Bug#26231 - select count(*) on myisam table returns wrong value

            when index is used

When the table contained TEXT columns with empty contents
('', zero length, but not NULL) _and_ strings starting with
control characters like tabulator or newline, the empty values
were not found in a "records in range" estimate. Hence count(*)
missed these records.

The reason was a different set of search flags used for key
insert and key range estimation.

I decided to fix the set of flags used in range estimation.
Otherwise millions of databases around the world would require
a repair after an upgrade.

The consequence is that the manual must be fixed, which claims
that TEXT columns are compared with "end space padding". This
is true for CHAR/VARCHAR but wrong for TEXT. See also bug 21335.

1 files changed, 35 insertions, 1 deletions

diff --git a/myisam/mi_range.c b/myisam/mi_range.c
index 4248fa7f04b..e9bec50f971 100644
--- a/myisam/mi_range.c
+++ b/myisam/mi_range.c
@@ -145,8 +145,42 @@ static ha_rows _mi_record_pos(MI_INFO *info, const byte *key, uint key_len,
   if (!(nextflag & (SEARCH_FIND | SEARCH_NO_FIND | SEARCH_LAST)))
     key_len=USE_WHOLE_KEY;
 
+  /*
+    my_handler.c:mi_compare_text() has a flag 'skip_end_space'.
+    This is set in my_handler.c:ha_key_cmp() in dependence on the
+    compare flags 'nextflag' and the column type.
+
+    TEXT columns are of type HA_KEYTYPE_VARTEXT. In this case the
+    condition is skip_end_space= ((nextflag & (SEARCH_FIND |
+    SEARCH_UPDATE)) == SEARCH_FIND).
+
+    SEARCH_FIND is used for an exact key search. The combination
+    SEARCH_FIND | SEARCH_UPDATE is used in write/update/delete
+    operations with a comment like "Not real duplicates", whatever this
+    means. From the condition above we can see that 'skip_end_space' is
+    always false for these operations. The result is that trailing space
+    counts in key comparison and hence, emtpy strings ('', string length
+    zero, but not NULL) compare less that strings starting with control
+    characters and these in turn compare less than strings starting with
+    blanks.
+
+    When estimating the number of records in a key range, we request an
+    exact search for the minimum key. This translates into a plain
+    SEARCH_FIND flag. Using this alone would lead to a 'skip_end_space'
+    compare. Empty strings would be expected above control characters.
+    Their keys would not be found because they are located below control
+    characters.
+
+    This is the reason that we add the SEARCH_UPDATE flag here. It makes
+    the key estimation compare in the same way like key write operations
+    do. Olny so we will find the keys where they have been inserted.
+
+    Adding the flag unconditionally does not hurt as it is used in the
+    above mentioned condition only. So it can safely be used together
+    with other flags.
+  */
   pos=_mi_search_pos(info,keyinfo,key_buff,key_len,
-		     nextflag | SEARCH_SAVE_BUFF,
+		     nextflag | SEARCH_SAVE_BUFF | SEARCH_UPDATE,
 		     info->s->state.key_root[inx]);
   if (pos >= 0.0)
   {