Bug#23831 deadlock not noticed
RBR bug in that when replicated msta (multi-statement-trans-action) deadlocks
with a local at write row event or gets timed-out, the event handler did not return
the correct error code.
Wrong error code stops slave sql thread instead of to proceed with
rollback and replay.
The correct code is typed in error log and stored for error handling rotine
to conduct rollback and replay of the transaction. The handling for the rbr
remains the same as for the sbr events.
Particularly, timed-out transaction still is rolled back - look at the related bugs.
Removing code to step the group log position and just stepping
the event log position. If the group log position were stepped
one time too many, it might be that the group starts at a position
that is not possible, e.g., at a Rows_log_event, or between an
Intvar_log_event and the following associated Query_log_event.
crash for, e.g., NDB):
Before, mysqlbinlog printed table map events as a separate statement, so
when executing the event, the opened table was subsequently closed
when the statement ended. Instead, the row-based events that make up
a statement are now printed as *one* BINLOG statement, which means
that the table maps and the following *_rows_log_event events are
executed fully before the statement ends.
Changing implementation of BINLOG statement to be able to read the
emitted format, which now consists of several chunks of BASE64-encoded
data.
We now reset the THD members related to auto_increment+binlog in
MYSQL_LOG::write(). This is better than in THD::cleanup_after_query(),
which was not able to distinguish between SELECT myfunc1(),myfunc2()
and INSERT INTO t SELECT myfunc1(),myfunc2() from a binlogging point
of view.
Rows_log_event::exec_event() now calls lex_start() instead of
mysql_init_query() because the latter now does too much (it resets
the binlog format).
this is a cleanup patch for our current auto_increment handling:
new names for auto_increment variables in THD, new methods to manipulate them
(see sql_class.h), some move into handler::, causing less backup/restore
work when executing substatements.
This makes the logic hopefully clearer, less work is is needed in
mysql_insert().
By cleaning up, using different variables for different purposes (instead
of one for 3 things...), we fix those bugs, which someone may want to fix
in 5.0 too:
BUG#20339 "stored procedure using LAST_INSERT_ID() does not replicate
statement-based"
BUG#20341 "stored function inserting into one auto_increment puts bad
data in slave"
BUG#19243 "wrong LAST_INSERT_ID() after ON DUPLICATE KEY UPDATE"
(now if a row is updated, LAST_INSERT_ID() will return its id)
and re-fixes:
BUG#6880 "LAST_INSERT_ID() value changes during multi-row INSERT"
(already fixed differently by Ramil in 4.1)
Test of documented behaviour of mysql_insert_id() (there was no test).
The behaviour changes introduced are:
- LAST_INSERT_ID() now returns "the first autogenerated auto_increment value
successfully inserted", instead of "the first autogenerated auto_increment
value if any row was successfully inserted", see auto_increment.test.
Same for mysql_insert_id(), see mysql_client_test.c.
- LAST_INSERT_ID() returns the id of the updated row if ON DUPLICATE KEY
UPDATE, see auto_increment.test. Same for mysql_insert_id(), see
mysql_client_test.c.
- LAST_INSERT_ID() does not change if no autogenerated value was successfully
inserted (it used to then be 0), see auto_increment.test.
- if in INSERT SELECT no autogenerated value was successfully inserted,
mysql_insert_id() now returns the id of the last inserted row (it already
did this for INSERT VALUES), see mysql_client_test.c.
- if INSERT SELECT uses LAST_INSERT_ID(X), mysql_insert_id() now returns X
(it already did this for INSERT VALUES), see mysql_client_test.c.
- NDB now behaves like other engines wrt SET INSERT_ID: with INSERT IGNORE,
the id passed in SET INSERT_ID is re-used until a row succeeds; SET INSERT_ID
influences not only the first row now.
Additionally, when unlocking a table we check that the thread is not keeping
a next_insert_id (as the table is unlocked that id is potentially out-of-date);
forgetting about this next_insert_id is done in a new
handler::ha_release_auto_increment().
Finally we prepare for engines capable of reserving finite-length intervals
of auto_increment values: we store such intervals in THD. The next step
(to be done by the replication team in 5.1) is to read those intervals from
THD and actually store them in the statement-based binary log. NDB
will be a good engine to test that.
before update trigger on NDB table".
Two main changes:
- We use TABLE::read_set/write_set bitmaps for marking fields used by
statement instead of Field::query_id in 5.1.
- Now when we mark columns used by statement we take into account columns
used by table's triggers instead of marking all columns as used if table
has triggers.
No test case as the bug is in an existing test case (rpl_trigger.test
when it is run under valgrind).
The warning was caused by memory corruption in replication slave: thd->db
was pointing at a stack address that was previously used by
sp_head::execute()::old_db. This happened because mysql_change_db
behaved differently in replication slave and did not make a copy of the
argument to assign to thd->db.
The solution is to always free the old value of thd->db and allocate a new
copy, regardless whether we're running in a replication slave or not.
Bug#19022 "Memory bug when switching db during trigger execution"
Bug#17199 "Problem when view calls function from another database."
Bug#18444 "Fully qualified stored function names don't work correctly in
SELECT statements"
Documentation note: this patch introduces a change in behaviour of prepared
statements.
This patch adds a few new invariants with regard to how THD::db should
be used. These invariants should be preserved in future:
- one should never refer to THD::db by pointer and always make a deep copy
(strmake, strdup)
- one should never compare two databases by pointer, but use strncmp or
my_strncasecmp
- TABLE_LIST object table->db should be always initialized in the parser or
by creator of the object.
For prepared statements it means that if the current database is changed
after a statement is prepared, the database that was current at prepare
remains active. This also means that you can not prepare a statement that
implicitly refers to the current database if the latter is not set.
This is not documented, and therefore needs documentation. This is NOT a
change in behavior for almost all SQL statements except:
- ALTER TABLE t1 RENAME t2
- OPTIMIZE TABLE t1
- ANALYZE TABLE t1
- TRUNCATE TABLE t1 --
until this patch t1 or t2 could be evaluated at the first execution of
prepared statement.
CURRENT_DATABASE() still works OK and is evaluated at every execution
of prepared statement.
Note, that in stored routines this is not an issue as the default
database is the database of the stored procedure and "use" statement
is prohibited in stored routines.
This patch makes obsolete the use of check_db_used (it was never used in the
old code too) and all other places that check for table->db and assign it
from THD::db if it's NULL, except the parser.
How this patch was created: THD::{db,db_length} were replaced with a
LEX_STRING, THD::db. All the places that refer to THD::{db,db_length} were
manually checked and:
- if the place uses thd->db by pointer, it was fixed to make a deep copy
- if a place compared two db pointers, it was fixed to compare them by value
(via strcmp/my_strcasecmp, whatever was approproate)
Then this intermediate patch was used to write a smaller patch that does the
same thing but without a rename.
TODO in 5.1:
- remove check_db_used
- deploy THD::set_db in mysql_change_db
See also comments to individual files.
Changes that requires code changes in other code of other storage engines.
(Note that all changes are very straightforward and one should find all issues
by compiling a --debug build and fixing all compiler errors and all
asserts in field.cc while running the test suite),
- New optional handler function introduced: reset()
This is called after every DML statement to make it easy for a handler to
statement specific cleanups.
(The only case it's not called is if force the file to be closed)
- handler::extra(HA_EXTRA_RESET) is removed. Code that was there before
should be moved to handler::reset()
- table->read_set contains a bitmap over all columns that are needed
in the query. read_row() and similar functions only needs to read these
columns
- table->write_set contains a bitmap over all columns that will be updated
in the query. write_row() and update_row() only needs to update these
columns.
The above bitmaps should now be up to date in all context
(including ALTER TABLE, filesort()).
The handler is informed of any changes to the bitmap after
fix_fields() by calling the virtual function
handler::column_bitmaps_signal(). If the handler does caching of
these bitmaps (instead of using table->read_set, table->write_set),
it should redo the caching in this code. as the signal() may be sent
several times, it's probably best to set a variable in the signal
and redo the caching on read_row() / write_row() if the variable was
set.
- Removed the read_set and write_set bitmap objects from the handler class
- Removed all column bit handling functions from the handler class.
(Now one instead uses the normal bitmap functions in my_bitmap.c instead
of handler dedicated bitmap functions)
- field->query_id is removed. One should instead instead check
table->read_set and table->write_set if a field is used in the query.
- handler::extra(HA_EXTRA_RETRIVE_ALL_COLS) and
handler::extra(HA_EXTRA_RETRIEVE_PRIMARY_KEY) are removed. One should now
instead use table->read_set to check for which columns to retrieve.
- If a handler needs to call Field->val() or Field->store() on columns
that are not used in the query, one should install a temporary
all-columns-used map while doing so. For this, we provide the following
functions:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->read_set);
field->val();
dbug_tmp_restore_column_map(table->read_set, old_map);
and similar for the write map:
my_bitmap_map *old_map= dbug_tmp_use_all_columns(table, table->write_set);
field->val();
dbug_tmp_restore_column_map(table->write_set, old_map);
If this is not done, you will sooner or later hit a DBUG_ASSERT
in the field store() / val() functions.
(For not DBUG binaries, the dbug_tmp_restore_column_map() and
dbug_tmp_restore_column_map() are inline dummy functions and should
be optimized away be the compiler).
- If one needs to temporary set the column map for all binaries (and not
just to avoid the DBUG_ASSERT() in the Field::store() / Field::val()
methods) one should use the functions tmp_use_all_columns() and
tmp_restore_column_map() instead of the above dbug_ variants.
- All 'status' fields in the handler base class (like records,
data_file_length etc) are now stored in a 'stats' struct. This makes
it easier to know what status variables are provided by the base
handler. This requires some trivial variable names in the extra()
function.
- New virtual function handler::records(). This is called to optimize
COUNT(*) if (handler::table_flags() & HA_HAS_RECORDS()) is true.
(stats.records is not supposed to be an exact value. It's only has to
be 'reasonable enough' for the optimizer to be able to choose a good
optimization path).
- Non virtual handler::init() function added for caching of virtual
constants from engine.
- Removed has_transactions() virtual method. Now one should instead return
HA_NO_TRANSACTIONS in table_flags() if the table handler DOES NOT support
transactions.
- The 'xxxx_create_handler()' function now has a MEM_ROOT_root argument
that is to be used with 'new handler_name()' to allocate the handler
in the right area. The xxxx_create_handler() function is also
responsible for any initialization of the object before returning.
For example, one should change:
static handler *myisam_create_handler(TABLE_SHARE *table)
{
return new ha_myisam(table);
}
->
static handler *myisam_create_handler(TABLE_SHARE *table, MEM_ROOT *mem_root)
{
return new (mem_root) ha_myisam(table);
}
- New optional virtual function: use_hidden_primary_key().
This is called in case of an update/delete when
(table_flags() and HA_PRIMARY_KEY_REQUIRED_FOR_DELETE) is defined
but we don't have a primary key. This allows the handler to take precisions
in remembering any hidden primary key to able to update/delete any
found row. The default handler marks all columns to be read.
- handler::table_flags() now returns a ulonglong (to allow for more flags).
- New/changed table_flags()
- HA_HAS_RECORDS Set if ::records() is supported
- HA_NO_TRANSACTIONS Set if engine doesn't support transactions
- HA_PRIMARY_KEY_REQUIRED_FOR_DELETE
Set if we should mark all primary key columns for
read when reading rows as part of a DELETE
statement. If there is no primary key,
all columns are marked for read.
- HA_PARTIAL_COLUMN_READ Set if engine will not read all columns in some
cases (based on table->read_set)
- HA_PRIMARY_KEY_ALLOW_RANDOM_ACCESS
Renamed to HA_PRIMARY_KEY_REQUIRED_FOR_POSITION.
- HA_DUPP_POS Renamed to HA_DUPLICATE_POS
- HA_REQUIRES_KEY_COLUMNS_FOR_DELETE
Set this if we should mark ALL key columns for
read when when reading rows as part of a DELETE
statement. In case of an update we will mark
all keys for read for which key part changed
value.
- HA_STATS_RECORDS_IS_EXACT
Set this if stats.records is exact.
(This saves us some extra records() calls
when optimizing COUNT(*))
- Removed table_flags()
- HA_NOT_EXACT_COUNT Now one should instead use HA_HAS_RECORDS if
handler::records() gives an exact count() and
HA_STATS_RECORDS_IS_EXACT if stats.records is exact.
- HA_READ_RND_SAME Removed (no one supported this one)
- Removed not needed functions ha_retrieve_all_cols() and ha_retrieve_all_pk()
- Renamed handler::dupp_pos to handler::dup_pos
- Removed not used variable handler::sortkey
Upper level handler changes:
- ha_reset() now does some overall checks and calls ::reset()
- ha_table_flags() added. This is a cached version of table_flags(). The
cache is updated on engine creation time and updated on open.
MySQL level changes (not obvious from the above):
- DBUG_ASSERT() added to check that column usage matches what is set
in the column usage bit maps. (This found a LOT of bugs in current
column marking code).
- In 5.1 before, all used columns was marked in read_set and only updated
columns was marked in write_set. Now we only mark columns for which we
need a value in read_set.
- Column bitmaps are created in open_binary_frm() and open_table_from_share().
(Before this was in table.cc)
- handler::table_flags() calls are replaced with handler::ha_table_flags()
- For calling field->val() you must have the corresponding bit set in
table->read_set. For calling field->store() you must have the
corresponding bit set in table->write_set. (There are asserts in
all store()/val() functions to catch wrong usage)
- thd->set_query_id is renamed to thd->mark_used_columns and instead
of setting this to an integer value, this has now the values:
MARK_COLUMNS_NONE, MARK_COLUMNS_READ, MARK_COLUMNS_WRITE
Changed also all variables named 'set_query_id' to mark_used_columns.
- In filesort() we now inform the handler of exactly which columns are needed
doing the sort and choosing the rows.
- The TABLE_SHARE object has a 'all_set' column bitmap one can use
when one needs a column bitmap with all columns set.
(This is used for table->use_all_columns() and other places)
- The TABLE object has 3 column bitmaps:
- def_read_set Default bitmap for columns to be read
- def_write_set Default bitmap for columns to be written
- tmp_set Can be used as a temporary bitmap when needed.
The table object has also two pointer to bitmaps read_set and write_set
that the handler should use to find out which columns are used in which way.
- count() optimization now calls handler::records() instead of using
handler->stats.records (if (table_flags() & HA_HAS_RECORDS) is true).
- Added extra argument to Item::walk() to indicate if we should also
traverse sub queries.
- Added TABLE parameter to cp_buffer_from_ref()
- Don't close tables created with CREATE ... SELECT but keep them in
the table cache. (Faster usage of newly created tables).
New interfaces:
- table->clear_column_bitmaps() to initialize the bitmaps for tables
at start of new statements.
- table->column_bitmaps_set() to set up new column bitmaps and signal
the handler about this.
- table->column_bitmaps_set_no_signal() for some few cases where we need
to setup new column bitmaps but don't signal the handler (as the handler
has already been signaled about these before). Used for the momement
only in opt_range.cc when doing ROR scans.
- table->use_all_columns() to install a bitmap where all columns are marked
as use in the read and the write set.
- table->default_column_bitmaps() to install the normal read and write
column bitmaps, but not signaling the handler about this.
This is mainly used when creating TABLE instances.
- table->mark_columns_needed_for_delete(),
table->mark_columns_needed_for_delete() and
table->mark_columns_needed_for_insert() to allow us to put additional
columns in column usage maps if handler so requires.
(The handler indicates what it neads in handler->table_flags())
- table->prepare_for_position() to allow us to tell handler that it
needs to read primary key parts to be able to store them in
future table->position() calls.
(This replaces the table->file->ha_retrieve_all_pk function)
- table->mark_auto_increment_column() to tell handler are going to update
columns part of any auto_increment key.
- table->mark_columns_used_by_index() to mark all columns that is part of
an index. It will also send extra(HA_EXTRA_KEYREAD) to handler to allow
it to quickly know that it only needs to read colums that are part
of the key. (The handler can also use the column map for detecting this,
but simpler/faster handler can just monitor the extra() call).
- table->mark_columns_used_by_index_no_reset() to in addition to other columns,
also mark all columns that is used by the given key.
- table->restore_column_maps_after_mark_index() to restore to default
column maps after a call to table->mark_columns_used_by_index().
- New item function register_field_in_read_map(), for marking used columns
in table->read_map. Used by filesort() to mark all used columns
- Maintain in TABLE->merge_keys set of all keys that are used in query.
(Simplices some optimization loops)
- Maintain Field->part_of_key_not_clustered which is like Field->part_of_key
but the field in the clustered key is not assumed to be part of all index.
(used in opt_range.cc for faster loops)
- dbug_tmp_use_all_columns(), dbug_tmp_restore_column_map()
tmp_use_all_columns() and tmp_restore_column_map() functions to temporally
mark all columns as usable. The 'dbug_' version is primarily intended
inside a handler when it wants to just call Field:store() & Field::val()
functions, but don't need the column maps set for any other usage.
(ie:: bitmap_is_set() is never called)
- We can't use compare_records() to skip updates for handlers that returns
a partial column set and the read_set doesn't cover all columns in the
write set. The reason for this is that if we have a column marked only for
write we can't in the MySQL level know if the value changed or not.
The reason this worked before was that MySQL marked all to be written
columns as also to be read. The new 'optimal' bitmaps exposed this 'hidden
bug'.
- open_table_from_share() does not anymore setup temporary MEM_ROOT
object as a thread specific variable for the handler. Instead we
send the to-be-used MEMROOT to get_new_handler().
(Simpler, faster code)
Bugs fixed:
- Column marking was not done correctly in a lot of cases.
(ALTER TABLE, when using triggers, auto_increment fields etc)
(Could potentially result in wrong values inserted in table handlers
relying on that the old column maps or field->set_query_id was correct)
Especially when it comes to triggers, there may be cases where the
old code would cause lost/wrong values for NDB and/or InnoDB tables.
- Split thd->options flag OPTION_STATUS_NO_TRANS_UPDATE to two flags:
OPTION_STATUS_NO_TRANS_UPDATE and OPTION_KEEP_LOG.
This allowed me to remove some wrong warnings about:
"Some non-transactional changed tables couldn't be rolled back"
- Fixed handling of INSERT .. SELECT and CREATE ... SELECT that wrongly reset
(thd->options & OPTION_STATUS_NO_TRANS_UPDATE) which caused us to loose
some warnings about
"Some non-transactional changed tables couldn't be rolled back")
- Fixed use of uninitialized memory in ha_ndbcluster.cc::delete_table()
which could cause delete_table to report random failures.
- Fixed core dumps for some tests when running with --debug
- Added missing FN_LIBCHAR in mysql_rm_tmp_tables()
(This has probably caused us to not properly remove temporary files after
crash)
- slow_logs was not properly initialized, which could maybe cause
extra/lost entries in slow log.
- If we get an duplicate row on insert, change column map to read and
write all columns while retrying the operation. This is required by
the definition of REPLACE and also ensures that fields that are only
part of UPDATE are properly handled. This fixed a bug in NDB and
REPLACE where REPLACE wrongly copied some column values from the replaced
row.
- For table handler that doesn't support NULL in keys, we would give an error
when creating a primary key with NULL fields, even after the fields has been
automaticly converted to NOT NULL.
- Creating a primary key on a SPATIAL key, would fail if field was not
declared as NOT NULL.
Cleanups:
- Removed not used condition argument to setup_tables
- Removed not needed item function reset_query_id_processor().
- Field->add_index is removed. Now this is instead maintained in
(field->flags & FIELD_IN_ADD_INDEX)
- Field->fieldnr is removed (use field->field_index instead)
- New argument to filesort() to indicate that it should return a set of
row pointers (not used columns). This allowed me to remove some references
to sql_command in filesort and should also enable us to return column
results in some cases where we couldn't before.
- Changed column bitmap handling in opt_range.cc to be aligned with TABLE
bitmap, which allowed me to use bitmap functions instead of looping over
all fields to create some needed bitmaps. (Faster and smaller code)
- Broke up found too long lines
- Moved some variable declaration at start of function for better code
readability.
- Removed some not used arguments from functions.
(setup_fields(), mysql_prepare_insert_check_table())
- setup_fields() now takes an enum instead of an int for marking columns
usage.
- For internal temporary tables, use handler::write_row(),
handler::delete_row() and handler::update_row() instead of
handler::ha_xxxx() for faster execution.
- Changed some constants to enum's and define's.
- Using separate column read and write sets allows for easier checking
of timestamp field was set by statement.
- Remove calls to free_io_cache() as this is now done automaticly in ha_reset()
- Don't build table->normalized_path as this is now identical to table->path
(after bar's fixes to convert filenames)
- Fixed some missed DBUG_PRINT(.."%lx") to use "0x%lx" to make it easier to
do comparision with the 'convert-dbug-for-diff' tool.
Things left to do in 5.1:
- We wrongly log failed CREATE TABLE ... SELECT in some cases when using
row based logging (as shown by testcase binlog_row_mix_innodb_myisam.result)
Mats has promised to look into this.
- Test that my fix for CREATE TABLE ... SELECT is indeed correct.
(I added several test cases for this, but in this case it's better that
someone else also tests this throughly).
Lars has promosed to do this.
