- Don't try to push down SELECTs that have a side effect
- In case the storage engine did support pushdown of SELECT with an INTO
clause, write the rows we've got from it into select->join->result,
and not thd->protocol. This way, SELECT ... INTO ... FROM
smart_engine_table will put the result into where instructed, and
NOT send it to the client.
Some fields were protected by log_sys.mutex, which adds quite some
overhead for readers. Some readers were submitting dirty reads.
log_t::lsn: Declare private and atomic. Add wrappers get_lsn()
and set_lsn() that will use relaxed memory access. Many accesses
to log_sys.lsn are still protected by log_sys.mutex; we avoid the
mutex for some readers.
log_t::flushed_to_disk_lsn: Declare private and atomic, and move
to the same cache line with log_t::lsn.
log_t::buf_free: Declare as size_t, and move to the same cache line
with log_t::lsn.
log_t::check_flush_or_checkpoint_: Declare private and atomic,
and move to the same cache line with log_t::lsn.
log_get_lsn(): Define as an alias of log_sys.get_lsn().
log_get_lsn_nowait(), log_peek_lsn(): Remove.
log_get_flush_lsn(): Define as an alias of log_sys.get_flush_lsn().
log_t::initiate_write(): Replaces log_buffer_sync_in_background().
O_DSYNC is faster than O_SYNC because it syncs as little as needed
(e.g. no timestamp changes)
This change is similar to change fsync() -> fdatasync() in MDEV-21382
The ut_crc32() function uses a hard-coded initial CRC-32C value of 0.
Replace it with ut_crc32_low(), which allows to specify the initial
checksum value, and provide an inlined compatibility wrapper ut_crc32().
Also, remove non-inlined wrapper functions on ARMv8 and POWER8,
and remove dead code (the generic implementation) on POWER8.
Note: The original AMD64 instruction set architecture in 2003 only
included SSE2. The CRC-32C instructions are part of the SSE4.2
instruction set extension for IA-32 and AMD64, with first processors
released in November 2007 (using the AMD Barcelona microarchitecture)
and November 2008 (Intel Nehalem microarchiteture). It might be safe
to assume that SSE4.2 is available on all currently used AMD64 based
systems, but we are not taking that step yet.
The configuration parameter innodb_scrub_log never really worked, as
reported in MDEV-13019 and MDEV-18370.
Because MDEV-14425 is changing the redo log format, the innodb_scrub_log
feature would have to be adjusted for it. Due to the known problems,
it is easier to remove the feature for now, and to ignore and deprecate
the parameters.
If old log contents should be kept secret, then enabling innodb_encrypt_log
or setting a smaller innodb_log_file_size could help.
MONITOR_PENDING_CHECKPOINT_WRITE and MONITOR_LOG_IO track
log_sys.n_pending_checkpoint_writes and log_sys.n_log_ios,
respectively. The MONITOR_INC calls are redundant, because
the values will be overwritten in srv_mon_process_existing_counter().
Compute MONITOR_LSN_CHECKPOINT_AGE on demand in
srv_mon_process_existing_counter().
This allows us to remove the overhead of MONITOR_SET
calls for the counter.
The function log_header_read() was only used during server startup,
and it will mostly be used only for reading checkpoint information
from pre-MDEV-14425 format redo log files.
Let us replace the function with more direct calls, so that
it is clearer what is going on. It is not strictly necessary to
hold any mutex during this operation, and because there will be
only a limited number of operations during early server startup,
it is not necessary to increment any I/O counters.
Simplify the logging of ALTER TABLE operations, by making use of the
TRX_UNDO_RENAME_TABLE undo log record that was introduced in
commit 0bc36758ba.
commit_try_rebuild(): Invoke row_rename_table_for_mysql() and
actually rename the files before committing the transaction.
fil_mtr_rename_log(), commit_cache_rebuild(),
log_append_on_checkpoint(), row_merge_rename_tables_dict(): Remove.
mtr_buf_copy_t, log_t::append_on_checkpoint: Remove.
row_rename_table_for_mysql(): If !use_fk, ignore missing foreign
keys. Remove a call to dict_table_rename_in_cache(), because
trx_rollback_to_savepoint() should invoke the function if needed.
In commit 138cbec5f2, we
computed an incorrect parameter to page_dir_split_slot(),
leading us to splitting the wrong directory slot, or
an out-of-bounds access when splitting the supremum slot.
This was once caught in the test innodb_gis.kill_server for
inserting records to a clustered index root page.
page_dir_split_slot(): Take the slot as a pointer, instead of
a numeric index.
page_apply_insert_redundant(), page_apply_insert_dynamic():
Rename slot to last_slot, and make owner_slot a pointer.
For undo log truncation, commit 055a3334ad
repurposed the MLOG_FILE_CREATE2 record with a nonzero page size
to indicate that an undo tablespace will be shrunk in size.
In commit 7ae21b18a6 the
MLOG_FILE_CREATE2 record was replaced by a FILE_CREATE record.
Now that the redo log encoding was changed, there is no actual need
to write a file name in the log record; it suffices to write the
page identifier of the first page that is not part of the file.
This TRIM_PAGES record could allow us to shrink any data files in the
future. For now, it will be limited to undo tablespaces.
mtr_t::log_file_op(): Remove the parameter first_page_no, because
it would always be 0 for file operations.
mtr_t::trim_pages(): Replaces fil_truncate_log().
mtr_t::log_write(): Avoid same_page encoding if !bpage&&!m_last.
fil_op_replay_rename(): Remove the constant parameter first_page_no=0.
This patch adds support of RENAME INDEX operation to the ALTER TABLE
statement. Code which determines if ALTER TABLE can be done in-place
for "simple" storage engines like MyISAM, Heap and etc. was updated to
handle ALTER TABLE ... RENAME INDEX as an in-place operation. Support
for in-place ALTER TABLE ... RENAME INDEX for InnoDB was covered by
MDEV-13301.
Syntax changes
==============
A new type of <alter_specification> is added:
<rename index clause> ::= RENAME ( INDEX | KEY ) <oldname> TO <newname>
Where <oldname> and <newname> are identifiers for old name and new
name of the index.
Semantic changes
================
The result of "ALTER TABLE t1 RENAME INDEX a TO b" is a table which
contents and structure are identical to the old version of 't1' with
the only exception index 'a' being called 'b'.
Neither <oldname> nor <newname> can be "primary". The index being
renamed should exist and its new name should not be occupied
by another index on the same table.
Related to: WL#6555, MDEV-13301
Introduce special synchronization primitive group_commit_lock
for more efficient synchronization of redo log writing and flushing.
The goal is to reduce CPU consumption on log_write_up_to, to reduce
the spurious wakeups, and improve the throughput in write-intensive
benchmarks.
XA specification doesn't permit empty gtrid. It is now enforced by this
patch. This solution was agreed in favour of fixing InnoDB, which doesn't
expect empty XID since early 10.5.
Also fixed wrong assertion (and added a test cases) that didn't permit
64 bytes gtrid + 64 bytes bqual.
Inserting a record into an index page involves updating multiple
fields in the page header as well as updating the next-record links
and potentially updating fields related to the sparse page directory.
Let us cover the insert operations by higher-level log records, to avoid
'redundant' logging about the writes.
The code for applying the high-level log records will check the
consistency of the page thoroughly, to avoid crashes during recovery.
We will refuse to replay the inserts if any inconsistency is detected.
With innodb_force_recovery=1, recovery will continue, but the affected
pages may be more inconsistent if some changes were omitted.
mrec_ext_t: Introduce the EXTENDED record subtypes
INSERT_HEAP_REDUNDANT, INSERT_REUSE_REDUNDANT,
INSERT_HEAP_DYNAMIC, INSERT_REUSE_DYNAMIC.
The record will explicitly identify the page type and whether
the space will be allocated from PAGE_HEAP_TOP or reused from
the PAGE_FREE list. It will also tell how many bytes to copy
from the preceding record header and payload, and how to
initialize the rest of the record header and payload.
mtr_t::page_insert(): Write the high-level log records.
log_phys_t::apply(): Parse the high-level log records.
page_apply_insert_redundant(), page_apply_insert_dynamic():
Apply the high-level log records.
page_dir_split_slot(): Introduce a variant that does not write log
nor deal with ROW_FORMAT=COMPRESSED pages.
page_mem_alloc_heap(): Remove the mtr_t parameter
page_cur_insert_rec_low(): Write log only via mtr_t::page_insert().
This is a follow-up to commit 572d20757b
where we introduced the EXTENDED log record subtypes
DELETE_ROW_FORMAT_REDUNDANT and DELETE_ROW_FORMAT_DYNAMIC.
log_phys_t::apply(): If corruption was noticed, stop applying the log
unless innodb_force_recovery is set.
This is a follow-up to commit 84e3f9ce84
that introduced the EXTENDED log record of UNDO_APPEND subtype.
mtr_t::undo_append(): Accurately enforce the mtr_buf_t::MAX_DATA_SIZE
limit. Also, replace mtr_buf_t::push() with simpler code, to append 1 byte
to the log.
log_phys_t::undo_append(): Return whether the page was found to
be in an inconsistent state.
log_phys_t::apply(): If corruption was noticed, stop applying log
unless innodb_force_recovery is set.
This bug was introduced in
commit 7ae21b18a6
(the main commit of MDEV-12353).
page_cur_insert_rec_low(): Before entering the comparison loop, make sure
that the range does not exceed c_end already at the start of the loop.
The loop is only comparing for pointer equality, and that condition
cdm == c_end would never hold if the end was already exceeded in
the beginning. Also, skip the comparison altogether if we could find
at most 2 equal bytes.
PageBulk::insertPage(): Apply a similar change. It seems that this
code was correct, because the loop checks for cdm < c_end.
mrec_ext_t: Introduce DELETE_ROW_FORMAT_REDUNDANT,
DELETE_ROW_FORMAT_DYNAMIC.
mtr_t::page_delete(): Write DELETE_ROW_FORMAT_REDUNDANT or
DELETE_ROW_FORMAT_DYNAMIC log records. We log the byte offset
of the preceding record, so that on recovery we can easily
find everything to update. For DELETE_ROW_FORMAT_DYNAMIC,
we must also write the header and data size of the record.
We will retain the physical logging for ROW_FORMAT=COMPRESSED pages.
page_zip_dir_balance_slot(): Renamed from page_dir_balance_slot(),
and specialized for ROW_FORMAT=COMPRESSED only.
page_rec_set_n_owned(), page_dir_slot_set_n_owned(),
page_dir_balance_slot(): New variants that do not write any log.
page_mem_free(): Take data_size, extra_size as parameters.
Always zerofill the record payload.
page_cur_delete_rec(): For other than ROW_FORMAT=COMPRESSED,
only write log by mtr_t::page_delete().
There is no reason for the dummy index object dict_ind_redundant
to exist any more. It was only being passed to btr_create().
btr_create(): If !index, assume that a ROW_FORMAT=REDUNDANT
table is being created.
We could pass ibuf.index, dict_sys.sys_tables->indexes.start
and so on, if those objects had been initialized before the
function btr_create() is called.
recv_sys_t opened redo log files along with log_sys_t. That's why I
removed file sharing logic from InnoDB
in 9ef2d29ff4
But it was actually used to ensure that only one MariaDB instance
will touch the same InnoDB files.
os0file.cc: revert some changes done previously
mapped_file_t::map(): now has arguments read_only, nvme
file_io::open(): now has argument read_only
class file_os_io: make final
log_file_t::open(): now has argument read_only
We introduce an EXTENDED log record for appending an undo log record
to an undo log page. This is equivalent to the MLOG_UNDO_INSERT record
that was removed in commit f802c989ec,
only using more compact encoding.
mtr_t::log_write(): Fix a bug that affects longer log
record writes in the !same_page && !have_offset case.
Similar code is already implemented for the have_offset code path.
The bug was unobservable before we started to write longer
EXTENDED records. All !have_offset records (FREE_PAGE, INIT_PAGE,
EXTENDED) that were written so far are short, and we never write
RESERVED or OPTION records.
mtr_t::undo_append(): Write an UNDO_APPEND record.
log_phys_t::undo_append(): Apply an UNDO_APPEND record.
trx_undo_page_set_next_prev_and_add(),
trx_undo_page_report_modify(),
trx_undo_page_report_rename():
Invoke mtr_t::undo_append() instead of emitting WRITE records.
We introduce an EXTENDED log record for initializing an undo log page.
The size of the record will be 2 bytes plus the optional page identifier.
The entire undo page will be initialized, except the space that is
already reserved for TRX_UNDO_SEG_HDR in trx_undo_seg_create().
mtr_t::undo_create(): Write the UNDO_INIT record.
trx_undo_page_init(): Initialize the undo page corresponding to the
UNDO_INIT record. Unlike the former MLOG_UNDO_INIT record, we will
initialize almost the entire page, including initializing the
TRX_UNDO_PAGE_NODE to an empty list node, so that the subsequent call
to flst_init() will avoid writing log for the undo page.
Now there can be only one log file instead of several which
logically work as a single file.
Possible names of redo log files: ib_logfile0,
ib_logfile101 (for just created one)
innodb_log_fiels_in_group: value of this variable is not used
by InnoDB. Possible values are still 1..100, to not break upgrade
LOG_FILE_NAME: add constant of value "ib_logfile0"
LOG_FILE_NAME_PREFIX: add constant of value "ib_logfile"
get_log_file_path(): convenience function that returns full
path of a redo log file
SRV_N_LOG_FILES_MAX: removed
srv_n_log_files: we can't remove this for compatibility reasons,
but now server doesn't use this variable
log_sys_t::file::fd: now just one, not std::vector
log_sys_t::log_capacity: removed word 'group'
find_and_check_log_file(): part of logic from huge srv_start()
moved here
recv_sys_t::files: file descriptors of redo log files.
There can be several of those in case we're upgrading
from older MariaDB version.
recv_sys_t::remove_extra_log_files: whether to remove
ib_logfile{1,2,3...} after successfull upgrade.
recv_sys_t::read(): open if needed and read from one
of several log files
recv_sys_t::files_size(): open if needed and return files count
redo_file_sizes_are_correct(): check that redo log files
sizes are equal. Just to log an error for a user.
Corresponding check was moved from srv0start.cc
namespace deprecated: put all deprecated variables here to
prevent usage of it by us, developers
We plan use the redo log record main type code 0x20 for
InnoDB specific index page operations.
mrec_type_t: Rename INIT_INDEX_PAGE to EXTENDED.
mrec_ext_t: The EXTENDED subtypes.
This is a non-functional change: the redo log record encoding
that was introduced in commit 7ae21b18a6
is not affected.
btr_page_reorganize_low(): Log only the changed data in the page.
TODO: Do not copy the entire changed payload to the redo log.
Emit a combination of MEMMOVE and WRITE records to reduce the log volume.
commit 08ba388713 of MDEV-12353
introduced an incorrect assumption, which was documented by
the failing assertion.
After instant ADD COLUMN, we can have a null (and in-place) UPDATE
of NULL to NULL. No data needs to be written for such updates.
For ROW_FORMAT=REDUNDANT, we reserve space for the NULL values,
and to be compatible with existing behaviour, we will zerofill
the unused data bytes when updating to NULL value.
trx_purge_free_segment(): In some cases (observed when running
the test innodb_zip.wl5522_debug_zip), there is no change to
the TRX_UNDO_NEEDS_PURGE field. Add mtr_t::OPT to disable a debug check.
The bogus debug check was introduced in
commit 56f6dab1d0.
We add FIXME comments and some sketch code for the following cases:
It is possible to write considerably less log for ROW_FORMAT=COMPRESSED
pages. For now, we will delete the records one by one.
It is also possible to treat 'deleting the last records' as a special
case that would involve shrinking PAGE_HEAP_TOP. That should reduce
the need of reorganizing pages.
page_mem_free(): When deleting the very last record of the page,
even if the record did not fully utilize all bytes in a
former PAGE_FREE record, truncate the PAGE_HEAP_TOP and reduce
PAGE_GARBAGE by the saved amount.
fsp_page_create(): Always initialize the page. The logic to
avoid initialization was made redundant and should have been removed
in mysql/mysql-server@ce0a1e85e2
(MySQL 5.7.5).
btr_store_big_rec_extern_fields(): Remove the redundant initialization
of FIL_PAGE_PREV and FIL_PAGE_NEXT. An INIT_PAGE record will have
been written already. Only write the ROW_FORMAT=COMPRESSED page payload
from FIL_PAGE_DATA onwards. We were unnecessarily writing from
FIL_PAGE_TYPE onwards, which caused an assertion failure on recovery:
recv_sys_t::alloc(size_t): Assertion 'len <= srv_page_size' failed
when running the following tests:
./mtr --no-reorder innodb_zip.blob,4k innodb_zip.bug56680,4k
trx_rseg_write_wsrep_checkpoint(): Add missing mtr_t::OPT,
and avoid an unnecessary call to mtr_t::memset().
This addresses a debug assertion failure in wsrep_info.plugin.
page_update_max_trx_id(), page_delete_rec_list_end(): Remove conditions
on recv_recovery_is_on(). These conditions should have been removed in
or before commit f8a9f90667
(removing the support for crash-upgrade).
The physical redo log based recovery will not call such high-level code.
page_mem_free(): When deleting the last record of a page,
do not add it to the PAGE_FREE list, but instead truncate the
PAGE_HEAP_TOP. Modify the page header fields by writing fewer
records.
page_cur_delete_rec(): Let page_mem_free() reset the PAGE_LAST_INSERT.
page_header_reset_last_insert(): Issue memset(), not memcpy(), for
the ROW_FORMAT=COMPRESSED page.
