Big endian checksum is needed to workaround 10+ years old bug, where checksum
was calculated incorrectly on big endian hardware. We can still checksum such
tablespaces using software implementation of CRC32.
- make hp_mask an inline function (short and called 16 times)
- Simplify hp_search. Biggest benefit is for doing key lookup
without a matching row. Matching rows may be a bit slower, but
is offseted by having hp_mask inlined.
The non-persistent UPDATE_TIME for InnoDB tables was not being
updated consistently at transaction commit.
If a transaction is partly rolled back so that in the end it will
not modify a table that it intended to modify, the update_time will
be updated nevertheless. This will also happen when InnoDB fails
to write an undo log record for the intended modification.
If a transaction is committed internally in InnoDB, instead of
being committed from the SQL interface, then the trx_t::mod_tables
will not be applied to the update_time of the tables.
trx_t::mod_tables: Replace the std::set<dict_table_t*>
with std::map<dict_table_t*,undo_no_t>, so that the very first
modification within the transaction is identified.
trx_undo_report_row_operation(): Update mod_tables for every operation
after the undo log record was successfully written.
trx_rollback_to_savepoint_low(): After partial rollback, erase from
trx_t::mod_tables any tables for which all changes were rolled back.
trx_commit_in_memory(): Tighten some assertions and simplify conditions.
Invoke trx_update_mod_tables_timestamp() if persistent tables were
affected.
trx_commit_for_mysql(): Remove the call to
trx_update_mod_tables_timestamp(), as it is now invoked at the
lower level, in trx_commit_in_memory().
trx_rollback_finish(): Clear mod_tables before invoking trx_commit(),
because the trx_commit_in_memory() would otherwise wrongly process
mod_tables after a full ROLLBACK.
TODO:
- Make get_thd_memroot() inline
- To do this, we need to reduce dependence of include files, especially
so that sql_class.h is not depending in item.h
The Debian packaging build was failing on Travis-CI due to timeouts.
These changes make it faster by skipping less relevant parts of the build,
like the test suite, RocksDB, TokuDB and Mroonga.
Also fix the usage of the 'make -j' flag to not have any spaces as
recommended in an example in the Debian policy.
As they fail on TRT schema check:
innodb.log_file
innodb.table_flags
innodb.row_format_redundant
encryption.innodb_encrypt_log_corruption
encryption.innodb_first_page
