The trx_t::undo_mutex covered both some main-memory data structures
(trx_undo_t) and access to undo pages. The trx_undo_t is only
accessed by the thread that is associated with a running transaction.
Likewise, each transaction has its private set of undo pages.
The thread that is associated with an active transaction may
lock multiple undo pages concurrently, but no other thread may
lock multiple pages of a foreign transaction.
Concurrent access to the undo logs of an active transaction is possible,
but trx_undo_get_undo_rec_low() only locks one undo page at a time,
without ever holding any undo_mutex.
It seems that the trx_t::undo_mutex would have been necessary if
multi-threaded execution or rollback of a single transaction
had been implemented in InnoDB.
There is only one lock_sys. Allocate it statically in order to avoid
dereferencing a pointer whenever accessing it. Also, align some
members to their own cache line in order to avoid false sharing.
lock_sys_t::create(): The deferred constructor.
lock_sys_t::close(): The early destructor.
This was done in, among other things:
- thd->db and thd->db_length
- TABLE_LIST tablename, db, alias and schema_name
- Audit plugin database name
- lex->db
- All db and table names in Alter_table_ctx
- st_select_lex db
Other things:
- Changed a lot of functions to take const LEX_CSTRING* as argument
for db, table_name and alias. See init_one_table() as an example.
- Changed some function arguments from LEX_CSTRING to const LEX_CSTRING
- Changed some lists from LEX_STRING to LEX_CSTRING
- threads_mysql.result changed because process list_db wasn't always
correctly updated
- New append_identifier() function that takes LEX_CSTRING* as arguments
- Added new element tmp_buff to Alter_table_ctx to separate temp name
handling from temporary space
- Ensure we store the length after my_casedn_str() of table/db names
- Removed not used version of rename_table_in_stat_tables()
- Changed Natural_join_column::table_name and db_name() to never return
NULL (used for print)
- thd->get_db() now returns db as a printable string (thd->db.str or "")
InnoDB RNG maintains global state, causing otherwise unnecessary bus
traffic. Even worse this is cross-mutex traffic. That is different
mutexes suffer from contention.
Fixed delay of 4 was verified to give best throughput by OLTP update
index and read-write benchmarks on Intel Broadwell (2/20/40) and
ARM (1/46/46).
trx_sys_t::rw_trx_set is implemented as std::set, which does a few quite
expensive operations under trx_sys_t::mutex protection: e.g. malloc/free
when adding/removing elements. Traversing b-tree is not that cheap either.
This has negative scalability impact, which is especially visible when running
oltp_update_index.lua benchmark on a ramdisk.
To reduce trx_sys_t::mutex contention std::set is replaced with LF_HASH. None
of LF_HASH operations require trx_sys_t::mutex (nor any other global mutex)
protection.
Another interesting issue observed with std::set is reproducible ~2% performance
decline after benchmark is ran for ~60 seconds. With LF_HASH results are stable.
All in all this patch optimises away one of three trx_sys->mutex locks per
oltp_update_index.lua query. The other two critical sections became smaller.
Relevant clean-ups:
Replaced rw_trx_set iteration at startup with local set. The latter is needed
because values inserted to rw_trx_list must be ordered by trx->id.
Removed redundant conditions from trx_reference(): it is (and even was) never
called with transactions that have trx->state == TRX_STATE_COMMITTED_IN_MEMORY.
do_ref_count doesn't (and probably even didn't) make any sense: now it is called
only when reference counter increment is actually requested.
Moved condition out of mutex in trx_erase_lists().
trx_rw_is_active(), trx_rw_is_active_low() and trx_get_rw_trx_by_id() were
greatly simplified and replaced by appropriate trx_rw_hash_t methods.
Compared to rw_trx_set, rw_trx_hash holds transactions only in PREPARED or
ACTIVE states. Transactions in COMMITTED state were required to be found
at InnoDB startup only. They are now looked up in the local set.
Removed unused trx_assert_recovered().
Removed unused innobase_get_trx() declaration.
Removed rather semantically incorrect trx_sys_rw_trx_add().
Moved information printout from trx_sys_init_at_db_start() to
trx_lists_init_at_db_start().
Relax memory barrier for lock_word.
rw_lock_lock_word_decr() - used to acquire rw-lock, thus we only need to issue
ACQUIRE when we succeed locking.
rw_lock_x_lock_func_nowait() - same as above, but used to attempt to acquire
X-lock.
rw_lock_s_unlock_func() - used to release S-lock, RELEASE is what we need here.
rw_lock_x_unlock_func() - used to release X-lock. Ideally we'd need only RELEASE
here, but due to mess with waiters (they must be loaded after lock_word is
stored) we have to issue both ACQUIRE and RELEASE.
rw_lock_sx_unlock_func() - same as above, but used to release SX-lock.
rw_lock_s_lock_spin(), rw_lock_x_lock_func(), rw_lock_sx_lock_func() -
fetch-and-store to waiters has to issue only ACQUIRE memory barrier, so that
waiters are stored before lock_word is loaded.
Note that there is violation of RELEASE-ACQUIRE protocol here, because we do
on lock:
my_atomic_fas32_explicit((int32*) &lock->waiters, 1, MY_MEMORY_ORDER_ACQUIRE);
my_atomic_load32_explicit(&lock->lock_word, MY_MEMORY_ORDER_RELAXED);
on unlock
my_atomic_add32_explicit(&lock->lock_word, X_LOCK_DECR, MY_MEMORY_ORDER_ACQ_REL);
my_atomic_load32_explicit((int32*) &lock->waiters, MY_MEMORY_ORDER_RELAXED);
That is we kind of synchronize ACQUIRE on lock_word with ACQUIRE on waiters.
It was there before this patch. Simple fix may have negative performance impact.
Proper fix requires refactoring of lock_word.
Remove volatile modifier from waiters: it's not supposed for inter-thread
communication, use appropriate atomic operations instead.
Changed waiters to int32_t, my_atomic friendly type.
Problem was that we could take page latches on different
order than wat is entitled with SX-lock. To follow the
latching order defined in WL#6326, acquire index->lock X-latch.
This entitles us to acquire page latches in any order for the index.
btr0btr.cc
Document latch rules before and after MariaDB 10.2.2
sync0rw.cc
Document latch compatibility rules better.
btr_defragment_merge_pages
Fix parameter value.
btr_defragment_thread
Acquire X-lock to dict_index_t::lock before restoring
cursor position and continuing defragmentation.
ha_innobase::optimize
Restore defragment feature.
Testing
Add GIS-index and FT-index to table being defragmented.
Defragmentation is not done to GIS-indexes and FT auxiliary
tables.
The parameter --innodb-sync-debug, which is disabled by default,
aims to find potential deadlocks in InnoDB.
When the parameter is enabled, lots of tests failed. Most of these
failures were due to bogus diagnostics. But, as part of this fix,
we are also fixing a bug in error handling code and removing dead
code, and fixing cases where an uninitialized mutex was being
locked and unlocked.
dict_create_foreign_constraints_low(): Remove an extraneous
mutex_exit() call that could cause corruption in an error handling
path. Also, do not unnecessarily acquire dict_foreign_err_mutex.
Its only purpose is to control concurrent access to
dict_foreign_err_file.
row_ins_foreign_trx_print(): Replace a redundant condition with a
debug assertion.
srv_dict_tmpfile, srv_dict_tmpfile_mutex: Remove. The
temporary file is never being written to or read from.
log_free_check(): Allow SYNC_FTS_CACHE (fts_cache_t::lock)
to be held.
ha_innobase::inplace_alter_table(), row_merge_insert_index_tuples():
Assert that no unexpected latches are being held.
sync_latch_meta_init(): Properly initialize dict_operation_lock_key
at SYNC_DICT_OPERATION. dict_sys->mutex is SYNC_DICT, and
the now-removed SRV_DICT_TMPFILE was wrongly registered at
SYNC_DICT_OPERATION.
buf_block_init(): Correctly register buf_block_t::debug_latch.
It was previously misleadingly reported as LATCH_ID_DICT_FOREIGN_ERR.
latch_level_t: Correct the relative latching order of
SYNC_IBUF_PESS_INSERT_MUTEX,SYNC_INDEX_TREE and
SYNC_FILE_FORMAT_TAG,SYNC_DICT_OPERATION to avoid bogus failures.
row_drop_table_for_mysql(): Avoid accessing btr_defragment_mutex
if the defragmentation thread has not been started. This is the
case during fts_drop_orphaned_tables() in recv_recovery_rollback_active().
fil_space_destroy_crypt_data(): Avoid acquiring fil_crypt_threads_mutex
when it is uninitialized. We may have created crypt_data before the
mutex was created, and the mutex creation would be skipped if
InnoDB startup failed or --innodb-read-only was specified.
buf_page_io_complete(): Do not test bpage for NULL, because
it is declared (and always passed) as nonnull.
buf_flush_batch(): Remove the constant local variable count=0.
fil_ibd_load(): Use magic comment to suppress -Wimplicit-fallthrough.
ut_stage_alter_t::inc(ulint): Disable references to an unused parameter.
lock_queue_validate(), sync_array_find_thread(), rbt_check_ordering():
Define only in debug builds.
When the btr_search_latch was split into an array of latches
in MySQL 5.7.8 as part of the Oracle Bug#20985298 fix, the "caching"
of the latch across storage engine API calls was removed, and
the field trx->has_search_latch would only be set during a short
time frame in the execution of row_search_mvcc(), which was
formerly called row_search_for_mysql().
This means that the column
INFORMATION_SCHEMA.INNODB_TRX.TRX_ADAPTIVE_HASH_LATCHED will always
report 0. That column cannot be removed in MariaDB 10.2, but it
can be removed in future releases.
trx_t::has_search_latch: Remove.
trx_assert_no_search_latch(): Remove.
row_sel_try_search_shortcut_for_mysql(): Remove a redundant condition
on trx->has_search_latch (it was always true).
sync_check_iterate(): Make the parameter const.
sync_check_functor_t: Make the operator() const, and remove result()
and the virtual destructor. There is no need to have mutable state
in the functors.
sync_checker<bool>: Replaces dict_sync_check and btrsea_sync_check.
sync_check: Replaces btrsea_sync_check.
dict_sync_check: Instantiated from sync_checker.
sync_allowed_latches: Use std::find() directly on the array.
Remove the std::vector.
TrxInInnoDB::enter(), TrxInInnoDB::exit(): Remove obviously redundant
debug assertions on trx->in_depth, and use equality comparison against 0
because it could be more efficient on some architectures.
The following options will be removed:
innodb_file_format
innodb_file_format_check
innodb_file_format_max
innodb_large_prefix
They have been deprecated in MySQL 5.7.7 (and MariaDB 10.2.2) in WL#7703.
The file_format column in two INFORMATION_SCHEMA tables will be removed:
innodb_sys_tablespaces
innodb_sys_tables
Code to update the file format tag at the end of page 0:5
(TRX_SYS_PAGE in the InnoDB system tablespace) will be removed.
When initializing a new database, the bytes will remain 0.
All references to the Barracuda file format will be removed.
Some references to the Antelope file format (meaning
ROW_FORMAT=REDUNDANT or ROW_FORMAT=COMPACT) will remain.
This basically ports WL#7704 from MySQL 8.0.0 to MariaDB 10.3.1:
commit 4a69dc2a95995501ed92d59a1de74414a38540c6
Author: Marko Mäkelä <marko.makela@oracle.com>
Date: Wed Mar 11 22:19:49 2015 +0200
Do not silence uncertain cases, or fix any bugs.
The only functional change should be that ha_federated::extra()
is not calling DBUG_PRINT to report an unhandled case for
HA_EXTRA_PREPARE_FOR_DROP.
rw_lock_free_func(): Assign lock->magic_n = 0 in debug builds,
because sometimes, the assignment in the explicit destructor call
is being optimized away.
This was verified when using GCC 6.3.0 with -O3 -g -mtune=native
on a Debian GNU/Linux 9.0 system on AMD64.
Issue:
======
Disabling macros such as UNIV_PFS_MUTEX/UNIV_PFS_RWLOCK/UNIV_PFS_THREAD
which are defined in InnoDB throws errors during compilation.
Fix:
====
Fix all the compilation errors.
RB: 14893
Reviewed-by: Jimmy Yang <Jimmy.Yang@oracle.com>
Reviewed-by: Satya Bodapati <satya.bodapati@oracle.com>
Problem: Some instantiations of std::map have discrepancies between
the value_type of the map and the value_type of the map's allocator.
On FreeBSD 11 this is detected by Clang, and an error is raised at
compilation time.
Fix: Specify the correct value_type for the allocators.
Also fix an unused variable warning in storage/innobase/os/os0file.cc.
Alias the InnoDB ulint and lint data types to size_t and ssize_t,
which are the standard names for the machine-word-width data types.
Correspondingly, define ULINTPF as "%zu" and introduce ULINTPFx as "%zx".
In this way, better compiler warnings for type mismatch are possible.
Furthermore, use PRIu64 for that 64-bit format, and define
the feature macro __STDC_FORMAT_MACROS to enable it on Red Hat systems.
Fix some errors in error messages, and replace some error messages
with assertions.
Most notably, an IMPORT TABLESPACE error message in InnoDB was
displaying the number of columns instead of the mismatching flags.
Define UNIV_WORD_SIZE as a simple alias to SIZEOF_SIZE_T.
In MariaDB 10.0 and 10.1, it was incorrectly defined as 4 on
64-bit Windows.
MONITOR_OS_PENDING_READS, MONITOR_OS_PENDING_WRITES: Enable by default.
os_n_pending_reads, os_n_pending_writes: Remove.
Use the monitor counters instead.
Allow 64-bit atomic operations on 32-bit systems,
only relying on HAVE_ATOMIC_BUILTINS_64, disregarding
the width of the register file.
Define UNIV_WORD_SIZE correctly on all systems, including Windows.
In MariaDB 10.0 and 10.1, it was incorrectly defined as 4 on
64-bit Windows.
Define HAVE_ATOMIC_BUILTINS_64 on Windows
(64-bit atomics are available on both 32-bit and 64-bit Windows
platforms; the operations were unnecessarily disabled even on
64-bit Windows).
MONITOR_OS_PENDING_READS, MONITOR_OS_PENDING_WRITES: Enable by default.
os_file_n_pending_preads, os_file_n_pending_pwrites,
os_n_pending_reads, os_n_pending_writes: Remove.
Use the monitor counters instead.
os_file_count_mutex: Remove. On a system that does not support
64-bit atomics, monitor_mutex will be used instead.
InnoDB defines some functions that are not called at all.
Other functions are called, but only from the same compilation unit.
Remove some function declarations and definitions, and add 'static'
keywords. Some symbols must be kept for separately compiled tools,
such as innochecksum.
Also, remove empty .ic files that were not removed by my MySQL commit.
Problem:
InnoDB used to support a compilation mode that allowed to choose
whether the function definitions in .ic files are to be inlined or not.
This stopped making sense when InnoDB moved to C++ in MySQL 5.6
(and ha_innodb.cc started to #include .ic files), and more so in
MySQL 5.7 when inline methods and functions were introduced
in .h files.
Solution:
Remove all references to UNIV_NONINL and UNIV_MUST_NOT_INLINE from
all files, assuming that the symbols are never defined.
Remove the files fut0fut.cc and ut0byte.cc which only mattered when
UNIV_NONINL was defined.
Also, implement MDEV-11027 a little differently from 5.5 and 10.0:
recv_apply_hashed_log_recs(): Change the return type back to void
(DB_SUCCESS was always returned).
Report progress also via systemd using sd_notifyf().
Define my_thread_id as an unsigned type, to avoid mismatch with
ulonglong. Change some parameters to this type.
Use size_t in a few more places.
Declare many flag constants as unsigned to avoid sign mismatch
when shifting bits or applying the unary ~ operator.
When applying the unary ~ operator to enum constants, explictly
cast the result to an unsigned type, because enum constants can
be treated as signed.
In InnoDB, change the source code line number parameters from
ulint to unsigned type. Also, make some InnoDB functions return
a narrower type (unsigned or uint32_t instead of ulint;
bool instead of ibool).
