MDEV-20371: Invalid reads at plan refinement stage: join->positions...bb-10.2-compatibility-build22257-opt-fixes

(Backported to 10.2-compatiblity, build 22257)

best_access_path() is called from two optimization phases:

1. Plan choice phase, in choose_plan(). Here, the join prefix being
   considered is in join->positions[]

2. Plan refinement stage, in fix_semijoin_strategies_for_picked_join_order
   Here, the join prefix is in join->best_positions[]

It used to access join->positions[] from stage #2. This didnt cause any
valgrind or asan failures (as join->positions[] has been written-to before)
but the effect was similar to that of reading the random data:
The join prefix we've picked (in join->best_positions) could have
nothing in common with the join prefix that was last to be considered
(in join->positions).

4 files changed, 52 insertions, 28 deletions

diff --git a/sql/opt_subselect.cc b/sql/opt_subselect.cc
index b4350551461..522aa299658 100644
--- a/sql/opt_subselect.cc
+++ b/sql/opt_subselect.cc
@@ -454,11 +454,6 @@ bool find_eq_ref_candidate(TABLE *table, table_map sj_inner_tables);
 static SJ_MATERIALIZATION_INFO *
 at_sjmat_pos(const JOIN *join, table_map remaining_tables, const JOIN_TAB *tab,
              uint idx, bool *loose_scan);
-void best_access_path(JOIN *join, JOIN_TAB *s, 
-                             table_map remaining_tables, uint idx, 
-                             bool disable_jbuf, double record_count,
-                             POSITION *pos, POSITION *loose_scan_pos);
-
 static Item *create_subq_in_equalities(THD *thd, SJ_MATERIALIZATION_INFO *sjm, 
                                 Item_in_subselect *subq_pred);
 static void remove_sj_conds(THD *thd, Item **tree);
@@ -2659,6 +2654,13 @@ void advance_sj_state(JOIN *join, table_map remaining_tables, uint idx,
     NULL,
   };
 
+#ifdef HAVE_valgrind
+  new (&pos->firstmatch_picker) Firstmatch_picker;
+  new (&pos->loosescan_picker) LooseScan_picker;
+  new (&pos->sjmat_picker) Sj_materialization_picker;
+  new (&pos->dups_weedout_picker) Duplicate_weedout_picker;
+#endif
+
   if (join->emb_sjm_nest)
   {
     /* 
@@ -2942,7 +2944,8 @@ bool Sj_materialization_picker::check_qep(JOIN *join,
     bool disable_jbuf= (join->thd->variables.join_cache_level == 0);
     for (i= first_tab + mat_info->tables; i <= idx; i++)
     {
-      best_access_path(join, join->positions[i].table, rem_tables, i,
+      best_access_path(join, join->positions[i].table, rem_tables,
+                       join->positions, i,
                        disable_jbuf, prefix_rec_count, &curpos, &dummy);
       prefix_rec_count *= curpos.records_read;
       prefix_cost += curpos.read_time;
@@ -3555,7 +3558,8 @@ void fix_semijoin_strategies_for_picked_join_order(JOIN *join)
       join->cur_sj_inner_tables= 0;
       for (i= first + sjm->tables; i <= tablenr; i++)
       {
-        best_access_path(join, join->best_positions[i].table, rem_tables, i, 
+        best_access_path(join, join->best_positions[i].table, rem_tables,
+                         join->best_positions, i,
                          FALSE, prefix_rec_count,
                          join->best_positions + i, &dummy);
         prefix_rec_count *= join->best_positions[i].records_read;
@@ -3587,8 +3591,9 @@ void fix_semijoin_strategies_for_picked_join_order(JOIN *join)
       {
         if (join->best_positions[idx].use_join_buffer)
         {
-           best_access_path(join, join->best_positions[idx].table, 
-                            rem_tables, idx, TRUE /* no jbuf */,
+           best_access_path(join, join->best_positions[idx].table,
+                            rem_tables, join->best_positions, idx,
+                            TRUE /* no jbuf */,
                             record_count, join->best_positions + idx, &dummy);
         }
         record_count *= join->best_positions[idx].records_read;
@@ -3618,7 +3623,8 @@ void fix_semijoin_strategies_for_picked_join_order(JOIN *join)
         if (join->best_positions[idx].use_join_buffer || (idx == first))
         {
            best_access_path(join, join->best_positions[idx].table,
-                            rem_tables, idx, TRUE /* no jbuf */,
+                            rem_tables, join->best_positions, idx,
+                            TRUE /* no jbuf */,
                             record_count, join->best_positions + idx,
                             &loose_scan_pos);
            if (idx==first)
diff --git a/sql/opt_subselect.h b/sql/opt_subselect.h
index 7954becfad4..65d08bef2eb 100644
--- a/sql/opt_subselect.h
+++ b/sql/opt_subselect.h
@@ -88,6 +88,7 @@ class Loose_scan_opt
   KEYUSE *best_loose_scan_start_key;
 
   uint best_max_loose_keypart;
+  table_map best_ref_depend_map;
 
 public:
   Loose_scan_opt():
@@ -250,13 +251,14 @@ public:
           best_loose_scan_records= records;
           best_max_loose_keypart= max_loose_keypart;
           best_loose_scan_start_key= start_key;
+          best_ref_depend_map= 0;
         }
       }
     }
   }
   
   void check_ref_access_part2(uint key, KEYUSE *start_key, double records, 
-                              double read_time)
+                              double read_time, table_map ref_depend_map_arg)
   {
     if (part1_conds_met && read_time < best_loose_scan_cost)
     {
@@ -266,6 +268,7 @@ public:
       best_loose_scan_records= records;
       best_max_loose_keypart= max_loose_keypart;
       best_loose_scan_start_key= start_key;
+      best_ref_depend_map= ref_depend_map_arg;
     }
   }
 
@@ -281,6 +284,7 @@ public:
       best_loose_scan_records= rows2double(quick->records);
       best_max_loose_keypart= quick_max_loose_keypart;
       best_loose_scan_start_key= NULL;
+      best_ref_depend_map= 0;
     }
   }
 
@@ -296,7 +300,7 @@ public:
       pos->loosescan_picker.loosescan_parts= best_max_loose_keypart + 1;
       pos->use_join_buffer= FALSE;
       pos->table=           tab;
-      // todo need ref_depend_map ?
+      pos->ref_depend_map=  best_ref_depend_map;
       DBUG_PRINT("info", ("Produced a LooseScan plan, key %s, %s",
                           tab->table->key_info[best_loose_scan_key].name,
                           best_loose_scan_start_key? "(ref access)":
diff --git a/sql/sql_select.cc b/sql/sql_select.cc
index be2aeb52618..9bc47569f1a 100644
--- a/sql/sql_select.cc
+++ b/sql/sql_select.cc
@@ -97,10 +97,6 @@ static int sort_keyuse(KEYUSE *a,KEYUSE *b);
 static bool are_tables_local(JOIN_TAB *jtab, table_map used_tables);
 static bool create_ref_for_key(JOIN *join, JOIN_TAB *j, KEYUSE *org_keyuse,
 			       bool allow_full_scan, table_map used_tables);
-void best_access_path(JOIN *join, JOIN_TAB *s, 
-                             table_map remaining_tables, uint idx, 
-                             bool disable_jbuf, double record_count,
-                             POSITION *pos, POSITION *loose_scan_pos);
 static void optimize_straight_join(JOIN *join, table_map join_tables);
 static bool greedy_search(JOIN *join, table_map remaining_tables,
                           uint depth, uint prune_level,
@@ -4573,6 +4569,13 @@ make_join_statistics(JOIN *join, List<TABLE_LIST> &tables_list,
     {
       if (choose_plan(join, all_table_map & ~join->const_table_map))
         goto error;
+
+#ifdef HAVE_valgrind
+      // JOIN::positions holds the current query plan. We've already
+      // made the plan choice, so we should only use JOIN::best_positions
+      for (uint k=join->const_tables; k < join->table_count; k++)
+        MEM_UNDEFINED(&join->positions[k], sizeof(join->positions[k]));
+#endif
     }
     else
     {
@@ -6288,6 +6291,7 @@ void
 best_access_path(JOIN      *join,
                  JOIN_TAB  *s,
                  table_map remaining_tables,
+                 const POSITION *join_positions,
                  uint      idx,
                  bool      disable_jbuf,
                  double    record_count,
@@ -6391,7 +6395,7 @@ best_access_path(JOIN      *join,
             if (!(keyuse->used_tables & ~join->const_table_map))
               const_part|= keyuse->keypart_map;
 
-            double tmp2= prev_record_reads(join->positions, idx,
+            double tmp2= prev_record_reads(join_positions, idx,
                                            (found_ref | keyuse->used_tables));
             if (tmp2 < best_prev_record_reads)
             {
@@ -6432,7 +6436,7 @@ best_access_path(JOIN      *join,
           Really, there should be records=0.0 (yes!)
           but 1.0 would be probably safer
         */
-        tmp= prev_record_reads(join->positions, idx, found_ref);
+        tmp= prev_record_reads(join_positions, idx, found_ref);
         records= 1.0;
       }
       else
@@ -6448,7 +6452,7 @@ best_access_path(JOIN      *join,
           if ((key_flags & (HA_NOSAME | HA_NULL_PART_KEY)) == HA_NOSAME ||
               MY_TEST(key_flags & HA_EXT_NOSAME))
           {
-            tmp = prev_record_reads(join->positions, idx, found_ref);
+            tmp = prev_record_reads(join_positions, idx, found_ref);
             records=1.0;
           }
           else
@@ -6692,7 +6696,8 @@ best_access_path(JOIN      *join,
         }
 
         tmp += s->startup_cost;
-        loose_scan_opt.check_ref_access_part2(key, start_key, records, tmp);
+        loose_scan_opt.check_ref_access_part2(key, start_key, records, tmp,
+                                              found_ref);
       } /* not ft_key */
       if (tmp + 0.0001 < best_time - records/(double) TIME_FOR_COMPARE)
       {
@@ -7364,7 +7369,8 @@ optimize_straight_join(JOIN *join, table_map join_tables)
   for (JOIN_TAB **pos= join->best_ref + idx ; (s= *pos) ; pos++)
   {
     /* Find the best access method from 's' to the current partial plan */
-    best_access_path(join, s, join_tables, idx, disable_jbuf, record_count,
+    best_access_path(join, s, join_tables, join->positions, idx,
+                     disable_jbuf, record_count,
                      join->positions + idx, &loose_scan_pos);
 
     /* compute the cost of the new plan extended with 's' */
@@ -8270,8 +8276,9 @@ best_extension_by_limited_search(JOIN      *join,
 
       /* Find the best access method from 's' to the current partial plan */
       POSITION loose_scan_pos;
-      best_access_path(join, s, remaining_tables, idx, disable_jbuf,
-                       record_count, join->positions + idx, &loose_scan_pos);
+      best_access_path(join, s, remaining_tables, join->positions, idx,
+                       disable_jbuf, record_count, join->positions + idx,
+                       &loose_scan_pos);
 
       /* Compute the cost of extending the plan with 's', avoid overflow */
       if (position->records_read < DBL_MAX / record_count)
@@ -8654,11 +8661,11 @@ cache_record_length(JOIN *join,uint idx)
 */
 
 double
-prev_record_reads(POSITION *positions, uint idx, table_map found_ref)
+prev_record_reads(const POSITION *positions, uint idx, table_map found_ref)
 {
   double found=1.0;
-  POSITION *pos_end= positions - 1;
-  for (POSITION *pos= positions + idx - 1; pos != pos_end; pos--)
+  const POSITION *pos_end= positions - 1;
+  for (const POSITION *pos= positions + idx - 1; pos != pos_end; pos--)
   {
     if (pos->table->table->map & found_ref)
     {
@@ -15560,7 +15567,8 @@ void optimize_wo_join_buffering(JOIN *join, uint first_tab, uint last_tab,
     if ((i == first_tab && first_alt) || join->positions[i].use_join_buffer)
     {
       /* Find the best access method that would not use join buffering */
-      best_access_path(join, rs, reopt_remaining_tables, i, 
+      best_access_path(join, rs, reopt_remaining_tables,
+                       join->positions, i,
                        TRUE, rec_count,
                        &pos, &loose_scan_pos);
     }
diff --git a/sql/sql_select.h b/sql/sql_select.h
index eb78f4216f9..98fbb1f4d5b 100644
--- a/sql/sql_select.h
+++ b/sql/sql_select.h
@@ -794,6 +794,7 @@ public:
   friend void best_access_path(JOIN      *join,
                                JOIN_TAB  *s,
                                table_map remaining_tables,
+                               const struct st_position *join_positions,
                                uint      idx,
                                bool      disable_jbuf,
                                double    record_count,
@@ -1970,6 +1971,11 @@ protected:
   }
 };
 
+void best_access_path(JOIN *join, JOIN_TAB *s,
+                      table_map remaining_tables,
+                      const POSITION *join_positions, uint idx,
+                      bool disable_jbuf, double record_count,
+                      POSITION *pos, POSITION *loose_scan_pos);
 bool cp_buffer_from_ref(THD *thd, TABLE *table, TABLE_REF *ref);
 bool error_if_full_join(JOIN *join);
 int report_error(TABLE *table, int error);
@@ -2334,7 +2340,7 @@ bool instantiate_tmp_table(TABLE *table, KEY *keyinfo,
                            ulonglong options);
 bool open_tmp_table(TABLE *table);
 void setup_tmp_table_column_bitmaps(TABLE *table, uchar *bitmaps);
-double prev_record_reads(POSITION *positions, uint idx, table_map found_ref);
+double prev_record_reads(const POSITION *positions, uint idx, table_map found_ref);
 void fix_list_after_tbl_changes(SELECT_LEX *new_parent, List<TABLE_LIST> *tlist);
 
 struct st_cond_statistic