SECONDARY INDEX IN INNODB
The patches for Bug#11751388 and Bug#11784056 enabled concurrent
reads while creating secondary indexes in InnoDB. However, they
introduced a regression. This regression occured if ALTER TABLE
failed after the index had been added, for example during the
lock upgrade needed to update .FRM. If this happened, InnoDB
and the server got out of sync with regards to which indexes
actually existed. Therefore the patch for Bug#11815600 again
disabled concurrent reads.
This patch re-enables concurrent reads. The original regression
is fixed by splitting the ADD INDEX operation into two parts.
First the new index is created but not made active. This is
done while concurrent reads are allowed. The second part of
the operation makes the index active (or reverts the change).
This is done after lock upgrade, which prevents the original
regression.
In order to implement this change, the patch changes the storage
API for in-place index creation. handler::add_index() is split
into two functions, handler_add_index() and
handler::final_add_index(). The former for creating indexes without
making them visible and the latter for commiting (i.e. making
visible) new indexes or reverting the changes.
Large parts of this patch were written by Marko Mäkelä.
Test case added to innodb_mysql_lock.test.
With this change, the index prefix column length lifted from 767 bytes
to 3072 bytes if "innodb_large_prefix" is set to "true".
rb://603 approved by Marko
mtr_start(): Declare the mtr memory area uninitialized in Valgrind
before initializing the fields.
mtr_commit(): Declare everything uninitialized except
mtr->start_lsn, mtr->end_lsn and mtr->state.
Fix a deadlock in the initial patch. lock_validate() must not hold the
lock system mutex while s-latching a block, because some functions,
such as lock_rec_convert_impl_to_expl(), may be already holding an x-latch
on the block that lock_validate() is interested in while attempting to
acquire the lock system mutex.
This deadlock was not caught by UNIV_SYNC_DEBUG because of
buf_block_dbg_add_level(block, SYNC_NO_ORDER_CHECK).
lock_clust_rec_some_has_impl(), row_get_rec_trx_id(),
lock_rec_queue_validate(), lock_table_other_has_incompatible(),
lock_table_has_to_wait_in_queue(), lock_table_queue_validate():
Add const qualifiers.
row_get_trx_id_offset(): Add const qualifiers. Keep the parameter rec
only in UNIV_DEBUG builds. Inline the function.
lock_rec_validate_page(): Take the buffer block as a parameter, to
avoid a buf_page_get_gen() call in most cases.
lock_rec_validate_page_low(): A version of lock_rec_validate_page()
that assumes that the lock system mutexes are already being held.
lock_rec_get_next_on_page_const(): A const variant of
lock_rec_get_next_on_page().
lock_validate(): Do not release the lock system mutex while
buffer-fixing the block for the lock_rec_validate_page() call.
Releasing the mutex apparently caused the assertion failure.
rb:665 approved by Sunny Bains
------------------------------------------------------------
revno 2876.244.305
revision id marko.makela@oracle.com-20110413082211-e6ouhjz5rmqxcqap
parent marko.makela@oracle.com-20110413075948-kvytmc37ye1nt7d9
committer Marko Mäkelä <marko.makela@oracle.com>
branch nick 5.6-innodb
timestamp Wed 2011-04-13 11:22:11 +0300
message:
Suppress the Bug #58815 (Bug #11765812) assertion failure.
buf_page_get_gen(): Introduce BUF_GET_POSSIBLY_FREED for suppressing the
check that the file page must not have been freed.
btr_estimate_n_rows_in_range_on_level(): Pass BUF_GET_POSSIBLY_FREED and
explain in the comments why this is needed and why it should be mostly
harmless to ignore the problem. If InnoDB had always initialized all
unused fields in data files, no problem would exist.
This change does not fix the bug, it just "shoots the messenger".
rb:647 approved by Jimmy Yang
or Oracle Bug #11766513.
trx_undo_free_prepared(): Do not acquire or release trx->rseg->mutex.
This code is invoked in the single-threaded part of shutdown, therefore
a mutex is not needed.
node->modification_counter == node->flush_counter
This bug (an assertion failure during a crash-like shutdown)
was introduced by the fix of
Bug#12323643 Clean up the InnoDB thread shutdown and assertions (WL#5136).
fil_node_close_file(): Relax the failing assertion
when innodb_fast_shutdown=2.
large prefix (>=768). Table with such large prefix index will not
be loaded into memory (for its metadata), unless innodb_force_recovery
is on.
rb://604 Approved by Marko
The innoDB global variable srv_lower_case_table_names is set to the value of lower_case_table_names declared in mysqld.h server in ha_innodb.cc. Since this variable can change at runtime, it is reset for each handler call to ::create, ::open, ::rename_table & ::delete_table.
But it is possible for tables to be implicitly opened before an explicit handler call is made when an engine is first started or restarted. I was able to reproduce that with the testcase in this patch on a version of InnoDB from 2 weeks ago. It seemed like the change buffer entries for the secondary key was getting put into pages after the restart. (But I am not sure, I did not write down the call stack while it was reproducing.) In the current code, the implicit open, which is actually a call to dict_load_foreigns(), does not occur with this testcase.
The change is to replace srv_lower_case_table_names by an interface function in innodb.cc that retrieves the server global variable when it is needed.
SRV_CONC_FORCE_EXIT_INNODB
This is a bogus UNIV_SYNC_DEBUG assertion failure that I introduced
when introducing assertions for checking that InnoDB is not holding
any mutexes or rw-locks when returning control to MySQL.
srv_suspend_mysql_thread(): Release dict_operation_lock before
invoking srv_conc_force_exit_innodb(), which would now check that the
thread is not holding any mutexes or rw-locks. After resuming, check
sync_thread_levels_nonempty_trx() and do srv_conc_force_enter_innodb()
before reacquiring the dict_operation_lock.
rb:646 approved by Sunny Bains
In ha_innobase::create(), we check some things while holding an
exclusive lock on the data dictionary. Defer the locking and the
creation of transactions until after the checks have passed. The
THDVAR could hang due to a mutex wait (see Bug #11750569 - 41163:
deadlock in mysqld: LOCK_global_system_variables and LOCK_open), and
we want to avoid waiting while holding InnoDB mutexes.
innobase_index_name_is_reserved(): Replace the parameter trx_t with
THD, so that the test can be performed before starting an InnoDB
transaction. We only needed trx->mysql_thd.
ha_innobase::create(): Create transaction and lock the data dictionary
only after passing the basic tests.
create_table_def(): Move the IS_MAGIC_TABLE_AND_USER_DENIED_ACCESS
check to ha_innobase::create(). Assign to srv_lower_case_table_names
while holding dict_sys->mutex.
ha_innobase::delete_table(), ha_innobase::rename_table(),
innobase_rename_table(): Assign srv_lower_case_table_names as late as
possible. Here, the variable is not necessarily protected by
dict_sys->mutex.
ha_innobase::add_index(): Invoke innobase_index_name_is_reserved() and
innobase_check_index_keys() before allocating anything.
rb:618 approved by Jimmy Yang
causes future shutdown hang
InnoDB would hang on shutdown if any XA transactions exist in the
system in the PREPARED state. This has been masked by the fact that
MySQL would roll back any PREPARED transaction on shutdown, in the
spirit of Bug #12161 Xa recovery and client disconnection.
[mysql-test-run] do_shutdown_server: Interpret --shutdown_server 0 as
a request to kill the server immediately without initiating a
shutdown procedure.
xid_cache_insert(): Initialize XID_STATE::rm_error in order to avoid a
bogus error message on XA ROLLBACK of a recovered PREPARED transaction.
innobase_commit_by_xid(), innobase_rollback_by_xid(): Free the InnoDB
transaction object after rolling back a PREPARED transaction.
trx_get_trx_by_xid(): Only consider transactions whose
trx->is_prepared flag is set. The MySQL layer seems to prevent
attempts to roll back connected transactions that are in the PREPARED
state from another connection, but it is better to play it safe. The
is_prepared flag was introduced in the InnoDB Plugin.
trx_n_prepared: A new counter, counting the number of InnoDB
transactions in the PREPARED state.
logs_empty_and_mark_files_at_shutdown(): On shutdown, allow
trx_n_prepared transactions to exist in the system.
trx_undo_free_prepared(), trx_free_prepared(): New functions, to free
the memory objects of PREPARED transactions on shutdown. This is not
needed in the built-in InnoDB, because it would collect all allocated
memory on shutdown. The InnoDB Plugin needs this because of
innodb_use_sys_malloc.
trx_sys_close(): Invoke trx_free_prepared() on all remaining
transactions.
On shutdown, do not exit threads in os_event_wait(). This method of
exiting was only used by the I/O handler threads. Exit them on a
higher level.
os_event_wait_low(), os_event_wait_time_low(): Do not exit on shutdown.
os_thread_exit(), ut_dbg_assertion_failed(), ut_print_timestamp(): Add
attribute cold, so that GCC knows that these functions are rarely
invoked and can be optimized for size.
os_aio_linux_collect(): Return on shutdown.
os_aio_linux_handle(), os_aio_simulated_handle(), os_aio_windows_handle():
Set *message1 = *message2 = NULL and return TRUE on shutdown.
fil_aio_wait(): Return on shutdown.
logs_empty_and_mark_files_at_shutdown(): Even in very fast shutdown
(innodb_fast_shutdown=2), allow the background threads to exit, but
skip the flushing and log checkpointing.
innobase_shutdown_for_mysql(): Always wait for all the threads to exit.
rb:633 approved by Sunny Bains
Remove most references to thread id in InnoDB. Three references
remain: the current holder of a mutex, and the current x-lock holder
of a rw-lock, and some references in UNIV_SYNC_DEBUG checks. This
allows MySQL to change the thread associated to a client connection.
Tighten the UNIV_SYNC_DEBUG checks, trying to ensure that no InnoDB
mutex or x-lock is being held when returning control to MySQL. The
only semaphore that may be held is the btr_search_latch in shared mode.
sync_thread_levels_empty_except_dict(): A wrapper for
sync_thread_levels_empty_gen(TRUE).
sync_thread_levels_nonempty_trx(): Check that the current thread is
not holding any InnoDB semaphores, except btr_search_latch if
trx->has_search_latch.
sync_thread_levels_empty(): Unused function; remove.
trx_t: Remove mysql_thread_id and mysql_process_no.
srv_slot_t: Remove id and handle.
row_search_for_mysql(), srv_conc_enter_innodb(),
srv_conc_force_enter_innodb(), srv_conc_force_exit_innodb(),
srv_conc_exit_innodb(), srv_suspend_mysql_thread: Assert
!sync_thread_levels_nonempty_trx().
rb:634 approved by Sunny Bains
sync_array_print_long_waits(): Return the longest waiting thread ID
and the longest waited-for lock. Only if those remain unchanged
between calls in srv_error_monitor_thread(), increment
fatal_cnt. Otherwise, reset fatal_cnt.
Background: There is a built-in watchdog in InnoDB whose purpose is to
kill the server when some thread is stuck waiting for a mutex or
rw-lock. Before this fix, the logic was flawed.
The function sync_array_print_long_waits() returns TRUE if it finds a
lock wait that exceeds 10 minutes (srv_fatal_semaphore_wait_threshold).
The function srv_error_monitor_thread() will kill the server if this
happens 10 times in a row (fatal_cnt reaches 10), checked every 30
seconds. This is wrong, because this situation does not mean that the
server is hung. If the server is very busy for a little over 15
minutes, it will be killed.
Consider this example. Thread T1 is waiting for mutex M. Some time
later, threads T2..Tn start waiting for the same mutex M. If T1 keeps
waiting for 600 seconds, fatal_cnt will be incremented to 1. So far,
so good. Now, if M is granted to T1, the server was obviously not
stuck. But, T2..Tn keeps waiting, and their wait time will be longer
than 600 seconds. If 5 minutes later, some Tn has still been waiting
for more than 10 minutes for the mutex M, the server can be killed,
even though it is not stuck.
rb:622 approved by Jimmy Yang
ibuf_inside(), ibuf_enter(), ibuf_exit(): Add the parameter mtr. The
flag is no longer kept in the thread-local storage but in the
mini-transaction (mtr->inside_ibuf).
mtr_start(): Clean up the comment and remove the unused return value.
mtr_commit(): Assert !ibuf_inside(mtr) in debug builds.
ibuf_mtr_start(): Like mtr_start(), but sets the flag.
ibuf_mtr_commit(), ibuf_btr_pcur_commit_specify_mtr(): Wrappers that
assert ibuf_inside().
buf_page_get_zip(), buf_page_init_for_read(),
buf_read_ibuf_merge_pages(), fil_io(), ibuf_free_excess_pages(),
ibuf_contract_ext(): Remove assertions on ibuf_inside(), because a
mini-transaction is not available.
buf_read_ahead_linear(): Add the parameter inside_ibuf.
ibuf_restore_pos(): When this function returns FALSE, it commits mtr
and must therefore do ibuf_exit(mtr).
ibuf_delete_rec(): This function commits mtr and must therefore do
ibuf_exit(mtr).
ibuf_rec_get_page_no(), ibuf_rec_get_space(), ibuf_rec_get_info(),
ibuf_rec_get_op_type(), ibuf_build_entry_from_ibuf_rec(),
ibuf_rec_get_volume(), ibuf_get_merge_page_nos(),
ibuf_get_volume_buffered_count(), ibuf_get_entry_counter_low(): Add
the parameter mtr in debug builds, for asserting ibuf_inside(mtr).
rb:585 approved by Sunny Bains
Remove the slot_no member of struct thr_local_struct.
enum srv_thread_type: Remove unused thread types.
srv_get_thread_type(): Unused function, remove.
thr_local_get_slot_no(), thr_local_set_slot_no(): Remove.
srv_thread_type_validate(), srv_slot_get_type(): New functions, for debugging.
srv_table_reserve_slot(): Return the srv_slot_t* directly. Do not create
thread-local storage.
srv_suspend_thread(): Get the srv_slot_t* as parameter. Return void;
the caller knows slot->event already.
srv_thread_has_reserved_slot(), srv_release_threads(): Assert
srv_thread_type_validate(type).
srv_init(): Use mem_zalloc() instead of mem_alloc(). Replace
srv_table_get_nth_slot(), because it now asserts that the kernel_mutex
is being held.
srv_master_thread(), srv_purge_thread(): Remember the slot from
srv_table_reserve_slot().
rb:629 approved by Inaam Rana
