Bug #20662 "Infinite loop in CREATE TABLE IF NOT EXISTS ... SELECT
with locked tables"
Bug #20903 "Crash when using CREATE TABLE .. SELECT and triggers"
Bug #24738 "CREATE TABLE ... SELECT is not isolated properly"
Bug #24508 "Inconsistent results of CREATE TABLE ... SELECT when
temporary table exists"
Deadlock occured when one tried to execute CREATE TABLE IF NOT
EXISTS ... SELECT statement under LOCK TABLES which held
read lock on target table.
Attempt to execute the same statement for already existing
target table with triggers caused server crashes.
Also concurrent execution of CREATE TABLE ... SELECT statement
and other statements involving target table suffered from
various races (some of which might've led to deadlocks).
Finally, attempt to execute CREATE TABLE ... SELECT in case
when a temporary table with same name was already present
led to the insertion of data into this temporary table and
creation of empty non-temporary table.
All above problems stemmed from the old implementation of CREATE
TABLE ... SELECT in which we created, opened and locked target
table without any special protection in a separate step and not
with the rest of tables used by this statement.
This underminded deadlock-avoidance approach used in server
and created window for races. It also excluded target table
from prelocking causing problems with trigger execution.
The patch solves these problems by implementing new approach to
handling of CREATE TABLE ... SELECT for base tables.
We try to open and lock table to be created at the same time as
the rest of tables used by this statement. If such table does not
exist at this moment we create and place in the table cache special
placeholder for it which prevents its creation or any other usage
by other threads.
We still use old approach for creation of temporary tables.
Also note that we decided to postpone introduction of some tests
for concurrent behaviour of CREATE TABLE ... SELECT till 5.1.
The main reason for this is absence in 5.0 ability to set @@debug
variable at runtime, which can be circumvented only by using several
test files with individual .opt files. Since the latter is likely
to slowdown test-suite unnecessary we chose not to push this tests
into 5.0, but run them manually for this version and later push
their optimized version into 5.1
Before this fix, the parser would sometime change where a token starts by
altering Lex_input_string::tok_start, which later confused the code in
sql_yacc.yy that needs to capture the source code of a SQL statement,
like to represent the body of a stored procedure.
This line of code in sql_lex.cc :
case MY_LEX_USER_VARIABLE_DELIMITER:
lip->tok_start= lip->ptr; // Skip first `
would <skip the first back quote> ... and cause the bug reported.
In general, the responsibility of sql_lex.cc is to *find* where token are
in the SQL text, but is *not* to make up fake or incomplete tokens.
With a quoted label like `my_label`, the token starts on the first quote.
Extracting the token value should not change that (it did).
With this fix, the lexical analysis has been cleaned up to not change
lip->tok_start (in the case found for this bug).
The functions get_token() and get_quoted_token() now have an extra
parameters, used when some characters from the beginning of the token need
to be skipped when extracting a token value, like when extracting 'AB' from
'0xAB', for example, for a HEX_NUM token.
This exposed a bad assumption in Item_hex_string and Item_bin_string,
which has been fixed:
The assumption was that the string given, 'AB', was in fact preceded in
memory by '0x', which might be false (it can be preceded by "x'" and
followed by "'" -- or not be preceded by valid memory at all)
If a name is needed for Item_hex_string or Item_bin_string, the name is
taken from the original and true source code ('0xAB'), and assigned in
the select_item rule, instead of relying on assumptions related to how
memory is used.
The issue found with bug 25411 is due to the function skip_rear_comments()
which damages the source code while implementing a work around.
The root cause of the problem is in the lexical analyser, which does not
process special comments properly.
For special comments like :
[1] aaa /*!50000 bbb */ ccc
since 5.0 is a version older that the current code, the parser is in lining
the content of the special comment, so that the query to process is
[2] aaa bbb ccc
However, the text of the query captured when processing a stored procedure,
stored function or trigger (or event in 5.1), can be after rebuilding it:
[3] aaa bbb */ ccc
which is wrong.
To fix bug 25411 properly, the lexical analyser needs to return [2] when
in lining special comments.
In order to implement this, some preliminary cleanup is required in the code,
which is implemented by this patch.
Before this change, the structure named LEX (or st_lex) contains attributes
that belong to lexical analysis, as well as attributes that represents the
abstract syntax tree (AST) of a statement.
Creating a new LEX structure for each statements (which makes sense for the
AST part) also re-initialized the lexical analysis phase each time, which
is conceptually wrong.
With this patch, the previous st_lex structure has been split in two:
- st_lex represents the Abstract Syntax Tree for a statement. The name "lex"
has not been changed to avoid a bigger impact in the code base.
- class lex_input_stream represents the internal state of the lexical
analyser, which by definition should *not* be reinitialized when parsing
multiple statements from the same input stream.
This change is a pre-requisite for bug 25411, since the implementation of
lex_input_stream will later improve to deal properly with special comments,
and this processing can not be done with the current implementation of
sp_head::reset_lex and sp_head::restore_lex, which interfere with the lexer.
This change set alone does not fix bug 25411.
- GRANT and REVOKE statments didn't have the "updating" flag set and
thus statements with a table specified would not replicate if
slave filtering rules where turned on.
For example "GRANT ... ON test.t1 TO ..." would not replicate.
- mysqldump executes a SHOW CREATE VIEW statement to generate the text
that it outputs. When the function name is retrieved it's database
name is unconditionally prepended. This change causes the function's
database name to be prepended only when it was used to define the
function.
INTO clause can be specified only for the last select of a UNION and it
receives the result of the whole query. But it was wrongly allowed in
non-last selects of a UNION which leads to a confusing query result.
Now INTO allowed only in the last select of a UNION.
Before this fix, the parser would accept illegal code in SQL exceptions
handlers, that later causes the runtime to crash when executing the code,
due to memory violations in the exception handler stack.
The root cause of the problem is instructions within an exception handler
that jumps to code located outside of the handler. This is illegal according
to the SQL 2003 standard, since labels located outside the handler are not
supposed to be visible (they are "out of scope"), so any instruction that
jumps to these labels, like ITERATE or LEAVE, should not parse.
The section of the standard that is relevant for this is :
SQL:2003 SQL/PSM (ISO/IEC 9075-4:2003)
section 13.1 <compound statement>,
syntax rule 4
<quote>
The scope of the <beginning label> is CS excluding every <SQL schema
statement> contained in CS and excluding every
<local handler declaration list> contained in CS. <beginning label> shall
not be equivalent to any other <beginning label>s within that scope.
</quote>
With this fix, the C++ class sp_pcontext, which represent the "parsing
context" tree (a.k.a symbol table) of a stored procedure, has been changed
as follows:
- constructors have been cleaned up, so that only building a root node for
the tree is public; building nodes inside a tree is not public.
- a new member, m_label_scope, indicates if a given syntactic context
belongs to a DECLARE HANDLER block,
- label resolution, in the method find_label(), has been changed to
implement the restriction of scope regarding labels used in a compound
statement.
The actions in the parser, when parsing the body of a SQL exception handler,
have been changed as follows:
- the implementation of an exception handler (DECLARE HANDLER) now creates
explicitly a new sp_pcontext, to isolate the code inside the handler from
the containing compound statement context.
- registering exception handlers as a result occurs in the parent context,
see the rule sp_hcond_element
- the code in sp_hcond_list has been cleaned up, to avoid code duplication
In addition, the flags IN_SIMPLE_CASE and IN_HANDLER, declared in sp_head.h
have been removed, since they are unused and broken by design (as seen with
Bug 19194 (Right recursion in parser for CASE causes excessive stack usage,
limitation), representing a stack in a single flag is not possible.
Tests in sp-error have been added to show that illegal constructs are now
rejected.
Tests in sp have been added for code coverage, to show that ITERATE or LEAVE
statements are legal when jumping to a label in scope, inside the body of
an exception handler.
fixes).
The legend: on a replication slave, in case a trigger creation
was filtered out because of application of replicate-do-table/
replicate-ignore-table rule, the parsed definition of a trigger was not
cleaned up properly. LEX::sphead member was left around and leaked
memory. Until the actual implementation of support of
replicate-ignore-table rules for triggers by the patch for Bug 24478 it
was never the case that "case SQLCOM_CREATE_TRIGGER"
was not executed once a trigger was parsed,
so the deletion of lex->sphead there worked and the memory did not leak.
The fix:
The real cause of the bug is that there is no 1 or 2 places where
we can clean up the main LEX after parse. And the reason we
can not have just one or two places where we clean up the LEX is
asymmetric behaviour of MYSQLparse in case of success or error.
One of the root causes of this behaviour is the code in Item::Item()
constructor. There, a newly created item adds itself to THD::free_list
- a single-linked list of Items used in a statement. Yuck. This code
is unaware that we may have more than one statement active at a time,
and always assumes that the free_list of the current statement is
located in THD::free_list. One day we need to be able to explicitly
allocate an item in a given Query_arena.
Thus, when parsing a definition of a stored procedure, like
CREATE PROCEDURE p1() BEGIN SELECT a FROM t1; SELECT b FROM t1; END;
we actually need to reset THD::mem_root, THD::free_list and THD::lex
to parse the nested procedure statement (SELECT *).
The actual reset and restore is implemented in semantic actions
attached to sp_proc_stmt grammar rule.
The problem is that in case of a parsing error inside a nested statement
Bison generated parser would abort immediately, without executing the
restore part of the semantic action. This would leave THD in an
in-the-middle-of-parsing state.
This is why we couldn't have had a single place where we clean up the LEX
after MYSQLparse - in case of an error we needed to do a clean up
immediately, in case of success a clean up could have been delayed.
This left the door open for a memory leak.
One of the following possibilities were considered when working on a fix:
- patch the replication logic to do the clean up. Rejected
as breaks module borders, replication code should not need to know the
gory details of clean up procedure after CREATE TRIGGER.
- wrap MYSQLparse with a function that would do a clean up.
Rejected as ideally we should fix the problem when it happens, not
adjust for it outside of the problematic code.
- make sure MYSQLparse cleans up after itself by invoking the clean up
functionality in the appropriate places before return. Implemented in
this patch.
- use %destructor rule for sp_proc_stmt to restore THD - cleaner
than the prevoius approach, but rejected
because needs a careful analysis of the side effects, and this patch is
for 5.0, and long term we need to use the next alternative anyway
- make sure that sp_proc_stmt doesn't juggle with THD - this is a
large work that will affect many modules.
Cleanup: move main_lex and main_mem_root from Statement to its
only two descendants Prepared_statement and THD. This ensures that
when a Statement instance was created for purposes of statement backup,
we do not involve LEX constructor/destructor, which is fairly expensive.
In order to track that the transformation produces equivalent
functionality please check the respective constructors and destructors
of Statement, Prepared_statement and THD - these members were
used only there.
This cleanup is unrelated to the patch.
This patch fixes problem that LOAD DATA could use different
character sets when loading files on master and on slave sides:
- Adding replication of thd->variables.collation_database
- Adding optional character set clause into LOAD DATA
Note, the second way, with explicit CHARACTER SET clause
should be the recommended way to load data using an alternative
character set.
The old way, using "SET @@character_set_database=xxx" should be
gradually depricated.
Post fix for bug#23800.
The Item_field constructor now increases the select_n_where_fields counter.
sql_yacc.yy:
Post fix for bug#23800.
Take into account fields that might be added by subselects.
sql_lex.h:
Post fix for bug#23800.
Added the select_n_where_fields variable to the st_select_lex class.
sql_lex.cc:
Post fix for bug#23800.
Initialization of the select_n_where_fields variable.
Several problems fixed:
1. There was a "catch-all" context initialization in setup_tables()
that was causing the table that we insert into to be visible in the
SELECT part of an INSERT .. SELECT .. statement with no tables in
its FROM clause. This was making sure all the under-initialized
contexts in various parts of the code are not left uninitialized.
Fixed by removing the "catch-all" statement and initializing the
context in the parser.
2. Incomplete name resolution context when resolving the right-hand
values in the ON DUPLICATE KEY UPDATE ... part of an INSERT ... SELECT ...
caused columns from NATURAL JOIN/JOIN USING table references in the
FROM clause of the select to be unavailable.
Fixed by establishing a proper name resolution context.
3. When setting up the special name resolution context for problem 2
there was no check for cases where an aggregate function without a
GROUP BY effectively takes the column from the SELECT part of an
INSERT ... SELECT unavailable for ON DUPLICATE KEY UPDATE.
Fixed by checking for that condition when setting up the name
resolution context.
operations)
Before this change, the boolean predicates:
- X IS TRUE,
- X IS NOT TRUE,
- X IS FALSE,
- X IS NOT FALSE
were implemented by expanding the Item tree in the parser, by using a
construct like:
Item_func_if(Item_func_ifnull(X, <value>), <value>, <value>)
Each <value> was a constant integer, either 0 or 1.
A bug in the implementation of the function IF(a, b, c), in
Item_func_if::fix_length_and_dec(), would cause the following :
When the arguments b and c are both unsigned, the result type of the
function was signed, instead of unsigned.
When the result of the if function is signed, space for the sign could be
counted twice (in the max() expression for a signed argument, and in the
total), causing the member max_length to be too high.
An effect of this is that the final type of IF(x, int(1), int(1)) would be
int(2) instead of int(1).
With this fix, the problems found in Item_func_if::fix_length_and_dec()
have been fixed.
While it's semantically correct to represent 'X IS TRUE' with
Item_func_if(Item_func_ifnull(X, <value>), <value>, <value>),
there are however more problems with this construct.
a)
Building the parse tree involves :
- creating 5 Item instances (3 ints, 1 ifnull, 1 if),
- creating each Item calls my_pthread_getspecific_ptr() once in the operator
new(size), and a second time in the Item::Item() constructor, resulting
in a total of 10 calls to get the current thread.
Evaluating the expression involves evaluating up to 4 nodes at runtime.
This representation could be greatly simplified and improved.
b)
Transforming the parse tree internally with if(ifnull(...)) is fine as long
as this transformation is internal to the server implementation.
With views however, the result of the parse tree is later exposed by the
::print() functions, and stored as part of the view definition.
Doing this has long term consequences:
1)
The original semantic 'X IS TRUE' is lost, and replaced by the
if(ifnull(...)) expression. As a result, SHOW CREATE VIEW does not restore
the original code.
2)
Should a future version of MySQL implement the SQL BOOLEAN data type for
example, views created today using 'X IS NULL' can be exported using
mysqldump, and imported again. Such views would be converted correctly and
automatically to use a BOOLEAN column in the future version.
With 'X IS TRUE' and the current implementations, views using these
"boolean" predicates would not be converted during the export/import, and
would use integer columns instead.
The difference traces back to how SHOW CREATE VIEW preserves 'X IS NULL' but
does not preserve the 'X IS TRUE' semantic.
With this fix, internal representation of 'X IS TRUE' booleans predicates
has changed, so that:
- dedicated Item classes are created for each predicate,
- only 1 Item is created to represent 1 predicate
- my_pthread_getspecific_ptr() is invoked 1 time instead of 10
- SHOW CREATE VIEW preserves the original semantic, and prints 'X IS TRUE'.
Note that, because of the fix in Item_func_if, views created before this fix
will:
- correctly use a int(1) type instead of int(2) for boolean predicates,
- incorrectly print the if(ifnull(...), ...) expression in SHOW CREATE VIEW,
since the original semantic (X IS TRUE) has been lost.
- except for the syntax used in SHOW CREATE VIEW, these views will operate
properly, no action is needed.
Views created after this fix will operate correctly, and will preserve the
original code semantic in SHOW CREATE VIEW.
Two problems here:
Problem 1:
While constructing the join columns list the optimizer does as follows:
1. Sets the join_using_fields/natural_join members of the right JOIN
operand.
2. Makes a "table reference" (TABLE_LIST) to parent the two tables.
3. Assigns the join_using_fields/is_natural_join of the wrapper table
using join_using_fields/natural_join of the rightmost table
4. Sets join_using_fields to NULL for the right JOIN operand.
5. Passes the parent table up to the same procedure on the upper
level.
Step 1 overrides the the join_using_fields that are set for a nested
join wrapping table in step 4.
Fixed by making a designated variable SELECT_LEX::prev_join_using to
pass the data from step 1 to step 4 without destroying the wrapping
table data.
Problem 2:
The optimizer checks for ambiguous columns while transforming
NATURAL JOIN/JOIN USING to JOIN ON. While doing that there was no
distinction between columns that are used in the generated join
condition (where ambiguity can be checked) and the other columns
(where ambiguity can be checked only when resolving references
coming from outside the JOIN construct itself).
Fixed by allowing the non-USING columns to be present in multiple
copies in both sides of the join and moving the ambiguity check
to the place where unqualified references to the join columns are
resolved (find_field_in_natural_join()).
Before this fix, a IN predicate of the form: "IN (( subselect ))", with two
parenthesis, would be evaluated as a single row subselect: if the subselect
returns more that 1 row, the statement would fail.
The SQL:2003 standard defines a special exception in the specification,
and mandates that this particular form of IN predicate shall be equivalent
to "IN ( subselect )", which involves a table subquery and works with more
than 1 row.
This fix implements "IN (( subselect ))", "IN ((( subselect )))" etc
as per the SQL:2003 requirement.
All the details related to the implementation of this change have been
commented in the code, and the relevant sections of the SQL:2003 spec
are given for reference, so they are not repeated here.
Having access to the spec is a requirement to review in depth this patch.
WL#3681 (ALTER TABLE ORDER BY)
Before this fix, the ALTER TABLE statement implemented an ORDER BY option
with the following characteristics :
1) The order by clause accepts a list of criteria, with optional ASC or
DESC keywords
2) Each criteria can be a general expression, involving operators,
native functions, stored functions, user defined functions, subselects ...
With this fix :
1) has been left unchanged, since it's a de-facto existing feature,
that was already present in the code base and partially covered in the test
suite. Code coverage for ASC and DESC was missing and has been improved.
2) has been changed to limit the kind of criteria that are permissible:
now only a column name is valid.
- Removed not used variables and functions
- Added #ifdef around code that is not used
- Renamed variables and functions to avoid conflicts
- Removed some not used arguments
Fixed some class/struct warnings in ndb
Added define IS_LONGDATA() to simplify code in libmysql.c
I did run gcov on the changes and added 'purecov' comments on almost all lines that was not just variable name changes
