The motivation of introducing the parameter
innodb_purge_rseg_truncate_frequency in
mysql/mysql-server@28bbd66ea5 and
mysql/mysql-server@8fc2120fed
seems to have been to avoid stalls due to freeing undo log pages
or truncating undo log tablespaces. In MariaDB Server,
innodb_undo_log_truncate=ON should be a much lighter operation
than in MySQL, because it will not involve any log checkpoint.
Another source of performance stalls should be
trx_purge_truncate_rseg_history(), which is shrinking the history list
by freeing the undo log pages whose undo records have been purged.
To alleviate that, we will introduce a purge_truncation_task that will
offload this from the purge_coordinator_task. In that way, the next
innodb_purge_batch_size pages may be parsed and purged while the pages
from the previous batch are being freed and the history list being shrunk.
The processing of innodb_undo_log_truncate=ON will still remain the
responsibility of the purge_coordinator_task.
purge_coordinator_state::count: Remove. We will ignore
innodb_purge_rseg_truncate_frequency, and act as if it had been
set to 1 (the maximum shrinking frequency).
purge_coordinator_state::do_purge(): Invoke an asynchronous task
purge_truncation_callback() to free the undo log pages.
purge_sys_t::iterator::free_history(): Free those undo log pages
that have been processed. This used to be a part of
trx_purge_truncate_history().
purge_sys_t::clone_end_view(): Take a new value of purge_sys.head
as a parameter, so that it will be updated while holding exclusive
purge_sys.latch. This is needed for race-free access to the field
in purge_truncation_callback().
Reviewed by: Vladislav Lesin
purge_sys_t::sees(): Wrapper for view.sees().
trx_purge_truncate_history(): Invoke purge_sys.sees() instead of
comparing to head.trx_no, to determine if undo pages can be safely freed.
The test innodb.cursor-restore-locking was adjusted by Vladislav Lesin,
as was the the debug instrumentation in row_purge_del_mark().
Reviewed by: Vladislav Lesin
It is not safe to invoke trx_purge_free_segment() or execute
innodb_undo_log_truncate=ON before all undo log records in
the rollback segment has been processed.
A prominent failure that would occur due to premature freeing of
undo log pages is that trx_undo_get_undo_rec() would crash when
trying to copy an undo log record to fetch the previous version
of a record.
If trx_undo_get_undo_rec() was not invoked in the unlucky time frame,
then the symptom would be that some committed transaction history is
never removed. This would be detected by CHECK TABLE...EXTENDED that
was impleented in commit ab0190101b0587e0e03b2d75a967050b9a85fd1b.
Such a garbage collection leak should be possible even when using
innodb_undo_log_truncate=OFF, just involving trx_purge_free_segment().
trx_rseg_t::needs_purge: Change the type from Boolean to a transaction
identifier, noting the most recent non-purged transaction, or 0 if
everything has been purged. On transaction start, we initialize this
to 1 more than the transaction start ID. On recovery, the field may be
adjusted to the transaction end ID (TRX_UNDO_TRX_NO) if it is larger.
The field TRX_UNDO_NEEDS_PURGE becomes write-only; only some debug
assertions that would validate the value. The field reflects the old
inaccurate Boolean field trx_rseg_t::needs_purge.
trx_undo_mem_create_at_db_start(), trx_undo_lists_init(),
trx_rseg_mem_restore(): Remove the parameter max_trx_id.
Instead, store the maximum in trx_rseg_t::needs_purge,
where trx_rseg_array_init() will find it.
trx_purge_free_segment(): Contiguously hold a lock on
trx_rseg_t to prevent any concurrent allocation of undo log.
trx_purge_truncate_rseg_history(): Only invoke trx_purge_free_segment()
if the rollback segment is empty and there are no pending transactions
associated with it.
trx_purge_truncate_history(): Only proceed with innodb_undo_log_truncate=ON
if trx_rseg_t::needs_purge indicates that all history has been purged.
Tested by: Matthias Leich
Before the fix next-key lock was requested only if a record was
delete-marked for locking unique search in RR isolation level.
There can be several delete-marked records for the same unique key,
that's why InnoDB scans the records until eighter non-delete-marked record
is reached or all delete-marked records with the same unique key are
scanned.
For range scan next-key locks are used for RR to protect scanned range from
inserting new records by other transactions. And this is the reason of why
next-key locks are used for delete-marked records for unique searches.
If a record is not delete-marked, the requested lock type was "not-gap".
When a record is not delete-marked during lock request by trx 1, and
some other transaction holds conflicting lock, trx 1 creates waiting
not-gap lock on the record and suspends. During trx 1 suspending the
record can be delete-marked. And when the lock is granted on conflicting
transaction commit or rollback, its type is still "not-gap". So we have
"not-gap" lock on delete-marked record for RR. And this let some other
transaction to insert some record with the same unique key when trx 1 is
not committed, what can cause isolation level violation.
The fix is to set next-key locks for both delete-marked and
non-delete-marked records for unique search in RR.
The issue is that trx_t::lock.was_chosen_as_deadlock_victim can be reset
before the transaction check it and set trx_t::error_state.
The fix is to reset trx_t::lock.was_chosen_as_deadlock_victim only in
trx_t::commit_in_memory(), which is invoked on full rollback. There is
also no need to have separate bit in
trx_t::lock.was_chosen_as_deadlock_victim to flag transaction it was
chosen as a victim of Galera conflict resolution, the same variable can be
used for both cases except debug build. For debug build we need to
distinguish deadlock and Galera's abort victims for debug checks. Also
there is no need to check for deadlock in lock_table_enqueue_waiting() for
Galera as the coresponding check presents in lock_wait().
Local variable "error_state" in lock_wait() was replaced with
trx->error_state, because before the replace
lock_sys_t::cancel<false>(trx, lock) and lock_sys.deadlock_check() could
change trx->error_state, which then could be overwritten with the local
"error_state" variable value.
The lock_wait_suspend_thread_enter DEBUG_SYNC point name is misleading,
because lock_wait_suspend_thread was eliminated in e71e613. It was renamed
to lock_wait_start.
Reviewed by: Marko Mäkelä, Jan Lindström.
Backported from 10.5 20e9e804c131c6522bc7c469e4863e8d1eaa3ee0 and
5948d7602ec7f61937c368dcb134e6ec226a2990.
sel_restore_position_for_mysql() moves forward persistent cursor
position after btr_pcur_restore_position() call if cursor relative position
is BTR_PCUR_ON and the cursor points to the record with NOT the same field
values as in a stored record(and some other not important for this case
conditions).
It was done because btr_pcur_restore_position() sets
page_cur_mode_t mode to PAGE_CUR_LE for cursor->rel_pos == BTR_PCUR_ON
before opening cursor. So we are searching for the record less or equal
to stored one. And if the found record is not equal to stored one, then
it is less and we need to move cursor forward.
But there can be a situation when the stored record was purged, but the
new one with the same key but different value was inserted while
row_search_mvcc() was suspended. In this case, when the thread is
awaken, it will invoke sel_restore_position_for_mysql(), which, in turns,
invoke btr_pcur_restore_position(), which will return false because found
record don't match stored record, and
sel_restore_position_for_mysql() will move forward cursor position.
The above can lead to the case when awaken row_search_mvcc() do not see
records inserted by other transactions while it slept. The mtr test case
shows the example how it can be.
The fix is to return special value from persistent cursor restoring
function which would notify its caller that uniq fields of restored
record and stored record are the same, and in this case
sel_restore_position_for_mysql() don't move cursor forward.
Delete-marked records are correctly processed in row_search_mvcc().
Non-unique secondary indexes are "uniquified" by adding the PK, the
index->n_uniq should then be index->n_fields. So there is no need in
additional checks in the fix.
If transaction's readview can't see the changes made in secondary index
record, it requests clustered index record in row_search_mvcc() to check
its transaction id and get the correspondent record version. After this
row_search_mvcc() commits mtr to preserve clustered index latching
order, and starts mtr. Between those mtr commit and start secondary
index pages are unlatched, and purge has the ability to remove stored in
the cursor record, what causes rows duplication in result set for
non-locking reads, as cursor position is restored to the previously
visited record.
To solve this the changes are just switched off for non-locking reads,
it's quite simple solution, besides the changes don't make sense for
non-locking reads.
The more complex and effective from performance perspective solution is
to create mtr savepoint before clustered record requesting and rolling
back to that savepoint after that. See MDEV-27557.
One more solution is to have per-record transaction id for secondary
indexes. See MDEV-17598.
If any of those is implemented, just remove select_lock_type argument in
sel_restore_position_for_mysql().
sel_restore_position_for_mysql() moves forward persistent cursor
position after btr_pcur_restore_position() call if cursor relative position
is BTR_PCUR_ON and the cursor points to the record with NOT the same field
values as in a stored record(and some other not important for this case
conditions).
It was done because btr_pcur_restore_position() sets
page_cur_mode_t mode to PAGE_CUR_LE for cursor->rel_pos == BTR_PCUR_ON
before opening cursor. So we are searching for the record less or equal
to stored one. And if the found record is not equal to stored one, then
it is less and we need to move cursor forward.
But there can be a situation when the stored record was purged, but the
new one with the same key but different value was inserted while
row_search_mvcc() was suspended. In this case, when the thread is
awaken, it will invoke sel_restore_position_for_mysql(), which, in turns,
invoke btr_pcur_restore_position(), which will return false because found
record don't match stored record, and
sel_restore_position_for_mysql() will move forward cursor position.
The above can lead to the case when awaken row_search_mvcc() do not see
records inserted by other transactions while it slept. The mtr test case
shows the example how it can be.
The fix is to return special value from persistent cursor restoring
function which would notify its caller that uniq fields of restored
record and stored record are the same, and in this case
sel_restore_position_for_mysql() don't move cursor forward.
Delete-marked records are correctly processed in row_search_mvcc().
Non-unique secondary indexes are "uniquified" by adding the PK, the
index->n_uniq should then be index->n_fields. So there is no need in
additional checks in the fix.
If transaction's readview can't see the changes made in secondary index
record, it requests clustered index record in row_search_mvcc() to check
its transaction id and get the correspondent record version. After this
row_search_mvcc() commits mtr to preserve clustered index latching
order, and starts mtr. Between those mtr commit and start secondary
index pages are unlatched, and purge has the ability to remove stored in
the cursor record, what causes rows duplication in result set for
non-locking reads, as cursor position is restored to the previously
visited record.
To solve this the changes are just switched off for non-locking reads,
it's quite simple solution, besides the changes don't make sense for
non-locking reads.
The more complex and effective from performance perspective solution is
to create mtr savepoint before clustered record requesting and rolling
back to that savepoint after that. See MDEV-27557.
One more solution is to have per-record transaction id for secondary
indexes. See MDEV-17598.
If any of those is implemented, just remove select_lock_type argument in
sel_restore_position_for_mysql().
