/***************************************************************************** Copyright (c) 1997, 2017, Oracle and/or its affiliates. All Rights Reserved. Copyright (c) 2008, Google Inc. Copyright (c) 2015, 2023, MariaDB Corporation. 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., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1335 USA *****************************************************************************/ /***************************************************//** @file row/row0sel.cc Select Created 12/19/1997 Heikki Tuuri *******************************************************/ #include "row0sel.h" #include "dict0dict.h" #include "dict0boot.h" #include "trx0undo.h" #include "trx0trx.h" #include "trx0purge.h" #include "trx0rec.h" #include "btr0btr.h" #include "btr0cur.h" #include "btr0sea.h" #include "gis0rtree.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 "buf0lru.h" #include "srv0srv.h" #include "srv0mon.h" #include "sql_error.h" #ifdef WITH_WSREP #include "mysql/service_wsrep.h" /* For wsrep_thd_skip_locking */ #endif /* 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 /********************************************************************//** 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 fields are compared with collation! @return whether the columns are equal */ static bool row_sel_sec_rec_is_for_blob( /*========================*/ ulint mtype, /*!< in: main type */ ulint prtype, /*!< in: precise type */ ulint mbminlen, /*!< in: minimum length of a character, in bytes */ ulint mbmaxlen, /*!< in: maximum length of a character, in bytes */ const byte* clust_field, /*!< in: the locally stored part of 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) */ 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 prefix_len, /*!< in: index column prefix length in bytes, or 0 for full column */ dict_table_t* table) /*!< in: table */ { ulint len; byte buf[REC_VERSION_56_MAX_INDEX_COL_LEN + 1]; /* This function should never be invoked on tables in ROW_FORMAT=REDUNDANT or ROW_FORMAT=COMPACT, because they should always contain enough prefix in the clustered index record. */ ut_ad(dict_table_has_atomic_blobs(table)); ut_a(clust_len >= BTR_EXTERN_FIELD_REF_SIZE); ut_ad(!prefix_len || prefix_len >= sec_len); ut_a(prefix_len <= sizeof buf); if (!memcmp(clust_field + clust_len - BTR_EXTERN_FIELD_REF_SIZE, field_ref_zero, BTR_EXTERN_FIELD_REF_SIZE)) { /* The externally stored field was not written yet. This record should only be seen by trx_rollback_recovered() or any TRX_ISO_READ_UNCOMMITTED transactions. */ return false; } len = btr_copy_externally_stored_field_prefix( buf, prefix_len ? prefix_len : sizeof buf, table->space->zip_size(), clust_field, clust_len); if (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; } if (prefix_len) { len = dtype_get_at_most_n_mbchars(prtype, mbminlen, mbmaxlen, prefix_len, len, reinterpret_cast (buf)); } else if (len >= sizeof buf) { ut_ad("too long column" == 0); return false; } return(!cmp_data_data(mtype, prtype, buf, len, sec_field, sec_len)); } /** Function to read the secondary spatial index, calculate the minimum bounding rectangle for clustered index record and secondary index record and compare it. @param sec_rec secondary index record @param sec_index spatial secondary index @param clust_rec clustered index record @param clust_index clustered index @retval DB_SUCCESS_LOCKED_REC if the secondary record is equal to the corresponding fields in the clustered record, when compared with collation; @retval DB_SUCCESS if not equal */ static dberr_t row_sel_spatial_sec_rec_is_for_clust_rec( const rec_t *sec_rec, const dict_index_t *sec_index, const rec_t *clust_rec, dict_index_t *clust_index) { mem_heap_t *heap= mem_heap_create(256); rec_offs clust_offsets_[REC_OFFS_NORMAL_SIZE]; rec_offs *clust_offs= clust_offsets_; ulint clust_len; rec_offs_init(clust_offsets_); ulint clust_pos= dict_col_get_clust_pos( dict_index_get_nth_col(sec_index, 0), clust_index); clust_offs= rec_get_offsets(clust_rec, clust_index, clust_offs, clust_index->n_core_fields, clust_pos + 1, &heap); ut_ad(sec_index->n_user_defined_cols == 1); const byte *clust_field= rec_get_nth_field(clust_rec, clust_offs, clust_pos, &clust_len); if (clust_len == UNIV_SQL_NULL || clust_len < GEO_DATA_HEADER_SIZE) { ut_ad("corrupted geometry column" == 0); err_exit: mem_heap_free(heap); return DB_SUCCESS; } /* For externally stored field, we need to get full geo data to generate the MBR for comparing. */ if (rec_offs_nth_extern(clust_offs, clust_pos)) { clust_field= btr_copy_externally_stored_field( &clust_len, clust_field, sec_index->table->space->zip_size(), clust_len, heap); if (clust_field == NULL) { ut_ad("corrupted geometry blob" == 0); goto err_exit; } } ut_ad(clust_len >= GEO_DATA_HEADER_SIZE); rtr_mbr_t tmp_mbr; rtr_mbr_t sec_mbr; rtree_mbr_from_wkb( clust_field + GEO_DATA_HEADER_SIZE, static_cast(clust_len - GEO_DATA_HEADER_SIZE), SPDIMS, reinterpret_cast(&tmp_mbr)); rtr_read_mbr(sec_rec, &sec_mbr); mem_heap_free(heap); return MBR_EQUAL_CMP(&sec_mbr, &tmp_mbr) ? DB_SUCCESS_LOCKED_REC : DB_SUCCESS; } /** 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 fields are compared with collation! @param[in] sec_rec secondary index record @param[in] sec_index secondary index @param[in] clust_rec clustered index record; must be protected by a page s-latch @param[in] clust_index clustered index @param[in] thr query thread @retval DB_COMPUTE_VALUE_FAILED in case of virtual column value computation failure. @retval DB_SUCCESS_LOCKED_REC if the secondary record is equal to the corresponding fields in the clustered record, when compared with collation; @retval DB_SUCCESS if not equal or if the clustered record has been marked for deletion */ static dberr_t row_sel_sec_rec_is_for_clust_rec( const rec_t* sec_rec, dict_index_t* sec_index, const rec_t* clust_rec, dict_index_t* clust_index, que_thr_t* thr) { if (rec_get_deleted_flag(clust_rec, dict_table_is_comp(clust_index->table))) { /* In delete-marked records, DB_TRX_ID must always refer to an existing undo log record. */ ut_ad(rec_get_trx_id(clust_rec, clust_index)); /* 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 DB_SUCCESS; } if (dict_index_is_spatial(sec_index)) { return row_sel_spatial_sec_rec_is_for_clust_rec( sec_rec, sec_index, clust_rec, clust_index); } const byte* sec_field; ulint sec_len; const byte* clust_field; ulint n; ulint i; mem_heap_t* heap = mem_heap_create(256); rec_offs clust_offsets_[REC_OFFS_NORMAL_SIZE]; rec_offs sec_offsets_[REC_OFFS_SMALL_SIZE]; rec_offs* clust_offs = clust_offsets_; rec_offs* sec_offs = sec_offsets_; rec_offs_init(clust_offsets_); rec_offs_init(sec_offsets_); ib_vcol_row vc(heap); clust_offs = rec_get_offsets(clust_rec, clust_index, clust_offs, clust_index->n_core_fields, ULINT_UNDEFINED, &heap); sec_offs = rec_get_offsets(sec_rec, sec_index, sec_offs, sec_index->n_fields, ULINT_UNDEFINED, &heap); n = dict_index_get_n_ordering_defined_by_user(sec_index); for (i = 0; i < n; i++) { const dict_field_t* ifield; const dict_col_t* col; ulint clust_pos = 0; ulint clust_len = 0; ulint len; ifield = dict_index_get_nth_field(sec_index, i); col = dict_field_get_col(ifield); sec_field = rec_get_nth_field(sec_rec, sec_offs, i, &sec_len); const bool is_virtual = col->is_virtual(); /* For virtual column, its value will need to be reconstructed from base column in cluster index */ if (is_virtual) { const dict_v_col_t* v_col; dfield_t* vfield; row_ext_t* ext; byte *record = vc.record(thr_get_trx(thr)->mysql_thd, clust_index, &thr->prebuilt->m_mysql_table); v_col = reinterpret_cast(col); dtuple_t* row = row_build( ROW_COPY_POINTERS, clust_index, clust_rec, clust_offs, NULL, NULL, NULL, &ext, heap); vfield = innobase_get_computed_value( row, v_col, clust_index, &heap, NULL, NULL, thr_get_trx(thr)->mysql_thd, thr->prebuilt->m_mysql_table, record, NULL, NULL, true); if (vfield == NULL) { innobase_report_computed_value_failed(row); return DB_COMPUTE_VALUE_FAILED; } len = clust_len = vfield->len; clust_field = static_cast(vfield->data); } else { clust_pos = dict_col_get_clust_pos(col, clust_index); clust_field = rec_get_nth_cfield( clust_rec, clust_index, clust_offs, clust_pos, &clust_len); if (clust_len == UNIV_SQL_NULL) { if (sec_len == UNIV_SQL_NULL) { continue; } return DB_SUCCESS; } if (sec_len == UNIV_SQL_NULL) { return DB_SUCCESS; } len = clust_len; ulint prefix_len = ifield->prefix_len; if (rec_offs_nth_extern(clust_offs, clust_pos)) { /* BLOB can contain prefix. */ len -= BTR_EXTERN_FIELD_REF_SIZE; if (!len) { goto compare_blobs; } } if (prefix_len) { len = dtype_get_at_most_n_mbchars( col->prtype, col->mbminlen, col->mbmaxlen, prefix_len, len, reinterpret_cast( clust_field)); if (len < sec_len) { goto check_for_blob; } } else { check_for_blob: if (rec_offs_nth_extern(clust_offs, clust_pos)) { compare_blobs: if (!row_sel_sec_rec_is_for_blob( col->mtype, col->prtype, col->mbminlen, col->mbmaxlen, clust_field, clust_len, sec_field, sec_len, prefix_len, clust_index->table)) { return DB_SUCCESS; } continue; } } } if (0 != cmp_data_data(col->mtype, col->prtype, clust_field, len, sec_field, sec_len)) { return DB_SUCCESS; } } return DB_SUCCESS_LOCKED_REC; } /*********************************************************************//** Creates a select node struct. @return own: select node struct */ sel_node_t* sel_node_create( /*============*/ mem_heap_t* heap) /*!< in: memory heap where created */ { sel_node_t* node; node = static_cast( mem_heap_alloc(heap, sizeof(sel_node_t))); node->common.type = QUE_NODE_SELECT; node->state = SEL_NODE_OPEN; node->plans = NULL; 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. */ void sel_node_free_private( /*==================*/ sel_node_t* node) /*!< in: select node struct */ { 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( /*=================*/ sel_node_t* node) /*!< in: select node */ { 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( /*=======================*/ sym_node_t* var, /*!< in: first variable in a list of variables */ sel_node_t* node) /*!< in: select node */ { que_node_t* exp; if (var == NULL) { return; } for (exp = node->select_list; var != 0; var = static_cast(que_node_get_next(var))) { ut_ad(exp); eval_node_copy_val(var->alias, exp); exp = que_node_get_next(exp); } } /*********************************************************************//** Resets the aggregate value totals in the select list of an aggregate type query. */ UNIV_INLINE void sel_reset_aggregate_vals( /*=====================*/ sel_node_t* node) /*!< in: select node */ { func_node_t* func_node; ut_ad(node->is_aggregate); for (func_node = static_cast(node->select_list); func_node != 0; func_node = static_cast( que_node_get_next(func_node))) { eval_node_set_int_val(func_node, 0); } 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( /*=============================*/ sel_node_t* node) /*!< in: select node */ { 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( /*==================*/ dict_index_t* index, /*!< in: record index */ const rec_t* rec, /*!< in: record in a clustered or non-clustered index; must be protected by a page latch */ const rec_offs* offsets,/*!< in: rec_get_offsets(rec, index) */ sym_node_t* column) /*!< in: first column in a column list, or NULL */ { dfield_t* val; ulint index_type; ulint field_no; const byte* data; ulint len; ut_ad(rec_offs_validate(rec, index, offsets)); if (dict_index_is_clust(index)) { index_type = SYM_CLUST_FIELD_NO; } else { index_type = SYM_SEC_FIELD_NO; } while (column) { mem_heap_t* heap = NULL; ibool needs_copy; field_no = column->field_nos[index_type]; if (field_no != ULINT_UNDEFINED) { if (UNIV_UNLIKELY(rec_offs_nth_extern( offsets, field_no) != 0)) { /* Copy an externally stored field to the temporary heap, if possible. */ heap = mem_heap_create(1); data = btr_rec_copy_externally_stored_field( rec, offsets, index->table->space->zip_size(), field_no, &len, heap); /* data == NULL means that the externally stored field was not written yet. This record should only be seen by trx_rollback_recovered() or any TRX_ISO_READ_UNCOMMITTED transactions. The InnoDB SQL parser (the sole caller of this function) does not implement READ UNCOMMITTED, and it is not involved during rollback. */ ut_a(data); ut_a(len != UNIV_SQL_NULL); needs_copy = TRUE; } else { data = rec_get_nth_cfield(rec, index, offsets, field_no, &len); needs_copy = column->copy_val; } if (needs_copy) { eval_node_copy_and_alloc_val(column, data, len); } else { val = que_node_get_val(column); dfield_set_data(val, data, len); } if (UNIV_LIKELY_NULL(heap)) { mem_heap_free(heap); } } 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( /*=======================*/ sym_node_t* column) /*!< in: symbol table node for a column */ { sel_buf_t* sel_buf; ulint i; ut_ad(que_node_get_type(column) == QUE_NODE_SYMBOL); column->prefetch_buf = static_cast( ut_malloc_nokey(SEL_MAX_N_PREFETCH * sizeof(sel_buf_t))); for (i = 0; i < SEL_MAX_N_PREFETCH; i++) { sel_buf = column->prefetch_buf + i; sel_buf->data = NULL; sel_buf->len = 0; sel_buf->val_buf_size = 0; } } /*********************************************************************//** Frees a prefetch buffer for a column, including the dynamically allocated memory for data stored there. */ void sel_col_prefetch_buf_free( /*======================*/ sel_buf_t* prefetch_buf) /*!< in, own: prefetch buffer */ { 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) { ut_free(sel_buf->data); } } ut_free(prefetch_buf); } /*********************************************************************//** 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_dequeue_prefetched_row( /*=======================*/ plan_t* plan) /*!< in: plan node for a table */ { sym_node_t* column; sel_buf_t* sel_buf; dfield_t* val; byte* data; ulint len; ulint val_buf_size; 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); ut_d(dfield_set_null(val)); goto next_col; } ut_ad(column->prefetch_buf); ut_ad(!dfield_is_ext(val)); 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 = static_cast(dfield_get_data(val)); sel_buf->len = dfield_get_len(val); sel_buf->val_buf_size = que_node_get_val_buf_size(column); 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_enqueue_prefetched_row( /*=======================*/ plan_t* plan) /*!< in: plan node for a table */ { 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: */ ut_ad(plan->first_prefetched == 0); } plan->n_rows_prefetched++; ut_ad(pos < SEL_MAX_N_PREFETCH); for (column = UT_LIST_GET_FIRST(plan->columns); column != 0; column = UT_LIST_GET_NEXT(col_var_list, 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! */ continue; } 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 = static_cast(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); sel_buf->data = data; sel_buf->len = len; sel_buf->val_buf_size = val_buf_size; } } /*********************************************************************//** Builds a previous version of a clustered index record for a consistent read @return DB_SUCCESS or error code */ static MY_ATTRIBUTE((nonnull, warn_unused_result)) dberr_t row_sel_build_prev_vers( /*====================*/ ReadView* read_view, /*!< in: read view */ dict_index_t* index, /*!< in: plan node for table */ rec_t* rec, /*!< in: record in a clustered index */ rec_offs** offsets, /*!< in/out: offsets returned by rec_get_offsets(rec, plan->index) */ mem_heap_t** offset_heap, /*!< in/out: memory heap from which the offsets are allocated */ mem_heap_t** old_vers_heap, /*!< out: old version heap to use */ rec_t** old_vers, /*!< out: old version, or NULL if the record does not exist in the view: i.e., it was freshly inserted afterwards */ mtr_t* mtr) /*!< in: mtr */ { dberr_t err; if (*old_vers_heap) { mem_heap_empty(*old_vers_heap); } else { *old_vers_heap = mem_heap_create(512); } err = row_vers_build_for_consistent_read( rec, mtr, index, offsets, read_view, offset_heap, *old_vers_heap, old_vers, NULL); return(err); } /*********************************************************************//** Builds the last committed version of a clustered index record for a semi-consistent read. */ static void row_sel_build_committed_vers_for_mysql( /*===================================*/ 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 */ rec_offs** offsets, /*!< in/out: offsets returned by rec_get_offsets(rec, clust_index) */ mem_heap_t** offset_heap, /*!< in/out: memory heap from which the offsets are allocated */ const rec_t** old_vers, /*!< out: old version, or NULL if the record does not exist in the view: i.e., it was freshly inserted afterwards */ dtuple_t** vrow, /*!< out: to be filled with old virtual column version if any */ mtr_t* mtr) /*!< in: mtr */ { if (prebuilt->old_vers_heap) { mem_heap_empty(prebuilt->old_vers_heap); } else { prebuilt->old_vers_heap = mem_heap_create( rec_offs_size(*offsets)); } row_vers_build_for_semi_consistent_read(prebuilt->trx, rec, mtr, clust_index, offsets, offset_heap, prebuilt->old_vers_heap, old_vers, vrow); } /*********************************************************************//** Tests the conditions which determine when the index segment we are searching through has been exhausted. @return TRUE if row passed the tests */ UNIV_INLINE ibool row_sel_test_end_conds( /*===================*/ plan_t* plan) /*!< in: plan for the table; the column values must 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 */ for (cond = UT_LIST_GET_FIRST(plan->end_conds); cond != 0; cond = UT_LIST_GET_NEXT(cond_list, cond)) { /* Evaluate the left side of the comparison, i.e., get the column value if there is an indirection */ eval_sym(static_cast(cond->args)); /* Do the comparison */ if (!eval_cmp(cond)) { return(FALSE); } } return(TRUE); } /*********************************************************************//** Tests the other conditions. @return TRUE if row passed the tests */ UNIV_INLINE ibool row_sel_test_other_conds( /*=====================*/ plan_t* plan) /*!< in: plan for the table; the column values must already have been retrieved */ { func_node_t* cond; 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); } /** Check that a clustered index record is visible in a consistent read view. @param rec clustered index record (in leaf page, or in memory) @param index clustered index @param offsets rec_get_offsets(rec, index) @param view consistent read view @retval DB_SUCCESS if rec is visible in view @retval DB_SUCCESS_LOCKED_REC if rec is not visible in view @retval DB_CORRUPTION if the DB_TRX_ID is corrupted */ static dberr_t row_sel_clust_sees(const rec_t *rec, const dict_index_t &index, const rec_offs *offsets, const ReadView &view) { ut_ad(index.is_primary()); ut_ad(page_rec_is_user_rec(rec)); ut_ad(rec_offs_validate(rec, &index, offsets)); ut_ad(!rec_is_metadata(rec, index)); ut_ad(!index.table->is_temporary()); const trx_id_t id= row_get_rec_trx_id(rec, &index, offsets); if (view.changes_visible(id)) return DB_SUCCESS; if (UNIV_LIKELY(id < view.low_limit_id() || id < trx_sys.get_max_trx_id())) return DB_SUCCESS_LOCKED_REC; ib::warn() << "A transaction id in a record of table " << index.table->name << " is newer than the system-wide maximum."; return DB_CORRUPTION; } /*********************************************************************//** Retrieves the clustered index record corresponding to a record in a non-clustered index. Does the necessary locking. @return DB_SUCCESS or error code */ static MY_ATTRIBUTE((nonnull, warn_unused_result)) dberr_t row_sel_get_clust_rec( /*==================*/ 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 it, NULL if the old version did not exist in the read view, i.e., it was a fresh inserted version */ mtr_t* mtr) /*!< in: mtr used to get access to the 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; mem_heap_t* heap = NULL; rec_offs offsets_[REC_OFFS_NORMAL_SIZE]; rec_offs* offsets = offsets_; rec_offs_init(offsets_); *out_rec = NULL; offsets = rec_get_offsets(rec, plan->pcur.index(), offsets, plan->pcur.index()->n_core_fields, ULINT_UNDEFINED, &heap); row_build_row_ref_fast(plan->clust_ref, plan->clust_map, rec, offsets); index = dict_table_get_first_index(plan->table); plan->clust_pcur.old_rec = nullptr; plan->clust_pcur.btr_cur.page_cur.index = index; dberr_t err = btr_pcur_open_with_no_init(plan->clust_ref, PAGE_CUR_LE, BTR_SEARCH_LEAF, &plan->clust_pcur, mtr); if (UNIV_UNLIKELY(err != DB_SUCCESS)) { goto err_exit; } 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) || btr_pcur_get_low_match(&(plan->clust_pcur)) < dict_index_get_n_unique(index)) { if (!node->read_view || !rec_get_deleted_flag(rec, plan->table->not_redundant())) { err = DB_CORRUPTION; } /* In a rare case it is possible that no clust rec is found for a delete-marked secondary index record: if in row0umod.cc 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 err_exit; } offsets = rec_get_offsets(clust_rec, index, offsets, index->n_core_fields, ULINT_UNDEFINED, &heap); if (!node->read_view) { /* Try to place a lock on the index record */ trx_t* trx = thr_get_trx(thr); /* At READ UNCOMMITTED or READ COMMITTED isolation level we lock only the record, i.e., next-key locking is not used. */ err = lock_clust_rec_read_check_and_lock( 0, btr_pcur_get_block(&plan->clust_pcur), clust_rec, index, offsets, node->row_lock_mode, trx->isolation_level <= TRX_ISO_READ_COMMITTED ? LOCK_REC_NOT_GAP : LOCK_ORDINARY, thr); switch (err) { case DB_SUCCESS: case DB_SUCCESS_LOCKED_REC: /* Declare the variable uninitialized. It should be set to DB_SUCCESS at func_exit. */ MEM_UNDEFINED(&err, sizeof err); break; default: goto err_exit; } } else { /* This is a non-locking consistent read: if necessary, fetch a previous version of the record */ old_vers = NULL; err = row_sel_clust_sees(clust_rec, *index, offsets, *node->read_view); switch (err) { default: goto err_exit; case DB_SUCCESS: break; case DB_SUCCESS_LOCKED_REC: err = row_sel_build_prev_vers( node->read_view, index, clust_rec, &offsets, &heap, &plan->old_vers_heap, &old_vers, mtr); if (err != DB_SUCCESS) { goto err_exit; } clust_rec = old_vers; if (clust_rec == NULL) { goto err_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. */ if (old_vers || rec_get_deleted_flag(rec, dict_table_is_comp( plan->table))) { err = row_sel_sec_rec_is_for_clust_rec(rec, plan->index, clust_rec, index, thr); if (err != DB_SUCCESS_LOCKED_REC) { goto err_exit; } } } /* 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(). */ ut_ad(!rec_get_deleted_flag(clust_rec, rec_offs_comp(offsets))); row_sel_fetch_columns(index, clust_rec, offsets, UT_LIST_GET_FIRST(plan->columns)); *out_rec = clust_rec; err = DB_SUCCESS; err_exit: if (UNIV_LIKELY_NULL(heap)) { mem_heap_free(heap); } return(err); } /*********************************************************************//** Sets a lock on a page of R-Tree record. This is all or none action, mostly due to we cannot reposition a record in R-Tree (with the nature of splitting) @return DB_SUCCESS, DB_SUCCESS_LOCKED_REC, or error code */ UNIV_INLINE dberr_t sel_set_rtr_rec_lock( /*=================*/ btr_pcur_t* pcur, /*!< in: cursor */ const rec_t* first_rec,/*!< in: record */ dict_index_t* index, /*!< in: index */ const rec_offs* offsets,/*!< in: rec_get_offsets(rec, index) */ unsigned mode, /*!< in: lock mode */ unsigned type, /*!< in: LOCK_ORDINARY, LOCK_GAP, or LOC_REC_NOT_GAP */ que_thr_t* thr, /*!< in: query thread */ mtr_t* mtr) /*!< in: mtr */ { matched_rec_t* match = pcur->btr_cur.rtr_info->matches; mem_heap_t* heap = NULL; dberr_t err = DB_SUCCESS; trx_t* trx = thr_get_trx(thr); buf_block_t* cur_block = btr_pcur_get_block(pcur); rec_offs offsets_[REC_OFFS_NORMAL_SIZE]; rec_offs* my_offsets = const_cast(offsets); rec_t* rec = const_cast(first_rec); rtr_rec_vector* match_rec; rtr_rec_vector::iterator end; rec_offs_init(offsets_); if (match->locked || page_rec_is_supremum(first_rec)) { return(DB_SUCCESS_LOCKED_REC); } ut_ad(page_align(first_rec) == cur_block->page.frame); ut_ad(match->valid); match->block.page.lock.x_lock(); retry: cur_block = btr_pcur_get_block(pcur); ut_ad(match->block.page.lock.have_x() || match->block.page.lock.have_s()); ut_ad(page_is_leaf(cur_block->page.frame)); err = lock_sec_rec_read_check_and_lock( 0, cur_block, rec, index, my_offsets, static_cast(mode), type, thr); if (err == DB_LOCK_WAIT) { re_scan: mtr->commit(); trx->error_state = err; thr->lock_state = QUE_THR_LOCK_ROW; if (row_mysql_handle_errors( &err, trx, thr, NULL)) { thr->lock_state = QUE_THR_LOCK_NOLOCK; mtr->start(); mysql_mutex_lock(&match->rtr_match_mutex); if (!match->valid && match->matched_recs->empty()) { mysql_mutex_unlock(&match->rtr_match_mutex); err = DB_RECORD_NOT_FOUND; goto func_end; } mysql_mutex_unlock(&match->rtr_match_mutex); /* MDEV-14059 FIXME: why re-latch the block? pcur is already positioned on it! */ cur_block = buf_page_get_gen( btr_pcur_get_block(pcur)->page.id(), btr_pcur_get_block(pcur)->zip_size(), RW_X_LATCH, NULL, BUF_GET, mtr, &err); if (!cur_block) { goto func_end; } } else { mtr->start(); goto func_end; } DEBUG_SYNC_C("rtr_set_lock_wait"); if (!match->valid) { /* Page got deleted */ mtr->commit(); mtr->start(); err = DB_RECORD_NOT_FOUND; goto func_end; } match->matched_recs->clear(); // FIXME: check for !cur_block rtr_cur_search_with_match( cur_block, index, pcur->btr_cur.rtr_info->search_tuple, pcur->btr_cur.rtr_info->search_mode, &pcur->btr_cur.page_cur, pcur->btr_cur.rtr_info); if (!page_is_leaf(buf_block_get_frame(cur_block))) { /* Page got splitted and promoted (only for root page it is possible). Release the page and ask for a re-search */ mtr->commit(); mtr->start(); err = DB_RECORD_NOT_FOUND; goto func_end; } rec = btr_pcur_get_rec(pcur); my_offsets = offsets_; my_offsets = rec_get_offsets(rec, index, my_offsets, index->n_fields, ULINT_UNDEFINED, &heap); /* No match record */ if (page_rec_is_supremum(rec) || !match->valid) { mtr->commit(); mtr->start(); err = DB_RECORD_NOT_FOUND; goto func_end; } goto retry; } my_offsets = offsets_; match_rec = match->matched_recs; end = match_rec->end(); for (rtr_rec_vector::iterator it = match_rec->begin(); it != end; ++it) { rtr_rec_t* rtr_rec = &(*it); my_offsets = rec_get_offsets( rtr_rec->r_rec, index, my_offsets, index->n_fields, ULINT_UNDEFINED, &heap); err = lock_sec_rec_read_check_and_lock( 0, &match->block, rtr_rec->r_rec, index, my_offsets, static_cast(mode), type, thr); if (err == DB_SUCCESS || err == DB_SUCCESS_LOCKED_REC) { rtr_rec->locked = true; } else if (err == DB_LOCK_WAIT) { goto re_scan; } else { goto func_end; } } match->locked = true; func_end: match->block.page.lock.x_unlock(); if (heap != NULL) { mem_heap_free(heap); } ut_ad(err != DB_LOCK_WAIT); return(err); } /*********************************************************************//** Sets a lock on a record. @return DB_SUCCESS, DB_SUCCESS_LOCKED_REC, or error code */ UNIV_INLINE dberr_t sel_set_rec_lock( /*=============*/ btr_pcur_t* pcur, /*!< in: cursor */ const rec_t* rec, /*!< in: record */ dict_index_t* index, /*!< in: index */ const rec_offs* offsets,/*!< in: rec_get_offsets(rec, index) */ unsigned mode, /*!< in: lock mode */ unsigned type, /*!< in: LOCK_ORDINARY, LOCK_GAP, or LOC_REC_NOT_GAP */ que_thr_t* thr, /*!< in: query thread */ mtr_t* mtr) /*!< in: mtr */ { trx_t* trx; dberr_t err = DB_SUCCESS; const buf_block_t* block; block = btr_pcur_get_block(pcur); trx = thr_get_trx(thr); if (UT_LIST_GET_LEN(trx->lock.trx_locks) > 10000 && buf_pool.running_out()) { return DB_LOCK_TABLE_FULL; } if (dict_index_is_clust(index)) { err = lock_clust_rec_read_check_and_lock( 0, block, rec, index, offsets, static_cast(mode), type, thr); } else { if (dict_index_is_spatial(index)) { if (type == LOCK_GAP || type == LOCK_ORDINARY) { ut_ad(0); ib::error() << "Incorrectly request GAP lock " "on RTree"; return(DB_SUCCESS); } err = sel_set_rtr_rec_lock(pcur, rec, index, offsets, mode, type, thr, mtr); } else { err = lock_sec_rec_read_check_and_lock( 0, block, rec, index, offsets, static_cast(mode), type, thr); } } return(err); } /*********************************************************************//** Opens a pcur to a table index. */ MY_ATTRIBUTE((warn_unused_result, nonnull)) static dberr_t row_sel_open_pcur( /*==============*/ plan_t* plan, /*!< in: table plan */ mtr_t* mtr) /*!< in/out: mini-transaction */ { dict_index_t* index; func_node_t* cond; que_node_t* exp; ulint n_fields; ulint i; ut_ad(!plan->n_rows_prefetched); ut_ad(!plan->n_rows_fetched); ut_ad(!plan->cursor_at_end); 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)); cond = UT_LIST_GET_NEXT(cond_list, cond); } plan->pcur.old_rec = nullptr; plan->pcur.btr_cur.page_cur.index = index; dberr_t err; if (plan->tuple) { n_fields = dtuple_get_n_fields(plan->tuple); if (plan->n_exact_match < n_fields) { /* There is a non-exact match field which must be evaluated separately */ eval_exp(plan->tuple_exps[n_fields - 1]); } for (i = 0; i < n_fields; i++) { exp = plan->tuple_exps[i]; dfield_copy_data(dtuple_get_nth_field(plan->tuple, i), que_node_get_val(exp)); } err = btr_pcur_open_with_no_init(plan->tuple, plan->mode, BTR_SEARCH_LEAF, &plan->pcur, mtr); } else { err = plan->pcur.open_leaf(plan->asc, index, BTR_SEARCH_LEAF, mtr); } plan->pcur_is_open = err == DB_SUCCESS; return err; } /*********************************************************************//** 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 */ static ibool row_sel_restore_pcur_pos( /*=====================*/ plan_t* plan, /*!< in: table plan */ mtr_t* mtr) /*!< in: mtr */ { ibool equal_position; ulint relative_position; ut_ad(!plan->cursor_at_end); relative_position = btr_pcur_get_rel_pos(&(plan->pcur)); equal_position = plan->pcur.restore_position(BTR_SEARCH_LEAF, mtr) == btr_pcur_t::SAME_ALL; /* If the cursor is traveling upwards, and relative_position is (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 || relative_position == BTR_PCUR_AFTER_LAST_IN_TREE); return(FALSE); } /* If the cursor is traveling downwards, and relative_position is (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 || relative_position == BTR_PCUR_BEFORE_FIRST_IN_TREE) { 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 || relative_position == BTR_PCUR_AFTER_LAST_IN_TREE); return(TRUE); } /*********************************************************************//** Resets a plan cursor to a closed state. */ UNIV_INLINE void plan_reset_cursor( /*==============*/ plan_t* plan) /*!< in: plan */ { plan->pcur_is_open = FALSE; plan->cursor_at_end = FALSE; plan->n_rows_fetched = 0; plan->n_rows_prefetched = 0; } #ifdef BTR_CUR_HASH_ADAPT /*********************************************************************//** Tries to do a shortcut to fetch a clustered index record with a unique key, using the hash index if possible (not always). @return SEL_FOUND, SEL_EXHAUSTED, SEL_RETRY */ static ulint row_sel_try_search_shortcut( /*========================*/ sel_node_t* node, /*!< in: select node for a consistent read */ plan_t* plan, /*!< in: plan for a unique search in clustered index */ mtr_t* mtr) /*!< in: mtr */ { dict_index_t* index = plan->index; ut_ad(!index->table->is_temporary()); ut_ad(node->read_view); ut_ad(node->read_view->is_open()); ut_ad(plan->unique_search); ut_ad(!plan->must_get_clust); if (row_sel_open_pcur(plan, mtr) != DB_SUCCESS) { return SEL_RETRY; } const rec_t* rec = btr_pcur_get_rec(&(plan->pcur)); if (!page_rec_is_user_rec(rec) || rec_is_metadata(rec, *index)) { return SEL_RETRY; } ut_ad(plan->mode == PAGE_CUR_GE); /* 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 tuple */ if (btr_pcur_get_up_match(&(plan->pcur)) < plan->n_exact_match) { return SEL_EXHAUSTED; } if (trx_id_t bulk_trx_id = index->table->bulk_trx_id) { /* See row_search_mvcc() for a comment on bulk_trx_id */ if (!node->read_view->changes_visible(bulk_trx_id)) { return SEL_EXHAUSTED; } } /* This is a non-locking consistent read: if necessary, fetch a previous version of the record */ mem_heap_t* heap = NULL; rec_offs offsets_[REC_OFFS_NORMAL_SIZE]; rec_offs* offsets = offsets_; rec_offs_init(offsets_); offsets = rec_get_offsets(rec, index, offsets, index->n_core_fields, ULINT_UNDEFINED, &heap); if (dict_index_is_clust(index)) { if (row_sel_clust_sees(rec, *index, offsets, *node->read_view) != DB_SUCCESS) { return SEL_RETRY; } } else if (!srv_read_only_mode) { trx_id_t trx_id = page_get_max_trx_id(page_align(rec)); ut_ad(trx_id); if (!node->read_view->sees(trx_id)) { return SEL_RETRY; } } if (rec_get_deleted_flag(rec, dict_table_is_comp(plan->table))) { return SEL_EXHAUSTED; } /* 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(). */ row_sel_fetch_columns(index, rec, offsets, UT_LIST_GET_FIRST(plan->columns)); /* Test the rest of search conditions */ if (!row_sel_test_other_conds(plan)) { return SEL_EXHAUSTED; } ut_ad(plan->pcur.latch_mode == BTR_SEARCH_LEAF); plan->n_rows_fetched++; if (UNIV_LIKELY_NULL(heap)) { mem_heap_free(heap); } return SEL_FOUND; } #endif /* BTR_CUR_HASH_ADAPT */ /*********************************************************************//** Performs a select step. @return DB_SUCCESS or error code */ static MY_ATTRIBUTE((warn_unused_result)) dberr_t row_sel( /*====*/ sel_node_t* node, /*!< in: select node */ que_thr_t* thr) /*!< in: query thread */ { dict_index_t* index; plan_t* plan; mtr_t mtr; ibool moved; rec_t* rec; rec_t* old_vers; rec_t* clust_rec; /* 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: */ ibool cons_read_requires_clust_rec = FALSE; ulint cost_counter = 0; ibool cursor_just_opened; ibool must_go_to_next; ibool mtr_has_extra_clust_latch = FALSE; /* 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 */ dberr_t err; mem_heap_t* heap = NULL; rec_offs offsets_[REC_OFFS_NORMAL_SIZE]; rec_offs* offsets = offsets_; rec_offs_init(offsets_); const trx_t* trx = thr_get_trx(thr); ut_ad(thr->run_node == node); ut_ad(!node->read_view || node->read_view == &trx->read_view); ut_ad(!node->read_view || node->read_view->is_open()); table_loop: /* TABLE LOOP ---------- 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_dequeue_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 */ mtr.start(); #ifdef BTR_CUR_HASH_ADAPT if (node->read_view && plan->unique_search && !plan->pcur_is_open && !plan->must_get_clust) { switch (row_sel_try_search_shortcut(node, plan, &mtr)) { case SEL_FOUND: goto next_table; case SEL_EXHAUSTED: goto table_exhausted; default: ut_ad(0); /* fall through */ case SEL_RETRY: break; } plan_reset_cursor(plan); mtr.commit(); mtr.start(); } #endif /* BTR_CUR_HASH_ADAPT */ if (!plan->pcur_is_open) { /* Evaluate the expressions to build the search tuple and open the cursor */ err = row_sel_open_pcur(plan, &mtr); if (UNIV_UNLIKELY(err != DB_SUCCESS)) { goto mtr_commit_exit; } cursor_just_opened = TRUE; /* A new search was made: increment the cost counter */ cost_counter++; } else { /* Restore pcur position to the index */ must_go_to_next = row_sel_restore_pcur_pos(plan, &mtr); cursor_just_opened = FALSE; if (must_go_to_next) { /* We have already processed the cursor record: move to the next */ goto next_rec; } } if (!node->read_view || trx->isolation_level == TRX_ISO_READ_UNCOMMITTED) { } else if (trx_id_t bulk_trx_id = index->table->bulk_trx_id) { /* See row_search_mvcc() for a comment on bulk_trx_id */ if (!trx->read_view.changes_visible(bulk_trx_id)) { goto table_exhausted; } } rec_loop: /* RECORD LOOP ----------- 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. */ ut_ad(mtr_has_extra_clust_latch == FALSE); rec = btr_pcur_get_rec(&(plan->pcur)); /* PHASE 1: Set a lock if specified */ if (!node->asc && cursor_just_opened && !page_rec_is_supremum(rec)) { /* Do not support "descending search" for Spatial index */ ut_ad(!dict_index_is_spatial(index)); /* 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. */ if (!node->read_view) { const rec_t* next_rec = page_rec_get_next_const(rec); if (UNIV_UNLIKELY(!next_rec)) { err = DB_CORRUPTION; goto lock_wait_or_error; } unsigned lock_type; offsets = rec_get_offsets(next_rec, index, offsets, index->n_core_fields, ULINT_UNDEFINED, &heap); /* At READ UNCOMMITTED or READ COMMITTED isolation level, we lock only the record, i.e., next-key locking is not used. */ if (trx->isolation_level <= TRX_ISO_READ_COMMITTED) { if (page_rec_is_supremum(next_rec)) { goto skip_lock; } lock_type = LOCK_REC_NOT_GAP; } else { lock_type = LOCK_ORDINARY; } err = sel_set_rec_lock(&plan->pcur, next_rec, index, offsets, node->row_lock_mode, lock_type, thr, &mtr); switch (err) { case DB_SUCCESS_LOCKED_REC: err = DB_SUCCESS; /* fall through */ case DB_SUCCESS: break; default: /* Note that in this case we will store in pcur the PREDECESSOR of the record we are waiting the lock for */ goto lock_wait_or_error; } } } skip_lock: 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 (rec_is_metadata(rec, *index)) { /* Skip the metadata pseudo-record. */ cost_counter++; goto next_rec; } if (!node->read_view) { /* Try to place a lock on the index record */ unsigned lock_type; offsets = rec_get_offsets(rec, index, offsets, index->n_core_fields, ULINT_UNDEFINED, &heap); /* At READ UNCOMMITTED or READ COMMITTED isolation level, we lock only the record, i.e., next-key locking is not used. */ if (trx->isolation_level <= TRX_ISO_READ_COMMITTED || dict_index_is_spatial(index)) { if (page_rec_is_supremum(rec)) { goto next_rec; } lock_type = LOCK_REC_NOT_GAP; } else { lock_type = LOCK_ORDINARY; } err = sel_set_rec_lock(&plan->pcur, rec, index, offsets, node->row_lock_mode, lock_type, thr, &mtr); switch (err) { case DB_SUCCESS_LOCKED_REC: err = DB_SUCCESS; /* fall through */ case DB_SUCCESS: break; default: 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) { /* 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; } /* PHASE 2: Check a mixed index mix id if needed */ if (plan->unique_search && cursor_just_opened) { ut_ad(plan->mode == PAGE_CUR_GE); /* 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 tuple */ if (btr_pcur_get_up_match(&(plan->pcur)) < plan->n_exact_match) { 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, index->n_core_fields, ULINT_UNDEFINED, &heap); if (node->read_view) { /* This is a non-locking consistent read: if necessary, fetch a previous version of the record */ if (dict_index_is_clust(index)) { const trx_id_t id = row_get_rec_trx_id( rec, index, offsets); if (!node->read_view->changes_visible(id)) { if (id >= node->read_view->low_limit_id() && id >= trx_sys.get_max_trx_id()) { err = DB_CORRUPTION; goto lock_wait_or_error; } err = row_sel_build_prev_vers( node->read_view, index, rec, &offsets, &heap, &plan->old_vers_heap, &old_vers, &mtr); if (err != DB_SUCCESS) { goto lock_wait_or_error; } if (old_vers == NULL) { /* The record does not exist in our read view. Skip it, but first attempt to determine whether the index segment we are searching through has been exhausted. */ offsets = rec_get_offsets( rec, index, offsets, index->n_core_fields, ULINT_UNDEFINED, &heap); /* 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(). */ row_sel_fetch_columns( index, rec, offsets, UT_LIST_GET_FIRST( plan->columns)); if (!row_sel_test_end_conds(plan)) { goto table_exhausted; } goto next_rec; } rec = old_vers; } } else if (!srv_read_only_mode) { trx_id_t trx_id = page_get_max_trx_id(page_align(rec)); ut_ad(trx_id); if (!node->read_view->sees(trx_id)) { cons_read_requires_clust_rec = TRUE; } } } /* PHASE 4: Test search end conditions and deleted flag */ /* 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(). */ row_sel_fetch_columns(index, rec, offsets, UT_LIST_GET_FIRST(plan->columns)); /* 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; } if (rec_get_deleted_flag(rec, dict_table_is_comp(plan->table)) && !cons_read_requires_clust_rec) { /* 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) { goto table_exhausted; } 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, &mtr); mtr_has_extra_clust_latch = TRUE; 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(node->read_view); goto next_rec; } if (rec_get_deleted_flag(clust_rec, dict_table_is_comp(plan->table))) { /* In delete-marked records, DB_TRX_ID must always refer to an existing update_undo log record. */ ut_ad(rec_get_trx_id(clust_rec, dict_table_get_first_index( plan->table))); /* 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); } } /* PHASE 6: Test the rest of search conditions */ if (!row_sel_test_other_conds(plan)) { if (plan->unique_search) { 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 */ plan->n_rows_fetched++; ut_ad(plan->pcur.latch_mode == BTR_SEARCH_LEAF); if ((plan->n_rows_fetched <= SEL_PREFETCH_LIMIT) || plan->unique_search || plan->no_prefetch) { /* No prefetch in operation: go to the next table */ goto next_table; } sel_enqueue_prefetched_row(plan); if (plan->n_rows_prefetched == SEL_MAX_N_PREFETCH) { /* The prefetch buffer is now full */ sel_dequeue_prefetched_row(plan); goto next_table; } next_rec: 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; } if (node->asc) { moved = btr_pcur_move_to_next(&(plan->pcur), &mtr); } else { moved = btr_pcur_move_to_prev(&(plan->pcur), &mtr); } if (!moved) { goto table_exhausted; } cursor_just_opened = FALSE; /* END OF RECORD LOOP ------------------ */ 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 */ ut_ad(btr_pcur_is_on_user_rec(&plan->pcur)); if (plan->unique_search && !node->can_get_updated) { plan->cursor_at_end = TRUE; } else { plan->stored_cursor_rec_processed = TRUE; btr_pcur_store_position(&(plan->pcur), &mtr); } mtr.commit(); 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 */ if (node->fetch_table + 1 == node->n_tables) { sel_eval_select_list(node); if (node->is_aggregate) { goto table_loop; } sel_assign_into_var_values(node->into_list, node); 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 */ plan_reset_cursor(sel_node_get_nth_plan(node, node->fetch_table)); goto table_loop; table_exhausted: /* The table cursor pcur reached the result set end: backtrack to the previous table in the join if we do not have cached prefetched rows */ plan->cursor_at_end = TRUE; mtr.commit(); mtr_has_extra_clust_latch = FALSE; if (plan->n_rows_prefetched > 0) { /* The table became exhausted during a prefetch */ sel_dequeue_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); } else { node->state = SEL_NODE_NO_MORE_ROWS; thr->run_node = que_node_get_parent(node); } 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, which would result in the phantom problem. */ plan->stored_cursor_rec_processed = FALSE; btr_pcur_store_position(&(plan->pcur), &mtr); err = DB_SUCCESS; goto mtr_commit_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 &mtr would not be committed and the latches released. */ plan->stored_cursor_rec_processed = TRUE; btr_pcur_store_position(&(plan->pcur), &mtr); mtr.commit(); mtr_has_extra_clust_latch = FALSE; goto table_loop; lock_wait_or_error: /* See the note at stop_for_a_while: the same holds for this case */ ut_ad(!btr_pcur_is_before_first_on_page(&plan->pcur) || !node->asc); plan->stored_cursor_rec_processed = FALSE; btr_pcur_store_position(&(plan->pcur), &mtr); mtr_commit_exit: mtr.commit(); func_exit: if (heap != NULL) { mem_heap_free(heap); } return(err); } /**********************************************************************//** Performs a select step. This is a high-level function used in SQL execution graphs. @return query thread to run next or NULL */ que_thr_t* row_sel_step( /*=========*/ que_thr_t* thr) /*!< in: query thread */ { sel_node_t* node; ut_ad(thr); node = static_cast(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_xa(thr_get_trx(thr), false); plan_reset_cursor(sel_node_get_nth_plan(node, 0)); if (node->consistent_read) { trx_t *trx = thr_get_trx(thr); /* Assign a read view for the query */ trx->read_view.open(trx); node->read_view = trx->read_view.is_open() ? &trx->read_view : NULL; } else { sym_node_t* table_node; lock_mode i_lock_mode; if (node->set_x_locks) { i_lock_mode = LOCK_IX; } else { i_lock_mode = LOCK_IS; } for (table_node = node->table_list; table_node != 0; table_node = static_cast( que_node_get_next(table_node))) { dberr_t err = lock_table( table_node->table, nullptr, i_lock_mode, thr); if (err != DB_SUCCESS) { trx_t* trx; trx = thr_get_trx(thr); trx->error_state = err; return(NULL); } } } /* 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) */ if (node->explicit_cursor && UT_LIST_GET_FIRST(node->copy_variables)) { row_sel_copy_input_variable_vals(node); } node->state = SEL_NODE_FETCH; node->fetch_table = 0; if (node->is_aggregate) { /* Reset the aggregate total values */ sel_reset_aggregate_vals(node); } } dberr_t 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: */ thr->graph->last_sel_node = node; if (err != DB_SUCCESS) { thr_get_trx(thr)->error_state = err; return(NULL); } return(thr); } /**********************************************************************//** Performs a fetch for a cursor. @return query thread to run next or NULL */ que_thr_t* fetch_step( /*=======*/ que_thr_t* thr) /*!< in: query thread */ { sel_node_t* sel_node; fetch_node_t* node; ut_ad(thr); node = static_cast(thr->run_node); sel_node = node->cursor_def; 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) { if (node->into_list) { sel_assign_into_var_values(node->into_list, sel_node); } else { ibool ret = (*node->func->func)( sel_node, node->func->arg); if (!ret) { sel_node->state = SEL_NODE_NO_MORE_ROWS; } } } 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 */ sel_node->common.parent = node; if (sel_node->state == SEL_NODE_CLOSED) { ib::error() << "fetch called on a closed cursor"; thr_get_trx(thr)->error_state = DB_ERROR; return(NULL); } thr->run_node = sel_node; return(thr); } /***********************************************************//** Prints a row in a select result. @return query thread to run next or NULL */ que_thr_t* row_printf_step( /*============*/ que_thr_t* thr) /*!< in: query thread */ { row_printf_node_t* node; sel_node_t* sel_node; que_node_t* arg; ut_ad(thr); node = static_cast(thr->run_node); 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)) { /* Reset the cursor */ sel_node->state = SEL_NODE_OPEN; /* Fetch next row to print */ thr->run_node = sel_node; 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); 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); } /****************************************************************//** 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. */ void row_sel_convert_mysql_key_to_innobase( /*==================================*/ dtuple_t* tuple, /*!< in/out: tuple where to build; NOTE: we assume that the type info in the tuple is already according to index! */ byte* buf, /*!< in: buffer to use in field conversions; NOTE that dtuple->data may end up pointing inside buf so do not discard that buffer while the tuple is being used. See row_mysql_store_col_in_innobase_format() in the case of DATA_INT */ 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 */ { byte* original_buf = buf; const byte* original_key_ptr = key_ptr; dict_field_t* field; dfield_t* dfield; ulint data_offset; ulint data_len; ulint data_field_len; ibool is_null; const byte* key_end; ulint n_fields = 0; /* 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): */ dtuple_set_n_fields(tuple, ULINT_MAX); dfield = dtuple_get_nth_field(tuple, 0); field = dict_index_get_nth_field(index, 0); if (UNIV_UNLIKELY(dfield_get_type(dfield)->mtype == DATA_SYS)) { /* 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); dtuple_set_n_fields(tuple, 1); return; } while (key_ptr < key_end) { ulint type = dfield_get_type(dfield)->mtype; ut_a(field->col->mtype == type); data_offset = 0; is_null = FALSE; if (!(dfield_get_type(dfield)->prtype & DATA_NOT_NULL)) { /* The first byte in the field tells if this is an SQL NULL value */ data_offset = 1; if (*key_ptr != 0) { dfield_set_null(dfield); is_null = TRUE; } } /* Calculate data length and data field total length */ if (DATA_LARGE_MTYPE(type) || DATA_GEOMETRY_MTYPE(type)) { /* For R-tree index, data length should be the total size of the wkb data.*/ if (dict_index_is_spatial(index)) { ut_ad(DATA_GEOMETRY_MTYPE(type)); data_len = key_len; data_field_len = data_offset + data_len; } else { /* 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 = ulint(key_ptr[data_offset]) | ulint(key_ptr[data_offset + 1]) << 8; 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? */ data_len = field->prefix_len; data_field_len = data_offset + data_len; } else { data_len = dfield_get_type(dfield)->len; data_field_len = data_offset + data_len; } if ((dtype_get_mysql_type(dfield_get_type(dfield)) == DATA_MYSQL_TRUE_VARCHAR) && (type != DATA_INT)) { /* 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 */ if (UNIV_LIKELY(!is_null)) { buf = 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)); ut_a(buf <= original_buf + buf_len); } key_ptr += data_field_len; if (UNIV_UNLIKELY(key_ptr > key_end)) { /* The last field in key was not a complete key field but a prefix of it. Print a warning about this! HA_READ_PREFIX_LAST does 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. */ ib::warn() << "Using a partial-field key prefix in" " search, index " << index->name << " of table " << index->table->name << ". Last data field length " << data_field_len << " bytes, key ptr now" " exceeds key end by " << (key_ptr - key_end) << " bytes. Key value in the MariaDB format:"; ut_print_buf(stderr, original_key_ptr, key_len); putc('\n', stderr); if (!is_null) { ulint len = dfield_get_len(dfield); dfield_set_len(dfield, len - (ulint) (key_ptr - key_end)); } ut_ad(0); } n_fields++; field++; dfield++; } ut_a(buf <= original_buf + buf_len); /* We set the length of tuple to n_fields: we assume that the memory area allocated for it is big enough (usually bigger than n_fields). */ dtuple_set_n_fields(tuple, n_fields); } /**************************************************************//** 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.cc. */ void row_sel_field_store_in_mysql_format_func( byte* dest, const mysql_row_templ_t* templ, #ifdef UNIV_DEBUG const dict_index_t* index, ulint field_no, #endif /* UNIV_DEBUG */ const byte* data, ulint len) { #ifdef UNIV_DEBUG const dict_field_t* field = templ->is_virtual ? NULL : dict_index_get_nth_field(index, field_no); #endif /* UNIV_DEBUG */ ut_ad(len != UNIV_SQL_NULL); MEM_CHECK_DEFINED(data, len); MEM_CHECK_ADDRESSABLE(dest, templ->mysql_col_len); MEM_UNDEFINED(dest, templ->mysql_col_len); byte* pad = dest + len; switch (templ->type) { const byte* field_end; case DATA_VARCHAR: case DATA_VARMYSQL: case DATA_BINARY: 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. */ dest = row_mysql_store_true_var_len( dest, len, templ->mysql_length_bytes); /* Copy the actual data. Leave the rest of the buffer uninitialized. */ memcpy(dest, data, len); break; } /* Copy the actual data */ memcpy(dest, data, len); /* Pad with trailing spaces. */ if (pad == field_end) { break; } if (UNIV_UNLIKELY(templ->type == DATA_FIXBINARY)) { memset(pad, 0, field_end - pad); break; } ut_ad(templ->mbminlen <= templ->mbmaxlen); /* We treat some Unicode charset strings specially. */ switch (templ->mbminlen) { case 4: /* InnoDB should never have stripped partial UTF-32 characters. */ ut_a(!(len & 3)); break; case 2: /* A space char is two bytes, 0x0020 in UCS2 and UTF-16 */ if (UNIV_UNLIKELY(len & 1)) { /* A 0x20 has been stripped from the column. Pad it back. */ if (pad < field_end) { *pad++ = 0x20; } } } row_mysql_pad_col(templ->mbminlen, pad, ulint(field_end - pad)); break; case DATA_BLOB: /* 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, len); break; case DATA_GEOMETRY: /* We store all geometry data as BLOB data at server layer. */ row_mysql_store_geometry(dest, templ->mysql_col_len, data, len); break; case DATA_MYSQL: memcpy(dest, data, len); ut_ad(templ->mysql_col_len >= len); ut_ad(templ->mbmaxlen >= templ->mbminlen); /* If field_no equals to templ->icp_rec_field_no, we are examining a row pointed by "icp_rec_field_no". There is possibility that icp_rec_field_no refers to a field in a secondary index while templ->rec_field_no points to field in a primary index. The length should still be equal, unless the field pointed by icp_rec_field_no has a prefix */ ut_ad(templ->mbmaxlen > templ->mbminlen || templ->mysql_col_len == len || (field_no == templ->icp_rec_field_no && field->prefix_len > 0)); /* The following assertion would fail for old tables containing UTF-8 ENUM columns due to Bug #9526. */ ut_ad(!templ->mbmaxlen || !(templ->mysql_col_len % templ->mbmaxlen)); ut_ad(len * templ->mbmaxlen >= templ->mysql_col_len || (field_no == templ->icp_rec_field_no && field->prefix_len > 0) || templ->rec_field_is_prefix); ut_ad(templ->is_virtual || !(field->prefix_len % templ->mbmaxlen)); if (templ->mbminlen == 1 && templ->mbmaxlen != 1) { /* Pad with spaces. This undoes the stripping done in row0mysql.cc, function row_mysql_store_col_in_innobase_format(). */ memset(pad, 0x20, templ->mysql_col_len - len); } break; default: #ifdef UNIV_DEBUG case DATA_SYS_CHILD: case DATA_SYS: /* These column types should never be shipped to MySQL. */ ut_ad(0); /* fall through */ case DATA_CHAR: case DATA_FIXBINARY: case DATA_FLOAT: case DATA_DOUBLE: case DATA_DECIMAL: #endif /* UNIV_DEBUG */ ut_ad((templ->is_virtual && !field) || (field && field->prefix_len ? field->prefix_len == len : templ->mysql_col_len == len)); memcpy(dest, data, len); break; case DATA_INT: /* Convert InnoDB big-endian integer to little-endian format, sign bit restored to 2's complement form */ DBUG_ASSERT(templ->mysql_col_len == len); byte* ptr = pad; do *--ptr = *data++; while (ptr != dest); if (!templ->is_unsigned) { pad[-1] ^= 0x80; } } } /** Convert a field in the Innobase format to a field in the MySQL format. @param[out] mysql_rec record in the MySQL format @param[in,out] prebuilt prebuilt struct @param[in] rec InnoDB record; must be protected by a page latch @param[in] index index of rec @param[in] offsets array returned by rec_get_offsets() @param[in] field_no templ->rec_field_no or templ->clust_rec_field_no or templ->icp_rec_field_no @param[in] templ row template */ static MY_ATTRIBUTE((warn_unused_result)) ibool row_sel_store_mysql_field( byte* mysql_rec, row_prebuilt_t* prebuilt, const rec_t* rec, const dict_index_t* index, const rec_offs* offsets, ulint field_no, const mysql_row_templ_t*templ) { DBUG_ENTER("row_sel_store_mysql_field_func"); const byte* data; ulint len; ut_ad(prebuilt->default_rec); ut_ad(templ); ut_ad(templ >= prebuilt->mysql_template); ut_ad(templ < &prebuilt->mysql_template[prebuilt->n_template]); ut_ad(field_no == templ->clust_rec_field_no || field_no == templ->rec_field_no || field_no == templ->icp_rec_field_no); ut_ad(rec_offs_validate(rec, index, offsets)); if (UNIV_UNLIKELY(rec_offs_nth_extern(offsets, field_no) != 0)) { mem_heap_t* heap; /* Copy an externally stored field to a temporary heap */ ut_ad(field_no == templ->clust_rec_field_no); if (DATA_LARGE_MTYPE(templ->type)) { if (prebuilt->blob_heap == NULL) { prebuilt->blob_heap = mem_heap_create( srv_page_size); } heap = prebuilt->blob_heap; } else { heap = mem_heap_create(srv_page_size); } /* NOTE: if we are retrieving a big BLOB, we may already run out of memory in the next call, which causes an assert */ data = btr_rec_copy_externally_stored_field( rec, offsets, prebuilt->table->space->zip_size(), field_no, &len, heap); if (UNIV_UNLIKELY(!data)) { /* The externally stored field was not written yet. This record should only be seen by trx_rollback_recovered() or any TRX_ISO_READ_UNCOMMITTED transactions. */ if (heap != prebuilt->blob_heap) { mem_heap_free(heap); } ut_a(prebuilt->trx->isolation_level == TRX_ISO_READ_UNCOMMITTED); DBUG_RETURN(FALSE); } ut_a(len != UNIV_SQL_NULL); row_sel_field_store_in_mysql_format( mysql_rec + templ->mysql_col_offset, templ, index, field_no, data, len); if (heap != prebuilt->blob_heap) { mem_heap_free(heap); } } else { /* The field is stored in the index record, or in the metadata for instant ADD COLUMN. */ data = rec_get_nth_cfield(rec, index, offsets, field_no, &len); if (len == UNIV_SQL_NULL) { /* MySQL assumes that the field for an SQL NULL value is set to the default value. */ ut_ad(templ->mysql_null_bit_mask); MEM_CHECK_DEFINED(prebuilt->default_rec + templ->mysql_col_offset, templ->mysql_col_len); #if defined __GNUC__ && !defined __clang__ && __GNUC__ < 6 # pragma GCC diagnostic push # pragma GCC diagnostic ignored "-Wconversion" /* GCC 5 may need this here */ #endif mysql_rec[templ->mysql_null_byte_offset] |= (byte) templ->mysql_null_bit_mask; #if defined __GNUC__ && !defined __clang__ && __GNUC__ < 6 # pragma GCC diagnostic pop #endif memcpy(mysql_rec + templ->mysql_col_offset, (const byte*) prebuilt->default_rec + templ->mysql_col_offset, templ->mysql_col_len); DBUG_RETURN(TRUE); } if (DATA_LARGE_MTYPE(templ->type) || DATA_GEOMETRY_MTYPE(templ->type)) { /* It is a BLOB field locally stored in the InnoDB record: we MUST copy its contents to prebuilt->blob_heap here because row_sel_field_store_in_mysql_format() stores a pointer to the data, and the data passed to us will be invalid as soon as the mini-transaction is committed and the page latch on the clustered index page is released. */ if (prebuilt->blob_heap == NULL) { prebuilt->blob_heap = mem_heap_create( srv_page_size); DBUG_PRINT("anna", ("blob_heap allocated: %p", prebuilt->blob_heap)); } data = static_cast( mem_heap_dup(prebuilt->blob_heap, data, len)); } row_sel_field_store_in_mysql_format( mysql_rec + templ->mysql_col_offset, templ, index, field_no, data, len); } ut_ad(len != UNIV_SQL_NULL); if (templ->mysql_null_bit_mask) { /* It is a nullable column with a non-NULL value */ mysql_rec[templ->mysql_null_byte_offset] &= static_cast(~templ->mysql_null_bit_mask); } DBUG_RETURN(TRUE); } /** 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 be needed in the query. @param[out] mysql_rec row in the MySQL format @param[in] prebuilt cursor @param[in] rec Innobase record in the index which was described in prebuilt's template, or in the clustered index; must be protected by a page latch @param[in] vrow virtual columns @param[in] rec_clust whether index must be the clustered index @param[in] index index of rec @param[in] offsets array returned by rec_get_offsets(rec) @retval true on success @retval false if not all columns could be retrieved */ MY_ATTRIBUTE((warn_unused_result)) static bool row_sel_store_mysql_rec( byte* mysql_rec, row_prebuilt_t* prebuilt, const rec_t* rec, const dtuple_t* vrow, bool rec_clust, const dict_index_t* index, const rec_offs* offsets) { DBUG_ENTER("row_sel_store_mysql_rec"); ut_ad(rec_clust || index == prebuilt->index); ut_ad(!rec_clust || dict_index_is_clust(index)); if (UNIV_LIKELY_NULL(prebuilt->blob_heap)) { row_mysql_prebuilt_free_blob_heap(prebuilt); } for (ulint i = 0; i < prebuilt->n_template; i++) { const mysql_row_templ_t*templ = &prebuilt->mysql_template[i]; if (templ->is_virtual && dict_index_is_clust(index)) { /* Skip virtual columns if it is not a covered search or virtual key read is not requested. */ if (!rec_clust || !prebuilt->index->has_virtual() || !prebuilt->read_just_key) { /* Initialize the NULL bit. */ if (templ->mysql_null_bit_mask) { #if defined __GNUC__ && !defined __clang__ && __GNUC__ < 6 # pragma GCC diagnostic push # pragma GCC diagnostic ignored "-Wconversion" /* GCC 5 may need this here */ #endif mysql_rec[templ->mysql_null_byte_offset] |= (byte) templ->mysql_null_bit_mask; #if defined __GNUC__ && !defined __clang__ && __GNUC__ < 6 # pragma GCC diagnostic pop #endif } continue; } dict_v_col_t* col; col = dict_table_get_nth_v_col( index->table, templ->clust_rec_field_no); ut_ad(vrow); const dfield_t* dfield = dtuple_get_nth_v_field( vrow, col->v_pos); if (dfield_get_type(dfield)->mtype == DATA_MISSING) { ut_ad("no ha_innopart in MariaDB" == 0); continue; } if (dfield->len == UNIV_SQL_NULL) { #if defined __GNUC__ && !defined __clang__ && __GNUC__ < 6 # pragma GCC diagnostic push # pragma GCC diagnostic ignored "-Wconversion" /* GCC 5 may need this here */ #endif mysql_rec[templ->mysql_null_byte_offset] |= (byte) templ->mysql_null_bit_mask; #if defined __GNUC__ && !defined __clang__ && __GNUC__ < 6 # pragma GCC diagnostic pop #endif memcpy(mysql_rec + templ->mysql_col_offset, (const byte*) prebuilt->default_rec + templ->mysql_col_offset, templ->mysql_col_len); } else { row_sel_field_store_in_mysql_format( mysql_rec + templ->mysql_col_offset, templ, index, templ->clust_rec_field_no, (const byte*)dfield->data, dfield->len); if (templ->mysql_null_bit_mask) { mysql_rec[ templ->mysql_null_byte_offset] &= static_cast (~templ->mysql_null_bit_mask); } } continue; } const ulint field_no = rec_clust ? templ->clust_rec_field_no : templ->rec_field_no; /* We should never deliver column prefixes to the SQL layer, except for evaluating handler_index_cond_check() or handler_rowid_filter_check(). */ /* ...actually, we do want to do this in order to support the prefix query optimization. ut_ad(dict_index_get_nth_field(index, field_no)->prefix_len == 0); ...so we disable this assert. */ if (!row_sel_store_mysql_field(mysql_rec, prebuilt, rec, index, offsets, field_no, templ)) { DBUG_RETURN(false); } } /* FIXME: We only need to read the doc_id if an FTS indexed column is being updated. NOTE, the record can be cluster or secondary index record. if secondary index is used then FTS_DOC_ID column should be part of this index. */ if (dict_table_has_fts_index(prebuilt->table)) { if (dict_index_is_clust(index) || prebuilt->fts_doc_id_in_read_set) { prebuilt->fts_doc_id = fts_get_doc_id_from_rec( rec, index, offsets); } } DBUG_RETURN(true); } static void row_sel_reset_old_vers_heap(row_prebuilt_t *prebuilt) { if (prebuilt->old_vers_heap) mem_heap_empty(prebuilt->old_vers_heap); else prebuilt->old_vers_heap= mem_heap_create(200); } /*********************************************************************//** Builds a previous version of a clustered index record for a consistent read @return DB_SUCCESS or error code */ static MY_ATTRIBUTE((warn_unused_result)) dberr_t row_sel_build_prev_vers_for_mysql( /*==============================*/ row_prebuilt_t* prebuilt, /*!< in/out: prebuilt struct */ dict_index_t* clust_index, /*!< in: clustered index */ const rec_t* rec, /*!< in: record in a clustered index */ rec_offs** offsets, /*!< in/out: offsets returned by rec_get_offsets(rec, clust_index) */ mem_heap_t** offset_heap, /*!< in/out: memory heap from which the offsets are allocated */ rec_t** old_vers, /*!< out: old version, or NULL if the record does not exist in the view: i.e., it was freshly inserted afterwards */ dtuple_t** vrow, /*!< out: dtuple to hold old virtual column data */ mtr_t* mtr) /*!< in: mtr */ { row_sel_reset_old_vers_heap(prebuilt); return row_vers_build_for_consistent_read( rec, mtr, clust_index, offsets, &prebuilt->trx->read_view, offset_heap, prebuilt->old_vers_heap, old_vers, vrow); } /** Helper class to cache clust_rec and old_vers */ class Row_sel_get_clust_rec_for_mysql { const rec_t *cached_clust_rec; rec_t *cached_old_vers; lsn_t cached_lsn; page_id_t cached_page_id; #ifdef UNIV_DEBUG void check_eq(const dict_index_t *index, const rec_offs *offsets) const { rec_offs vers_offs[REC_OFFS_HEADER_SIZE + MAX_REF_PARTS]; rec_offs_init(vers_offs); mem_heap_t *heap= nullptr; ut_ad(rec_offs_validate(cached_clust_rec, index, offsets)); ut_ad(index->first_user_field() <= rec_offs_n_fields(offsets)); ut_ad(vers_offs == rec_get_offsets(cached_old_vers, index, vers_offs, index->n_core_fields, index->db_trx_id(), &heap)); ut_ad(!heap); for (auto n= index->db_trx_id(); n--; ) { const dict_col_t *col= dict_index_get_nth_col(index, n); ulint len1, len2; const byte *b1= rec_get_nth_field(cached_clust_rec, offsets, n, &len1); const byte *b2= rec_get_nth_field(cached_old_vers, vers_offs, n, &len2); ut_ad(!cmp_data_data(col->mtype, col->prtype, b1, len1, b2, len2)); } } #endif public: Row_sel_get_clust_rec_for_mysql() : cached_clust_rec(NULL), cached_old_vers(NULL), cached_lsn(0), cached_page_id(page_id_t(0,0)) {} dberr_t operator()(row_prebuilt_t *prebuilt, dict_index_t *sec_index, const rec_t *rec, que_thr_t *thr, const rec_t **out_rec, rec_offs **offsets, mem_heap_t **offset_heap, dtuple_t **vrow, mtr_t *mtr); }; /*********************************************************************//** Retrieves the clustered index record corresponding to a record in a non-clustered index. Does the necessary locking. Used in the MySQL interface. @return DB_SUCCESS, DB_SUCCESS_LOCKED_REC, or error code */ dberr_t Row_sel_get_clust_rec_for_mysql::operator()( /*============================*/ 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 this is a locking read, then rec is not allowed to be delete-marked, and that would not make sense either */ que_thr_t* thr, /*!< in: query thread */ const rec_t** out_rec,/*!< out: clustered record or an old version of it, NULL if the old version did not exist in the read view, i.e., it was a fresh inserted version */ rec_offs** offsets,/*!< in: offsets returned by rec_get_offsets(rec, sec_index); out: offsets returned by rec_get_offsets(out_rec, clust_index) */ mem_heap_t** offset_heap,/*!< in/out: memory heap from which the offsets are allocated */ dtuple_t** vrow, /*!< out: virtual column to fill */ mtr_t* mtr) /*!< in: mtr used to get access to the non-clustered record; the same mtr is used to access the clustered index */ { dict_index_t* clust_index; rec_t* old_vers; trx_t* trx; prebuilt->clust_pcur->old_rec = nullptr; *out_rec = NULL; trx = thr_get_trx(thr); srv_stats.n_sec_rec_cluster_reads.inc( thd_get_thread_id(trx->mysql_thd)); row_build_row_ref_in_tuple(prebuilt->clust_ref, rec, sec_index, *offsets); clust_index = dict_table_get_first_index(sec_index->table); prebuilt->clust_pcur->btr_cur.page_cur.index = clust_index; dberr_t err = btr_pcur_open_with_no_init(prebuilt->clust_ref, PAGE_CUR_LE, BTR_SEARCH_LEAF, prebuilt->clust_pcur, mtr); if (UNIV_UNLIKELY(err != DB_SUCCESS)) { return err; } const rec_t* 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) || btr_pcur_get_low_match(prebuilt->clust_pcur) < dict_index_get_n_unique(clust_index)) { btr_cur_t* btr_cur = btr_pcur_get_btr_cur(prebuilt->pcur); /* If this is a spatial index scan, and we are reading from a shadow buffer, the record could be already deleted (due to rollback etc.). So get the original page and verify that */ if (dict_index_is_spatial(sec_index) && btr_cur->rtr_info->matches && (page_align(rec) == btr_cur->rtr_info->matches->block.page.frame || rec != btr_pcur_get_rec(prebuilt->pcur))) { #ifdef UNIV_DEBUG rtr_info_t* rtr_info = btr_cur->rtr_info; mysql_mutex_lock(&rtr_info->matches->rtr_match_mutex); /* The page could be deallocated (by rollback etc.) */ if (!rtr_info->matches->valid) { mysql_mutex_unlock(&rtr_info->matches->rtr_match_mutex); clust_rec = NULL; goto func_exit; } mysql_mutex_unlock(&rtr_info->matches->rtr_match_mutex); if (rec_get_deleted_flag(rec, dict_table_is_comp(sec_index->table)) && prebuilt->select_lock_type == LOCK_NONE) { clust_rec = NULL; goto func_exit; } if (rec != btr_pcur_get_rec(prebuilt->pcur)) { clust_rec = NULL; goto func_exit; } /* FIXME: Why is this block not the same as btr_pcur_get_block(prebuilt->pcur), and is it not unsafe to use RW_NO_LATCH here? */ buf_block_t* block = buf_page_get_gen( btr_pcur_get_block(prebuilt->pcur)->page.id(), btr_pcur_get_block(prebuilt->pcur)->zip_size(), RW_NO_LATCH, NULL, BUF_GET, mtr, &err); ut_ad(block); // FIXME: avoid crash mem_heap_t* heap = mem_heap_create(256); dtuple_t* tuple = dict_index_build_data_tuple( rec, sec_index, true, sec_index->n_fields, heap); page_cur_t page_cursor; page_cursor.block = block; page_cursor.index = sec_index; ulint up_match = 0, low_match = 0; ut_ad(!page_cur_search_with_match(tuple, PAGE_CUR_LE, &up_match, &low_match, &page_cursor, nullptr)); ut_ad(low_match < dtuple_get_n_fields_cmp(tuple)); mem_heap_free(heap); err = DB_SUCCESS; #endif /* UNIV_DEBUG */ } else if (!rec_get_deleted_flag(rec, dict_table_is_comp(sec_index->table)) || prebuilt->select_lock_type != LOCK_NONE) { /* In a rare case it is possible that no clust rec is found for a delete-marked secondary index record: if row_undo_mod_clust() has 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. */ ib::error() << "Clustered record for sec rec not found" " index " << sec_index->name << " of table " << sec_index->table->name; fputs("InnoDB: sec index record ", stderr); rec_print(stderr, rec, sec_index); fputs("\n" "InnoDB: clust index record ", stderr); rec_print(stderr, clust_rec, clust_index); err = DB_CORRUPTION; } clust_rec = NULL; goto func_exit; } *offsets = rec_get_offsets(clust_rec, clust_index, *offsets, clust_index->n_core_fields, ULINT_UNDEFINED, offset_heap); 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 */ err = lock_clust_rec_read_check_and_lock( 0, btr_pcur_get_block(prebuilt->clust_pcur), clust_rec, clust_index, *offsets, prebuilt->select_lock_type, LOCK_REC_NOT_GAP, thr); switch (err) { case DB_SUCCESS: case DB_SUCCESS_LOCKED_REC: break; default: return err; } } else { /* This is a non-locking consistent read: if necessary, fetch a previous version of the record */ old_vers = NULL; if (trx->isolation_level == TRX_ISO_READ_UNCOMMITTED || clust_index->table->is_temporary()) { } else { /* If the isolation level allows reading of uncommitted data, then we never look for an earlier version */ err = row_sel_clust_sees(clust_rec, *clust_index, *offsets, trx->read_view); } switch (err) { default: return err; case DB_SUCCESS: break; case DB_SUCCESS_LOCKED_REC: const buf_page_t& bpage = btr_pcur_get_block( prebuilt->clust_pcur)->page; const lsn_t lsn = mach_read_from_8( page_align(clust_rec) + FIL_PAGE_LSN); if (lsn != cached_lsn || bpage.id() != cached_page_id || clust_rec != cached_clust_rec) { /* The following call returns 'offsets' associated with 'old_vers' */ err = row_sel_build_prev_vers_for_mysql( prebuilt, clust_index, clust_rec, offsets, offset_heap, &old_vers, vrow, mtr); if (UNIV_UNLIKELY(err != DB_SUCCESS)) { return err; } cached_lsn = lsn; cached_page_id = bpage.id(); cached_clust_rec = clust_rec; cached_old_vers = old_vers; } else { err = DB_SUCCESS; old_vers = cached_old_vers; /* The offsets need not be same for the latest version of clust_rec and its old version old_vers. Re-calculate the offsets for old_vers. */ if (old_vers) { ut_d(check_eq(clust_index, *offsets)); *offsets = rec_get_offsets( old_vers, clust_index, *offsets, clust_index->n_core_fields, ULINT_UNDEFINED, offset_heap); } } if (old_vers == NULL) { return err; } 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. */ /* And for spatial index, since the rec is from shadow buffer, so we need to check if it's exactly match the clust_rec. */ if (clust_rec && (old_vers || trx->isolation_level <= TRX_ISO_READ_UNCOMMITTED || dict_index_is_spatial(sec_index) || rec_get_deleted_flag(rec, dict_table_is_comp( sec_index->table)))) { err = row_sel_sec_rec_is_for_clust_rec(rec, sec_index, clust_rec, clust_index, thr); switch (err) { case DB_SUCCESS: clust_rec = NULL; break; case DB_SUCCESS_LOCKED_REC: break; default: return err; } } err = DB_SUCCESS; } func_exit: *out_rec = clust_rec; if (prebuilt->select_lock_type != LOCK_NONE) { /* We may use the cursor in update or in unlock_row(): store its position */ btr_pcur_store_position(prebuilt->clust_pcur, mtr); } 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. Then we may have to move the cursor one step up or down. @param[out] same_user_rec true if we were able to restore the cursor on a user record with the same ordering prefix in in the B-tree index @param[in] latch_mode latch mode wished in restoration @param[in] pcur cursor whose position has been stored @param[in] moves_up true if the cursor moves up in the index @param[in,out] mtr mtr; CAUTION: may commit mtr temporarily! @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) */ static bool sel_restore_position_for_mysql(bool *same_user_rec, btr_latch_mode latch_mode, btr_pcur_t *pcur, bool moves_up, mtr_t *mtr) { auto status = pcur->restore_position(latch_mode, mtr); *same_user_rec = status == btr_pcur_t::SAME_ALL; ut_ad(!*same_user_rec || pcur->rel_pos == BTR_PCUR_ON); #ifdef UNIV_DEBUG if (pcur->pos_state == BTR_PCUR_IS_POSITIONED_OPTIMISTIC) { ut_ad(pcur->rel_pos == BTR_PCUR_BEFORE || pcur->rel_pos == BTR_PCUR_AFTER); } else { ut_ad(pcur->pos_state == BTR_PCUR_IS_POSITIONED); ut_ad((pcur->rel_pos == BTR_PCUR_ON) == btr_pcur_is_on_user_rec(pcur)); } #endif /* UNIV_DEBUG */ /* The position may need be adjusted for rel_pos and moves_up. */ switch (pcur->rel_pos) { case BTR_PCUR_ON: if (!*same_user_rec && moves_up) { if (status == btr_pcur_t::SAME_UNIQ) return true; next: if (btr_pcur_move_to_next(pcur, mtr) && rec_is_metadata(btr_pcur_get_rec(pcur), *pcur->index())) { btr_pcur_move_to_next(pcur, mtr); } return true; } return(!*same_user_rec); case BTR_PCUR_AFTER_LAST_IN_TREE: case BTR_PCUR_BEFORE_FIRST_IN_TREE: return true; case BTR_PCUR_AFTER: /* positioned to record after pcur->old_rec. */ pcur->pos_state = BTR_PCUR_IS_POSITIONED; prev: if (btr_pcur_is_on_user_rec(pcur) && !moves_up && !rec_is_metadata(btr_pcur_get_rec(pcur), *pcur->index())) { if (!btr_pcur_move_to_prev(pcur, mtr)) { return true; } } return true; case BTR_PCUR_BEFORE: /* For non optimistic restoration: The position is now set to the record before pcur->old_rec. For optimistic restoration: The position also needs to take the previous search_mode into consideration. */ switch (pcur->pos_state) { case BTR_PCUR_IS_POSITIONED_OPTIMISTIC: pcur->pos_state = BTR_PCUR_IS_POSITIONED; if (pcur->search_mode == PAGE_CUR_GE) { /* Positioned during Greater or Equal search with BTR_PCUR_BEFORE. Optimistic restore to the same record. If scanning for lower then we must move to previous record. This can happen with: HANDLER READ idx a = (const); HANDLER READ idx PREV; */ goto prev; } return true; case BTR_PCUR_IS_POSITIONED: if (moves_up && btr_pcur_is_on_user_rec(pcur)) { goto next; } return true; case BTR_PCUR_WAS_POSITIONED: case BTR_PCUR_NOT_POSITIONED: break; } } ut_ad(0); return true; } /********************************************************************//** Copies a cached field for MySQL from the fetch cache. */ static void row_sel_copy_cached_field_for_mysql( /*================================*/ byte* buf, /*!< in/out: row buffer */ const byte* cache, /*!< in: cached row */ const mysql_row_templ_t*templ) /*!< in: column template */ { ulint len; buf += templ->mysql_col_offset; cache += templ->mysql_col_offset; MEM_CHECK_ADDRESSABLE(buf, templ->mysql_col_len); if (templ->mysql_type == DATA_MYSQL_TRUE_VARCHAR && (templ->type != DATA_INT)) { /* Check for != DATA_INT to make sure we do not treat MySQL ENUM or SET as a true VARCHAR! Find the actual length of the true VARCHAR field. */ row_mysql_read_true_varchar( &len, cache, templ->mysql_length_bytes); len += templ->mysql_length_bytes; MEM_UNDEFINED(buf, templ->mysql_col_len); } else { len = templ->mysql_col_len; } memcpy(buf, cache, len); } /** Copy used fields from cached row. Copy cache record field by field, don't touch fields that are not covered by current key. @param[out] buf Where to copy the MySQL row. @param[in] cached_rec What to copy (in MySQL row format). @param[in] prebuilt prebuilt struct. */ void row_sel_copy_cached_fields_for_mysql( byte* buf, const byte* cached_rec, row_prebuilt_t* prebuilt) { const mysql_row_templ_t*templ; ulint i; for (i = 0; i < prebuilt->n_template; i++) { templ = prebuilt->mysql_template + i; /* Skip virtual columns */ if (templ->is_virtual) { continue; } row_sel_copy_cached_field_for_mysql( buf, cached_rec, templ); /* Copy NULL bit of the current field from cached_rec to buf */ if (templ->mysql_null_bit_mask) { #if defined __GNUC__ && !defined __clang__ && __GNUC__ < 6 # pragma GCC diagnostic push # pragma GCC diagnostic ignored "-Wconversion" /* GCC 5 may need this here */ #endif buf[templ->mysql_null_byte_offset] ^= (buf[templ->mysql_null_byte_offset] ^ cached_rec[templ->mysql_null_byte_offset]) & (byte) templ->mysql_null_bit_mask; #if defined __GNUC__ && !defined __clang__ && __GNUC__ < 6 # pragma GCC diagnostic pop #endif } } } /********************************************************************//** Pops a cached row for MySQL from the fetch cache. */ UNIV_INLINE void row_sel_dequeue_cached_row_for_mysql( /*=================================*/ byte* buf, /*!< in/out: buffer where to copy the row */ row_prebuilt_t* prebuilt) /*!< in: prebuilt struct */ { ulint i; const mysql_row_templ_t*templ; const byte* cached_rec; ut_ad(prebuilt->n_fetch_cached > 0); ut_ad(prebuilt->mysql_prefix_len <= prebuilt->mysql_row_len); MEM_CHECK_ADDRESSABLE(buf, prebuilt->mysql_row_len); cached_rec = prebuilt->fetch_cache[prebuilt->fetch_cache_first]; if (UNIV_UNLIKELY(prebuilt->keep_other_fields_on_keyread)) { row_sel_copy_cached_fields_for_mysql(buf, cached_rec, prebuilt); } else if (prebuilt->mysql_prefix_len > 63) { /* The record is long. Copy it field by field, in case there are some long VARCHAR column of which only a small length is being used. */ MEM_UNDEFINED(buf, prebuilt->mysql_prefix_len); /* First copy the NULL bits. */ memcpy(buf, cached_rec, prebuilt->null_bitmap_len); /* Then copy the requested fields. */ for (i = 0; i < prebuilt->n_template; i++) { templ = prebuilt->mysql_template + i; /* Skip virtual columns */ if (templ->is_virtual && !(dict_index_has_virtual(prebuilt->index) && prebuilt->read_just_key)) { continue; } row_sel_copy_cached_field_for_mysql( buf, cached_rec, templ); } } else { memcpy(buf, cached_rec, prebuilt->mysql_prefix_len); } prebuilt->n_fetch_cached--; prebuilt->fetch_cache_first++; if (prebuilt->n_fetch_cached == 0) { prebuilt->fetch_cache_first = 0; } } /********************************************************************//** Initialise the prefetch cache. */ UNIV_INLINE void row_sel_prefetch_cache_init( /*========================*/ row_prebuilt_t* prebuilt) /*!< in/out: prebuilt struct */ { ulint i; ulint sz; byte* ptr; /* Reserve space for the magic number. */ sz = UT_ARR_SIZE(prebuilt->fetch_cache) * (prebuilt->mysql_row_len + 8); ptr = static_cast(ut_malloc_nokey(sz)); for (i = 0; i < UT_ARR_SIZE(prebuilt->fetch_cache); i++) { /* A user has reported memory corruption in these buffers in Linux. Put magic numbers there to help to track a possible bug. */ mach_write_to_4(ptr, ROW_PREBUILT_FETCH_MAGIC_N); ptr += 4; prebuilt->fetch_cache[i] = ptr; ptr += prebuilt->mysql_row_len; mach_write_to_4(ptr, ROW_PREBUILT_FETCH_MAGIC_N); ptr += 4; } } /********************************************************************//** Get the last fetch cache buffer from the queue. @return pointer to buffer. */ UNIV_INLINE byte* row_sel_fetch_last_buf( /*===================*/ row_prebuilt_t* prebuilt) /*!< in/out: prebuilt struct */ { ut_ad(!prebuilt->templ_contains_blob); ut_ad(prebuilt->n_fetch_cached < MYSQL_FETCH_CACHE_SIZE); if (prebuilt->fetch_cache[0] == NULL) { /* Allocate memory for the fetch cache */ ut_ad(prebuilt->n_fetch_cached == 0); row_sel_prefetch_cache_init(prebuilt); } ut_ad(prebuilt->fetch_cache_first == 0); MEM_UNDEFINED(prebuilt->fetch_cache[prebuilt->n_fetch_cached], prebuilt->mysql_row_len); return(prebuilt->fetch_cache[prebuilt->n_fetch_cached]); } /********************************************************************//** Pushes a row for MySQL to the fetch cache. */ UNIV_INLINE void row_sel_enqueue_cache_row_for_mysql( /*================================*/ byte* mysql_rec, /*!< in/out: MySQL record */ row_prebuilt_t* prebuilt) /*!< in/out: prebuilt struct */ { /* For non ICP code path the row should already exist in the next fetch cache slot. */ if (prebuilt->pk_filter || prebuilt->idx_cond) { memcpy(row_sel_fetch_last_buf(prebuilt), mysql_rec, prebuilt->mysql_row_len); } ++prebuilt->n_fetch_cached; } #ifdef BTR_CUR_HASH_ADAPT /*********************************************************************//** 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, btr search latch has been locked in S-mode if AHI is enabled. @return SEL_FOUND, SEL_EXHAUSTED, SEL_RETRY */ static ulint row_sel_try_search_shortcut_for_mysql( /*==================================*/ const rec_t** out_rec,/*!< out: record if found */ row_prebuilt_t* prebuilt,/*!< in: prebuilt struct */ rec_offs** 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 */ { dict_index_t* index = prebuilt->index; const dtuple_t* search_tuple = prebuilt->search_tuple; btr_pcur_t* pcur = prebuilt->pcur; trx_t* trx = prebuilt->trx; const rec_t* rec; ut_ad(index->is_primary()); ut_ad(!index->table->is_temporary()); ut_ad(!prebuilt->templ_contains_blob); ut_ad(trx->read_view.is_open()); pcur->old_rec = nullptr; if (btr_pcur_open_with_no_init(search_tuple, PAGE_CUR_GE, BTR_SEARCH_LEAF, pcur, mtr) != DB_SUCCESS) { return SEL_RETRY; } rec = btr_pcur_get_rec(pcur); if (!page_rec_is_user_rec(rec) || rec_is_metadata(rec, *index)) { 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 fields in the user record matched to the search tuple */ if (btr_pcur_get_up_match(pcur) < dtuple_get_n_fields(search_tuple)) { return SEL_EXHAUSTED; } if (trx->isolation_level == TRX_ISO_READ_UNCOMMITTED) { } else if (trx_id_t bulk_trx_id = index->table->bulk_trx_id) { /* See row_search_mvcc() for a comment on bulk_trx_id */ if (!trx->read_view.changes_visible(bulk_trx_id)) { 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, index->n_core_fields, ULINT_UNDEFINED, heap); if (row_sel_clust_sees(rec, *index, *offsets, trx->read_view) != DB_SUCCESS) { return SEL_RETRY; } if (rec_get_deleted_flag(rec, dict_table_is_comp(index->table))) { /* In delete-marked records, DB_TRX_ID must always refer to an existing undo log record. */ ut_ad(row_get_rec_trx_id(rec, index, *offsets)); return SEL_EXHAUSTED; } *out_rec = rec; return SEL_FOUND; } #endif /* BTR_CUR_HASH_ADAPT */ /*********************************************************************//** Check a pushed-down index condition. @return CHECK_ABORTED_BY_USER, CHECK_NEG, CHECK_POS, or CHECK_OUT_OF_RANGE */ static check_result_t row_search_idx_cond_check( /*======================*/ byte* mysql_rec, /*!< out: record in MySQL format (invalid unless prebuilt->idx_cond!=NULL and we return ICP_MATCH) */ row_prebuilt_t* prebuilt, /*!< in/out: prebuilt struct for the table handle */ const rec_t* rec, /*!< in: InnoDB record */ const rec_offs* offsets) /*!< in: rec_get_offsets() */ { ulint i; ut_ad(rec_offs_validate(rec, prebuilt->index, offsets)); if (!prebuilt->idx_cond) { if (!handler_rowid_filter_is_active(prebuilt->pk_filter)) { return(CHECK_POS); } } else { MONITOR_INC(MONITOR_ICP_ATTEMPTS); } /* Convert to MySQL format those fields that are needed for evaluating the index condition. */ if (UNIV_LIKELY_NULL(prebuilt->blob_heap)) { mem_heap_empty(prebuilt->blob_heap); } for (i = 0; i < prebuilt->idx_cond_n_cols; i++) { const mysql_row_templ_t*templ = &prebuilt->mysql_template[i]; /* Skip virtual columns */ if (templ->is_virtual) { continue; } if (!row_sel_store_mysql_field(mysql_rec, prebuilt, rec, prebuilt->index, offsets, templ->icp_rec_field_no, templ)) { return(CHECK_NEG); } } /* We assume that the index conditions on case-insensitive columns are case-insensitive. The case of such columns may be wrong in a secondary index, if the case of the column has been updated in the past, or a record has been deleted and a record inserted in a different case. */ check_result_t result = prebuilt->idx_cond ? handler_index_cond_check(prebuilt->idx_cond) : CHECK_POS; switch (result) { case CHECK_POS: if (handler_rowid_filter_is_active(prebuilt->pk_filter)) { ut_ad(!prebuilt->index->is_primary()); if (prebuilt->clust_index_was_generated) { ulint len; dict_index_t* index = prebuilt->index; const byte* data = rec_get_nth_field( rec, offsets, index->n_fields - 1, &len); ut_ad(dict_index_get_nth_col(index, index->n_fields - 1) ->prtype == (DATA_ROW_ID | DATA_NOT_NULL)); ut_ad(len == DATA_ROW_ID_LEN); memcpy(prebuilt->row_id, data, DATA_ROW_ID_LEN); } result = handler_rowid_filter_check(prebuilt->pk_filter); switch (result) { case CHECK_NEG: MONITOR_INC(MONITOR_ICP_NO_MATCH); return(result); case CHECK_OUT_OF_RANGE: MONITOR_INC(MONITOR_ICP_OUT_OF_RANGE); return(result); case CHECK_POS: break; default: return(result); } } /* Convert the remaining fields to MySQL format. If this is a secondary index record, we must defer this until we have fetched the clustered index record. */ if (!prebuilt->need_to_access_clustered || dict_index_is_clust(prebuilt->index)) { if (!row_sel_store_mysql_rec( mysql_rec, prebuilt, rec, NULL, false, prebuilt->index, offsets)) { ut_ad(dict_index_is_clust(prebuilt->index)); return(CHECK_NEG); } } MONITOR_INC(MONITOR_ICP_MATCH); return(result); case CHECK_NEG: MONITOR_INC(MONITOR_ICP_NO_MATCH); return(result); case CHECK_OUT_OF_RANGE: MONITOR_INC(MONITOR_ICP_OUT_OF_RANGE); return(result); case CHECK_ERROR: case CHECK_ABORTED_BY_USER: return(result); } ut_error; return(result); } /** Extract virtual column data from a virtual index record and fill a dtuple @param[in] rec the virtual (secondary) index record @param[in] index the virtual index @param[in,out] vrow the dtuple where data extract to @param[in] heap memory heap to allocate memory */ static void row_sel_fill_vrow( const rec_t* rec, dict_index_t* index, dtuple_t** vrow, mem_heap_t* heap) { rec_offs offsets_[REC_OFFS_NORMAL_SIZE]; rec_offs* offsets = offsets_; rec_offs_init(offsets_); ut_ad(!(*vrow)); ut_ad(heap); ut_ad(!dict_index_is_clust(index)); ut_ad(!index->is_instant()); ut_ad(page_rec_is_leaf(rec)); offsets = rec_get_offsets(rec, index, offsets, index->n_core_fields, ULINT_UNDEFINED, &heap); *vrow = dtuple_create_with_vcol( heap, 0, dict_table_get_n_v_cols(index->table)); /* Initialize all virtual row's mtype to DATA_MISSING */ dtuple_init_v_fld(*vrow); for (ulint i = 0; i < dict_index_get_n_fields(index); i++) { const dict_field_t* field; const dict_col_t* col; field = dict_index_get_nth_field(index, i); col = dict_field_get_col(field); if (col->is_virtual()) { const byte* data; ulint len; data = rec_get_nth_field(rec, offsets, i, &len); const dict_v_col_t* vcol = reinterpret_cast< const dict_v_col_t*>(col); dfield_t* dfield = dtuple_get_nth_v_field( *vrow, vcol->v_pos); dfield_set_data(dfield, data, len); dict_col_copy_type(col, dfield_get_type(dfield)); } } } /** Return the record field length in characters. @param[in] col table column of the field @param[in] field_no field number @param[in] rec physical record @param[in] offsets field offsets in the physical record @return field length in characters. */ static size_t rec_field_len_in_chars( const dict_col_t* col, const ulint field_no, const rec_t* rec, const rec_offs* offsets) { const ulint cset = dtype_get_charset_coll(col->prtype); const CHARSET_INFO* cs = all_charsets[cset]; ulint rec_field_len; const char* rec_field = reinterpret_cast( rec_get_nth_field( rec, offsets, field_no, &rec_field_len)); if (UNIV_UNLIKELY(!cs)) { ib::warn() << "Missing collation " << cset; return SIZE_T_MAX; } return cs->numchars(rec_field, rec_field + rec_field_len); } /** Avoid the clustered index lookup if all the following conditions are true: 1) all columns are in secondary index 2) all values for columns that are prefix-only indexes are shorter than the prefix size. This optimization can avoid many IOs for certain schemas. @return true, to avoid clustered index lookup. */ static bool row_search_with_covering_prefix( row_prebuilt_t* prebuilt, const rec_t* rec, const rec_offs* offsets) { const dict_index_t* index = prebuilt->index; ut_ad(!dict_index_is_clust(index)); if (dict_index_is_spatial(index)) { return false; } if (!srv_prefix_index_cluster_optimization) { return false; } /** Optimization only applicable if there the number of secondary index fields are greater than or equal to number of clustered index fields. */ if (prebuilt->n_template > index->n_fields) { return false; } /* We can avoid a clustered index lookup if all of the following hold: (1) all columns are in the secondary index (2) all values for columns that are prefix-only indexes are shorter than the prefix size This optimization can avoid many IOs for certain schemas. */ for (ulint i = 0; i < prebuilt->n_template; i++) { mysql_row_templ_t* templ = prebuilt->mysql_template + i; ulint j = templ->rec_prefix_field_no; ut_ad(!templ->mbminlen == !templ->mbmaxlen); /** Condition (1) : is the field in the index. */ if (j == ULINT_UNDEFINED) { return false; } /** Condition (2): If this is a prefix index then row's value size shorter than prefix length. */ if (!templ->rec_field_is_prefix || rec_offs_nth_sql_null(offsets, j)) { continue; } const dict_field_t* field = dict_index_get_nth_field(index, j); if (!field->prefix_len) { continue; } const ulint rec_size = rec_offs_nth_size(offsets, j); if (rec_size >= field->prefix_len) { /* Shortest representation string by the byte length of the record is longer than the maximum possible index prefix. */ return false; } if (templ->mbminlen != templ->mbmaxlen && rec_field_len_in_chars(field->col, j, rec, offsets) >= field->prefix_len / templ->mbmaxlen) { /* No of chars to store the record exceeds the index prefix character length. */ return false; } } /* If prefix index optimization condition satisfied then for all columns above, use rec_prefix_field_no instead of rec_field_no, and skip the clustered lookup below. */ for (ulint i = 0; i < prebuilt->n_template; i++) { mysql_row_templ_t* templ = prebuilt->mysql_template + i; templ->rec_field_no = templ->rec_prefix_field_no; ut_a(templ->rec_field_no != ULINT_UNDEFINED); } srv_stats.n_sec_rec_cluster_reads_avoided.inc(); return true; } /** Searches for rows in the database using cursor. Function is mainly used for tables that are shared across connections and so it employs technique that can help re-construct the rows that transaction is suppose to see. It also has optimization such as pre-caching the rows, using AHI, etc. @param[out] buf buffer for the fetched row in MySQL format @param[in] mode search mode PAGE_CUR_L @param[in,out] prebuilt prebuilt struct for the table handler; this contains the info to search_tuple, index; if search tuple contains 0 field then we position the cursor at start or the end of index, depending on 'mode' @param[in] match_mode 0 or ROW_SEL_EXACT or ROW_SEL_EXACT_PREFIX @param[in] direction 0 or ROW_SEL_NEXT or ROW_SEL_PREV; Note: if this is != 0, then prebuilt must has a pcur with stored position! In opening of a cursor 'direction' should be 0. @return DB_SUCCESS or error code */ dberr_t row_search_mvcc( byte* buf, page_cur_mode_t mode, row_prebuilt_t* prebuilt, ulint match_mode, ulint direction) { DBUG_ENTER("row_search_mvcc"); DBUG_ASSERT(prebuilt->index->table == prebuilt->table); dict_index_t* index = prebuilt->index; ibool comp = dict_table_is_comp(prebuilt->table); const dtuple_t* search_tuple = prebuilt->search_tuple; btr_pcur_t* pcur = prebuilt->pcur; trx_t* trx = prebuilt->trx; dict_index_t* clust_index; que_thr_t* thr; const rec_t* UNINIT_VAR(rec); dtuple_t* vrow = NULL; const rec_t* result_rec = NULL; const rec_t* clust_rec; Row_sel_get_clust_rec_for_mysql row_sel_get_clust_rec_for_mysql; ibool unique_search = FALSE; ulint mtr_extra_clust_savepoint = 0; bool moves_up = false; /* if the returned record was locked and we did a semi-consistent read (fetch the newest committed version), then this is set to TRUE */ ulint next_offs; bool same_user_rec; ibool table_lock_waited = FALSE; byte* next_buf = 0; bool spatial_search = false; ut_ad(index && pcur && search_tuple); ut_a(prebuilt->magic_n == ROW_PREBUILT_ALLOCATED); ut_a(prebuilt->magic_n2 == ROW_PREBUILT_ALLOCATED); /* We don't support FTS queries from the HANDLER interfaces, because we implemented FTS as reversed inverted index with auxiliary tables. So anything related to traditional index query would not apply to it. */ if (prebuilt->index->type & DICT_FTS) { DBUG_RETURN(DB_END_OF_INDEX); } if (!prebuilt->table->space) { DBUG_RETURN(DB_TABLESPACE_DELETED); } else if (!prebuilt->table->is_readable()) { if (fil_space_crypt_t* crypt_data = prebuilt->table->space->crypt_data) { if (crypt_data->should_encrypt()) { DBUG_RETURN(DB_DECRYPTION_FAILED); } } DBUG_RETURN(DB_CORRUPTION); } else if (!prebuilt->index_usable) { DBUG_RETURN(DB_MISSING_HISTORY); } else if (prebuilt->index->is_corrupted()) { DBUG_RETURN(DB_CORRUPTION); } pcur->btr_cur.page_cur.index = index; /* We need to get the virtual column values stored in secondary index key, if this is covered index scan or virtual key read is requested. */ bool need_vrow = prebuilt->read_just_key && prebuilt->index->has_virtual(); /* Reset the new record lock info if READ UNCOMMITTED or READ COMMITED isolation level is used. Then we are able to remove the record locks set here on an individual row. */ prebuilt->new_rec_locks = 0; /*-------------------------------------------------------------*/ /* PHASE 1: Try to pop the row from the prefetch cache */ if (UNIV_UNLIKELY(direction == 0)) { trx->op_info = "starting index read"; 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! */ } 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_dequeue_cached_row_for_mysql(buf, prebuilt); prebuilt->n_rows_fetched++; trx->op_info = ""; DBUG_RETURN(DB_SUCCESS); } if (prebuilt->fetch_cache_first > 0 && prebuilt->fetch_cache_first < MYSQL_FETCH_CACHE_SIZE) { early_not_found: /* 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 */ trx->op_info = ""; DBUG_RETURN(DB_RECORD_NOT_FOUND); } 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 can use a LOCK_REC_NOT_GAP type record lock when locking a non-delete-marked matching record. 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. */ if (match_mode == ROW_SEL_EXACT && dict_index_is_unique(index) && dtuple_get_n_fields(search_tuple) == dict_index_get_n_unique(index) && (dict_index_is_clust(index) || !dtuple_contains_null(search_tuple))) { /* 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. */ if (UNIV_UNLIKELY(direction != 0 && !prebuilt->used_in_HANDLER)) { goto early_not_found; } } /* We don't support sequencial scan for Rtree index, because it is no meaning to do so. */ if (dict_index_is_spatial(index) && !RTREE_SEARCH_MODE(mode)) { trx->op_info = ""; DBUG_RETURN(DB_END_OF_INDEX); } /* if the query is a plain locking SELECT, and the isolation level is <= TRX_ISO_READ_COMMITTED, then this is set to FALSE */ bool did_semi_consistent_read = false; mtr_t mtr; mtr.start(); mem_heap_t* heap = NULL; rec_offs offsets_[REC_OFFS_NORMAL_SIZE]; rec_offs* offsets = offsets_; rec_offs_init(offsets_); #ifdef BTR_CUR_HASH_ADAPT /*-------------------------------------------------------------*/ /* 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) && unique_search && btr_search_enabled && dict_index_is_clust(index) && !index->table->is_temporary() && !prebuilt->templ_contains_blob && !prebuilt->used_in_HANDLER && (prebuilt->mysql_row_len < srv_page_size / 8)) { mode = PAGE_CUR_GE; if (prebuilt->select_lock_type == LOCK_NONE && trx->isolation_level > TRX_ISO_READ_UNCOMMITTED && trx->read_view.is_open()) { /* 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. */ dberr_t err = DB_SUCCESS; switch (row_sel_try_search_shortcut_for_mysql( &rec, prebuilt, &offsets, &heap, &mtr)) { case SEL_FOUND: /* 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(). */ ut_ad(!rec_get_deleted_flag(rec, comp)); if (prebuilt->pk_filter || prebuilt->idx_cond) { switch (row_search_idx_cond_check( buf, prebuilt, rec, offsets)) { case CHECK_ABORTED_BY_USER: goto aborted; case CHECK_NEG: case CHECK_OUT_OF_RANGE: case CHECK_ERROR: err = DB_RECORD_NOT_FOUND; goto shortcut_done; case CHECK_POS: goto shortcut_done; } ut_ad("incorrect code" == 0); aborted: err = DB_INTERRUPTED; goto shortcut_done; } if (!row_sel_store_mysql_rec( buf, prebuilt, rec, NULL, false, index, offsets)) { /* Only fresh inserts may contain incomplete externally stored columns. Pretend that such records do not exist. Such records may only be accessed at the READ UNCOMMITTED isolation level or when rolling back a recovered transaction. Rollback happens at a lower level, not here. */ /* Proceed as in case SEL_RETRY. */ break; } goto shortcut_done; case SEL_EXHAUSTED: err = DB_RECORD_NOT_FOUND; shortcut_done: mtr.commit(); /* NOTE that we do NOT store the cursor position */ trx->op_info = ""; ut_ad(!did_semi_consistent_read); if (UNIV_LIKELY_NULL(heap)) { mem_heap_free(heap); } DBUG_RETURN(err); case SEL_RETRY: break; default: ut_ad(0); } mtr.commit(); mtr.start(); } } #endif /* BTR_CUR_HASH_ADAPT */ /*-------------------------------------------------------------*/ /* PHASE 3: Open or restore index cursor position */ spatial_search = dict_index_is_spatial(index) && mode >= PAGE_CUR_CONTAIN; #ifdef UNIV_DEBUG /* The state of a running trx can only be changed by the thread that is currently serving the transaction. Because we are that thread, we can read trx->state without holding any mutex. */ switch (trx->state) { case TRX_STATE_ACTIVE: break; case TRX_STATE_NOT_STARTED: ut_ad(prebuilt->sql_stat_start || prebuilt->table->no_rollback()); break; default: ut_ad("invalid trx->state" == 0); } #endif ut_ad(prebuilt->sql_stat_start || prebuilt->select_lock_type != LOCK_NONE || trx->read_view.is_open() || prebuilt->table->no_rollback() || srv_read_only_mode); /* Do not lock gaps at READ UNCOMMITTED or READ COMMITTED isolation level */ const bool set_also_gap_locks = prebuilt->select_lock_type != LOCK_NONE && trx->isolation_level > TRX_ISO_READ_COMMITTED #ifdef WITH_WSREP && !wsrep_thd_skip_locking(trx->mysql_thd) #endif /* WITH_WSREP */ ; /* Note that if the search mode was GE or G, then the cursor naturally moves upward (in fetch next) in alphabetical order, otherwise downward */ if (UNIV_UNLIKELY(direction == 0)) { if (mode == PAGE_CUR_GE || mode == PAGE_CUR_G || mode >= PAGE_CUR_CONTAIN) { moves_up = true; } } else if (direction == ROW_SEL_NEXT) { moves_up = true; } thr = que_fork_get_first_thr(prebuilt->sel_graph); clust_index = dict_table_get_first_index(prebuilt->table); dberr_t err = DB_SUCCESS; /* Do some start-of-statement preparations */ if (prebuilt->table->no_rollback()) { /* NO_ROLLBACK tables do not support MVCC or locking. */ prebuilt->select_lock_type = LOCK_NONE; prebuilt->sql_stat_start = FALSE; } else if (!prebuilt->sql_stat_start) { /* No need to set an intention lock or assign a read view */ ut_a(prebuilt->select_lock_type != LOCK_NONE || srv_read_only_mode || trx->read_view.is_open()); } else { prebuilt->sql_stat_start = FALSE; trx_start_if_not_started(trx, false); if (prebuilt->select_lock_type == LOCK_NONE) { trx->read_view.open(trx); } else { wait_table_again: err = lock_table(prebuilt->table, nullptr, prebuilt->select_lock_type == LOCK_S ? LOCK_IS : LOCK_IX, thr); if (err != DB_SUCCESS) { table_lock_waited = TRUE; goto lock_table_wait; } } } /* Open or restore index cursor position */ if (UNIV_LIKELY(direction != 0)) { if (spatial_search) { /* R-Tree access does not need to do cursor position and resposition */ goto next_rec; } bool need_to_process = sel_restore_position_for_mysql( &same_user_rec, BTR_SEARCH_LEAF, pcur, moves_up, &mtr); if (UNIV_UNLIKELY(need_to_process)) { if (UNIV_UNLIKELY(!btr_pcur_get_rec(pcur))) { mtr.commit(); trx->op_info = ""; if (UNIV_LIKELY_NULL(heap)) { mem_heap_free(heap); } return DB_CORRUPTION; } if (UNIV_UNLIKELY(prebuilt->row_read_type == ROW_READ_DID_SEMI_CONSISTENT)) { /* 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 != ROW_READ_DID_SEMI_CONSISTENT)) { /* 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. */ goto next_rec_after_check; } } else if (dtuple_get_n_fields(search_tuple) > 0) { pcur->btr_cur.thr = thr; pcur->old_rec = nullptr; if (index->is_spatial()) { if (!prebuilt->rtr_info) { prebuilt->rtr_info = rtr_create_rtr_info( set_also_gap_locks, true, btr_pcur_get_btr_cur(pcur), index); prebuilt->rtr_info->search_tuple = search_tuple; prebuilt->rtr_info->search_mode = mode; rtr_info_update_btr(btr_pcur_get_btr_cur(pcur), prebuilt->rtr_info); } else { rtr_info_reinit_in_cursor( btr_pcur_get_btr_cur(pcur), index, set_also_gap_locks); prebuilt->rtr_info->search_tuple = search_tuple; prebuilt->rtr_info->search_mode = mode; } err = rtr_search_leaf(pcur, search_tuple, mode, &mtr); } else { err = btr_pcur_open_with_no_init(search_tuple, mode, BTR_SEARCH_LEAF, pcur, &mtr); } if (err != DB_SUCCESS) { page_corrupted: rec = NULL; goto page_read_error; } pcur->trx_if_known = trx; rec = btr_pcur_get_rec(pcur); ut_ad(page_rec_is_leaf(rec)); if (!moves_up && set_also_gap_locks && !page_rec_is_supremum(rec) && !dict_index_is_spatial(index)) { /* Try to place a gap lock on the next index record to prevent phantoms in ORDER BY ... DESC queries */ const rec_t* next_rec = page_rec_get_next_const(rec); if (UNIV_UNLIKELY(!next_rec)) { err = DB_CORRUPTION; goto page_corrupted; } offsets = rec_get_offsets(next_rec, index, offsets, index->n_core_fields, ULINT_UNDEFINED, &heap); err = sel_set_rec_lock(pcur, next_rec, index, offsets, prebuilt->select_lock_type, LOCK_GAP, thr, &mtr); switch (err) { case DB_SUCCESS_LOCKED_REC: err = DB_SUCCESS; /* fall through */ case DB_SUCCESS: break; default: goto lock_wait_or_error; } } } else if (mode == PAGE_CUR_G || mode == PAGE_CUR_L) { err = pcur->open_leaf(mode == PAGE_CUR_G, index, BTR_SEARCH_LEAF, &mtr); if (err != DB_SUCCESS) { if (err == DB_DECRYPTION_FAILED) { btr_decryption_failed(*index); } rec = NULL; goto page_read_error; } } /* Check if the table is supposed to be empty for our read view. If we read bulk_trx_id as an older transaction ID, it is not incorrect to check here whether that transaction should be visible to us. If bulk_trx_id is not visible to us, the table must have been empty at an earlier point of time, also in our read view. An INSERT would only update bulk_trx_id in row_ins_clust_index_entry_low() if the table really was empty (everything had been purged), when holding a leaf page latch in the clustered index (actually, the root page is the only leaf page in that case). We are already holding a leaf page latch here, either in a secondary index or in a clustered index. If we are holding a clustered index page latch, there clearly is no potential for race condition with a concurrent INSERT: such INSERT would be blocked by us. If we are holding a secondary index page latch, then we are not directly blocking a concurrent INSERT that might update bulk_trx_id to something that does not exist in our read view. But, in that case, the entire table (all indexes) must have been empty. So, even if our read below missed the update of index->table->bulk_trx_id, we can safely proceed to reading the empty secondary index page. Our latch will prevent the INSERT from proceeding to that page. It will first modify the clustered index. Also, we may only look up something in the clustered index if the secondary index page is not empty to begin with. So, only if the table is corrupted (the clustered index is empty but the secondary index is not) we could return corrupted results. */ if (trx->isolation_level == TRX_ISO_READ_UNCOMMITTED || !trx->read_view.is_open()) { } else if (trx_id_t bulk_trx_id = index->table->bulk_trx_id) { /* InnoDB should allow the transaction to read all the rows when InnoDB intends to do any locking on the record */ if (prebuilt->select_lock_type == LOCK_NONE && !trx->read_view.changes_visible(bulk_trx_id)) { trx->op_info = ""; err = DB_END_OF_INDEX; goto normal_return; } } rec_loop: DEBUG_SYNC_C("row_search_rec_loop"); if (trx_is_interrupted(trx)) { if (!spatial_search) { btr_pcur_store_position(pcur, &mtr); } err = DB_INTERRUPTED; goto normal_return; } /*-------------------------------------------------------------*/ /* PHASE 4: Look for matching records in a loop */ rec = btr_pcur_get_rec(pcur); ut_ad(!!page_rec_is_comp(rec) == comp); ut_ad(page_rec_is_leaf(rec)); 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 && !dict_index_is_spatial(index)) { /* Try to place a lock on the index record */ /* If the transaction isolation level is READ UNCOMMITTED or READ COMMITTED, we do not lock gaps. Supremum record is really a gap and therefore we do not set locks there. */ offsets = rec_get_offsets(rec, index, offsets, index->n_core_fields, ULINT_UNDEFINED, &heap); err = sel_set_rec_lock(pcur, rec, index, offsets, prebuilt->select_lock_type, LOCK_ORDINARY, thr, &mtr); switch (err) { case DB_SUCCESS_LOCKED_REC: err = DB_SUCCESS; /* fall through */ case DB_SUCCESS: break; default: 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 */ goto next_rec; } /*-------------------------------------------------------------*/ /* Do sanity checks in case our cursor has bumped into page corruption */ if (comp) { if (rec_get_info_bits(rec, true) & REC_INFO_MIN_REC_FLAG) { /* Skip the metadata pseudo-record. */ ut_ad(index->is_instant()); goto next_rec; } next_offs = rec_get_next_offs(rec, TRUE); if (UNIV_UNLIKELY(next_offs < PAGE_NEW_SUPREMUM)) { goto wrong_offs; } } else { if (rec_get_info_bits(rec, false) & REC_INFO_MIN_REC_FLAG) { /* Skip the metadata pseudo-record. */ ut_ad(index->is_instant()); goto next_rec; } next_offs = rec_get_next_offs(rec, FALSE); if (UNIV_UNLIKELY(next_offs < PAGE_OLD_SUPREMUM)) { goto wrong_offs; } } if (UNIV_UNLIKELY(next_offs >= srv_page_size - PAGE_DIR)) { wrong_offs: if (srv_force_recovery == 0 || moves_up == false) { ib::error() << "Rec address " << static_cast(rec) << ", buf block fix count " << btr_pcur_get_block(pcur)->page .buf_fix_count(); ib::error() << "Index corruption: rec offs " << page_offset(rec) << " next offs " << next_offs << btr_pcur_get_block(pcur)->page.id() << ", index " << index->name << " of table " << index->table->name << ". Run CHECK TABLE. You may need to" " restore from a backup, or dump + drop +" " reimport the table."; ut_ad(0); err = DB_CORRUPTION; goto page_read_error; } else { /* The user may be dumping a corrupt table. Jump over the corruption to recover as much as possible. */ ib::info() << "Index corruption: rec offs " << page_offset(rec) << " next offs " << next_offs << btr_pcur_get_block(pcur)->page.id() << ", index " << index->name << " of table " << index->table->name << ". We try to skip the rest of the page."; page_cur_set_after_last(btr_pcur_get_block(pcur), btr_pcur_get_page_cur(pcur)); pcur->old_rec = nullptr; goto next_rec; } } /*-------------------------------------------------------------*/ /* Calculate the 'offsets' associated with 'rec' */ ut_ad(fil_page_index_page_check(btr_pcur_get_page(pcur))); ut_ad(btr_page_get_index_id(btr_pcur_get_page(pcur)) == index->id); offsets = rec_get_offsets(rec, index, offsets, index->n_core_fields, ULINT_UNDEFINED, &heap); if (UNIV_UNLIKELY(srv_force_recovery > 0)) { if (!rec_validate(rec, offsets) || !btr_index_rec_validate(rec, index, FALSE)) { ib::error() << "Index corruption: rec offs " << page_offset(rec) << " next offs " << next_offs << btr_pcur_get_block(pcur)->page.id() << ", index " << index->name << " of table " << index->table->name << ". We try to skip the record."; 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); */ if (0 != cmp_dtuple_rec(search_tuple, rec, offsets)) { if (set_also_gap_locks && !dict_index_is_spatial(index)) { err = sel_set_rec_lock( pcur, rec, index, offsets, prebuilt->select_lock_type, LOCK_GAP, thr, &mtr); switch (err) { case DB_SUCCESS_LOCKED_REC: case DB_SUCCESS: break; default: goto lock_wait_or_error; } } btr_pcur_store_position(pcur, &mtr); /* The found record was not a match, but may be used as NEXT record (index_next). Set the relative position to BTR_PCUR_BEFORE, to reflect that the position of the persistent cursor is before the found/stored row (pcur->old_rec). */ ut_ad(pcur->rel_pos == BTR_PCUR_ON); pcur->rel_pos = BTR_PCUR_BEFORE; err = DB_RECORD_NOT_FOUND; goto normal_return; } } else if (match_mode == ROW_SEL_EXACT_PREFIX) { if (!cmp_dtuple_is_prefix_of_rec(search_tuple, rec, offsets)) { if (set_also_gap_locks && !dict_index_is_spatial(index)) { err = sel_set_rec_lock( pcur, rec, index, offsets, prebuilt->select_lock_type, LOCK_GAP, thr, &mtr); switch (err) { case DB_SUCCESS_LOCKED_REC: case DB_SUCCESS: break; default: goto lock_wait_or_error; } } btr_pcur_store_position(pcur, &mtr); /* The found record was not a match, but may be used as NEXT record (index_next). Set the relative position to BTR_PCUR_BEFORE, to reflect that the position of the persistent cursor is before the found/stored row (pcur->old_rec). */ ut_ad(pcur->rel_pos == BTR_PCUR_ON); pcur->rel_pos = BTR_PCUR_BEFORE; err = DB_RECORD_NOT_FOUND; 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. */ unsigned lock_type; if (trx->isolation_level <= TRX_ISO_READ_COMMITTED) { /* At READ COMMITTED or READ UNCOMMITTED isolation levels, do not lock committed delete-marked records. */ if (!rec_get_deleted_flag(rec, comp)) { goto no_gap_lock; } /* At most one transaction can be active for temporary table. */ if (clust_index->table->is_temporary()) { goto no_gap_lock; } if (index == clust_index) { trx_id_t trx_id = row_get_rec_trx_id( rec, index, offsets); /* In delete-marked records, DB_TRX_ID must always refer to an existing undo log record. */ ut_ad(trx_id); if (!trx_sys.is_registered(trx, trx_id)) { /* The clustered index record was delete-marked in a committed transaction. Ignore the record. */ goto locks_ok_del_marked; } } else if (trx_t* t = row_vers_impl_x_locked( trx, rec, index, offsets)) { /* The record belongs to an active transaction. We must acquire a lock. */ t->release_reference(); } else { /* The secondary index record does not point to a delete-marked clustered index record that belongs to an active transaction. Ignore the secondary index record, because it is not locked. */ goto next_rec; } goto no_gap_lock; } #ifdef WITH_WSREP if (UNIV_UNLIKELY(!set_also_gap_locks)) { ut_ad(wsrep_thd_skip_locking(trx->mysql_thd)); goto no_gap_lock; } #else /* WITH_WSREP */ ut_ad(set_also_gap_locks); #endif /* WITH_WSREP */ /* Set next-key lock both for delete- and non-delete-marked records for unique search, because non-delete-marked record can be marked as deleted while transaction suspends. */ if (index->is_spatial()) { goto no_gap_lock; } /* 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 && 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)) { no_gap_lock: lock_type = LOCK_REC_NOT_GAP; } else { lock_type = LOCK_ORDINARY; } err = sel_set_rec_lock(pcur, rec, index, offsets, prebuilt->select_lock_type, lock_type, thr, &mtr); switch (err) { const rec_t* old_vers; case DB_SUCCESS_LOCKED_REC: if (trx->isolation_level <= TRX_ISO_READ_COMMITTED) { /* Note that a record of prebuilt->index was locked. */ prebuilt->new_rec_locks = 1; } err = DB_SUCCESS; /* fall through */ case DB_SUCCESS: break; case DB_LOCK_WAIT: /* Lock wait for R-tree should already be handled in sel_set_rtr_rec_lock() */ ut_ad(!dict_index_is_spatial(index)); /* Never unlock rows that were part of a conflict. */ prebuilt->new_rec_locks = 0; if (UNIV_LIKELY(prebuilt->row_read_type != ROW_READ_TRY_SEMI_CONSISTENT) || unique_search || index != clust_index) { if (!prebuilt->skip_locked) { goto lock_wait_or_error; } } else { /* The following call returns 'offsets' associated with 'old_vers' */ row_sel_build_committed_vers_for_mysql( clust_index, prebuilt, rec, &offsets, &heap, &old_vers, need_vrow ? &vrow : NULL, &mtr); } /* Check whether it was a deadlock or not, if not a deadlock and the transaction had to wait then release the lock it is waiting on. */ err = lock_trx_handle_wait(trx); switch (err) { case DB_SUCCESS: ut_ad( !trx->lock.was_chosen_as_deadlock_victim); /* 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, index->n_core_fields, ULINT_UNDEFINED, &heap); goto locks_ok; case DB_DEADLOCK: goto lock_wait_or_error; case DB_LOCK_WAIT: ut_ad(!dict_index_is_spatial(index)); err = DB_SUCCESS; if (prebuilt->skip_locked) { goto next_rec; } break; case DB_LOCK_WAIT_TIMEOUT: if (prebuilt->skip_locked) { err = DB_SUCCESS; goto next_rec; } /* fall through */ default: ut_error; } if (old_vers == NULL) { /* The row was not yet committed */ goto next_rec; } did_semi_consistent_read = true; rec = old_vers; break; case DB_RECORD_NOT_FOUND: if (dict_index_is_spatial(index)) { goto next_rec; } else { goto lock_wait_or_error; } break; case DB_LOCK_WAIT_TIMEOUT: if (prebuilt->skip_locked) { err = DB_SUCCESS; goto next_rec; } /* fall through */ default: 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 || prebuilt->table->is_temporary() || prebuilt->table->no_rollback()) { /* Do nothing: we let a non-locking SELECT read the latest version of the record */ } else if (index == clust_index) { /* 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 */ err = row_sel_clust_sees(rec, *index, offsets, trx->read_view); switch (err) { default: goto lock_wait_or_error; case DB_SUCCESS: break; case DB_SUCCESS_LOCKED_REC: ut_ad(srv_force_recovery < SRV_FORCE_NO_UNDO_LOG_SCAN); rec_t* old_vers; /* The following call returns 'offsets' associated with 'old_vers' */ err = row_sel_build_prev_vers_for_mysql( prebuilt, clust_index, rec, &offsets, &heap, &old_vers, need_vrow ? &vrow : nullptr, &mtr); 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; } } else { /* 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(!dict_index_is_clust(index)); if (!srv_read_only_mode) { trx_id_t trx_id = page_get_max_trx_id( page_align(rec)); ut_ad(trx_id); if (trx->read_view.sees(trx_id)) { goto locks_ok; } /* We should look at the clustered index. However, as this is a non-locking read, we can skip the clustered index lookup if the condition does not match the secondary index entry. */ switch (row_search_idx_cond_check( buf, prebuilt, rec, offsets)) { case CHECK_NEG: goto next_rec; case CHECK_ABORTED_BY_USER: err = DB_INTERRUPTED; goto idx_cond_failed; case CHECK_OUT_OF_RANGE: case CHECK_ERROR: err = DB_RECORD_NOT_FOUND; goto idx_cond_failed; case CHECK_POS: goto requires_clust_rec; } ut_error; } } } locks_ok: /* 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 (rec_get_deleted_flag(rec, comp)) { locks_ok_del_marked: /* In delete-marked records, DB_TRX_ID must always refer to an existing undo log record. */ ut_ad(index != clust_index || row_get_rec_trx_id(rec, index, offsets)); /* The record is delete-marked: we can skip it */ /* 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 except for HANDLER queries because HANDLER allows NEXT and PREV even in unique search on clustered index. 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 && !prebuilt->used_in_HANDLER) { err = DB_RECORD_NOT_FOUND; goto normal_return; } goto next_rec; } /* Check if the record matches the index condition. */ switch (row_search_idx_cond_check(buf, prebuilt, rec, offsets)) { case CHECK_NEG: if (did_semi_consistent_read) { row_unlock_for_mysql(prebuilt, TRUE); } goto next_rec; case CHECK_ABORTED_BY_USER: err = DB_INTERRUPTED; goto idx_cond_failed; case CHECK_OUT_OF_RANGE: case CHECK_ERROR: err = DB_RECORD_NOT_FOUND; goto idx_cond_failed; case CHECK_POS: break; } if (index != clust_index && prebuilt->need_to_access_clustered) { if (row_search_with_covering_prefix(prebuilt, rec, offsets)) { goto use_covering_index; } requires_clust_rec: ut_ad(index != clust_index); /* 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_extra_clust_savepoint = mtr.get_savepoint(); ut_ad(!vrow); /* The following call returns 'offsets' associated with 'clust_rec'. Note that 'clust_rec' can be an old version built for a consistent read. */ err = row_sel_get_clust_rec_for_mysql(prebuilt, index, rec, thr, &clust_rec, &offsets, &heap, need_vrow ? &vrow : NULL, &mtr); if (err == DB_LOCK_WAIT && prebuilt->skip_locked) { err = lock_trx_handle_wait(trx); } switch (err) { case DB_SUCCESS: if (clust_rec == NULL) { /* The record did not exist in the read view */ ut_ad(prebuilt->select_lock_type == LOCK_NONE || dict_index_is_spatial(index)); goto next_rec; } break; case DB_SUCCESS_LOCKED_REC: ut_a(clust_rec != NULL); if (trx->isolation_level <= TRX_ISO_READ_COMMITTED) { /* Note that the clustered index record was locked. */ prebuilt->new_rec_locks = 2; } err = DB_SUCCESS; break; case DB_LOCK_WAIT_TIMEOUT: case DB_LOCK_WAIT: if (prebuilt->skip_locked) { err = DB_SUCCESS; goto next_rec; } /* fall through */ default: vrow = NULL; goto lock_wait_or_error; } if (rec_get_deleted_flag(clust_rec, comp)) { /* The record is delete marked: we can skip it */ if (trx->isolation_level <= TRX_ISO_READ_COMMITTED && prebuilt->select_lock_type != LOCK_NONE) { /* 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; } if (need_vrow && !vrow) { if (!heap) { heap = mem_heap_create(100); } row_sel_fill_vrow(rec, index, &vrow, heap); } result_rec = clust_rec; ut_ad(rec_offs_validate(result_rec, clust_index, offsets)); if (prebuilt->pk_filter || prebuilt->idx_cond) { /* Convert the record to MySQL format. We were unable to do this in row_search_idx_cond_check(), because the condition is on the secondary index and the requested column is in the clustered index. We convert all fields, including those that may have been used in ICP, because the secondary index may contain a column prefix rather than the full column. Also, as noted in Bug #56680, the column in the secondary index may be in the wrong case, and the authoritative case is in result_rec, the appropriate version of the clustered index record. */ if (!row_sel_store_mysql_rec( buf, prebuilt, result_rec, vrow, true, clust_index, offsets)) { goto next_rec; } } } else { use_covering_index: 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)); ut_ad(!rec_get_deleted_flag(result_rec, comp)); /* Decide whether to prefetch extra rows. 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(). */ if ((match_mode == ROW_SEL_EXACT || 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->in_fts_query) { /* 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. */ ut_a(prebuilt->n_fetch_cached < MYSQL_FETCH_CACHE_SIZE); /* We only convert from InnoDB row format to MySQL row format when ICP is disabled. */ if (!prebuilt->pk_filter && !prebuilt->idx_cond) { /* We use next_buf to track the allocation of buffers where we store and enqueue the buffers for our pre-fetch optimisation. If next_buf == 0 then we store the converted record directly into the MySQL record buffer (buf). If it is != 0 then we allocate a pre-fetch buffer and store the converted record there. If the conversion fails and the MySQL record buffer was not written to then we reset next_buf so that we can re-use the MySQL record buffer in the next iteration. */ next_buf = next_buf ? row_sel_fetch_last_buf(prebuilt) : buf; if (!row_sel_store_mysql_rec( next_buf, prebuilt, result_rec, vrow, result_rec != rec, result_rec != rec ? clust_index : index, offsets)) { if (next_buf == buf) { ut_a(prebuilt->n_fetch_cached == 0); next_buf = 0; } /* Only fresh inserts may contain incomplete externally stored columns. Pretend that such records do not exist. Such records may only be accessed at the READ UNCOMMITTED isolation level or when rolling back a recovered transaction. Rollback happens at a lower level, not here. */ goto next_rec; } if (next_buf != buf) { row_sel_enqueue_cache_row_for_mysql( next_buf, prebuilt); } } else { row_sel_enqueue_cache_row_for_mysql(buf, prebuilt); } if (prebuilt->n_fetch_cached < MYSQL_FETCH_CACHE_SIZE) { goto next_rec; } } else { if (!prebuilt->pk_filter && !prebuilt->idx_cond) { /* The record was not yet converted to MySQL format. */ if (!row_sel_store_mysql_rec( buf, prebuilt, result_rec, vrow, result_rec != rec, result_rec != rec ? clust_index : index, offsets)) { /* Only fresh inserts may contain incomplete externally stored columns. Pretend that such records do not exist. Such records may only be accessed at the READ UNCOMMITTED isolation level or when rolling back a recovered transaction. Rollback happens at a lower level, not here. */ goto next_rec; } } if (!prebuilt->clust_index_was_generated) { } else if (result_rec != rec || index->is_primary()) { memcpy(prebuilt->row_id, result_rec, DATA_ROW_ID_LEN); } else { ulint len; const byte* data = rec_get_nth_field( result_rec, offsets, index->n_fields - 1, &len); ut_ad(dict_index_get_nth_col(index, index->n_fields - 1) ->prtype == (DATA_ROW_ID | DATA_NOT_NULL)); ut_ad(len == DATA_ROW_ID_LEN); memcpy(prebuilt->row_id, data, DATA_ROW_ID_LEN); } } /* From this point on, 'offsets' are invalid. */ /* 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 return 'end of file'. Exceptions are locking reads and the MySQL HANDLER command where the user can move the cursor with PREV or NEXT even after a unique search. */ err = DB_SUCCESS; idx_cond_failed: if (!unique_search || !dict_index_is_clust(index) || direction != 0 || prebuilt->select_lock_type != LOCK_NONE || prebuilt->used_in_HANDLER) { /* Inside an update always store the cursor position */ if (!spatial_search) { btr_pcur_store_position(pcur, &mtr); } } goto normal_return; next_rec: /* Reset the old and new "did semi-consistent read" flags. */ if (UNIV_UNLIKELY(prebuilt->row_read_type == ROW_READ_DID_SEMI_CONSISTENT)) { prebuilt->row_read_type = ROW_READ_TRY_SEMI_CONSISTENT; } next_rec_after_check: did_semi_consistent_read = false; prebuilt->new_rec_locks = 0; vrow = NULL; /*-------------------------------------------------------------*/ /* PHASE 5: Move the cursor to the next index record */ /* NOTE: For moves_up==FALSE, the mini-transaction will be committed and restarted every time when switching b-tree pages. For moves_up==TRUE in index condition pushdown, we can scan an entire secondary index tree within a single mini-transaction. As long as the prebuilt->idx_cond does not match, we do not need to consult the clustered index or return records to MySQL, and thus we can avoid repositioning the cursor. What prevents us from buffer-fixing all leaf pages within the mini-transaction is the btr_leaf_page_release() call in btr_pcur_move_to_next_page(). Only the leaf page where the cursor is positioned will remain buffer-fixed. For R-tree spatial search, we also commit the mini-transaction each time */ if (spatial_search) { /* No need to do store restore for R-tree */ mtr.rollback_to_savepoint(0); } else if (mtr_extra_clust_savepoint) { /* We must release any clustered index latches 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. */ mtr.rollback_to_savepoint(mtr_extra_clust_savepoint); } mtr_extra_clust_savepoint = 0; if (moves_up) { if (UNIV_UNLIKELY(spatial_search)) { if (rtr_pcur_move_to_next( search_tuple, mode, pcur, 0, &mtr)) { goto rec_loop; } } else { /* This is based on btr_pcur_move_to_next() */ ut_ad(pcur->pos_state == BTR_PCUR_IS_POSITIONED); ut_ad(pcur->latch_mode != BTR_NO_LATCHES); pcur->old_rec = nullptr; if (btr_pcur_is_after_last_on_page(pcur)) { if (btr_pcur_is_after_last_in_tree(pcur)) { goto not_moved; } err = btr_pcur_move_to_next_page(pcur, &mtr); if (err != DB_SUCCESS) { goto lock_wait_or_error; } } else if (!btr_pcur_move_to_next_on_page(pcur)) { goto corrupted; } goto rec_loop; } } else { if (btr_pcur_move_to_prev(pcur, &mtr)) { goto rec_loop; } if (UNIV_UNLIKELY(!btr_pcur_get_rec(pcur))) { corrupted: err = DB_CORRUPTION; goto normal_return; } } not_moved: if (!spatial_search) { btr_pcur_store_position(pcur, &mtr); } err = match_mode ? DB_RECORD_NOT_FOUND : DB_END_OF_INDEX; goto normal_return; lock_wait_or_error: if (!dict_index_is_spatial(index)) { btr_pcur_store_position(pcur, &mtr); } page_read_error: /* Reset the old and new "did semi-consistent read" flags. */ if (UNIV_UNLIKELY(prebuilt->row_read_type == ROW_READ_DID_SEMI_CONSISTENT)) { prebuilt->row_read_type = ROW_READ_TRY_SEMI_CONSISTENT; } did_semi_consistent_read = false; lock_table_wait: mtr.commit(); mtr_extra_clust_savepoint = 0; trx->error_state = err; 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(); /* Table lock waited, go try to obtain table lock again */ if (table_lock_waited) { table_lock_waited = FALSE; goto wait_table_again; } if (!dict_index_is_spatial(index)) { sel_restore_position_for_mysql( &same_user_rec, BTR_SEARCH_LEAF, pcur, moves_up, &mtr); } if (trx->isolation_level <= TRX_ISO_READ_COMMITTED && !same_user_rec) { /* 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.cc 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. */ prebuilt->new_rec_locks = 0; } mode = pcur->search_mode; goto rec_loop; } thr->lock_state = QUE_THR_LOCK_NOLOCK; goto func_exit; normal_return: mtr.commit(); DEBUG_SYNC_C("row_search_for_mysql_before_return"); if (prebuilt->pk_filter || prebuilt->idx_cond) { /* When ICP is active we don't write to the MySQL buffer directly, only to buffers that are enqueued in the pre-fetch queue. We need to dequeue the first buffer and copy the contents to the record buffer that was passed in by MySQL. */ if (prebuilt->n_fetch_cached > 0) { row_sel_dequeue_cached_row_for_mysql(buf, prebuilt); err = DB_SUCCESS; } } else if (next_buf != 0) { /* We may or may not have enqueued some buffers to the pre-fetch queue, but we definitely wrote to the record buffer passed to use by MySQL. */ DEBUG_SYNC_C("row_search_cached_row"); err = DB_SUCCESS; } #ifdef UNIV_DEBUG if (dict_index_is_spatial(index) && err != DB_SUCCESS && err != DB_END_OF_INDEX && err != DB_INTERRUPTED) { rtr_node_path_t* path = pcur->btr_cur.rtr_info->path; ut_ad(path->empty()); } #endif func_exit: trx->op_info = ""; if (UNIV_LIKELY_NULL(heap)) { mem_heap_free(heap); } /* 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 || !did_semi_consistent_read); if (prebuilt->row_read_type != ROW_READ_WITH_LOCKS) { if (did_semi_consistent_read) { prebuilt->row_read_type = ROW_READ_DID_SEMI_CONSISTENT; } else { prebuilt->row_read_type = ROW_READ_TRY_SEMI_CONSISTENT; } } DEBUG_SYNC_C("innodb_row_search_for_mysql_exit"); DBUG_RETURN(err); } /********************************************************************//** Count rows in a R-Tree leaf level. @return DB_SUCCESS if successful */ dberr_t row_count_rtree_recs( /*=================*/ row_prebuilt_t* prebuilt, /*!< in: prebuilt struct for the 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' */ ulint* n_rows) /*!< out: number of entries seen in the consistent read */ { dict_index_t* index = prebuilt->index; dberr_t ret = DB_SUCCESS; mtr_t mtr; mem_heap_t* heap; dtuple_t* entry; dtuple_t* search_entry = prebuilt->search_tuple; ulint entry_len; ulint i; byte* buf; ut_a(dict_index_is_spatial(index)); *n_rows = 0; heap = mem_heap_create(256); /* Build a search tuple. */ entry_len = dict_index_get_n_fields(index); entry = dtuple_create(heap, entry_len); for (i = 0; i < entry_len; i++) { const dict_field_t* ind_field = dict_index_get_nth_field(index, i); const dict_col_t* col = ind_field->col; dfield_t* dfield = dtuple_get_nth_field(entry, i); if (i == 0) { double* mbr; double tmp_mbr[SPDIMS * 2]; dfield->type.mtype = DATA_GEOMETRY; dfield->type.prtype |= DATA_GIS_MBR; /* Allocate memory for mbr field */ mbr = static_cast (mem_heap_alloc(heap, DATA_MBR_LEN)); /* Set mbr field data. */ dfield_set_data(dfield, mbr, DATA_MBR_LEN); for (uint j = 0; j < SPDIMS; j++) { tmp_mbr[j * 2] = DBL_MAX; tmp_mbr[j * 2 + 1] = -DBL_MAX; } dfield_write_mbr(dfield, tmp_mbr); continue; } dfield->type.mtype = col->mtype; dfield->type.prtype = col->prtype; } prebuilt->search_tuple = entry; ulint bufsize = std::max(srv_page_size, prebuilt->mysql_row_len); buf = static_cast(ut_malloc_nokey(bufsize)); ulint direction = 0; loop: ret = row_search_mvcc(buf, PAGE_CUR_WITHIN, prebuilt, 0, direction); direction = ROW_SEL_NEXT; switch (ret) { case DB_SUCCESS: break; case DB_DEADLOCK: case DB_LOCK_TABLE_FULL: case DB_LOCK_WAIT_TIMEOUT: case DB_INTERRUPTED: goto func_exit; default: /* fall through (this error is ignored by CHECK TABLE) */ case DB_END_OF_INDEX: ret = DB_SUCCESS; func_exit: prebuilt->search_tuple = search_entry; ut_free(buf); mem_heap_free(heap); return(ret); } ++*n_rows; goto loop; } /** Check if a version of a clustered index record and a secondary index record match. @param prebuilt index and transaction @param clust_rec a version of a clustered index record @param clust_index clustered index @param clust_offsets rec_get_offsets(clust_rec, clust_index) @param rec secondary index leaf page record @param offsets rec_get_offsets(rec, index) @return an error code @retval DB_SUCCESS if rec matches clust_rec @retval DB_SUCCESS_LOCKED_REC if rec does not match clust_rec */ static dberr_t row_check_index_match(row_prebuilt_t *prebuilt, const rec_t *clust_rec, const dict_index_t *clust_index, const rec_offs *clust_offsets, const rec_t *rec, const dict_index_t *index, const rec_offs *offsets) { ut_ad(index == prebuilt->index); ib_vcol_row vc(index->has_virtual() ? mem_heap_create(256) : nullptr); const uint16_t n= index->n_user_defined_cols; for (uint16_t i= 0; i < n; i++) { ulint pos= 0; ulint len, sec_len; const dict_field_t &ifield= index->fields[i]; const byte *sec_field= rec_get_nth_field(rec, offsets, i, &sec_len); const byte *field; if (ifield.col->is_virtual()) { /* Virtual column values must be reconstructed from the base columns. */ row_ext_t *ext; byte *record= vc.record(prebuilt->trx->mysql_thd, clust_index, &prebuilt->m_mysql_table); const dict_v_col_t *v_col= reinterpret_cast (ifield.col); dtuple_t *row= row_build(ROW_COPY_POINTERS, clust_index, clust_rec, clust_offsets, nullptr, nullptr, nullptr, &ext, vc.heap); if (dfield_t *vfield= innobase_get_computed_value(row, v_col, clust_index, &vc.heap, nullptr, nullptr, prebuilt->trx->mysql_thd, prebuilt->m_mysql_table, record, nullptr, nullptr)) { len= vfield->len; field= static_cast(vfield->data); } else { innobase_report_computed_value_failed(row); return DB_COMPUTE_VALUE_FAILED; } } else { pos= dict_col_get_clust_pos(ifield.col, clust_index); field= rec_get_nth_cfield(clust_rec, clust_index, clust_offsets, pos, &len); if (len == UNIV_SQL_NULL) { if (sec_len == UNIV_SQL_NULL) continue; return DB_SUCCESS_LOCKED_REC; } if (sec_len == UNIV_SQL_NULL) return DB_SUCCESS_LOCKED_REC; if (rec_offs_nth_extern(clust_offsets, pos)) { if (len == BTR_EXTERN_FIELD_REF_SIZE) goto compare_blobs; len-= BTR_EXTERN_FIELD_REF_SIZE; } if (ifield.prefix_len) { len= dtype_get_at_most_n_mbchars(ifield.col->prtype, ifield.col->mbminlen, ifield.col->mbmaxlen, ifield.prefix_len, len, reinterpret_cast(field)); if (len < sec_len) goto check_for_blob; } else { check_for_blob: if (rec_offs_nth_extern(clust_offsets, pos)) { compare_blobs: if (!row_sel_sec_rec_is_for_blob(ifield.col->mtype, ifield.col->prtype, ifield.col->mbminlen, ifield.col->mbmaxlen, field, len, sec_field, sec_len, ifield.prefix_len, clust_index->table)) return DB_SUCCESS_LOCKED_REC; continue; } } } if (cmp_data_data(ifield.col->mtype, ifield.col->prtype, field, len, sec_field, sec_len)) return DB_SUCCESS_LOCKED_REC; } return DB_SUCCESS; } /** Check the index records in CHECK TABLE. The index must contain entries in an ascending order, unique constraint must not be violated by duplicated keys, and the number of index entries is counted in according to the current read view. @param prebuilt index and transaction @param n_rows number of records counted @return error code @retval DB_SUCCESS if no error was found */ dberr_t row_check_index(row_prebuilt_t *prebuilt, ulint *n_rows) { rec_offs offsets_[REC_OFFS_NORMAL_SIZE]; rec_offs_init(offsets_); *n_rows= 0; dict_index_t *const index= prebuilt->index; if (!index->is_btree()) return DB_CORRUPTION; mem_heap_t *heap= mem_heap_create(100); dtuple_t *prev_entry= nullptr; mtr_t mtr; mtr.start(); dict_index_t *clust_index= dict_table_get_first_index(prebuilt->table); prebuilt->clust_pcur->btr_cur.page_cur.index = clust_index; dberr_t err= prebuilt->pcur->open_leaf(true, index, BTR_SEARCH_LEAF, &mtr); if (UNIV_UNLIKELY(err != DB_SUCCESS)) { func_exit: mtr.commit(); mem_heap_free(heap); return err; } if (const trx_id_t bulk_trx_id= index->table->bulk_trx_id) if (!prebuilt->trx->read_view.changes_visible(bulk_trx_id)) goto func_exit; ReadView check_table_extended_view; ReadView &view= prebuilt->need_to_access_clustered && !prebuilt->table->is_temporary() && prebuilt->trx->isolation_level != TRX_ISO_READ_UNCOMMITTED ? check_table_extended_view : prebuilt->trx->read_view; if (&view == &check_table_extended_view) check_table_extended_view.set_creator_trx_id(prebuilt->trx->id); page_loop: if (&view == &check_table_extended_view) /* In CHECK TABLE...EXTENDED, we make a copy of purge_sys.end_view while holding a shared latch on the index leaf page. Should a currently active purge batch desire to remove any further records from this page, it would be blocked by our page latch. We will consult check_table_extended_view to determine if a clustered index record corresponding to a secondary index record is visible to the current purge batch. Right after we have made our copy, purge_sys.end_view is free to be changed again. If we have an orphan secondary index record, we may attempt to request a clustered index record version that cannot be retrieved any more because the undo log records may have been freed (according to the purge_sys.end_view). In such a case, trx_undo_get_undo_rec() would cause trx_undo_prev_version_build() and trx_undo_prev_version_build() to return DB_MISSING_HISTORY. */ static_cast(check_table_extended_view)= purge_sys_t::end_view_guard{}.view(); rec_loop: ut_ad(err == DB_SUCCESS); if (!btr_pcur_move_to_next_on_page(prebuilt->pcur)) { err= DB_CORRUPTION; goto func_exit; } const rec_t *rec= btr_pcur_get_rec(prebuilt->pcur); rec_offs *offsets= offsets_; if (page_rec_is_supremum(rec)) { next_page: if (btr_pcur_is_after_last_in_tree(prebuilt->pcur)) goto func_exit; err= btr_pcur_move_to_next_page(prebuilt->pcur, &mtr); if (err == DB_SUCCESS && trx_is_interrupted(prebuilt->trx)) err= DB_INTERRUPTED; if (UNIV_UNLIKELY(err != DB_SUCCESS)) goto func_exit; goto page_loop; } offsets= rec_get_offsets(rec, index, offsets, index->n_core_fields, ULINT_UNDEFINED, &heap); const auto info_bits= rec_get_info_bits(rec, prebuilt->table->not_redundant()); const bool rec_deleted= info_bits & REC_INFO_DELETED_FLAG; if (UNIV_UNLIKELY(info_bits & REC_INFO_MIN_REC_FLAG)) { if (*n_rows || !index->is_instant()) { push_warning_printf(prebuilt->trx->mysql_thd, Sql_condition::WARN_LEVEL_WARN, ER_NOT_KEYFILE, "InnoDB: invalid record encountered"); prebuilt->autoinc_error= DB_INDEX_CORRUPT; } goto next_rec; } if (prebuilt->table->is_temporary()) { count_or_not: if (rec_deleted) goto next_rec; } else if (index->is_clust()) { if (prebuilt->trx->isolation_level == TRX_ISO_READ_UNCOMMITTED) goto count_or_not; trx_id_t rec_trx_id= row_get_rec_trx_id(rec, index, offsets); if (rec_trx_id >= prebuilt->trx->read_view.low_limit_id() && UNIV_UNLIKELY(rec_trx_id >= trx_sys.get_max_trx_id())) { invalid_trx_id: if (prebuilt->autoinc_error == DB_SUCCESS) push_warning_printf(prebuilt->trx->mysql_thd, Sql_condition::WARN_LEVEL_WARN, ER_NOT_KEYFILE, "InnoDB: DB_TRX_ID=" TRX_ID_FMT " exceeds the system-wide maximum", rec_trx_id); prebuilt->autoinc_error= DB_CORRUPTION; goto next_rec; } if (!prebuilt->trx->read_view.changes_visible(rec_trx_id)) { ut_ad(srv_force_recovery < SRV_FORCE_NO_UNDO_LOG_SCAN); rec_t *old_vers; /* The following call returns 'offsets' associated with 'old_vers' */ err= row_sel_build_prev_vers_for_mysql(prebuilt, index, rec, &offsets, &heap, &old_vers, nullptr, &mtr); if (err != DB_SUCCESS) goto func_exit; if (old_vers) { rec= old_vers; rec_trx_id= row_get_rec_trx_id(rec, index, offsets); if (rec_trx_id >= prebuilt->trx->read_view.low_limit_id() && UNIV_UNLIKELY(rec_trx_id >= trx_sys.get_max_trx_id())) goto invalid_trx_id; if (!rec_get_deleted_flag(rec, prebuilt->table->not_redundant())) goto count_row; } else offsets= rec_get_offsets(rec, index, offsets, index->n_core_fields, ULINT_UNDEFINED, &heap); goto next_rec; } else if (!rec_deleted && !rec_trx_id); else if (!check_table_extended_view.changes_visible(rec_trx_id)); else if (prebuilt->autoinc_error == DB_SUCCESS) { const char *msg= rec_deleted ? "Unpurged clustered index record" : "Clustered index record with stale history"; ib::warn w; w << msg << " in table " << index->table->name << ": " << rec_offsets_print(rec, offsets); prebuilt->autoinc_error= DB_MISSING_HISTORY; push_warning_printf(prebuilt->trx->mysql_thd, Sql_condition::WARN_LEVEL_WARN, ER_NOT_KEYFILE, "InnoDB: %s", w.m_oss.str().c_str()); } goto count_or_not; } else if (const trx_id_t page_trx_id= page_get_max_trx_id(page_align(rec))) { if (page_trx_id >= trx_sys.get_max_trx_id()) goto invalid_PAGE_MAX_TRX_ID; if (prebuilt->trx->isolation_level == TRX_ISO_READ_UNCOMMITTED); else if (&view == &check_table_extended_view || rec_deleted || !view.sees(page_trx_id)) { bool got_extended_match= &view == &check_table_extended_view; const auto savepoint= mtr.get_savepoint(); row_build_row_ref_in_tuple(prebuilt->clust_ref, rec, index, offsets); err= btr_pcur_open_with_no_init(prebuilt->clust_ref, PAGE_CUR_LE, BTR_SEARCH_LEAF, prebuilt->clust_pcur, &mtr); if (err != DB_SUCCESS) goto func_exit; const rec_t *clust_rec= btr_pcur_get_rec(prebuilt->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) || btr_pcur_get_low_match(prebuilt->clust_pcur) < clust_index->n_uniq) { if (!rec_deleted) { not_found: /* MDEV-29823 FIXME: There is a race condition between rollback, purge, and possibly other SQL connections that are creating and releasing read views. At the time row_undo_mod_del_mark_or_remove_sec_low() is executing rollback on a secondary index record, purge_sys.view may not allow it to delete the record, and it will be delete-marked. Eventually purge_sys.view would advance, but the delete-marked record could never be removed, because no undo log record was ever added to the purge queue by trx_purge_add_undo_to_history(). For now, we will not flag an error about orphan secondary index records that are delete-marked; we will only warn about them. */ if (!rec_deleted || prebuilt->autoinc_error == DB_SUCCESS) { ib::error_or_warn w(!rec_deleted); w << "Clustered index record not found for index " << index->name << " of table " << index->table->name << ": " << rec_offsets_print(rec, offsets); push_warning_printf(prebuilt->trx->mysql_thd, Sql_condition::WARN_LEVEL_WARN, ER_NOT_KEYFILE, "InnoDB: %s", w.m_oss.str().c_str()); } if (prebuilt->autoinc_error == DB_SUCCESS) prebuilt->autoinc_error= rec_deleted ? DB_MISSING_HISTORY : DB_CORRUPTION; } else if (&view == &check_table_extended_view) extended_not_found: if (view.changes_visible(page_trx_id)) goto not_found; did_not_find: mtr.rollback_to_savepoint(savepoint); goto next_rec; } rec_offs *clust_offsets; trx_id_t rec_trx_id; rec_t *old_vers= nullptr; bool found_in_view= false; trx_id_t visible_trx_id= ~0ULL; if (ulint trx_id_offset= clust_index->trx_id_offset) { clust_offsets= nullptr; read_trx_id: rec_trx_id= trx_read_trx_id(clust_rec + trx_id_offset); if (clust_rec[trx_id_offset + DATA_TRX_ID_LEN] & 0x80) { if (UNIV_UNLIKELY (rec_get_deleted_flag(clust_rec, prebuilt->table->not_redundant()))) { err= DB_CORRUPTION; goto func_exit; } /* This is the oldest available record version (fresh insert). */ if (!view.changes_visible(rec_trx_id)) { if (rec_trx_id >= view.low_limit_id() && UNIV_UNLIKELY(rec_trx_id >= trx_sys.get_max_trx_id())) goto invalid_rec_trx_id; if (got_extended_match) goto check_latest_version; goto did_not_find; } } } else { clust_offsets= rec_get_offsets(clust_rec, clust_index, nullptr, clust_index->n_core_fields, ULINT_UNDEFINED, &heap); ulint trx_id_pos= clust_index->n_uniq ? clust_index->n_uniq : 1; ulint len; trx_id_offset= rec_get_nth_field_offs(clust_offsets, trx_id_pos, &len); ut_ad(len == DATA_TRX_ID_LEN); goto read_trx_id; } if (got_extended_match) { check_latest_version: /* In CHECK TABLE...EXTENDED, always check if the secondary index record matches the latest clustered index record version, no matter if it is visible in our own read view. If the latest clustered index version is delete-marked and purgeable, it is not safe to fetch any BLOBs for column prefix indexes because they may already have been freed. */ if (rec_trx_id && rec_get_deleted_flag(clust_rec, prebuilt->table->not_redundant()) && purge_sys.is_purgeable(rec_trx_id)) goto did_not_find; if (!clust_offsets) clust_offsets= rec_get_offsets(clust_rec, clust_index, nullptr, clust_index->n_core_fields, ULINT_UNDEFINED, &heap); err= row_check_index_match(prebuilt, clust_rec, clust_index, clust_offsets, rec, index, offsets); switch (err) { default: goto func_exit; case DB_SUCCESS_LOCKED_REC: case DB_SUCCESS: break; } got_extended_match= err == DB_SUCCESS; err= DB_SUCCESS; if (!prebuilt->trx->read_view.changes_visible(rec_trx_id)) /* While CHECK TABLE ... EXTENDED checks for a matching clustered index record version for each secondary index record, it must count only those records that belong to its own read view. If the latest version of clust_rec matches rec but is not in our read view, there may still be an older version of clust_rec that not only matches rec but is in our view. We must evaluate old versions before deciding whether rec should be counted. */ goto check_old_vers; /* Remember that this is the visible clust_rec for rec, and whether it matches rec. */ visible_trx_id= rec_trx_id; found_in_view= got_extended_match && !rec_get_deleted_flag(clust_rec, prebuilt->table->not_redundant()); if (!got_extended_match) goto check_old_vers; if (!found_in_view) goto did_not_find; found_match: mtr.rollback_to_savepoint(savepoint); goto count_row; } else if (!view.changes_visible(rec_trx_id)) { check_old_vers: if (rec_trx_id >= view.low_limit_id() && UNIV_UNLIKELY(rec_trx_id >= trx_sys.get_max_trx_id())) { invalid_rec_trx_id: if (prebuilt->autoinc_error == DB_SUCCESS) push_warning_printf(prebuilt->trx->mysql_thd, Sql_condition::WARN_LEVEL_WARN, ER_NOT_KEYFILE, "InnoDB: DB_TRX_ID=" TRX_ID_FMT " exceeds the system-wide maximum", rec_trx_id); goto not_found; } if (!clust_offsets) clust_offsets= rec_get_offsets(clust_rec, clust_index, nullptr, clust_index->n_core_fields, ULINT_UNDEFINED, &heap); row_sel_reset_old_vers_heap(prebuilt); /* The following is adapted from row_vers_build_for_consistent_read() because when using check_table_extended_view, we must consider every available version of the clustered index record. */ mem_heap_t *vers_heap= nullptr; for (;;) { mem_heap_t *prev_heap= vers_heap; vers_heap= mem_heap_create(1024); err= trx_undo_prev_version_build(clust_rec, clust_index, clust_offsets, vers_heap, &old_vers, nullptr, nullptr, 0); if (prev_heap) mem_heap_free(prev_heap); if (err != DB_SUCCESS) { old_vers_err: mem_heap_free(vers_heap); if (err == DB_MISSING_HISTORY) { err= DB_SUCCESS; if (got_extended_match) goto did_not_find; goto not_found; } goto func_exit; } if (UNIV_UNLIKELY(!old_vers)) { mem_heap_free(vers_heap); /* We did not find a matching clustered index record version for the secondary index record. Normal CHECK TABLE will simply not count the secondary index record; CHECK TABLE ... EXTENDED will flag such orphan records if appropriate. A secondary index record may may be "temporarily orphan" if purge is in progress. We will only flag them if everything up to PAGE_MAX_TRX_ID has been fully purged. "Temporary orphans" may be produced when row_undo_mod_clust() resets the DB_TRX_ID of the latest clust_rec version or when trx_undo_prev_version_build() encounters a BLOB that may have been freed according to purge_sys.view (not purge_sys.end_view). */ if (&view == &check_table_extended_view && !got_extended_match) goto extended_not_found; goto did_not_find; } clust_rec= old_vers; clust_offsets= rec_get_offsets(clust_rec, clust_index, clust_offsets, clust_index->n_core_fields, ULINT_UNDEFINED, &heap); rec_trx_id= row_get_rec_trx_id(clust_rec, clust_index, clust_offsets); if (UNIV_UNLIKELY(rec_trx_id >= prebuilt->trx->read_view.low_limit_id() && rec_trx_id >= trx_sys.get_max_trx_id())) { mem_heap_free(vers_heap); goto invalid_rec_trx_id; } const bool rec_visible= prebuilt->trx->read_view.changes_visible(rec_trx_id); const bool clust_rec_deleted= rec_get_deleted_flag(clust_rec, prebuilt->table->not_redundant()); if (&view != &prebuilt->trx->read_view) { /* It is not safe to fetch BLOBs of committed delete-marked records that may have been freed in purge. */ err= clust_rec_deleted && rec_trx_id && purge_sys.is_purgeable(rec_trx_id) ? DB_SUCCESS_LOCKED_REC : row_check_index_match(prebuilt, clust_rec, clust_index, clust_offsets, rec, index, offsets); switch (err) { default: goto old_vers_err; case DB_SUCCESS_LOCKED_REC: if (rec_visible && !~visible_trx_id) visible_trx_id= rec_trx_id; continue; case DB_SUCCESS: got_extended_match= true; if (!rec_visible) continue; if (!~visible_trx_id) { visible_trx_id= rec_trx_id; found_in_view= !clust_rec_deleted; } mem_heap_free(vers_heap); if (!found_in_view) goto did_not_find; goto found_match; } } else if (rec_visible) { if (!clust_rec_deleted) { clust_rec= rec_copy(mem_heap_alloc(heap, rec_offs_size(clust_offsets)), clust_rec, clust_offsets); rec_offs_make_valid(clust_rec, clust_index, true, clust_offsets); } mem_heap_free(vers_heap); if (clust_rec_deleted) goto did_not_find; goto check_match; } } } else if (rec_get_deleted_flag(clust_rec, prebuilt->table->not_redundant())) goto did_not_find; ut_ad(clust_rec); ut_ad(&view != &check_table_extended_view); /* 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. */ if (rec_deleted) { if (!clust_offsets) clust_offsets= rec_get_offsets(clust_rec, clust_index, nullptr, clust_index->n_core_fields, ULINT_UNDEFINED, &heap); check_match: /* This clustered index record version exists in prebuilt->trx->read_view and is not delete-marked. By design, any BLOBs in it are not allowed to be freed in the purge of committed transaction history. */ err= row_check_index_match(prebuilt, clust_rec, clust_index, clust_offsets, rec, index, offsets); switch (err) { case DB_SUCCESS: break; case DB_SUCCESS_LOCKED_REC: err= DB_SUCCESS; goto did_not_find; default: goto func_exit; } } mtr.rollback_to_savepoint(savepoint); } } else { invalid_PAGE_MAX_TRX_ID: if (UNIV_LIKELY(srv_force_recovery < SRV_FORCE_NO_UNDO_LOG_SCAN)) { push_warning_printf(prebuilt->trx->mysql_thd, Sql_condition::WARN_LEVEL_WARN, ER_NOT_KEYFILE, "InnoDB: Invalid PAGE_MAX_TRX_ID=%llu" " in index '%-.200s'", page_trx_id, index->name()); prebuilt->autoinc_error= DB_INDEX_CORRUPT; } goto next_rec; } count_row: ++*n_rows; if (prev_entry) { ulint matched_fields= 0; int cmp= cmp_dtuple_rec_with_match(prev_entry, rec, offsets, &matched_fields); const char* msg; if (UNIV_LIKELY(cmp < 0)); else if (cmp > 0) { prebuilt->autoinc_error= DB_INDEX_CORRUPT; msg= "index records in a wrong order in "; not_ok: ib::error() << msg << index->name << " of table " << index->table->name << ": " << *prev_entry << ", " << rec_offsets_print(rec, offsets); } else if (index->is_unique() && matched_fields >= dict_index_get_n_ordering_defined_by_user(index)) { /* NULL values in unique indexes are considered not to be duplicates */ for (ulint i= 0; i < dict_index_get_n_ordering_defined_by_user(index); i++) if (dfield_is_null(dtuple_get_nth_field(prev_entry, i))) goto next_rec; if (prebuilt->autoinc_error == DB_SUCCESS) prebuilt->autoinc_error= DB_DUPLICATE_KEY; msg= "duplicate key in "; goto not_ok; } } next_rec: ut_ad(err == DB_SUCCESS); { mem_heap_t *tmp_heap= nullptr; /* Empty the heap on each round. But preserve offsets[] for the row_rec_to_index_entry() call, by copying them into a separate memory heap when needed. */ if (UNIV_UNLIKELY(offsets != offsets_)) { ulint size= rec_offs_get_n_alloc(offsets) * sizeof *offsets; tmp_heap= mem_heap_create(size); offsets= static_cast(mem_heap_dup(tmp_heap, offsets, size)); } mem_heap_empty(heap); prev_entry= row_rec_to_index_entry(rec, index, offsets, heap); if (UNIV_LIKELY_NULL(tmp_heap)) mem_heap_free(tmp_heap); } if (btr_pcur_is_after_last_on_page(prebuilt->pcur)) goto next_page; goto rec_loop; } /*******************************************************************//** Read the AUTOINC column from the current row. If the value is less than 0 and the type is not unsigned then we reset the value to 0. @return value read from the column */ static ib_uint64_t row_search_autoinc_read_column( /*===========================*/ dict_index_t* index, /*!< in: index to read from */ const rec_t* rec, /*!< in: current rec */ ulint col_no, /*!< in: column number */ ulint mtype, /*!< in: column main type */ ibool unsigned_type) /*!< in: signed or unsigned flag */ { ulint len; const byte* data; ib_uint64_t value; mem_heap_t* heap = NULL; rec_offs offsets_[REC_OFFS_NORMAL_SIZE]; rec_offs* offsets = offsets_; rec_offs_init(offsets_); ut_ad(page_rec_is_leaf(rec)); offsets = rec_get_offsets(rec, index, offsets, index->n_core_fields, col_no + 1, &heap); if (rec_offs_nth_sql_null(offsets, col_no)) { /* There is no non-NULL value in the auto-increment column. */ value = 0; goto func_exit; } data = rec_get_nth_field(rec, offsets, col_no, &len); value = row_parse_int(data, len, mtype, unsigned_type); func_exit: if (UNIV_LIKELY_NULL(heap)) { mem_heap_free(heap); } return(value); } /** Get the maximum and non-delete-marked record in an index. @param[in] index index tree @param[in,out] mtr mini-transaction (may be committed and restarted) @return maximum record, page s-latched in mtr @retval NULL if there are no records, or if all of them are delete-marked */ static const rec_t* row_search_get_max_rec( dict_index_t* index, mtr_t* mtr) { btr_pcur_t pcur; const rec_t* rec; /* Open at the high/right end (false), and init cursor */ if (pcur.open_leaf(false, index, BTR_SEARCH_LEAF, mtr) != DB_SUCCESS) { return nullptr; } do { const page_t* page; page = btr_pcur_get_page(&pcur); rec = page_find_rec_max_not_deleted(page); if (page_rec_is_user_rec(rec)) { break; } else { rec = NULL; } btr_pcur_move_before_first_on_page(&pcur); } while (btr_pcur_move_to_prev(&pcur, mtr)); ut_ad(!rec || !(rec_get_info_bits(rec, dict_table_is_comp(index->table)) & (REC_INFO_MIN_REC_FLAG | REC_INFO_DELETED_FLAG))); return(rec); } /** Read the max AUTOINC value from an index. @param[in] index index starting with an AUTO_INCREMENT column @return the largest AUTO_INCREMENT value @retval 0 if no records were found */ ib_uint64_t row_search_max_autoinc(dict_index_t* index) { const dict_field_t* dfield = dict_index_get_nth_field(index, 0); ib_uint64_t value = 0; mtr_t mtr; mtr.start(); if (const rec_t* rec = row_search_get_max_rec(index, &mtr)) { value = row_search_autoinc_read_column( index, rec, 0, dfield->col->mtype, dfield->col->prtype & DATA_UNSIGNED); } mtr.commit(); return(value); }