On several buildbot machines, the test fails like this:
CURRENT_TEST: wsrep.binlog_format
mysqltest: At line 12: query 'SET binlog_format=STATEMENT' failed:
1231: Variable 'binlog_format' can't be set to the value of 'STATEMENT'
ANALYSIS
This is regression caused due to worklog 6742 which
implemented ha_innobase::records() which always
uses clustered index to get the row count. Previously
optimizer chose secondary index which was smaller in
size of clustered index to scan for rows and resulted in
a quicker scan.
FIX
After discussion it was decided to remove this feature in 5.7.
[#rb14040 Approved by Kevin and Oystein ]
- Allow the server to start if innodb force recovery is set to 6
even though change buffer is not empty
Reviewed-by: Jimmy Yang <jimmy.yang@oracle.com>
Problem:
========
- Drop table assert if innodb_force_recovery is set to 5 or 6.
For innodb_force_recovery 5 and 6, InnoDB doesn't scan the undo log
and it makes the redo rollback segment as NULL. There is no way for
transaction to write any undo log.
- If innodb_force_recovery is set to 6 then InnoDB does not do the
redo log roll-forward in connection with recovery. In this case,
log_sys will be initalized only and it will not have latest
checkpoint information. Checkpoint is done during shutdown even
innodb_force_recovery is set to 6. So it leads to incorrect
information update in checkpoint header.
Solution:
========
1) Allow drop table only if innodb_force_recovery < 5.
2) Make innodb as read-only if innodb_force_recovery is set to 6.
3) During shutdown, remove the checkpoint if innodb_force_recovery
is set to 6.
Reviewed-by: Jimmy Yang <jimmy.yang@oracle.com>
RB: 15075
Problem:
========
During checkpoint, we are writing all MLOG_FILE_NAME records in one mtr
and parse buffer can't be processed till MLOG_MULTI_REC_END. Eventually parse
buffer exceeds the RECV_PARSING_BUF_SIZE and eventually it overflows.
Fix:
===
1) Break the large mtr if it exceeds LOG_CHECKPOINT_FREE_PER_THREAD into multiple mtr during checkpoint.
2) Move the parsing buffer if we are encountering only MLOG_FILE_NAME
records. So that it will never exceed the RECV_PARSING_BUF_SIZE.
Reviewed-by: Debarun Bannerjee <debarun.bannerjee@oracle.com>
Reviewed-by: Rahul M Malik <rahul.m.malik@oracle.com>
RB: 14743
Description:
===========
Add my_thread_init() and my_thread_exit() for background threads which
initializes and frees the st_my_thread_var structure.
Reviewed-by: Jimmy Yang<jimmy.yang@oracle.com>
RB: 15003
Problem:
During read head, wrong page size is used to calcuate the tablespace size.
Fix:
Use physical page size to calculate tablespace size
Reveiwed-By: Satya Bodapati
RB: 14993
Split the test case so that a server restart is not needed.
Reduce the test cases and use a simpler mechanism for triggering
and waiting for purge.
fil_table_accessible(): Check if a table can be accessed without
enjoying MDL protection.
PROBLEM
When truncating single tablespace tables, we need to scan the entire
buffer pool to remove the pages of the table from the buffer pool.
During this scan and removal dict_sys->mutex is being held ,causing
stalls in other DDL operations.
FIX
Release the dict_sys->mutex during the scan and reacquire it after the
scan. Make sure that purge thread doesn't purge the records of the table
being truncated and background stats collection thread skips the updation
of stats for the table being truncated.
[#rb 14564 Approved by Jimmy and satya ]
srv_sys_t::n_threads_active[]: Protect writes by both the mutex and
by atomic memory access.
srv_active_wake_master_thread_low(): Reliably wake up the master
thread if there is work to do. The trick is to atomically read
srv_sys->n_threads_active[].
srv_wake_purge_thread_if_not_active(): Atomically read
srv_sys->n_threads_active[] (and trx_sys->rseg_history_len),
so that the purge should always be triggered when there is work to do.
trx_commit_in_memory(): Invoke srv_wake_purge_thread_if_not_active()
whenever a transaction is committed. Purge could have been prevented by
the read view of the currently committing transaction, even if it is
a read-only transaction.
trx_purge_add_update_undo_to_history(): Do not wake up the purge.
This is only called by trx_undo_update_cleanup(), as part of
trx_write_serialisation_history(), which in turn is only called by
trx_commit_low() which will always call trx_commit_in_memory().
Thus, the added call in trx_commit_in_memory() will cover also
this use case where a committing read-write transaction added
some update_undo log to the purge queue.
trx_rseg_mem_restore(): Atomically modify trx_sys->rseg_history_len.
Problem:
=======
Concurrent update dml statement doesn't reflect in virtual index during
inplace table rebuild. It results mismatch value in virutal index and
clustered index. Deleting the table content tries to search the mismatch
value in virtual index but it can't find the value. During log update
apply phase, virtual information is being ignored while constructing
the new entry.
Solution:
=========
In row_log_update_apply phase, build the entry with virtual column
information. So that it can reflect in newly constructed virtual index.
Reviewed-by: Jimmy Yang<jimmy.yang@oracle.com>
RB: 14974
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>
Issue:
======
The issue is that if a fts index is present in a table the space size is
incorrectly calculated in the case of truncate which results in a invalid
read.
Fix:
====
Have a different space size calculation in truncate if fts indexes are
present.
RB:14755
Reviewed-by: Shaohua Wang <shaohua.wang@oracle.com>
Analysis:
========
There was missing bracket for IF conditon in dict_stats_analyze_index_level()
and it leads to wrong result.
Fix:
====
Fix the IF condition in dict_stats_analyze_index_level() so that it satisfied
the if condtion only if level is zero.
Reviewed-by : Jimmy Yang <jimmy.yang@oracle.com>
Analysis:
========
Field name comparison happens while filling the virtual columns
affected by foreign constraint. But field name is NULL in virtual
index for the newly added virtual column.
Fix:
===
Ignore the index if it has newly added virtual column. Foreign
key affected virtual column information is filled during
loading operation.
Reviewed-by: Jimmy Yang <jimmy.yang@oracle.com>
RB: 14895
Problem :
---------
Information_Schema.referential_constraints (UNIQUE_CONSTRAINT_NAME)
shows NULL for a foreign key constraint after restarting the server.
If any dml or query (select/insert/update/delete) is done on
referenced table, then the constraint name is correctly shown.
Solution :
----------
UNIQUE_CONSTRAINT_NAME column is the key name of the referenced table.
In innodb, FK reference is stored as a list of columns in referenced
table in INNODB_SYS_FOREIGN and INNODB_SYS_FOREIGN_COLS. The referenced
column must have at least one index/key with the referenced column as
prefix but the key name itself is not included in FK metadata. For this
reason, the UNIQUE_CONSTRAINT_NAME is only filled up when the
referenced table is actually loaded in innodb dictionary cache.
The information_schema view calls handler::get_foreign_key_list() on
foreign key table to read the FK metadata. The UNIQUE_CONSTRAINT_NAME
information shows NULL based on whether the referenced table is
already loaded or not.
One way to fix this issue is to load the referenced table while reading
the FK metadata information, if needed.
Reviewed-by: Sunny Bains <sunny.bains@oracle.com>
RB: 14654
Problem :
---------
This bug is filed from the base replication bug#25040331 where the
slave thread times out while INSERT operation waits on GAP lock taken
during Foreign Key validation.
The primary reason for the lock wait is because the statements are
getting replayed in different order. However, we also observed
two things ...
1. The slave thread could always use "Read Committed" isolation for
row level replication.
2. It is not necessary to have GAP locks in "READ Committed" isolation
level in innodb.
This bug is filed to address point(2) to avoid taking GAP locks during
Foreign Key validation.
Solution :
----------
Innodb is primarily designed for "Repeatable Read" and the GAP lock
behaviour is default. For "Read Committed" isolation, we have special
handling in row_search_mvcc to avoid taking the GAP lock while
scanning records.
While looking for Foreign Key, the code is following the default
behaviour taking GAP locks. The suggested fix is to avoid GAP
locking during FK validation similar to normal search operation
(row_search_mvcc) for "Read Committed" isolation level.
Reviewed-by: Sunny Bains <sunny.bains@oracle.com>
RB: 14526
Problem :
---------
1. delete_all_rows() and rnd_init() are not returning error
after async rollback in 5.7. This results in assert in
innodb in next call.
2. High priority transaction is rolling back prepared transaction.
This is because TRX_FORCE_ROLLBACK_DISABLE is getting set only for
first entry [TrxInInnoDB].
Solution :
----------
1. return DB_FORCED_ABORT error after rollback.
2. check and disable rollback in TrxInInnodb::enter always.
Reviewed-by: Sunny Bains <sunny.bains@oracle.com>
RB: 13777
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.
Prevent GCC from moving a mach_read_from_4() before we have checked that
we have 4 bytes to read. The pointer may only point to a 1, 2 or 3
bytes in which case the code should not read 4 bytes. This is a
workaround to a GCC bug:
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77673
Patch submitted by: Laurynas Biveinis <laurynas.biveinis@gmail.com>
RB: 14135
Reviewed by: Pawel Olchawa <pawel.olchawa@oracle.com>
Description
===========
Under heavy load, the aysnchronous Windows file IO API can return a
failure code that is handled in MySQL Server by retrying the file IO operation.
A cast necessary for the correct operation of the retry path in a 64 bit
build is missing, leading to the file IO retry result being misinterpreted
and ultimately the report of the OS error number 995
(ERROR_OPERATION_ABORTED) in the MySQL error log.
Fix
===
Supply the missing cast.
Reviewed-by: Sunny Bains <sunny.bains@oracle.com>
RB: 14109
Avoid detaching and exiting from threads that may finish before the
caller has returned from pthread_create(). Only exit from such threads,
without detach and join with them later.
Patch submitted by: Laurynas Biveinis <laurynas.biveinis@gmail.com>
RB: 13983
Reviewed by: Sunny Bains <sunny.bains@oracle.com>
Simplify the tests that are present in MySQL 5.7. Make the table
smaller while generating enough undo log. Do not unnecessarily
drop tables.
trx_purge_initiate_truncate(): Remove two crash injection points
(before and after normal redo log checkpoint), because they are
not adding any value. Clarify some messages.
trx_sys_create_rsegs(): Display the number of active undo tablespaces.
srv_undo_tablespaces_init(): When initializing the data files, do not
leave srv_undo_tablespaces_active at 0.
Do not display that number; let trx_sys_create_rsegs() display it once
the final number is known.
innodb_params_adjust(): Adjust parameters after startup.
innobase_init(): Do not allow innodb_max_undo_size to be less
than SRV_UNDO_TABLESPACE_SIZE_IN_PAGES. This avoids unnecessary
repeated truncation of undo tablespaces when using
innodb_page_size=32k or innodb_page_size=64k.
Problem:
========
If we increase innodb_undo_logs value during startup. New rollback
segment might get allocated during trx_sys_create_rseg(). Essentially, it
would make these tablesapces active with transaction undo data and purge.
But it doesn't come in active undo tablespaces. so these tablespace would
never get truncated.
Fix:
===
Increase the number of active undo tablespace when we are assigning the
undo tablespace to the newly assigned rollback segment.
Reviewed-by: Kevin Lewis <kevin.lewis@oracle.com>
Reviewed-by: Debarun Banerjee <debarun.banerjee@oracle.com>
RB: 13746
MySQL 5.7 reduced the maximum number of innodb_undo_tablespaces
from 126 to 95 when it reserved 32 persistent rollback segments
for the temporary undo logs. Since MDEV-12289 restored all 128
persistent rollback segments for persistent undo logs, the
reasonable maximum value of innodb_undo_tablespaces is 127
(not 126 or 95). This is because out of the 128 rollback segments,
the first one will always be created in the system tablespace
and the remaining ones can be created in dedicated undo tablespaces.
Silence the Valgrind warnings on instrumented builds (-DWITH_VALGRIND).
assert_block_ahi_empty_on_init(): A variant of
assert_block_ahi_empty() that declares n_pointers initialized and then
asserts that n_pointers==0.
In Valgrind-instrumented builds, InnoDB declares allocated memory
uninitialized.
btr_search_drop_page_hash_index(): Do not return before ensuring
that block->index=NULL, even if !btr_search_enabled. We would
typically still skip acquiring the AHI latch when the AHI is
disabled, because block->index would already be NULL. Only if the AHI
is in the process of being disabled, we would wait for the AHI latch
and then notice that block->index=NULL and return.
The above bug was a regression caused in MySQL 5.7.9 by the fix of
Bug#21407023: DISABLING AHI SHOULD AVOID TAKING AHI LATCH
The rest of this patch improves diagnostics by adding assertions.
assert_block_ahi_valid(): A debug predicate for checking that
block->n_pointers!=0 implies block->index!=NULL.
assert_block_ahi_empty(): A debug predicate for checking that
block->n_pointers==0.
buf_block_init(): Instead of assigning block->n_pointers=0,
assert_block_ahi_empty(block).
buf_pool_clear_hash_index(): Clarify comments, and assign
block->n_pointers=0 before assigning block->index=NULL.
The wrong ordering could make block->n_pointers appear incorrect in
debug assertions. This bug was introduced in MySQL 5.1.52 by
Bug#13006367 62487: INNODB TAKES 3 MINUTES TO CLEAN UP THE
ADAPTIVE HASH INDEX AT SHUTDOWN
i_s_innodb_buffer_page_get_info(): Add a comment that
the IS_HASHED column in the INFORMATION_SCHEMA views
INNODB_BUFFER_POOL_PAGE and INNODB_BUFFER_PAGE_LRU may
show false positives (there may be no pointers after all.)
ha_insert_for_fold_func(), ha_delete_hash_node(),
ha_search_and_update_if_found_func(): Use atomics for
updating buf_block_t::n_pointers. While buf_block_t::index is
always protected by btr_search_x_lock(index), in
ha_insert_for_fold_func() the n_pointers-- may belong to
another dict_index_t whose btr_search_latches[] we are not holding.
RB: 13879
Reviewed-by: Jimmy Yang <jimmy.yang@oracle.com>
Analysis:
========
A foreign key constraint cannot reference a secondary index defined
on a generated virtual column. While adding new index/drop existing
column, server internally drops the internal foreign key index and
it leads to choose the virtual secondary index as foreign key index.
But innodb doesn't allow foreign key constraint reference to
secondary virtual index.
Fix:
===
Allow foreign key constraint refer to secondary index defined on
a generated virutal column.
Reviewed-by: Jimmy Yang<jimmy.yang@oracle.com>
RB: 13586
This bug was introduced in MySQL 5.6.8 with WL#6255.
When an error occurs while rebuilding a table that only has a
hidden GEN_CLUST_INDEX inside InnoDB, ha_alter_info->key_info_buffer
would be invalid and should not be dereferenced.
get_error_key_name(): Get the name of an erroneous key.
Avoid dereferencing ha_alter_info->key_info_buffer when no keys
exist in the SQL layer.
ha_innobase::inplace_alter_table(),
ha_innobase::commit_try_rebuild(): Invoke get_error_key_name()
for reporting ER_INNODB_ONLINE_LOG_TOO_BIG or ER_INDEX_CORRUPT.
RB: 13834
Reviewed-by: Jimmy Yang <jimmy.yang@oracle.com>
When the rows produced on the current iteration are sent to the
temporary table T of the UNION type created for CTE the rows
that were not there simultaneously are sent to the temporary
table D that contains rows for the next iteration. The test
whether a row was in T checks the return code of writing into T.
If just a HEAP table is used for T then the return code is
HA_ERR_FOUND_DUPP_KEY, but if an ARIA table is used for T then
the return code is HA_ERR_FOUND_DUPP_UNIQUE.
The implementation of select_union_recursive::send_data()
erroneously checked only for the first return code. So if an Aria
table was used for T then all rows produced by the current iteration
went to D and and in most cases D grew with each iteration.
Whether T has reached stabilization is detected by
checking whether D is empty. So as a result, the iterations were
never stopped unless a limit for them was set.
Fixed by checking for both HA_ERR_FOUND_DUPP_KEY and
HA_ERR_FOUND_DUPP_UNIQUE as return codes returned by
the function writing a row into the temporary table T.
When WSREP(thd) is not true we will use my_error(...) to print error. This
will set thd->is_error() to true and we wont be getting generic error.
Signed-off-by: Sachin Setiya <sachin.setiya@mariadb.com>
In the current code temporary tables we identified and opened before
other tables. CTE tables are identified in the same procedure as
regular tables. When a temporary table and a CTE table have the same
name T any reference to T that is in the scope of the CTE declaration
must be associated with this CTE. Yet it was not done properly.
When a reference to T was found in the scope of the declaration
of CTE T a pointer to this CTE was set in the reference. No check
that the reference had been already associated with a temporary table
was done. As a result, if the temporary table T had been created then
the reference to T was considered simultaneously as reference to the CTE
named T and as a reference to the temporary table named T. This
confused the code that were executed later and caused a crash of
the server.
Now when a table reference is associated with a CTE any previous
association with a temporary table is dropped.
This problem could be easily avoided if the temporary tables were
not identified prematurely.
as reference to CTE named T and
Here's what started happening after the patch that fixed
the bug mdev-10454 with query reported for the bug
SELECT * FROM t t1 right JOIN t t2 ON (t2.pk = t1.pk)
WHERE (t2.i, t2.pk) NOT IN ( SELECT t3.i, t3.i FROM t t3, t t4 )
AND t1.c = 'foo';
The patch added an implementation of propagate_equal_fields() for
the class Item_row and thus opened the possibility of equal fields
substitutions.
At the prepare stage after setup_conds() called for WHERE condition
had completed the flag of maybe_null of the Item_row object created
for (t2.i, t2.pk) was set to false, because the maybe_null flags of
both elements were set to false. However the flag of maybe_null for
t1.pk from the ON condition were set to true, because t1 was an inner
table of an outer join.
At the optimization stage the outer join was converted to inner join,
but the maybe_null flags were not corrected and remained the same.
So after the substitution t2.pk/t1.pk. the maybe_null flag for the
row remained false while the maybe_flag for the second element of
the row was true. As a result, when the in-to_exists transformation
was performed for the NOT IN predicate the guards variables were
not created for the elements of the row, but a guard object for
the second element was created. The object were not valid because
it referred to NULL as a guard variable. This ultimately caused
a crash when the expression with the guard was evaluated at the
execution stage.
The patch made sure that the guard objects are not created without
guard variables.
Yet it does not resolve the problem of inconsistent maybe_null flags.
and it might be that the problem will pop op in other pieces of code.
The resolution of this problem is not easy, but the problem should
be resolved in future versions.
