fix MDEV-18750: failed to flashback large-size binlog file
fix mysqlbinlog flashback failure caused by reading io_cache without MY_FULL_IO flag
fix MDEV-18750: mysqlbinlog flashback failure on large binlog
Problem:
========
The mysqlbinlog tool is leaking memory, causing failures in various tests when
compiling and testing with AddressSanitizer or LeakSanitizer like this:
cmake -DCMAKE_BUILD_TYPE=Debug -DWITH_ASAN:BOOL=ON /path/to/source
make -j$(nproc)
cd mysql-test
ASAN_OPTIONS=abort_on_error=1 ./mtr --parallel=auto
Analysis:
=========
Two types of leaks were observed during above execution.
1) Leak in Log_event::read_log_event(char const*, unsigned int, char const**,
Format_description_log_event const*, char)
File: sql/log_event.cc:2150
For all row based replication events the memory which is allocated during
read_log_event is not freed after the event is processed. The event specific
memory has to be retained only when flashback option is enabled with
mysqlbinlog tool. In this case all the events are retained till the end
statement is received and they are processed in reverse order and they are
destroyed. But in the existing code all events are retained irrespective of
flashback mode. Hence the memory leaks are observed.
2) read_remote_annotate_event(unsigned char*, unsigned long, char const**)
File: client/mysqlbinlog.cc:194
In general the Annotate event is not printed immediately because all
subsequent rbr-events can be filtered away. Instead it will be printed
together with the first not filtered away Table map or the last rbr will be
processed. While reading remote annotate events memory is allocated for event
buffer and event's temp_buf is made to point to the allocated buffer as shown
below. The TRUE flag is used for doing proper cleanup using free_temp_buf().
i.e at the time of deletion of annotate event its destructor takes care of
clearing the temp_buf.
/*
Ensure the event->temp_buf is pointing to the allocated buffer.
(TRUE = free temp_buf on the event deletion)
*/
event->register_temp_buf((char*)event_buf, TRUE);
But existing code does the following when it receives a remote annotate_event.
if (remote_opt)
ev->temp_buf= 0;
That is code immediately sets temp_buf=0, because of which free_temp_buf()
call will return empty handed as it has lost the reference to the allocated
temporary buffer. This results in memory leak
Fix:
====
1) If not in flashback mode, destroy the memory for events once they are
processed.
2) Remove the ev->temp_buf=0 code for remote option. Let the proper cleanup to
be done as part of free_temp_buf().
On some systems with 10,000+ binlogs, show binary logs could block
log rotation for more than 10 seconds.
This patch fixes this by first caching all binary log names and
releases all mutexes while calculating the sizes of the binary logs.
Other things:
- Ensure that reinit_io_cache() sets end_of_file when moving to read_cache.
This ensures that external changes of the underlying file is known to
the cache.
- get_binlog_list() is made more efficent and show_binlogs() is changed
to call get_binlog_list()
Reviewed by Andrei Elkin
ignore FK-prelocked tables when looking for write-prelocked tables
with auto-increment to complain about "Statement is unsafe because
it invokes a trigger or a stored function that inserts into an
AUTO_INCREMENT column"
The problem was originally stated in
http://bugs.mysql.com/bug.php?id=82212
The size of an base64-encoded Rows_log_event exceeds its
vanilla byte representation in 4/3 times.
When a binlogged event size is about 1GB mysqlbinlog generates
a BINLOG query that can't be send out due to its size.
It is fixed with fragmenting the BINLOG argument C-string into
(approximate) halves when the base64 encoded event is over 1GB size.
The mysqlbinlog in such case puts out
SET @binlog_fragment_0='base64-encoded-fragment_0';
SET @binlog_fragment_1='base64-encoded-fragment_1';
BINLOG @binlog_fragment_0, @binlog_fragment_1;
to represent a big BINLOG.
For prompt memory release BINLOG handler is made to reset the BINLOG argument
user variables in the middle of processing, as if @binlog_fragment_{0,1} = NULL
is assigned.
Notice the 2 fragments are enough, though the client and server still may
need to tweak their @@max_allowed_packet to satisfy to the fragment
size (which they would have to do anyway with greater number of
fragments, should that be desired).
On the lower level the following changes are made:
Log_event::print_base64()
remains to call encoder and store the encoded data into a cache but
now *without* doing any formatting. The latter is left for time
when the cache is copied to an output file (e.g mysqlbinlog output).
No formatting behavior is also reflected by the change in the meaning
of the last argument which specifies whether to cache the encoded data.
Rows_log_event::print_helper()
is made to invoke a specialized fragmented cache-to-file copying function
which is
copy_cache_to_file_wrapped()
that takes care of fragmenting also optionally wraps encoded
strings (fragments) into SQL stanzas.
my_b_copy_to_file()
is refactored to into my_b_copy_all_to_file(). The former function
is generalized
to accepts more a limit argument to constraint the copying and does
not reinitialize anymore the cache into reading mode.
The limit does not do any effect on the fully read cache.
main.derived_cond_pushdown: Move all 10.3 tests to the end,
trim trailing white space, and add an "End of 10.3 tests" marker.
Add --sorted_result to tests where the ordering is not deterministic.
main.win_percentile: Add --sorted_result to tests where the
ordering is no longer deterministic.
The test and also rpl_gtid_delete_domain failed on PPC64 platform
due to an incorrectly specified actual key for searching
in a gtid domain system hash. While the correct size is 32 bits
the supplied value was 8 bytes of long int size on the platform.
The problem became evident thanks to the big endiness which
cut off the *least* significant part of the value field.
Fixed with correcting a dynamic array initialization to hold
now uint32 values as well as the values extraction for
searching in the gtid domain system hash.
A new added test ensures no overflowed values are accepted
for deletion which prevents inadvertent action. Notice though
MariaDB [test]> set @@session.gtid_domain_id=(1 << 32) + 1;
MariaDB [test]> show warnings;
+---------+------+--------------------------------------------------------+
| Level | Code | Message |
+---------+------+--------------------------------------------------------+
| Warning | 1292 | Truncated incorrect gtid_domain_id value: '4294967297' |
+---------+------+--------------------------------------------------------+
MariaDB [test]> select @@session.gtid_domain_id;
+--------------------------+
| @@session.gtid_domain_id |
+--------------------------+
| 4294967295 |
+--------------------------+
Problem:- Create/drop index was logged into binlog.
Goal:- Operation on temporary table should not be binlog when binlog format
is row.
Solution:-
We should add CF_FORCE_ORIGINAL_BINLOG_FORMAT when there is ddl on temp
table.
For optimize, analyze, repair we wont change anything ,Then will
be logged in binlog , But they also dont throw any error if operation fails
Since slave wont be having any temp table , but these operation on tmp
table will be processed without breaking replication.
For rename we need a different logic MDEV-16728 will solve it.
The code passing positions in the query to constructors of
Rewritable_query_parameter descendants (e.g. Item_splocal)
was not reliable. It used various Lex_input_stream methods:
- get_tok_start()
- get_tok_start_prev()
- get_tok_end()
- get_ptr()
to find positions of the recently scanned tokens.
The challenge was mostly to choose between get_tok_start()
and get_tok_start_prev(), taking into account to the current
grammar (depending if lookahead takes place before
or after we read the positions in every particular rule).
But this approach did not work at all in combination
with token contractions, when MYSQLlex() translates
two tokens into one token ID, for example:
WITH ROLLUP -> WITH_ROLLUP_SYM
As a result, the tokenizer is already one more token ahead.
So in query fragment:
"GROUP BY d, spvar WITH ROLLUP"
get_tok_start() points to "ROLLUP".
get_tok_start_prev() points to "WITH".
As a result, it was "WITH" who was erroneously replaced
to NAME_CONST() instead of "spvar".
This patch modifies the code to do it a different way.
Changes:
1. For keywords and identifiers, the tokenizer now
returns LEX_CTRING pointing directly to the query
fragment. So query positions are now just available using:
- $1.str - for the beginning of a token
- $1.str+$1.length - for the end of a token
2. Identifiers are not allocated on the THD memory root
in the tokenizer any more. Allocation is now done
on later stages, in methods like LEX::create_item_ident().
3. Two LEX_CSTRING based structures were added:
- Lex_ident_cli_st - used to store the "client side"
identifier representation, pointing to the
query fragment. Note, these identifiers
are encoded in @@character_set_client
and can have broken byte sequences.
- Lex_ident_sys_st - used to store the "server side"
identifier representation, pointing to the
THD allocated memory. This representation
guarantees that the identifier was checked
for being well-formed, and is encoded in utf8.
4. To distinguish between two identifier types
in the grammar, two Bison types were added:
<ident_cli> and <ident_sys>
5. All non-reserved keywords were marked as
being of the type <ident_cli>.
All reserved keywords are still of the type NONE.
6. All curly brackets in rules collecting
non-reserved keywords into non-terminal
symbols were removed, e.g.:
Was:
keyword_sp_data_type:
BIT_SYM {}
| BOOLEAN_SYM {}
Now:
keyword_sp_data_type:
BIT_SYM
| BOOLEAN_SYM
This is important NOT to have brackets here!!!!
This is needed to make sure that the underlying
Lex_ident_cli_ststructure correctly passes up to
the calling rule.
6. The code to scan identifiers and keywords
was moved from lex_one_token() into new
Lex_input_stream methods:
scan_ident_sysvar()
scan_ident_start()
scan_ident_middle()
scan_ident_delimited()
This was done to:
- get rid of enormous amount of references to &yylval->lex_str
- and remove a lot of references like lip->xxx
7. The allocating functionality which puts identifiers on the
THD memory root now resides in methods of Lex_ident_sys_st,
and in THD::to_ident_sys_alloc().
get_quoted_token() was removed.
8. Cleanup: check_simple_select() was moved as a method to LEX.
9. Cleanup: Some more functionality was moved from *.yy
to new methods were added to LEX:
make_item_colon_ident_ident()
make_item_func_call_generic()
create_item_qualified_asterisk()
Problem was timing between the thread that was killed and reading the
binary log.
Updated the test to wait until the killed thread was properly terminated
before checking what's in the binary log.
To make check safe, I changed "threads_connected" to be updated after
thd::cleanup() is done, to ensure that all binary logs updates are done
before the variable is changed. This was mainly done to get the
test deterministic and have now other real influence in how the server
works.
Main problem was that no log-event print function checked for disk
full error on the IO_CACHE.
All changes in this patch only affects mysqlbinlog, not the server!
- Changed all log-event print functions to return 1 on error
- Fixed memory usage when not using --flashback.
- Added printing of number of rows in row events. Can be disabled with
--print-row-count=0
- Print annotated rows when using mysqlbinlog --short-form
- Fixed that mysqlbinlog --debug works
- Fixed create_drop_binlog.test test failure
- Reorganized fields in PRINT_EVENT_INFO to be according to size to
optimize storage
- Don't change print_row_event_position or print_row_counts if set by user
- Remove some testing of argument to my_free is 0
- base64-output=never is now supported and works in all context
- Updated help information for --base64-output and --short-form
- print_row_count is now on by default. Reset automatically if --short-form
is used
- Removed obsolote warning for mysql 5.6.0
- More DBUG_PRINT for mysqltest.cc
- my_b_write_byte() now checks for flush failures. This fixed a memory
overrun on disk full
- my_b_printf() now returns 1 on failure, 0 on ok. This simplifies code
and no old code was using the old return value of my_b_printf().
- my_b_Write_backtick_quote() now returns 1 on failure and 0 on ok
- Fixed some error conditions in log printing that was not previously
handled.
- Slave_rows_error_report() can now handle longlong positions
- Write_on_release_cache() rewritten so that we can detect errors
on flush. Not depending on automatic release anymore.
- Changed types for Pos and End_log_pos to 64 bit in SHOW BINLOG EVENTS
- Fixed that copy_event_cache_to_string_and_reinit() works with strings
longer than 4G (Changed to use LEX_STRING instead of String)
- Restricted binlog_rows_event_max_size to UINT32_MAX-1 as String's are
anyway restricted to UINT32_MAX
- Fixed bug in rpl_binlog_state::write_to_iocache() which hide write
failures (duplicate variable name)
- Fixed bug in String::append if original string was not allocated
- Stop mysqlbinlog output at once if there is an error.
- Before printing error message, flush result file. This ensures that
the error message is printed last. (Easier to find)
