read_view_t*	read_view,	/*!< in: read view */
	dict_index_t*	clust_index,	/*!< in: clustered index */
	row_prebuilt_t*	prebuilt,	/*!< in: prebuilt struct */
	const rec_t*	rec,		/*!< in: record in a clustered index */
	ulint**		offsets,	/*!< in/out: offsets returned by
osku's avatar
osku committed
2799
					rec_get_offsets(rec, clust_index) */
2800
	mem_heap_t**	offset_heap,	/*!< in/out: memory heap from which
osku's avatar
osku committed
2801
					the offsets are allocated */
2802
	rec_t**		old_vers,	/*!< out: old version, or NULL if the
osku's avatar
osku committed
2803 2804 2805
					record does not exist in the view:
					i.e., it was freshly inserted
					afterwards */
2806
	mtr_t*		mtr)		/*!< in: mtr */
osku's avatar
osku committed
2807 2808 2809 2810 2811 2812 2813 2814
{
	ulint	err;

	if (prebuilt->old_vers_heap) {
		mem_heap_empty(prebuilt->old_vers_heap);
	} else {
		prebuilt->old_vers_heap = mem_heap_create(200);
	}
2815

2816 2817 2818
	err = row_vers_build_for_consistent_read(
		rec, mtr, clust_index, offsets, read_view, offset_heap,
		prebuilt->old_vers_heap, old_vers);
osku's avatar
osku committed
2819 2820 2821 2822 2823 2824
	return(err);
}

/*************************************************************************
Retrieves the clustered index record corresponding to a record in a
non-clustered index. Does the necessary locking. Used in the MySQL
2825 2826
interface.
@return	DB_SUCCESS or error code */
osku's avatar
osku committed
2827 2828 2829 2830
static
ulint
row_sel_get_clust_rec_for_mysql(
/*============================*/
2831 2832 2833
	row_prebuilt_t*	prebuilt,/*!< in: prebuilt struct in the handle */
	dict_index_t*	sec_index,/*!< in: secondary index where rec resides */
	const rec_t*	rec,	/*!< in: record in a non-clustered index; if
osku's avatar
osku committed
2834 2835 2836
				this is a locking read, then rec is not
				allowed to be delete-marked, and that would
				not make sense either */
2837 2838
	que_thr_t*	thr,	/*!< in: query thread */
	const rec_t**	out_rec,/*!< out: clustered record or an old version of
osku's avatar
osku committed
2839 2840 2841
				it, NULL if the old version did not exist
				in the read view, i.e., it was a fresh
				inserted version */
2842
	ulint**		offsets,/*!< in: offsets returned by
2843 2844
				rec_get_offsets(rec, sec_index);
				out: offsets returned by
osku's avatar
osku committed
2845
				rec_get_offsets(out_rec, clust_index) */
2846
	mem_heap_t**	offset_heap,/*!< in/out: memory heap from which
osku's avatar
osku committed
2847
				the offsets are allocated */
2848
	mtr_t*		mtr)	/*!< in: mtr used to get access to the
osku's avatar
osku committed
2849 2850 2851 2852
				non-clustered record; the same mtr is used to
				access the clustered index */
{
	dict_index_t*	clust_index;
2853
	const rec_t*	clust_rec;
osku's avatar
osku committed
2854 2855 2856 2857 2858 2859
	rec_t*		old_vers;
	ulint		err;
	trx_t*		trx;

	*out_rec = NULL;
	trx = thr_get_trx(thr);
2860

2861 2862
	row_build_row_ref_in_tuple(prebuilt->clust_ref, rec,
				   sec_index, *offsets, trx);
osku's avatar
osku committed
2863 2864

	clust_index = dict_table_get_first_index(sec_index->table);
2865

osku's avatar
osku committed
2866
	btr_pcur_open_with_no_init(clust_index, prebuilt->clust_ref,
2867 2868
				   PAGE_CUR_LE, BTR_SEARCH_LEAF,
				   prebuilt->clust_pcur, 0, mtr);
osku's avatar
osku committed
2869 2870 2871 2872 2873 2874 2875 2876 2877

	clust_rec = btr_pcur_get_rec(prebuilt->clust_pcur);

	prebuilt->clust_pcur->trx_if_known = trx;

	/* Note: only if the search ends up on a non-infimum record is the
	low_match value the real match to the search tuple */

	if (!page_rec_is_user_rec(clust_rec)
2878 2879
	    || btr_pcur_get_low_match(prebuilt->clust_pcur)
	    < dict_index_get_n_unique(clust_index)) {
2880

osku's avatar
osku committed
2881 2882 2883 2884 2885 2886 2887 2888 2889
		/* In a rare case it is possible that no clust rec is found
		for a delete-marked secondary index record: if in row0umod.c
		in row_undo_mod_remove_clust_low() we have already removed
		the clust rec, while purge is still cleaning and removing
		secondary index records associated with earlier versions of
		the clustered index record. In that case we know that the
		clustered index record did not exist in the read view of
		trx. */

2890
		if (!rec_get_deleted_flag(rec,
2891 2892
					  dict_table_is_comp(sec_index->table))
		    || prebuilt->select_lock_type != LOCK_NONE) {
2893
			ut_print_timestamp(stderr);
2894
			fputs("  InnoDB: error clustered record"
2895 2896
			      " for sec rec not found\n"
			      "InnoDB: ", stderr);
osku's avatar
osku committed
2897 2898
			dict_index_name_print(stderr, trx, sec_index);
			fputs("\n"
2899
			      "InnoDB: sec index record ", stderr);
osku's avatar
osku committed
2900 2901
			rec_print(stderr, rec, sec_index);
			fputs("\n"
2902
			      "InnoDB: clust index record ", stderr);
osku's avatar
osku committed
2903 2904 2905 2906 2907
			rec_print(stderr, clust_rec, clust_index);
			putc('\n', stderr);
			trx_print(stderr, trx, 600);

			fputs("\n"
2908 2909
			      "InnoDB: Submit a detailed bug report"
			      " to http://bugs.mysql.com\n", stderr);
osku's avatar
osku committed
2910 2911 2912 2913 2914 2915 2916 2917
		}

		clust_rec = NULL;

		goto func_exit;
	}

	*offsets = rec_get_offsets(clust_rec, clust_index, *offsets,
2918
				   ULINT_UNDEFINED, offset_heap);
osku's avatar
osku committed
2919 2920 2921 2922 2923

	if (prebuilt->select_lock_type != LOCK_NONE) {
		/* Try to place a lock on the index record; we are searching
		the clust rec with a unique condition, hence
		we set a LOCK_REC_NOT_GAP type lock */
2924

2925
		err = lock_clust_rec_read_check_and_lock(
2926 2927
			0, btr_pcur_get_block(prebuilt->clust_pcur),
			clust_rec, clust_index, *offsets,
2928
			prebuilt->select_lock_type, LOCK_REC_NOT_GAP, thr);
osku's avatar
osku committed
2929 2930 2931 2932 2933 2934 2935 2936 2937 2938 2939 2940 2941 2942
		if (err != DB_SUCCESS) {

			goto err_exit;
		}
	} else {
		/* This is a non-locking consistent read: if necessary, fetch
		a previous version of the record */

		old_vers = NULL;

		/* If the isolation level allows reading of uncommitted data,
		then we never look for an earlier version */

		if (trx->isolation_level > TRX_ISO_READ_UNCOMMITTED
2943 2944 2945
		    && !lock_clust_rec_cons_read_sees(
			    clust_rec, clust_index, *offsets,
			    trx->read_view)) {
2946

osku's avatar
osku committed
2947 2948
			/* The following call returns 'offsets' associated with
			'old_vers' */
2949 2950 2951 2952
			err = row_sel_build_prev_vers_for_mysql(
				trx->read_view, clust_index, prebuilt,
				clust_rec, offsets, offset_heap, &old_vers,
				mtr);
2953

2954
			if (err != DB_SUCCESS || old_vers == NULL) {
osku's avatar
osku committed
2955 2956 2957 2958 2959 2960 2961 2962 2963 2964 2965 2966 2967 2968 2969 2970 2971 2972 2973

				goto err_exit;
			}

			clust_rec = old_vers;
		}

		/* If we had to go to an earlier version of row or the
		secondary index record is delete marked, then it may be that
		the secondary index record corresponding to clust_rec
		(or old_vers) is not rec; in that case we must ignore
		such row because in our snapshot rec would not have existed.
		Remember that from rec we cannot see directly which transaction
		id corresponds to it: we have to go to the clustered index
		record. A query where we want to fetch all rows where
		the secondary index value is in some interval would return
		a wrong result if we would not drop rows which we come to
		visit through secondary index records that would not really
		exist in our snapshot. */
2974

2975 2976 2977 2978
		if (clust_rec
		    && (old_vers
			|| rec_get_deleted_flag(rec, dict_table_is_comp(
							sec_index->table)))
2979 2980
		    && !row_sel_sec_rec_is_for_clust_rec(
			    rec, sec_index, clust_rec, clust_index)) {
osku's avatar
osku committed
2981 2982
			clust_rec = NULL;
#ifdef UNIV_SEARCH_DEBUG
2983
		} else {
2984
			ut_a(clust_rec == NULL
2985 2986
			     || row_sel_sec_rec_is_for_clust_rec(
				     rec, sec_index, clust_rec, clust_index));
2987
#endif
osku's avatar
osku committed
2988 2989 2990 2991 2992 2993
		}
	}

func_exit:
	*out_rec = clust_rec;

2994 2995 2996
	if (prebuilt->select_lock_type != LOCK_NONE) {
		/* We may use the cursor in update or in unlock_row():
		store its position */
2997

osku's avatar
osku committed
2998 2999 3000 3001 3002 3003 3004 3005 3006 3007 3008
		btr_pcur_store_position(prebuilt->clust_pcur, mtr);
	}

	err = DB_SUCCESS;
err_exit:
	return(err);
}

/************************************************************************
Restores cursor position after it has been stored. We have to take into
account that the record cursor was positioned on may have been deleted.
3009 3010
Then we may have to move the cursor one step up or down.
@return	TRUE if we may need to process the record the cursor is now positioned on (i.e. we should not go to the next record yet) */
osku's avatar
osku committed
3011 3012 3013 3014
static
ibool
sel_restore_position_for_mysql(
/*===========================*/
3015
	ibool*		same_user_rec,	/*!< out: TRUE if we were able to restore
osku's avatar
osku committed
3016 3017 3018
					the cursor on a user record with the
					same ordering prefix in in the
					B-tree index */
3019
	ulint		latch_mode,	/*!< in: latch mode wished in
osku's avatar
osku committed
3020
					restoration */
3021
	btr_pcur_t*	pcur,		/*!< in: cursor whose position
osku's avatar
osku committed
3022
					has been stored */
3023
	ibool		moves_up,	/*!< in: TRUE if the cursor moves up
osku's avatar
osku committed
3024
					in the index */
3025
	mtr_t*		mtr)		/*!< in: mtr; CAUTION: may commit
osku's avatar
osku committed
3026 3027 3028 3029 3030 3031
					mtr temporarily! */
{
	ibool	success;
	ulint	relative_position;

	relative_position = pcur->rel_pos;
3032

osku's avatar
osku committed
3033 3034 3035 3036 3037 3038 3039 3040 3041 3042 3043 3044 3045 3046 3047 3048 3049
	success = btr_pcur_restore_position(latch_mode, pcur, mtr);

	*same_user_rec = success;

	if (relative_position == BTR_PCUR_ON) {
		if (success) {
			return(FALSE);
		}

		if (moves_up) {
			btr_pcur_move_to_next(pcur, mtr);
		}

		return(TRUE);
	}

	if (relative_position == BTR_PCUR_AFTER
3050
	    || relative_position == BTR_PCUR_AFTER_LAST_IN_TREE) {
osku's avatar
osku committed
3051 3052 3053 3054

		if (moves_up) {
			return(TRUE);
		}
3055

3056
		if (btr_pcur_is_on_user_rec(pcur)) {
osku's avatar
osku committed
3057 3058 3059 3060 3061 3062 3063
			btr_pcur_move_to_prev(pcur, mtr);
		}

		return(TRUE);
	}

	ut_ad(relative_position == BTR_PCUR_BEFORE
3064
	      || relative_position == BTR_PCUR_BEFORE_FIRST_IN_TREE);
3065

3066
	if (moves_up && btr_pcur_is_on_user_rec(pcur)) {
osku's avatar
osku committed
3067 3068 3069 3070 3071 3072 3073 3074 3075 3076 3077 3078
		btr_pcur_move_to_next(pcur, mtr);
	}

	return(TRUE);
}

/************************************************************************
Pops a cached row for MySQL from the fetch cache. */
UNIV_INLINE
void
row_sel_pop_cached_row_for_mysql(
/*=============================*/
3079
	byte*		buf,		/*!< in/out: buffer where to copy the
osku's avatar
osku committed
3080
					row */
3081
	row_prebuilt_t*	prebuilt)	/*!< in: prebuilt struct */
osku's avatar
osku committed
3082 3083 3084 3085
{
	ulint			i;
	mysql_row_templ_t*	templ;
	byte*			cached_rec;
3086
	ut_ad(prebuilt->n_fetch_cached > 0);
osku's avatar
osku committed
3087
	ut_ad(prebuilt->mysql_prefix_len <= prebuilt->mysql_row_len);
3088

3089
	if (UNIV_UNLIKELY(prebuilt->keep_other_fields_on_keyread)) {
3090
		/* Copy cache record field by field, don't touch fields that
osku's avatar
osku committed
3091
		are not covered by current key */
3092 3093
		cached_rec = prebuilt->fetch_cache[
			prebuilt->fetch_cache_first];
osku's avatar
osku committed
3094 3095 3096

		for (i = 0; i < prebuilt->n_template; i++) {
			templ = prebuilt->mysql_template + i;
3097 3098 3099
			ut_memcpy(buf + templ->mysql_col_offset,
				  cached_rec + templ->mysql_col_offset,
				  templ->mysql_col_len);
3100
			/* Copy NULL bit of the current field from cached_rec
osku's avatar
osku committed
3101
			to buf */
3102
			if (templ->mysql_null_bit_mask) {
3103 3104
				buf[templ->mysql_null_byte_offset]
					^= (buf[templ->mysql_null_byte_offset]
3105
					    ^ cached_rec[templ->mysql_null_byte_offset])
3106
					& (byte)templ->mysql_null_bit_mask;
osku's avatar
osku committed
3107 3108 3109
			}
		}
	}
3110
	else {
3111 3112 3113
		ut_memcpy(buf,
			  prebuilt->fetch_cache[prebuilt->fetch_cache_first],
			  prebuilt->mysql_prefix_len);
osku's avatar
osku committed
3114 3115 3116 3117 3118 3119 3120 3121 3122 3123 3124 3125 3126 3127 3128
	}
	prebuilt->n_fetch_cached--;
	prebuilt->fetch_cache_first++;

	if (prebuilt->n_fetch_cached == 0) {
		prebuilt->fetch_cache_first = 0;
	}
}

/************************************************************************
Pushes a row for MySQL to the fetch cache. */
UNIV_INLINE
void
row_sel_push_cache_row_for_mysql(
/*=============================*/
3129 3130
	row_prebuilt_t*	prebuilt,	/*!< in: prebuilt struct */
	const rec_t*	rec,		/*!< in: record to push; must
3131
					be protected by a page latch */
3132
	const ulint*	offsets)	/*!< in: rec_get_offsets() */
osku's avatar
osku committed
3133 3134 3135 3136 3137 3138 3139 3140 3141 3142 3143 3144 3145 3146 3147 3148
{
	byte*	buf;
	ulint	i;

	ut_ad(prebuilt->n_fetch_cached < MYSQL_FETCH_CACHE_SIZE);
	ut_ad(rec_offs_validate(rec, NULL, offsets));
	ut_a(!prebuilt->templ_contains_blob);

	if (prebuilt->fetch_cache[0] == NULL) {
		/* Allocate memory for the fetch cache */

		for (i = 0; i < MYSQL_FETCH_CACHE_SIZE; i++) {

			/* A user has reported memory corruption in these
			buffers in Linux. Put magic numbers there to help
			to track a possible bug. */
3149

osku's avatar
osku committed
3150 3151 3152
			buf = mem_alloc(prebuilt->mysql_row_len + 8);

			prebuilt->fetch_cache[i] = buf + 4;
3153

osku's avatar
osku committed
3154 3155 3156 3157 3158 3159 3160 3161
			mach_write_to_4(buf, ROW_PREBUILT_FETCH_MAGIC_N);
			mach_write_to_4(buf + 4 + prebuilt->mysql_row_len,
					ROW_PREBUILT_FETCH_MAGIC_N);
		}
	}

	ut_ad(prebuilt->fetch_cache_first == 0);

3162 3163 3164 3165
	if (UNIV_UNLIKELY(!row_sel_store_mysql_rec(
				  prebuilt->fetch_cache[
					  prebuilt->n_fetch_cached],
				  prebuilt, rec, offsets))) {
osku's avatar
osku committed
3166 3167 3168 3169 3170 3171 3172 3173 3174 3175
		ut_error;
	}

	prebuilt->n_fetch_cached++;
}

/*************************************************************************
Tries to do a shortcut to fetch a clustered index record with a unique key,
using the hash index if possible (not always). We assume that the search
mode is PAGE_CUR_GE, it is a consistent read, there is a read view in trx,
3176 3177
btr search latch has been locked in S-mode.
@return	SEL_FOUND, SEL_EXHAUSTED, SEL_RETRY */
osku's avatar
osku committed
3178 3179 3180 3181
static
ulint
row_sel_try_search_shortcut_for_mysql(
/*==================================*/
3182 3183 3184 3185 3186
	const rec_t**	out_rec,/*!< out: record if found */
	row_prebuilt_t*	prebuilt,/*!< in: prebuilt struct */
	ulint**		offsets,/*!< in/out: for rec_get_offsets(*out_rec) */
	mem_heap_t**	heap,	/*!< in/out: heap for rec_get_offsets() */
	mtr_t*		mtr)	/*!< in: started mtr */
osku's avatar
osku committed
3187 3188
{
	dict_index_t*	index		= prebuilt->index;
marko's avatar
marko committed
3189
	const dtuple_t*	search_tuple	= prebuilt->search_tuple;
osku's avatar
osku committed
3190 3191
	btr_pcur_t*	pcur		= prebuilt->pcur;
	trx_t*		trx		= prebuilt->trx;
3192
	const rec_t*	rec;
3193

3194
	ut_ad(dict_index_is_clust(index));
osku's avatar
osku committed
3195
	ut_ad(!prebuilt->templ_contains_blob);
3196

osku's avatar
osku committed
3197
	btr_pcur_open_with_no_init(index, search_tuple, PAGE_CUR_GE,
3198
				   BTR_SEARCH_LEAF, pcur,
osku's avatar
osku committed
3199
#ifndef UNIV_SEARCH_DEBUG
3200
				   RW_S_LATCH,
osku's avatar
osku committed
3201
#else
3202
				   0,
osku's avatar
osku committed
3203
#endif
3204
				   mtr);
osku's avatar
osku committed
3205
	rec = btr_pcur_get_rec(pcur);
3206

osku's avatar
osku committed
3207 3208 3209 3210 3211 3212 3213
	if (!page_rec_is_user_rec(rec)) {

		return(SEL_RETRY);
	}

	/* As the cursor is now placed on a user record after a search with
	the mode PAGE_CUR_GE, the up_match field in the cursor tells how many
3214
	fields in the user record matched to the search tuple */
osku's avatar
osku committed
3215 3216 3217 3218 3219 3220 3221 3222 3223 3224

	if (btr_pcur_get_up_match(pcur) < dtuple_get_n_fields(search_tuple)) {

		return(SEL_EXHAUSTED);
	}

	/* This is a non-locking consistent read: if necessary, fetch
	a previous version of the record */

	*offsets = rec_get_offsets(rec, index, *offsets,
3225
				   ULINT_UNDEFINED, heap);
osku's avatar
osku committed
3226 3227

	if (!lock_clust_rec_cons_read_sees(rec, index,
3228
					   *offsets, trx->read_view)) {
osku's avatar
osku committed
3229 3230 3231 3232

		return(SEL_RETRY);
	}

3233
	if (rec_get_deleted_flag(rec, dict_table_is_comp(index->table))) {
osku's avatar
osku committed
3234 3235 3236 3237 3238

		return(SEL_EXHAUSTED);
	}

	*out_rec = rec;
3239

osku's avatar
osku committed
3240 3241 3242 3243 3244 3245 3246 3247
	return(SEL_FOUND);
}

/************************************************************************
Searches for rows in the database. This is used in the interface to
MySQL. This function opens a cursor, and also implements fetch next
and fetch prev. NOTE that if we do a search with a full key value
from a unique index (ROW_SEL_EXACT), then we will not store the cursor
3248 3249
position and fetch next or fetch prev must not be tried to the cursor!
@return	DB_SUCCESS, DB_RECORD_NOT_FOUND, DB_END_OF_INDEX, DB_DEADLOCK, DB_LOCK_TABLE_FULL, DB_CORRUPTION, or DB_TOO_BIG_RECORD */
3250
UNIV_INTERN
osku's avatar
osku committed
3251 3252 3253
ulint
row_search_for_mysql(
/*=================*/
3254
	byte*		buf,		/*!< in/out: buffer for the fetched
osku's avatar
osku committed
3255
					row in the MySQL format */
3256 3257
	ulint		mode,		/*!< in: search mode PAGE_CUR_L, ... */
	row_prebuilt_t*	prebuilt,	/*!< in: prebuilt struct for the
osku's avatar
osku committed
3258 3259 3260 3261 3262 3263
					table handle; this contains the info
					of search_tuple, index; if search
					tuple contains 0 fields then we
					position the cursor at the start or
					the end of the index, depending on
					'mode' */
3264
	ulint		match_mode,	/*!< in: 0 or ROW_SEL_EXACT or
3265
					ROW_SEL_EXACT_PREFIX */
3266
	ulint		direction)	/*!< in: 0 or ROW_SEL_NEXT or
osku's avatar
osku committed
3267 3268 3269 3270 3271 3272
					ROW_SEL_PREV; NOTE: if this is != 0,
					then prebuilt must have a pcur
					with stored position! In opening of a
					cursor 'direction' should be 0. */
{
	dict_index_t*	index		= prebuilt->index;
3273
	ibool		comp		= dict_table_is_comp(index->table);
marko's avatar
marko committed
3274
	const dtuple_t*	search_tuple	= prebuilt->search_tuple;
osku's avatar
osku committed
3275 3276 3277 3278
	btr_pcur_t*	pcur		= prebuilt->pcur;
	trx_t*		trx		= prebuilt->trx;
	dict_index_t*	clust_index;
	que_thr_t*	thr;
3279 3280 3281
	const rec_t*	rec;
	const rec_t*	result_rec;
	const rec_t*	clust_rec;
osku's avatar
osku committed
3282 3283 3284
	ulint		err				= DB_SUCCESS;
	ibool		unique_search			= FALSE;
	ibool		unique_search_from_clust_index	= FALSE;
3285 3286
	ibool		mtr_has_extra_clust_latch	= FALSE;
	ibool		moves_up			= FALSE;
osku's avatar
osku committed
3287
	ibool		set_also_gap_locks		= TRUE;
3288 3289
	/* if the query is a plain locking SELECT, and the isolation level
	is <= TRX_ISO_READ_COMMITTED, then this is set to FALSE */
3290
	ibool		did_semi_consistent_read	= FALSE;
3291 3292 3293
	/* if the returned record was locked and we did a semi-consistent
	read (fetch the newest committed version), then this is set to
	TRUE */
osku's avatar
osku committed
3294 3295 3296 3297 3298 3299 3300 3301 3302 3303
#ifdef UNIV_SEARCH_DEBUG
	ulint		cnt				= 0;
#endif /* UNIV_SEARCH_DEBUG */
	ulint		next_offs;
	ibool		same_user_rec;
	mtr_t		mtr;
	mem_heap_t*	heap				= NULL;
	ulint		offsets_[REC_OFFS_NORMAL_SIZE];
	ulint*		offsets				= offsets_;

3304
	rec_offs_init(offsets_);
osku's avatar
osku committed
3305 3306 3307 3308 3309

	ut_ad(index && pcur && search_tuple);
	ut_ad(trx->mysql_thread_id == os_thread_get_curr_id());

	if (UNIV_UNLIKELY(prebuilt->table->ibd_file_missing)) {
3310 3311
		ut_print_timestamp(stderr);
		fprintf(stderr, "  InnoDB: Error:\n"
3312 3313 3314 3315 3316 3317 3318 3319
			"InnoDB: MySQL is trying to use a table handle"
			" but the .ibd file for\n"
			"InnoDB: table %s does not exist.\n"
			"InnoDB: Have you deleted the .ibd file"
			" from the database directory under\n"
			"InnoDB: the MySQL datadir, or have you used"
			" DISCARD TABLESPACE?\n"
			"InnoDB: Look from\n"
3320
			"InnoDB: " REFMAN "innodb-troubleshooting.html\n"
3321 3322
			"InnoDB: how you can resolve the problem.\n",
			prebuilt->table->name);
osku's avatar
osku committed
3323 3324 3325 3326

		return(DB_ERROR);
	}

3327 3328 3329 3330 3331
	if (UNIV_UNLIKELY(!prebuilt->index_usable)) {

		return(DB_MISSING_HISTORY);
	}

osku's avatar
osku committed
3332 3333
	if (UNIV_UNLIKELY(prebuilt->magic_n != ROW_PREBUILT_ALLOCATED)) {
		fprintf(stderr,
3334 3335 3336
			"InnoDB: Error: trying to free a corrupt\n"
			"InnoDB: table handle. Magic n %lu, table name ",
			(ulong) prebuilt->magic_n);
3337
		ut_print_name(stderr, trx, TRUE, prebuilt->table->name);
osku's avatar
osku committed
3338 3339
		putc('\n', stderr);

3340
		mem_analyze_corruption(prebuilt);
osku's avatar
osku committed
3341 3342 3343 3344

		ut_error;
	}

3345 3346 3347 3348 3349
#if 0
	/* August 19, 2005 by Heikki: temporarily disable this error
	print until the cursor lock count is done correctly.
	See bugs #12263 and #12456!*/

osku's avatar
osku committed
3350
	if (trx->n_mysql_tables_in_use == 0
3351
	    && UNIV_UNLIKELY(prebuilt->select_lock_type == LOCK_NONE)) {
osku's avatar
osku committed
3352 3353 3354 3355 3356
		/* Note that if MySQL uses an InnoDB temp table that it
		created inside LOCK TABLES, then n_mysql_tables_in_use can
		be zero; in that case select_lock_type is set to LOCK_X in
		::start_stmt. */

3357 3358 3359 3360
		fputs("InnoDB: Error: MySQL is trying to perform a SELECT\n"
		      "InnoDB: but it has not locked"
		      " any tables in ::external_lock()!\n",
		      stderr);
osku's avatar
osku committed
3361
		trx_print(stderr, trx, 600);
3362
		fputc('\n', stderr);
osku's avatar
osku committed
3363
	}
3364
#endif
osku's avatar
osku committed
3365

3366 3367 3368
#if 0
	fprintf(stderr, "Match mode %lu\n search tuple ",
		(ulong) match_mode);
osku's avatar
osku committed
3369
	dtuple_print(search_tuple);
3370 3371 3372
	fprintf(stderr, "N tables locked %lu\n",
		(ulong) trx->mysql_n_tables_locked);
#endif
osku's avatar
osku committed
3373 3374 3375 3376
	/*-------------------------------------------------------------*/
	/* PHASE 0: Release a possible s-latch we are holding on the
	adaptive hash index latch if there is someone waiting behind */

3377
	if (UNIV_UNLIKELY(rw_lock_get_writer(&btr_search_latch) != RW_LOCK_NOT_LOCKED)
3378
	    && trx->has_search_latch) {
osku's avatar
osku committed
3379 3380 3381 3382 3383 3384 3385 3386 3387 3388 3389

		/* There is an x-latch request on the adaptive hash index:
		release the s-latch to reduce starvation and wait for
		BTR_SEA_TIMEOUT rounds before trying to keep it again over
		calls from MySQL */

		rw_lock_s_unlock(&btr_search_latch);
		trx->has_search_latch = FALSE;

		trx->search_latch_timeout = BTR_SEA_TIMEOUT;
	}
3390

3391
	/* Reset the new record lock info if srv_locks_unsafe_for_binlog
3392 3393 3394
	is set or session is using a READ COMMITED isolation level. Then
	we are able to remove the record locks set here on an individual
	row. */
3395
	prebuilt->new_rec_locks = 0;
osku's avatar
osku committed
3396 3397 3398 3399 3400 3401

	/*-------------------------------------------------------------*/
	/* PHASE 1: Try to pop the row from the prefetch cache */

	if (UNIV_UNLIKELY(direction == 0)) {
		trx->op_info = "starting index read";
3402

osku's avatar
osku committed
3403 3404 3405 3406 3407 3408 3409 3410 3411 3412 3413 3414 3415 3416 3417 3418 3419 3420 3421 3422 3423 3424 3425
		prebuilt->n_rows_fetched = 0;
		prebuilt->n_fetch_cached = 0;
		prebuilt->fetch_cache_first = 0;

		if (prebuilt->sel_graph == NULL) {
			/* Build a dummy select query graph */
			row_prebuild_sel_graph(prebuilt);
		}
	} else {
		trx->op_info = "fetching rows";

		if (prebuilt->n_rows_fetched == 0) {
			prebuilt->fetch_direction = direction;
		}

		if (UNIV_UNLIKELY(direction != prebuilt->fetch_direction)) {
			if (UNIV_UNLIKELY(prebuilt->n_fetch_cached > 0)) {
				ut_error;
				/* TODO: scrollable cursor: restore cursor to
				the place of the latest returned row,
				or better: prevent caching for a scroll
				cursor! */
			}
3426

osku's avatar
osku committed
3427 3428 3429 3430 3431 3432 3433 3434 3435 3436 3437 3438 3439 3440 3441
			prebuilt->n_rows_fetched = 0;
			prebuilt->n_fetch_cached = 0;
			prebuilt->fetch_cache_first = 0;

		} else if (UNIV_LIKELY(prebuilt->n_fetch_cached > 0)) {
			row_sel_pop_cached_row_for_mysql(buf, prebuilt);

			prebuilt->n_rows_fetched++;

			srv_n_rows_read++;
			err = DB_SUCCESS;
			goto func_exit;
		}

		if (prebuilt->fetch_cache_first > 0
3442
		    && prebuilt->fetch_cache_first < MYSQL_FETCH_CACHE_SIZE) {
osku's avatar
osku committed
3443

3444 3445 3446
			/* The previous returned row was popped from the fetch
			cache, but the cache was not full at the time of the
			popping: no more rows can exist in the result set */
osku's avatar
osku committed
3447 3448 3449 3450

			err = DB_RECORD_NOT_FOUND;
			goto func_exit;
		}
3451

osku's avatar
osku committed
3452 3453 3454 3455 3456 3457 3458 3459 3460 3461 3462
		prebuilt->n_rows_fetched++;

		if (prebuilt->n_rows_fetched > 1000000000) {
			/* Prevent wrap-over */
			prebuilt->n_rows_fetched = 500000000;
		}

		mode = pcur->search_mode;
	}

	/* In a search where at most one record in the index may match, we
3463 3464
	can use a LOCK_REC_NOT_GAP type record lock when locking a
	non-delete-marked matching record.
osku's avatar
osku committed
3465

3466 3467 3468 3469
	Note that in a unique secondary index there may be different
	delete-marked versions of a record where only the primary key
	values differ: thus in a secondary index we must use next-key
	locks when locking delete-marked records. */
3470

osku's avatar
osku committed
3471
	if (match_mode == ROW_SEL_EXACT
3472
	    && dict_index_is_unique(index)
3473 3474 3475 3476
	    && dtuple_get_n_fields(search_tuple)
	    == dict_index_get_n_unique(index)
	    && (dict_index_is_clust(index)
		|| !dtuple_contains_null(search_tuple))) {
osku's avatar
osku committed
3477 3478 3479 3480 3481 3482 3483 3484 3485 3486 3487 3488 3489 3490

		/* Note above that a UNIQUE secondary index can contain many
		rows with the same key value if one of the columns is the SQL
		null. A clustered index under MySQL can never contain null
		columns because we demand that all the columns in primary key
		are non-null. */

		unique_search = TRUE;

		/* Even if the condition is unique, MySQL seems to try to
		retrieve also a second row if a primary key contains more than
		1 column. Return immediately if this is not a HANDLER
		command. */

3491 3492
		if (UNIV_UNLIKELY(direction != 0
				  && !prebuilt->used_in_HANDLER)) {
3493

osku's avatar
osku committed
3494 3495 3496 3497 3498 3499 3500 3501 3502 3503 3504 3505 3506 3507 3508 3509 3510
			err = DB_RECORD_NOT_FOUND;
			goto func_exit;
		}
	}

	mtr_start(&mtr);

	/*-------------------------------------------------------------*/
	/* PHASE 2: Try fast adaptive hash index search if possible */

	/* Next test if this is the special case where we can use the fast
	adaptive hash index to try the search. Since we must release the
	search system latch when we retrieve an externally stored field, we
	cannot use the adaptive hash index in a search in the case the row
	may be long and there may be externally stored fields */

	if (UNIV_UNLIKELY(direction == 0)
3511 3512 3513 3514 3515
	    && unique_search
	    && dict_index_is_clust(index)
	    && !prebuilt->templ_contains_blob
	    && !prebuilt->used_in_HANDLER
	    && (prebuilt->mysql_row_len < UNIV_PAGE_SIZE / 8)) {
osku's avatar
osku committed
3516 3517 3518 3519 3520 3521

		mode = PAGE_CUR_GE;

		unique_search_from_clust_index = TRUE;

		if (trx->mysql_n_tables_locked == 0
3522 3523 3524
		    && prebuilt->select_lock_type == LOCK_NONE
		    && trx->isolation_level > TRX_ISO_READ_UNCOMMITTED
		    && trx->read_view) {
osku's avatar
osku committed
3525 3526 3527 3528 3529 3530 3531 3532 3533 3534 3535 3536 3537

			/* This is a SELECT query done as a consistent read,
			and the read view has already been allocated:
			let us try a search shortcut through the hash
			index.
			NOTE that we must also test that
			mysql_n_tables_locked == 0, because this might
			also be INSERT INTO ... SELECT ... or
			CREATE TABLE ... SELECT ... . Our algorithm is
			NOT prepared to inserts interleaved with the SELECT,
			and if we try that, we can deadlock on the adaptive
			hash index semaphore! */

3538
#ifndef UNIV_SEARCH_DEBUG
osku's avatar
osku committed
3539 3540 3541 3542 3543
			if (!trx->has_search_latch) {
				rw_lock_s_lock(&btr_search_latch);
				trx->has_search_latch = TRUE;
			}
#endif
3544 3545 3546
			switch (row_sel_try_search_shortcut_for_mysql(
					&rec, prebuilt, &offsets, &heap,
					&mtr)) {
osku's avatar
osku committed
3547 3548 3549
			case SEL_FOUND:
#ifdef UNIV_SEARCH_DEBUG
				ut_a(0 == cmp_dtuple_rec(search_tuple,
3550
							 rec, offsets));
3551
#endif
3552 3553 3554 3555 3556 3557
				/* At this point, rec is protected by
				a page latch that was acquired by
				row_sel_try_search_shortcut_for_mysql().
				The latch will not be released until
				mtr_commit(&mtr). */

osku's avatar
osku committed
3558
				if (!row_sel_store_mysql_rec(buf, prebuilt,
3559
							     rec, offsets)) {
3560
					err = DB_TOO_BIG_RECORD;
osku's avatar
osku committed
3561 3562 3563

					/* We let the main loop to do the
					error handling */
3564
					goto shortcut_fails_too_big_rec;
osku's avatar
osku committed
3565
				}
3566 3567

				mtr_commit(&mtr);
osku's avatar
osku committed
3568 3569 3570 3571 3572

				/* ut_print_name(stderr, index->name);
				fputs(" shortcut\n", stderr); */

				srv_n_rows_read++;
3573

osku's avatar
osku committed
3574
				err = DB_SUCCESS;
3575
				goto release_search_latch_if_needed;
osku's avatar
osku committed
3576 3577

			case SEL_EXHAUSTED:
3578
				mtr_commit(&mtr);
osku's avatar
osku committed
3579 3580 3581 3582

				/* ut_print_name(stderr, index->name);
				fputs(" record not found 2\n", stderr); */

3583 3584
				err = DB_RECORD_NOT_FOUND;
release_search_latch_if_needed:
osku's avatar
osku committed
3585
				if (trx->search_latch_timeout > 0
3586
				    && trx->has_search_latch) {
osku's avatar
osku committed
3587 3588 3589

					trx->search_latch_timeout--;

3590
					rw_lock_s_unlock(&btr_search_latch);
osku's avatar
osku committed
3591 3592 3593 3594 3595 3596
					trx->has_search_latch = FALSE;
				}

				/* NOTE that we do NOT store the cursor
				position */
				goto func_exit;
3597 3598 3599 3600 3601 3602

			case SEL_RETRY:
				break;

			default:
				ut_ad(0);
osku's avatar
osku committed
3603 3604 3605 3606 3607 3608 3609 3610 3611 3612 3613 3614 3615
			}
shortcut_fails_too_big_rec:
			mtr_commit(&mtr);
			mtr_start(&mtr);
		}
	}

	/*-------------------------------------------------------------*/
	/* PHASE 3: Open or restore index cursor position */

	if (trx->has_search_latch) {
		rw_lock_s_unlock(&btr_search_latch);
		trx->has_search_latch = FALSE;
3616
	}
osku's avatar
osku committed
3617 3618 3619 3620

	trx_start_if_not_started(trx);

	if (trx->isolation_level <= TRX_ISO_READ_COMMITTED
3621
	    && prebuilt->select_lock_type != LOCK_NONE
3622
	    && trx->mysql_thd != NULL
3623 3624 3625
	    && thd_is_select(trx->mysql_thd)) {
		/* It is a plain locking SELECT and the isolation
		level is low: do not lock gaps */
osku's avatar
osku committed
3626

3627
		set_also_gap_locks = FALSE;
osku's avatar
osku committed
3628
	}
3629

osku's avatar
osku committed
3630 3631 3632
	/* Note that if the search mode was GE or G, then the cursor
	naturally moves upward (in fetch next) in alphabetical order,
	otherwise downward */
3633

osku's avatar
osku committed
3634 3635 3636 3637 3638 3639 3640 3641 3642 3643 3644 3645 3646 3647 3648
	if (UNIV_UNLIKELY(direction == 0)) {
		if (mode == PAGE_CUR_GE || mode == PAGE_CUR_G) {
			moves_up = TRUE;
		}
	} else if (direction == ROW_SEL_NEXT) {
		moves_up = TRUE;
	}

	thr = que_fork_get_first_thr(prebuilt->sel_graph);

	que_thr_move_to_run_state_for_mysql(thr, trx);

	clust_index = dict_table_get_first_index(index->table);

	if (UNIV_LIKELY(direction != 0)) {
3649 3650 3651
		ibool	need_to_process = sel_restore_position_for_mysql(
			&same_user_rec, BTR_SEARCH_LEAF,
			pcur, moves_up, &mtr);
3652 3653 3654

		if (UNIV_UNLIKELY(need_to_process)) {
			if (UNIV_UNLIKELY(prebuilt->row_read_type
3655
					  == ROW_READ_DID_SEMI_CONSISTENT)) {
3656 3657 3658 3659 3660 3661 3662
				/* We did a semi-consistent read,
				but the record was removed in
				the meantime. */
				prebuilt->row_read_type
					= ROW_READ_TRY_SEMI_CONSISTENT;
			}
		} else if (UNIV_LIKELY(prebuilt->row_read_type
3663
				       != ROW_READ_DID_SEMI_CONSISTENT)) {
3664 3665 3666 3667 3668 3669 3670

			/* The cursor was positioned on the record
			that we returned previously.  If we need
			to repeat a semi-consistent read as a
			pessimistic locking read, the record
			cannot be skipped. */

osku's avatar
osku committed
3671 3672 3673 3674 3675 3676
			goto next_rec;
		}

	} else if (dtuple_get_n_fields(search_tuple) > 0) {

		btr_pcur_open_with_no_init(index, search_tuple, mode,
3677 3678
					   BTR_SEARCH_LEAF,
					   pcur, 0, &mtr);
osku's avatar
osku committed
3679 3680

		pcur->trx_if_known = trx;
3681 3682 3683 3684 3685 3686 3687 3688 3689 3690 3691 3692

		rec = btr_pcur_get_rec(pcur);

		if (!moves_up
		    && !page_rec_is_supremum(rec)
		    && set_also_gap_locks
		    && !(srv_locks_unsafe_for_binlog
			 || trx->isolation_level == TRX_ISO_READ_COMMITTED)
		    && prebuilt->select_lock_type != LOCK_NONE) {

			/* Try to place a gap lock on the next index record
			to prevent phantoms in ORDER BY ... DESC queries */
3693
			const rec_t*	next = page_rec_get_next_const(rec);
3694

3695
			offsets = rec_get_offsets(next, index, offsets,
3696 3697
						  ULINT_UNDEFINED, &heap);
			err = sel_set_rec_lock(btr_pcur_get_block(pcur),
3698
					       next, index, offsets,
3699 3700 3701 3702 3703 3704 3705 3706
					       prebuilt->select_lock_type,
					       LOCK_GAP, thr);

			if (err != DB_SUCCESS) {

				goto lock_wait_or_error;
			}
		}
osku's avatar
osku committed
3707 3708
	} else {
		if (mode == PAGE_CUR_G) {
3709 3710 3711
			btr_pcur_open_at_index_side(
				TRUE, index, BTR_SEARCH_LEAF, pcur, FALSE,
				&mtr);
osku's avatar
osku committed
3712
		} else if (mode == PAGE_CUR_L) {
3713 3714 3715
			btr_pcur_open_at_index_side(
				FALSE, index, BTR_SEARCH_LEAF, pcur, FALSE,
				&mtr);
osku's avatar
osku committed
3716 3717 3718 3719 3720 3721 3722
		}
	}

	if (!prebuilt->sql_stat_start) {
		/* No need to set an intention lock or assign a read view */

		if (trx->read_view == NULL
3723
		    && prebuilt->select_lock_type == LOCK_NONE) {
osku's avatar
osku committed
3724

3725 3726 3727 3728
			fputs("InnoDB: Error: MySQL is trying to"
			      " perform a consistent read\n"
			      "InnoDB: but the read view is not assigned!\n",
			      stderr);
osku's avatar
osku committed
3729
			trx_print(stderr, trx, 600);
3730
			fputc('\n', stderr);
osku's avatar
osku committed
3731 3732 3733
			ut_a(0);
		}
	} else if (prebuilt->select_lock_type == LOCK_NONE) {
3734
		/* This is a consistent read */
osku's avatar
osku committed
3735 3736 3737 3738 3739 3740
		/* Assign a read view for the query */

		trx_assign_read_view(trx);
		prebuilt->sql_stat_start = FALSE;
	} else {
		ulint	lock_mode;
3741
		if (prebuilt->select_lock_type == LOCK_S) {
osku's avatar
osku committed
3742 3743 3744 3745 3746 3747 3748 3749 3750 3751 3752 3753 3754 3755 3756 3757
			lock_mode = LOCK_IS;
		} else {
			lock_mode = LOCK_IX;
		}
		err = lock_table(0, index->table, lock_mode, thr);

		if (err != DB_SUCCESS) {

			goto lock_wait_or_error;
		}
		prebuilt->sql_stat_start = FALSE;
	}

rec_loop:
	/*-------------------------------------------------------------*/
	/* PHASE 4: Look for matching records in a loop */
3758

osku's avatar
osku committed
3759 3760 3761
	rec = btr_pcur_get_rec(pcur);
	ut_ad(!!page_rec_is_comp(rec) == comp);
#ifdef UNIV_SEARCH_DEBUG
3762
	/*
osku's avatar
osku committed
3763 3764 3765
	fputs("Using ", stderr);
	dict_index_name_print(stderr, index);
	fprintf(stderr, " cnt %lu ; Page no %lu\n", cnt,
3766
	page_get_page_no(page_align(rec)));
osku's avatar
osku committed
3767
	rec_print(rec);
3768
	*/
osku's avatar
osku committed
3769 3770 3771 3772 3773 3774 3775 3776 3777 3778 3779 3780 3781 3782
#endif /* UNIV_SEARCH_DEBUG */

	if (page_rec_is_infimum(rec)) {

		/* The infimum record on a page cannot be in the result set,
		and neither can a record lock be placed on it: we skip such
		a record. */

		goto next_rec;
	}

	if (page_rec_is_supremum(rec)) {

		if (set_also_gap_locks
3783 3784 3785
		    && !(srv_locks_unsafe_for_binlog
			 || trx->isolation_level == TRX_ISO_READ_COMMITTED)
		    && prebuilt->select_lock_type != LOCK_NONE) {
osku's avatar
osku committed
3786 3787 3788

			/* Try to place a lock on the index record */

3789 3790 3791
			/* If innodb_locks_unsafe_for_binlog option is used
			or this session is using a READ COMMITTED isolation
			level we do not lock gaps. Supremum record is really
osku's avatar
osku committed
3792
			a gap and therefore we do not set locks there. */
3793

osku's avatar
osku committed
3794
			offsets = rec_get_offsets(rec, index, offsets,
3795
						  ULINT_UNDEFINED, &heap);
3796 3797
			err = sel_set_rec_lock(btr_pcur_get_block(pcur),
					       rec, index, offsets,
3798 3799
					       prebuilt->select_lock_type,
					       LOCK_ORDINARY, thr);
osku's avatar
osku committed
3800 3801 3802 3803 3804 3805 3806 3807

			if (err != DB_SUCCESS) {

				goto lock_wait_or_error;
			}
		}
		/* A page supremum record cannot be in the result set: skip
		it now that we have placed a possible lock on it */
3808

osku's avatar
osku committed
3809 3810 3811 3812 3813 3814
		goto next_rec;
	}

	/*-------------------------------------------------------------*/
	/* Do sanity checks in case our cursor has bumped into page
	corruption */
3815

osku's avatar
osku committed
3816 3817 3818 3819 3820 3821 3822 3823 3824 3825 3826 3827 3828 3829 3830 3831 3832 3833 3834
	if (comp) {
		next_offs = rec_get_next_offs(rec, TRUE);
		if (UNIV_UNLIKELY(next_offs < PAGE_NEW_SUPREMUM)) {

			goto wrong_offs;
		}
	} else {
		next_offs = rec_get_next_offs(rec, FALSE);
		if (UNIV_UNLIKELY(next_offs < PAGE_OLD_SUPREMUM)) {

			goto wrong_offs;
		}
	}

	if (UNIV_UNLIKELY(next_offs >= UNIV_PAGE_SIZE - PAGE_DIR)) {

wrong_offs:
		if (srv_force_recovery == 0 || moves_up == FALSE) {
			ut_print_timestamp(stderr);
3835
			buf_page_print(page_align(rec), 0);
osku's avatar
osku committed
3836
			fprintf(stderr,
3837
				"\nInnoDB: rec address %p,"
3838
				" buf block fix count %lu\n",
3839
				(void*) rec, (ulong)
3840
				btr_cur_get_block(btr_pcur_get_btr_cur(pcur))
3841
				->page.buf_fix_count);
osku's avatar
osku committed
3842
			fprintf(stderr,
3843 3844 3845
				"InnoDB: Index corruption: rec offs %lu"
				" next offs %lu, page no %lu,\n"
				"InnoDB: ",
3846
				(ulong) page_offset(rec),
osku's avatar
osku committed
3847
				(ulong) next_offs,
3848
				(ulong) page_get_page_no(page_align(rec)));
osku's avatar
osku committed
3849 3850
			dict_index_name_print(stderr, trx, index);
			fputs(". Run CHECK TABLE. You may need to\n"
3851 3852 3853
			      "InnoDB: restore from a backup, or"
			      " dump + drop + reimport the table.\n",
			      stderr);
3854

osku's avatar
osku committed
3855 3856 3857 3858 3859 3860 3861 3862
			err = DB_CORRUPTION;

			goto lock_wait_or_error;
		} else {
			/* The user may be dumping a corrupt table. Jump
			over the corruption to recover as much as possible. */

			fprintf(stderr,
3863 3864 3865
				"InnoDB: Index corruption: rec offs %lu"
				" next offs %lu, page no %lu,\n"
				"InnoDB: ",
3866
				(ulong) page_offset(rec),
3867
				(ulong) next_offs,
3868
				(ulong) page_get_page_no(page_align(rec)));
osku's avatar
osku committed
3869 3870
			dict_index_name_print(stderr, trx, index);
			fputs(". We try to skip the rest of the page.\n",
3871
			      stderr);
osku's avatar
osku committed
3872 3873 3874 3875 3876 3877 3878 3879 3880 3881 3882 3883 3884 3885

			btr_pcur_move_to_last_on_page(pcur, &mtr);

			goto next_rec;
		}
	}
	/*-------------------------------------------------------------*/

	/* Calculate the 'offsets' associated with 'rec' */

	offsets = rec_get_offsets(rec, index, offsets, ULINT_UNDEFINED, &heap);

	if (UNIV_UNLIKELY(srv_force_recovery > 0)) {
		if (!rec_validate(rec, offsets)
3886
		    || !btr_index_rec_validate(rec, index, FALSE)) {
osku's avatar
osku committed
3887
			fprintf(stderr,
3888 3889 3890
				"InnoDB: Index corruption: rec offs %lu"
				" next offs %lu, page no %lu,\n"
				"InnoDB: ",
3891
				(ulong) page_offset(rec),
3892
				(ulong) next_offs,
3893
				(ulong) page_get_page_no(page_align(rec)));
osku's avatar
osku committed
3894 3895
			dict_index_name_print(stderr, trx, index);
			fputs(". We try to skip the record.\n",
3896
			      stderr);
osku's avatar
osku committed
3897 3898 3899 3900 3901 3902 3903 3904 3905 3906 3907 3908 3909 3910 3911

			goto next_rec;
		}
	}

	/* Note that we cannot trust the up_match value in the cursor at this
	place because we can arrive here after moving the cursor! Thus
	we have to recompare rec and search_tuple to determine if they
	match enough. */

	if (match_mode == ROW_SEL_EXACT) {
		/* Test if the index record matches completely to search_tuple
		in prebuilt: if not, then we return with DB_RECORD_NOT_FOUND */

		/* fputs("Comparing rec and search tuple\n", stderr); */
3912

osku's avatar
osku committed
3913 3914 3915
		if (0 != cmp_dtuple_rec(search_tuple, rec, offsets)) {

			if (set_also_gap_locks
3916 3917 3918 3919
			    && !(srv_locks_unsafe_for_binlog
				 || trx->isolation_level
				 == TRX_ISO_READ_COMMITTED)
			    && prebuilt->select_lock_type != LOCK_NONE) {
osku's avatar
osku committed
3920

3921
				/* Try to place a gap lock on the index
osku's avatar
osku committed
3922
				record only if innodb_locks_unsafe_for_binlog
3923 3924
				option is not set or this session is not
				using a READ COMMITTED isolation level. */
osku's avatar
osku committed
3925

3926
				err = sel_set_rec_lock(
3927
					btr_pcur_get_block(pcur),
3928 3929 3930
					rec, index, offsets,
					prebuilt->select_lock_type, LOCK_GAP,
					thr);
osku's avatar
osku committed
3931 3932 3933 3934 3935 3936 3937 3938 3939 3940 3941 3942

				if (err != DB_SUCCESS) {

					goto lock_wait_or_error;
				}
			}

			btr_pcur_store_position(pcur, &mtr);

			err = DB_RECORD_NOT_FOUND;
			/* ut_print_name(stderr, index->name);
			fputs(" record not found 3\n", stderr); */
3943

osku's avatar
osku committed
3944 3945 3946 3947 3948 3949
			goto normal_return;
		}

	} else if (match_mode == ROW_SEL_EXACT_PREFIX) {

		if (!cmp_dtuple_is_prefix_of_rec(search_tuple, rec, offsets)) {
3950

osku's avatar
osku committed
3951
			if (set_also_gap_locks
3952 3953 3954 3955
			    && !(srv_locks_unsafe_for_binlog
				 || trx->isolation_level
				 == TRX_ISO_READ_COMMITTED)
			    && prebuilt->select_lock_type != LOCK_NONE) {
osku's avatar
osku committed
3956

3957
				/* Try to place a gap lock on the index
osku's avatar
osku committed
3958
				record only if innodb_locks_unsafe_for_binlog
3959 3960
				option is not set or this session is not
				using a READ COMMITTED isolation level. */
osku's avatar
osku committed
3961

3962
				err = sel_set_rec_lock(
3963
					btr_pcur_get_block(pcur),
3964 3965 3966
					rec, index, offsets,
					prebuilt->select_lock_type, LOCK_GAP,
					thr);
osku's avatar
osku committed
3967 3968 3969 3970 3971 3972 3973 3974 3975 3976 3977 3978 3979 3980 3981 3982 3983 3984 3985 3986 3987 3988 3989 3990 3991 3992

				if (err != DB_SUCCESS) {

					goto lock_wait_or_error;
				}
			}

			btr_pcur_store_position(pcur, &mtr);

			err = DB_RECORD_NOT_FOUND;
			/* ut_print_name(stderr, index->name);
			fputs(" record not found 4\n", stderr); */

			goto normal_return;
		}
	}

	/* We are ready to look at a possible new index entry in the result
	set: the cursor is now placed on a user record */

	if (prebuilt->select_lock_type != LOCK_NONE) {
		/* Try to place a lock on the index record; note that delete
		marked records are a special case in a unique search. If there
		is a non-delete marked record, then it is enough to lock its
		existence with LOCK_REC_NOT_GAP. */

3993 3994 3995
		/* If innodb_locks_unsafe_for_binlog option is used
		or this session is using a READ COMMITED isolation
		level we lock only the record, i.e., next-key locking is
osku's avatar
osku committed
3996 3997 3998 3999 4000
		not used. */

		ulint	lock_type;

		if (!set_also_gap_locks
4001 4002 4003
		    || srv_locks_unsafe_for_binlog
		    || trx->isolation_level == TRX_ISO_READ_COMMITTED
		    || (unique_search
4004
			&& !UNIV_UNLIKELY(rec_get_deleted_flag(rec, comp)))) {
osku's avatar
osku committed
4005 4006 4007 4008 4009 4010 4011 4012 4013 4014 4015 4016 4017 4018 4019 4020 4021 4022

			goto no_gap_lock;
		} else {
			lock_type = LOCK_ORDINARY;
		}

		/* If we are doing a 'greater or equal than a primary key
		value' search from a clustered index, and we find a record
		that has that exact primary key value, then there is no need
		to lock the gap before the record, because no insert in the
		gap can be in our search range. That is, no phantom row can
		appear that way.

		An example: if col1 is the primary key, the search is WHERE
		col1 >= 100, and we find a record where col1 = 100, then no
		need to lock the gap before that record. */

		if (index == clust_index
4023 4024 4025 4026 4027
		    && mode == PAGE_CUR_GE
		    && direction == 0
		    && dtuple_get_n_fields_cmp(search_tuple)
		    == dict_index_get_n_unique(index)
		    && 0 == cmp_dtuple_rec(search_tuple, rec, offsets)) {
osku's avatar
osku committed
4028 4029 4030 4031
no_gap_lock:
			lock_type = LOCK_REC_NOT_GAP;
		}

4032 4033
		err = sel_set_rec_lock(btr_pcur_get_block(pcur),
				       rec, index, offsets,
4034 4035
				       prebuilt->select_lock_type,
				       lock_type, thr);
osku's avatar
osku committed
4036

4037
		switch (err) {
4038
			const rec_t*	old_vers;
4039
		case DB_SUCCESS:
4040 4041 4042 4043 4044 4045
			if (srv_locks_unsafe_for_binlog
			    || trx->isolation_level == TRX_ISO_READ_COMMITTED) {
				/* Note that a record of
				prebuilt->index was locked. */
				prebuilt->new_rec_locks = 1;
			}
4046 4047 4048
			break;
		case DB_LOCK_WAIT:
			if (UNIV_LIKELY(prebuilt->row_read_type
4049
					!= ROW_READ_TRY_SEMI_CONSISTENT)
4050
			    || index != clust_index) {
4051 4052 4053 4054 4055 4056

				goto lock_wait_or_error;
			}

			/* The following call returns 'offsets'
			associated with 'old_vers' */
4057 4058 4059
			err = row_sel_build_committed_vers_for_mysql(
				clust_index, prebuilt, rec,
				&offsets, &heap, &old_vers, &mtr);
4060 4061 4062 4063 4064 4065 4066 4067 4068

			if (err != DB_SUCCESS) {

				goto lock_wait_or_error;
			}

			mutex_enter(&kernel_mutex);
			if (trx->was_chosen_as_deadlock_victim) {
				mutex_exit(&kernel_mutex);
4069
				err = DB_DEADLOCK;
4070 4071 4072 4073

				goto lock_wait_or_error;
			}
			if (UNIV_LIKELY(trx->wait_lock != NULL)) {
4074 4075
				lock_cancel_waiting_and_release(
					trx->wait_lock);
4076
				prebuilt->new_rec_locks = 0;
4077 4078 4079 4080 4081 4082 4083 4084
			} else {
				mutex_exit(&kernel_mutex);

				/* The lock was granted while we were
				searching for the last committed version.
				Do a normal locking read. */

				offsets = rec_get_offsets(rec, index, offsets,
4085 4086
							  ULINT_UNDEFINED,
							  &heap);
4087
				err = DB_SUCCESS;
4088 4089 4090
				/* Note that a record of
				prebuilt->index was locked. */
				prebuilt->new_rec_locks = 1;
4091 4092 4093 4094 4095 4096 4097 4098 4099 4100 4101 4102 4103 4104
				break;
			}
			mutex_exit(&kernel_mutex);

			if (old_vers == NULL) {
				/* The row was not yet committed */

				goto next_rec;
			}

			did_semi_consistent_read = TRUE;
			rec = old_vers;
			break;
		default:
osku's avatar
osku committed
4105 4106 4107 4108 4109 4110 4111 4112 4113 4114 4115

			goto lock_wait_or_error;
		}
	} else {
		/* This is a non-locking consistent read: if necessary, fetch
		a previous version of the record */

		if (trx->isolation_level == TRX_ISO_READ_UNCOMMITTED) {

			/* Do nothing: we let a non-locking SELECT read the
			latest version of the record */
4116

osku's avatar
osku committed
4117
		} else if (index == clust_index) {
4118

osku's avatar
osku committed
4119 4120 4121 4122 4123 4124
			/* Fetch a previous version of the row if the current
			one is not visible in the snapshot; if we have a very
			high force recovery level set, we try to avoid crashes
			by skipping this lookup */

			if (UNIV_LIKELY(srv_force_recovery < 5)
4125 4126
			    && !lock_clust_rec_cons_read_sees(
				    rec, index, offsets, trx->read_view)) {
osku's avatar
osku committed
4127

4128
				rec_t*	old_vers;
osku's avatar
osku committed
4129 4130
				/* The following call returns 'offsets'
				associated with 'old_vers' */
4131 4132 4133 4134
				err = row_sel_build_prev_vers_for_mysql(
					trx->read_view, clust_index,
					prebuilt, rec, &offsets, &heap,
					&old_vers, &mtr);
4135

osku's avatar
osku committed
4136 4137 4138 4139 4140 4141 4142 4143 4144 4145 4146 4147 4148 4149
				if (err != DB_SUCCESS) {

					goto lock_wait_or_error;
				}

				if (old_vers == NULL) {
					/* The row did not exist yet in
					the read view */

					goto next_rec;
				}

				rec = old_vers;
			}
4150
		} else if (!lock_sec_rec_cons_read_sees(rec, trx->read_view)) {
osku's avatar
osku committed
4151 4152 4153 4154 4155 4156 4157 4158 4159 4160 4161 4162 4163 4164 4165 4166 4167 4168 4169 4170 4171
			/* We are looking into a non-clustered index,
			and to get the right version of the record we
			have to look also into the clustered index: this
			is necessary, because we can only get the undo
			information via the clustered index record. */

			ut_ad(index != clust_index);

			goto requires_clust_rec;
		}
	}

	/* NOTE that at this point rec can be an old version of a clustered
	index record built for a consistent read. We cannot assume after this
	point that rec is on a buffer pool page. Functions like
	page_rec_is_comp() cannot be used! */

	if (UNIV_UNLIKELY(rec_get_deleted_flag(rec, comp))) {

		/* The record is delete-marked: we can skip it */

4172
		if ((srv_locks_unsafe_for_binlog
4173 4174 4175
		     || trx->isolation_level == TRX_ISO_READ_COMMITTED)
		    && prebuilt->select_lock_type != LOCK_NONE
		    && !did_semi_consistent_read) {
osku's avatar
osku committed
4176 4177 4178 4179 4180 4181

			/* No need to keep a lock on a delete-marked record
			if we do not want to use next-key locking. */

			row_unlock_for_mysql(prebuilt, TRUE);
		}
4182

4183 4184 4185 4186 4187 4188 4189 4190 4191 4192 4193 4194 4195 4196 4197 4198 4199
		/* This is an optimization to skip setting the next key lock
		on the record that follows this delete-marked record. This
		optimization works because of the unique search criteria
		which precludes the presence of a range lock between this
		delete marked record and the record following it.

		For now this is applicable only to clustered indexes while
		doing a unique search. There is scope for further optimization
		applicable to unique secondary indexes. Current behaviour is
		to widen the scope of a lock on an already delete marked record
		if the same record is deleted twice by the same transaction */
		if (index == clust_index && unique_search) {
			err = DB_RECORD_NOT_FOUND;

			goto normal_return;
		}

osku's avatar
osku committed
4200 4201 4202 4203 4204 4205 4206 4207 4208 4209 4210 4211 4212 4213 4214 4215 4216 4217 4218 4219 4220 4221 4222
		goto next_rec;
	}

	/* Get the clustered index record if needed, if we did not do the
	search using the clustered index. */

	if (index != clust_index && prebuilt->need_to_access_clustered) {

requires_clust_rec:
		/* We use a 'goto' to the preceding label if a consistent
		read of a secondary index record requires us to look up old
		versions of the associated clustered index record. */

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

		/* It was a non-clustered index and we must fetch also the
		clustered index record */

		mtr_has_extra_clust_latch = TRUE;

		/* The following call returns 'offsets' associated with
		'clust_rec'. Note that 'clust_rec' can be an old version
		built for a consistent read. */
4223

osku's avatar
osku committed
4224
		err = row_sel_get_clust_rec_for_mysql(prebuilt, index, rec,
4225 4226
						      thr, &clust_rec,
						      &offsets, &heap, &mtr);
osku's avatar
osku committed
4227 4228 4229 4230 4231 4232 4233 4234 4235 4236 4237 4238
		if (err != DB_SUCCESS) {

			goto lock_wait_or_error;
		}

		if (clust_rec == NULL) {
			/* The record did not exist in the read view */
			ut_ad(prebuilt->select_lock_type == LOCK_NONE);

			goto next_rec;
		}

4239 4240 4241 4242 4243 4244 4245 4246 4247
		if ((srv_locks_unsafe_for_binlog
		     || trx->isolation_level == TRX_ISO_READ_COMMITTED)
		    && prebuilt->select_lock_type != LOCK_NONE) {
			/* Note that both the secondary index record
			and the clustered index record were locked. */
			ut_ad(prebuilt->new_rec_locks == 1);
			prebuilt->new_rec_locks = 2;
		}

osku's avatar
osku committed
4248 4249 4250 4251
		if (UNIV_UNLIKELY(rec_get_deleted_flag(clust_rec, comp))) {

			/* The record is delete marked: we can skip it */

4252
			if ((srv_locks_unsafe_for_binlog
4253 4254
			     || trx->isolation_level == TRX_ISO_READ_COMMITTED)
			    && prebuilt->select_lock_type != LOCK_NONE) {
osku's avatar
osku committed
4255 4256 4257 4258 4259 4260 4261 4262 4263 4264

				/* No need to keep a lock on a delete-marked
				record if we do not want to use next-key
				locking. */

				row_unlock_for_mysql(prebuilt, TRUE);
			}

			goto next_rec;
		}
4265

osku's avatar
osku committed
4266 4267
		if (prebuilt->need_to_access_clustered) {

4268
			result_rec = clust_rec;
osku's avatar
osku committed
4269 4270

			ut_ad(rec_offs_validate(result_rec, clust_index,
4271
						offsets));
osku's avatar
osku committed
4272 4273 4274 4275
		} else {
			/* We used 'offsets' for the clust rec, recalculate
			them for 'rec' */
			offsets = rec_get_offsets(rec, index, offsets,
4276
						  ULINT_UNDEFINED, &heap);
osku's avatar
osku committed
4277 4278 4279 4280 4281 4282 4283 4284 4285 4286 4287 4288 4289
			result_rec = rec;
		}
	} else {
		result_rec = rec;
	}

	/* We found a qualifying record 'result_rec'. At this point,
	'offsets' are associated with 'result_rec'. */

	ut_ad(rec_offs_validate(result_rec,
				result_rec != rec ? clust_index : index,
				offsets));

4290 4291 4292 4293
	/* At this point, the clustered index record is protected
	by a page latch that was acquired when pcur was positioned.
	The latch will not be released until mtr_commit(&mtr). */

osku's avatar
osku committed
4294
	if ((match_mode == ROW_SEL_EXACT
4295 4296 4297 4298 4299 4300 4301
	     || prebuilt->n_rows_fetched >= MYSQL_FETCH_CACHE_THRESHOLD)
	    && prebuilt->select_lock_type == LOCK_NONE
	    && !prebuilt->templ_contains_blob
	    && !prebuilt->clust_index_was_generated
	    && !prebuilt->used_in_HANDLER
	    && prebuilt->template_type
	    != ROW_MYSQL_DUMMY_TEMPLATE) {
osku's avatar
osku committed
4302 4303 4304 4305 4306 4307 4308 4309 4310 4311 4312

		/* Inside an update, for example, we do not cache rows,
		since we may use the cursor position to do the actual
		update, that is why we require ...lock_type == LOCK_NONE.
		Since we keep space in prebuilt only for the BLOBs of
		a single row, we cannot cache rows in the case there
		are BLOBs in the fields to be fetched. In HANDLER we do
		not cache rows because there the cursor is a scrollable
		cursor. */

		row_sel_push_cache_row_for_mysql(prebuilt, result_rec,
4313
						 offsets);
osku's avatar
osku committed
4314
		if (prebuilt->n_fetch_cached == MYSQL_FETCH_CACHE_SIZE) {
4315

osku's avatar
osku committed
4316 4317 4318 4319 4320 4321 4322
			goto got_row;
		}

		goto next_rec;
	} else {
		if (prebuilt->template_type == ROW_MYSQL_DUMMY_TEMPLATE) {
			memcpy(buf + 4, result_rec
4323 4324
			       - rec_offs_extra_size(offsets),
			       rec_offs_size(offsets));
osku's avatar
osku committed
4325 4326 4327 4328
			mach_write_to_4(buf,
					rec_offs_extra_size(offsets) + 4);
		} else {
			if (!row_sel_store_mysql_rec(buf, prebuilt,
4329
						     result_rec, offsets)) {
osku's avatar
osku committed
4330 4331 4332 4333 4334 4335 4336 4337
				err = DB_TOO_BIG_RECORD;

				goto lock_wait_or_error;
			}
		}

		if (prebuilt->clust_index_was_generated) {
			if (result_rec != rec) {
4338 4339 4340
				offsets = rec_get_offsets(
					rec, index, offsets, ULINT_UNDEFINED,
					&heap);
osku's avatar
osku committed
4341 4342
			}
			row_sel_store_row_id_to_prebuilt(prebuilt, rec,
4343
							 index, offsets);
osku's avatar
osku committed
4344 4345 4346 4347 4348 4349 4350 4351 4352
		}
	}

	/* From this point on, 'offsets' are invalid. */

got_row:
	/* We have an optimization to save CPU time: if this is a consistent
	read on a unique condition on the clustered index, then we do not
	store the pcur position, because any fetch next or prev will anyway
4353 4354
	return 'end of file'. Exceptions are locking reads and the MySQL
	HANDLER command where the user can move the cursor with PREV or NEXT
4355
	even after a unique search. */
osku's avatar
osku committed
4356 4357

	if (!unique_search_from_clust_index
4358 4359
	    || prebuilt->select_lock_type != LOCK_NONE
	    || prebuilt->used_in_HANDLER) {
osku's avatar
osku committed
4360 4361 4362 4363 4364 4365 4366 4367 4368 4369 4370

		/* Inside an update always store the cursor position */

		btr_pcur_store_position(pcur, &mtr);
	}

	err = DB_SUCCESS;

	goto normal_return;

next_rec:
4371 4372
	/* Reset the old and new "did semi-consistent read" flags. */
	if (UNIV_UNLIKELY(prebuilt->row_read_type
4373
			  == ROW_READ_DID_SEMI_CONSISTENT)) {
4374 4375 4376
		prebuilt->row_read_type = ROW_READ_TRY_SEMI_CONSISTENT;
	}
	did_semi_consistent_read = FALSE;
4377
	prebuilt->new_rec_locks = 0;
4378

osku's avatar
osku committed
4379 4380 4381 4382 4383 4384 4385 4386 4387 4388 4389 4390 4391
	/*-------------------------------------------------------------*/
	/* PHASE 5: Move the cursor to the next index record */

	if (UNIV_UNLIKELY(mtr_has_extra_clust_latch)) {
		/* We must commit mtr if we are moving to the next
		non-clustered index record, because we could break the
		latching order if we would access a different clustered
		index page right away without releasing the previous. */

		btr_pcur_store_position(pcur, &mtr);

		mtr_commit(&mtr);
		mtr_has_extra_clust_latch = FALSE;
4392

osku's avatar
osku committed
4393 4394
		mtr_start(&mtr);
		if (sel_restore_position_for_mysql(&same_user_rec,
4395 4396
						   BTR_SEARCH_LEAF,
						   pcur, moves_up, &mtr)) {
osku's avatar
osku committed
4397 4398 4399 4400 4401 4402 4403 4404
#ifdef UNIV_SEARCH_DEBUG
			cnt++;
#endif /* UNIV_SEARCH_DEBUG */

			goto rec_loop;
		}
	}

4405
	if (moves_up) {
osku's avatar
osku committed
4406 4407 4408 4409 4410 4411 4412 4413 4414 4415 4416 4417 4418 4419 4420 4421 4422 4423 4424 4425 4426 4427 4428 4429 4430
		if (UNIV_UNLIKELY(!btr_pcur_move_to_next(pcur, &mtr))) {
not_moved:
			btr_pcur_store_position(pcur, &mtr);

			if (match_mode != 0) {
				err = DB_RECORD_NOT_FOUND;
			} else {
				err = DB_END_OF_INDEX;
			}

			goto normal_return;
		}
	} else {
		if (UNIV_UNLIKELY(!btr_pcur_move_to_prev(pcur, &mtr))) {
			goto not_moved;
		}
	}

#ifdef UNIV_SEARCH_DEBUG
	cnt++;
#endif /* UNIV_SEARCH_DEBUG */

	goto rec_loop;

lock_wait_or_error:
4431 4432
	/* Reset the old and new "did semi-consistent read" flags. */
	if (UNIV_UNLIKELY(prebuilt->row_read_type
4433
			  == ROW_READ_DID_SEMI_CONSISTENT)) {
4434 4435 4436 4437
		prebuilt->row_read_type = ROW_READ_TRY_SEMI_CONSISTENT;
	}
	did_semi_consistent_read = FALSE;

osku's avatar
osku committed
4438 4439 4440 4441 4442 4443
	/*-------------------------------------------------------------*/

	btr_pcur_store_position(pcur, &mtr);

	mtr_commit(&mtr);
	mtr_has_extra_clust_latch = FALSE;
4444

osku's avatar
osku committed
4445 4446 4447 4448 4449 4450 4451 4452 4453 4454 4455 4456 4457 4458 4459
	trx->error_state = err;

	/* The following is a patch for MySQL */

	que_thr_stop_for_mysql(thr);

	thr->lock_state = QUE_THR_LOCK_ROW;

	if (row_mysql_handle_errors(&err, trx, thr, NULL)) {
		/* It was a lock wait, and it ended */

		thr->lock_state = QUE_THR_LOCK_NOLOCK;
		mtr_start(&mtr);

		sel_restore_position_for_mysql(&same_user_rec,
4460 4461
					       BTR_SEARCH_LEAF, pcur,
					       moves_up, &mtr);
4462 4463

		if ((srv_locks_unsafe_for_binlog
4464 4465
		     || trx->isolation_level == TRX_ISO_READ_COMMITTED)
		    && !same_user_rec) {
4466

osku's avatar
osku committed
4467 4468 4469 4470 4471 4472 4473 4474 4475 4476 4477 4478 4479 4480 4481 4482
			/* Since we were not able to restore the cursor
			on the same user record, we cannot use
			row_unlock_for_mysql() to unlock any records, and
			we must thus reset the new rec lock info. Since
			in lock0lock.c we have blocked the inheriting of gap
			X-locks, we actually do not have any new record locks
			set in this case.

			Note that if we were able to restore on the 'same'
			user record, it is still possible that we were actually
			waiting on a delete-marked record, and meanwhile
			it was removed by purge and inserted again by some
			other user. But that is no problem, because in
			rec_loop we will again try to set a lock, and
			new_rec_lock_info in trx will be right at the end. */

4483
			prebuilt->new_rec_locks = 0;
osku's avatar
osku committed
4484
		}
4485

osku's avatar
osku committed
4486 4487 4488 4489 4490 4491 4492 4493
		mode = pcur->search_mode;

		goto rec_loop;
	}

	thr->lock_state = QUE_THR_LOCK_NOLOCK;

#ifdef UNIV_SEARCH_DEBUG
4494
	/*	fputs("Using ", stderr);
osku's avatar
osku committed
4495 4496 4497 4498 4499 4500 4501 4502 4503 4504 4505 4506 4507 4508 4509 4510 4511 4512
	dict_index_name_print(stderr, index);
	fprintf(stderr, " cnt %lu ret value %lu err\n", cnt, err); */
#endif /* UNIV_SEARCH_DEBUG */
	goto func_exit;

normal_return:
	/*-------------------------------------------------------------*/
	que_thr_stop_for_mysql_no_error(thr, trx);

	mtr_commit(&mtr);

	if (prebuilt->n_fetch_cached > 0) {
		row_sel_pop_cached_row_for_mysql(buf, prebuilt);

		err = DB_SUCCESS;
	}

#ifdef UNIV_SEARCH_DEBUG
4513
	/*	fputs("Using ", stderr);
osku's avatar
osku committed
4514 4515 4516 4517 4518 4519 4520 4521 4522 4523 4524 4525
	dict_index_name_print(stderr, index);
	fprintf(stderr, " cnt %lu ret value %lu err\n", cnt, err); */
#endif /* UNIV_SEARCH_DEBUG */
	if (err == DB_SUCCESS) {
		srv_n_rows_read++;
	}

func_exit:
	trx->op_info = "";
	if (UNIV_LIKELY_NULL(heap)) {
		mem_heap_free(heap);
	}
4526 4527 4528 4529 4530

	/* Set or reset the "did semi-consistent read" flag on return.
	The flag did_semi_consistent_read is set if and only if
	the record being returned was fetched with a semi-consistent read. */
	ut_ad(prebuilt->row_read_type != ROW_READ_WITH_LOCKS
4531
	      || !did_semi_consistent_read);
4532 4533 4534 4535 4536 4537 4538 4539

	if (UNIV_UNLIKELY(prebuilt->row_read_type != ROW_READ_WITH_LOCKS)) {
		if (UNIV_UNLIKELY(did_semi_consistent_read)) {
			prebuilt->row_read_type = ROW_READ_DID_SEMI_CONSISTENT;
		} else {
			prebuilt->row_read_type = ROW_READ_TRY_SEMI_CONSISTENT;
		}
	}
osku's avatar
osku committed
4540 4541 4542 4543 4544
	return(err);
}

/***********************************************************************
Checks if MySQL at the moment is allowed for this table to retrieve a
4545 4546
consistent read result, or store it to the query cache.
@return	TRUE if storing or retrieving from the query cache is permitted */
4547
UNIV_INTERN
osku's avatar
osku committed
4548 4549 4550
ibool
row_search_check_if_query_cache_permitted(
/*======================================*/
4551 4552
	trx_t*		trx,		/*!< in: transaction object */
	const char*	norm_name)	/*!< in: concatenation of database name,
osku's avatar
osku committed
4553 4554 4555
					'/' char, table name */
{
	dict_table_t*	table;
4556
	ibool		ret	= FALSE;
osku's avatar
osku committed
4557

4558
	table = dict_table_get(norm_name, FALSE);
osku's avatar
osku committed
4559 4560 4561 4562 4563 4564 4565 4566 4567 4568 4569 4570 4571 4572 4573 4574 4575 4576

	if (table == NULL) {

		return(FALSE);
	}

	mutex_enter(&kernel_mutex);

	/* Start the transaction if it is not started yet */

	trx_start_if_not_started_low(trx);

	/* If there are locks on the table or some trx has invalidated the
	cache up to our trx id, then ret = FALSE.
	We do not check what type locks there are on the table, though only
	IX type locks actually would require ret = FALSE. */

	if (UT_LIST_GET_LEN(table->locks) == 0
4577 4578
	    && ut_dulint_cmp(trx->id,
			     table->query_cache_inv_trx_id) >= 0) {
osku's avatar
osku committed
4579 4580

		ret = TRUE;
4581

osku's avatar
osku committed
4582 4583 4584 4585
		/* If the isolation level is high, assign a read view for the
		transaction if it does not yet have one */

		if (trx->isolation_level >= TRX_ISO_REPEATABLE_READ
4586
		    && !trx->read_view) {
osku's avatar
osku committed
4587

4588 4589
			trx->read_view = read_view_open_now(
				trx->id, trx->global_read_view_heap);
osku's avatar
osku committed
4590 4591 4592
			trx->global_read_view = trx->read_view;
		}
	}
4593

osku's avatar
osku committed
4594 4595 4596 4597
	mutex_exit(&kernel_mutex);

	return(ret);
}
4598 4599

/***********************************************************************
4600
Read the AUTOINC column from the current row. If the value is less than
4601 4602
0 and the type is not unsigned then we reset the value to 0.
@return	value read from the column */
4603
static
4604
ib_uint64_t
4605 4606
row_search_autoinc_read_column(
/*===========================*/
4607 4608 4609 4610
	dict_index_t*	index,		/*!< in: index to read from */
	const rec_t*	rec,		/*!< in: current rec */
	ulint		col_no,		/*!< in: column number */
	ibool		unsigned_type)	/*!< in: signed or unsigned flag */
4611 4612 4613
{
	ulint		len;
	const byte*	data;
4614
	ib_uint64_t	value;
4615 4616 4617 4618
	mem_heap_t*	heap = NULL;
	ulint		offsets_[REC_OFFS_NORMAL_SIZE];
	ulint*		offsets	= offsets_;

4619
	rec_offs_init(offsets_);
4620

4621
	offsets = rec_get_offsets(rec, index, offsets, ULINT_UNDEFINED, &heap);
4622

4623
	data = rec_get_nth_field(rec, offsets, col_no, &len);
4624 4625 4626 4627

	ut_a(len != UNIV_SQL_NULL);
	ut_a(len <= sizeof value);

vasil's avatar
vasil committed
4628
	/* we assume AUTOINC value cannot be negative */
4629
	value = mach_read_int_type(data, len, unsigned_type);
4630 4631 4632 4633 4634

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

4635
	if (!unsigned_type && (ib_int64_t) value < 0) {
4636 4637
		value = 0;
	}
4638 4639 4640 4641 4642

	return(value);
}

/***********************************************************************
4643 4644
Get the last row.
@return	current rec or NULL */
4645 4646 4647 4648
static
const rec_t*
row_search_autoinc_get_rec(
/*=======================*/
4649 4650
	btr_pcur_t*	pcur,		/*!< in: the current cursor */
	mtr_t*		mtr)		/*!< in: mini transaction */
4651 4652 4653 4654 4655 4656 4657 4658 4659 4660 4661 4662 4663
{
	do {
		const rec_t* rec = btr_pcur_get_rec(pcur);

		if (page_rec_is_user_rec(rec)) {
			return(rec);
		}
	} while (btr_pcur_move_to_prev(pcur, mtr));

	return(NULL);
}

/***********************************************************************
4664 4665
Read the max AUTOINC value from an index.
@return	DB_SUCCESS if all OK else error code, DB_RECORD_NOT_FOUND if column name can't be found in index */
4666
UNIV_INTERN
4667 4668 4669
ulint
row_search_max_autoinc(
/*===================*/
4670 4671 4672
	dict_index_t*	index,		/*!< in: index to search */
	const char*	col_name,	/*!< in: name of autoinc column */
	ib_uint64_t*	value)		/*!< out: AUTOINC value read */
4673 4674 4675 4676 4677 4678 4679 4680 4681 4682 4683 4684 4685 4686 4687 4688 4689 4690 4691 4692 4693 4694 4695 4696 4697 4698 4699 4700 4701 4702 4703 4704 4705 4706 4707 4708 4709 4710 4711 4712 4713 4714 4715 4716 4717 4718 4719 4720 4721 4722 4723 4724 4725 4726
{
	ulint		i;
	ulint		n_cols;
	dict_field_t*	dfield = NULL;
	ulint		error = DB_SUCCESS;

	n_cols = dict_index_get_n_ordering_defined_by_user(index);

	/* Search the index for the AUTOINC column name */
	for (i = 0; i < n_cols; ++i) {
		dfield = dict_index_get_nth_field(index, i);

		if (strcmp(col_name, dfield->name) == 0) {
			break;
		}
	}

	*value = 0;

	/* Must find the AUTOINC column name */
	if (i < n_cols && dfield) {
		mtr_t		mtr;
		btr_pcur_t	pcur;

		mtr_start(&mtr);

		/* Open at the high/right end (FALSE), and INIT
		cursor (TRUE) */
		btr_pcur_open_at_index_side(
			FALSE, index, BTR_SEARCH_LEAF, &pcur, TRUE, &mtr);

		if (page_get_n_recs(btr_pcur_get_page(&pcur)) > 0) {
			const rec_t*	rec;

			rec = row_search_autoinc_get_rec(&pcur, &mtr);

			if (rec != NULL) {
				ibool unsigned_type = (
					dfield->col->prtype & DATA_UNSIGNED);

				*value = row_search_autoinc_read_column(
					index, rec, i, unsigned_type);
			}
		}

		btr_pcur_close(&pcur);

		mtr_commit(&mtr);
	} else {
		error = DB_RECORD_NOT_FOUND;
	}

	return(error);
}