The issue here is that end_of_file for encrypted temporary IO_CACHE (used by filesort) is updated
using lseek.
Encryption adds storage overhead and hides it from the caller by recalculating offsets and lengths.
Two different IO_CACHE cannot possibly modify the same file
because the encryption key is randomly generated and stored in the IO_CACHE.
So when the tempfiles are encrypted DO NOT use lseek to change end_of_file.
Further observations about updating end_of_file using lseek
1) The end_of_file update is only used for binlog index files
2) The whole point is to update file length when the file was modified via a different file descriptor.
3) The temporary IO_CACHE files can never be modified via a different file descriptor.
4) For encrypted temporary IO_CACHE, end_of_file should not be updated with lseek
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
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.
Users expect window functions to produce a certain ordering of rows in
the final result set. Although the standard does not require this, we
already have the filesort result done for when we computed the window
function. If there is no ORDER BY attached to the query, just keep it
till the SELECT is completely evaluated and use that to print the
result.
Update test cases as many did not take care to guarantee a stable
result.
According to logs analysis the Dump thread attempted to read again data which
was already sent. The reason of regressed read turns out in an _my_b_cache_read()
early exit branch which missed to distinguish between total zero size read (e.g
ineffective read when Count argument is zero) from a case when the
requested amount of data is fully read out by sole accessing the cache's
file. In the latter case such then *effective* reading was not
reflected in the cache's state to screw the cache's state.
Fixed with a check introduced of whether the file reading was effective prior to
early exit. When this is the case conduct standard cache state change to
account the actual read size.
Notice the bug can show up also as an error to read binlog event e.g
through BINLOG_GTID_POS() (of MDEV-16886).
Handle string length as size_t, consistently (almost always:))
Change function prototypes to accept size_t, where in the past
ulong or uint were used. change local/member variables to size_t
when appropriate.
This fix excludes rocksdb, spider,spider, sphinx and connect for now.
The reason for adding this was that while testing mysqlbinlog on
a replication event with 3G event output, Linux failed reading
the whole file in memory with one read (only got 2G on first read
even if file had just been written).
- Don't reset info->error on write error in IO_CACHE.
- In case of write_error in IO_CACHE , always return -1
- Fixed wrong result from my_read when using MY_FULL_IO. Also don't give
an error in case of retry.
- Fix win64 pointer truncation warnings
(usually coming from misusing 0x%lx and long cast in DBUG)
- Also fix printf-format warnings
Make the above mentioned warnings fatal.
- fix pthread_join on Windows to set return value.
These self references were previously used to avoid having to check the
IO_CACHE's type. However, a benchmark shows that on x86 5930k stock,
the type comparison is marginally faster than the double pointer dereference.
For 40 billion my_b_tell calls, the difference is .1 seconds in favor of performing the
type check. (Basically there is no measurable difference)
To prevent bugs from copying the structure using the equals(=) operator,
and having to do the bookkeeping manually, remove these "convenience"
variables.
The reason for this is that stop slave takes LOCK_active_mi over the
whole operation while some slave operations will also need LOCK_active_mi
which causes deadlocks.
Fixed by introducing object counting for Master_info and not taking
LOCK_active_mi over stop slave or even stop_all_slaves()
Another benefit of this approach is that it allows:
- Multiple threads can run SHOW SLAVE STATUS at the same time
- START/STOP/RESET/SLAVE STATUS on a slave will not block other slaves
- Simpler interface for handling get_master_info()
- Added some missing unlock of 'log_lock' in error condtions
- Moved rpl_parallel_inactivate_pool(&global_rpl_thread_pool) to end
of stop_slave() to not have to use LOCK_active_mi inside
terminate_slave_threads()
- Changed argument for remove_master_info() to Master_info, as we always
have this available
- Fixed core dump when doing FLUSH TABLES WITH READ LOCK and parallel
replication. Problem was that waiting for pause_for_ftwrl was not done
when deleting rpt->current_owner after a force_abort.
The bug was caused by several issues.
2 problems in seek_io_cache. Due to wrong offsets used, we would end up
seeking way too much (first change), or over the intended seek point
(second change). Fixing it requires correctly detecting available data
in buffer (first change), and not using "IO_SIZE alligned" reads. The
second is needed because _my_b_cache_read adjusts the pos_in_file itself
based on read_pos and read_end. Pretending buffer is empty when we want
to force a read will aleviate this problem.
Secondly, the big-table cursors didn't repect the interface definitions
of always returning the rownumber that Table_read_cursor::fetch() would activate.
At the same time, next(), prev() and move_to() should not perform any
row activation.
update info->write_end and info->write_pos together, with no
"return on error" in between, otherwise write_end might end up being
smaller than write_pos
Add support for having multiple IO_CACHEs with type=READ_CACHE to share
the file they are reading from.
Each IO_CACHE keeps its own in-memory buffer. When doing a read or seek
operation on the file, it notifies other IO_CACHEs that the file position
has been changed.
Make Rowid_seq_cursor use cloned IO_CACHE when reading filesort result.
Instead of encrypt(src, dst, key, iv) that encrypts all
data in one go, now we have encrypt_init(key,iv),
encrypt_update(src,dst), and encrypt_finish(dst).
This also causes collateral changes in the internal my_crypt.cc
encryption functions and in the encryption service.
There are wrappers to provide the old all-at-once encryption
functionality. But binlog events are often written piecewise,
they'll need the new api.
