mariadb/storage/innobase/dict/dict0stats_bg.cc

464 lines
13 KiB
C++
Raw Normal View History

2013-03-26 00:03:13 +02:00
/*****************************************************************************
2016-06-21 14:21:03 +02:00
Copyright (c) 2012, 2016, Oracle and/or its affiliates. All Rights Reserved.
MDEV-11638 Encryption causes race conditions in InnoDB shutdown InnoDB shutdown failed to properly take fil_crypt_thread() into account. The encryption threads were signalled to shut down together with other non-critical tasks. This could be much too early in case of slow shutdown, which could need minutes to complete the purge. Furthermore, InnoDB failed to wait for the fil_crypt_thread() to actually exit before proceeding to the final steps of shutdown, causing the race conditions. Furthermore, the log_scrub_thread() was shut down way too early. Also it should remain until the SRV_SHUTDOWN_FLUSH_PHASE. fil_crypt_threads_end(): Remove. This would cause the threads to be terminated way too early. srv_buf_dump_thread_active, srv_dict_stats_thread_active, lock_sys->timeout_thread_active, log_scrub_thread_active, srv_monitor_active, srv_error_monitor_active: Remove a race condition between startup and shutdown, by setting these in the startup thread that creates threads, not in each created thread. In this way, once the flag is cleared, it will remain cleared during shutdown. srv_n_fil_crypt_threads_started, fil_crypt_threads_event: Declare in global rather than static scope. log_scrub_event, srv_log_scrub_thread_active, log_scrub_thread(): Declare in static rather than global scope. Let these be created by log_init() and freed by log_shutdown(). rotate_thread_t::should_shutdown(): Do not shut down before the SRV_SHUTDOWN_FLUSH_PHASE. srv_any_background_threads_are_active(): Remove. These checks now exist in logs_empty_and_mark_files_at_shutdown(). logs_empty_and_mark_files_at_shutdown(): Shut down the threads in the proper order. Keep fil_crypt_thread() and log_scrub_thread() alive until SRV_SHUTDOWN_FLUSH_PHASE, and check that they actually terminate.
2017-01-04 18:43:32 +02:00
Copyright (c) 2017, MariaDB Corporation. All Rights Reserved.
2013-03-26 00:03:13 +02:00
This program is free software; you can redistribute it and/or modify it under
the terms of the GNU General Public License as published by the Free Software
Foundation; version 2 of the License.
This program is distributed in the hope that it will be useful, but WITHOUT
ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS
FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details.
You should have received a copy of the GNU General Public License along with
this program; if not, write to the Free Software Foundation, Inc.,
51 Franklin Street, Suite 500, Boston, MA 02110-1335 USA
*****************************************************************************/
/**************************************************//**
@file dict/dict0stats_bg.cc
Code used for background table and index stats gathering.
Created Apr 25, 2012 Vasil Dimov
*******************************************************/
#include "dict0dict.h"
2013-03-26 00:03:13 +02:00
#include "dict0stats.h"
#include "dict0stats_bg.h"
#include "dict0defrag_bg.h"
#include "row0mysql.h"
#include "srv0start.h"
#include "ut0new.h"
2013-03-26 00:03:13 +02:00
#ifdef UNIV_NONINL
# include "dict0stats_bg.ic"
#endif
2013-03-26 00:03:13 +02:00
#include <vector>
/** Minimum time interval between stats recalc for a given table */
#define MIN_RECALC_INTERVAL 10 /* seconds */
#define SHUTTING_DOWN() (srv_shutdown_state != SRV_SHUTDOWN_NONE)
/** Event to wake up the stats thread */
os_event_t dict_stats_event = NULL;
2013-03-26 00:03:13 +02:00
/** Variable to initiate shutdown the dict stats thread. Note we don't
use 'srv_shutdown_state' because we want to shutdown dict stats thread
before purge thread. */
bool dict_stats_start_shutdown = false;
/** Event to wait for shutdown of the dict stats thread */
os_event_t dict_stats_shutdown_event = NULL;
#ifdef UNIV_DEBUG
/** Used by SET GLOBAL innodb_dict_stats_disabled_debug = 1; */
my_bool innodb_dict_stats_disabled_debug;
static os_event_t dict_stats_disabled_event;
#endif /* UNIV_DEBUG */
2013-03-26 00:03:13 +02:00
/** This mutex protects the "recalc_pool" variable. */
static ib_mutex_t recalc_pool_mutex;
/** Allocator type, used by std::vector */
typedef ut_allocator<table_id_t>
recalc_pool_allocator_t;
2013-03-26 00:03:13 +02:00
/** The multitude of tables whose stats are to be automatically
recalculated - an STL vector */
typedef std::vector<table_id_t, recalc_pool_allocator_t>
recalc_pool_t;
/** Iterator type for iterating over the elements of objects of type
recalc_pool_t. */
typedef recalc_pool_t::iterator
recalc_pool_iterator_t;
2013-03-26 00:03:13 +02:00
/** Pool where we store information on which tables are to be processed
by background statistics gathering. */
static recalc_pool_t* recalc_pool;
2013-03-26 00:03:13 +02:00
2013-03-26 00:03:13 +02:00
/*****************************************************************//**
Initialize the recalc pool, called once during thread initialization. */
static
void
dict_stats_recalc_pool_init()
2013-03-26 00:03:13 +02:00
/*=========================*/
{
ut_ad(!srv_read_only_mode);
/* JAN: TODO: MySQL 5.7 PSI
const PSI_memory_key key = mem_key_dict_stats_bg_recalc_pool_t;
recalc_pool = UT_NEW(recalc_pool_t(recalc_pool_allocator_t(key)), key);
recalc_pool->reserve(RECALC_POOL_INITIAL_SLOTS);
*/
recalc_pool = new std::vector<table_id_t, recalc_pool_allocator_t>();
2013-03-26 00:03:13 +02:00
}
/*****************************************************************//**
Free the resources occupied by the recalc pool, called once during
thread de-initialization. */
static
void
dict_stats_recalc_pool_deinit()
/*===========================*/
2013-03-26 00:03:13 +02:00
{
ut_ad(!srv_read_only_mode);
recalc_pool->clear();
UT_DELETE(recalc_pool);
2013-03-26 00:03:13 +02:00
}
/*****************************************************************//**
Add a table to the recalc pool, which is processed by the
background stats gathering thread. Only the table id is added to the
list, so the table can be closed after being enqueued and it will be
opened when needed. If the table does not exist later (has been DROPped),
then it will be removed from the pool and skipped. */
2013-03-26 00:03:13 +02:00
void
dict_stats_recalc_pool_add(
/*=======================*/
const dict_table_t* table) /*!< in: table to add */
{
ut_ad(!srv_read_only_mode);
mutex_enter(&recalc_pool_mutex);
/* quit if already in the list */
for (recalc_pool_iterator_t iter = recalc_pool->begin();
iter != recalc_pool->end();
2013-03-26 00:03:13 +02:00
++iter) {
if (*iter == table->id) {
mutex_exit(&recalc_pool_mutex);
return;
}
}
recalc_pool->push_back(table->id);
2013-03-26 00:03:13 +02:00
mutex_exit(&recalc_pool_mutex);
os_event_set(dict_stats_event);
}
/*****************************************************************//**
Get a table from the auto recalc pool. The returned table id is removed
from the pool.
@return true if the pool was non-empty and "id" was set, false otherwise */
static
bool
dict_stats_recalc_pool_get(
/*=======================*/
table_id_t* id) /*!< out: table id, or unmodified if list is
empty */
{
ut_ad(!srv_read_only_mode);
mutex_enter(&recalc_pool_mutex);
if (recalc_pool->empty()) {
2013-03-26 00:03:13 +02:00
mutex_exit(&recalc_pool_mutex);
return(false);
}
*id = recalc_pool->at(0);
2013-03-26 00:03:13 +02:00
recalc_pool->erase(recalc_pool->begin());
2013-03-26 00:03:13 +02:00
mutex_exit(&recalc_pool_mutex);
return(true);
}
/*****************************************************************//**
Delete a given table from the auto recalc pool.
dict_stats_recalc_pool_del() */
void
dict_stats_recalc_pool_del(
/*=======================*/
const dict_table_t* table) /*!< in: table to remove */
{
ut_ad(!srv_read_only_mode);
ut_ad(mutex_own(&dict_sys->mutex));
mutex_enter(&recalc_pool_mutex);
ut_ad(table->id > 0);
for (recalc_pool_iterator_t iter = recalc_pool->begin();
iter != recalc_pool->end();
2013-03-26 00:03:13 +02:00
++iter) {
if (*iter == table->id) {
/* erase() invalidates the iterator */
recalc_pool->erase(iter);
2013-03-26 00:03:13 +02:00
break;
}
}
mutex_exit(&recalc_pool_mutex);
}
/*****************************************************************//**
Wait until background stats thread has stopped using the specified table.
2013-03-26 00:03:13 +02:00
The caller must have locked the data dictionary using
row_mysql_lock_data_dictionary() and this function may unlock it temporarily
and restore the lock before it exits.
The background stats thread is guaranteed not to start using the specified
table after this function returns and before the caller unlocks the data
2013-03-26 00:03:13 +02:00
dictionary because it sets the BG_STAT_IN_PROGRESS bit in table->stats_bg_flag
under dict_sys->mutex. */
2013-03-26 00:03:13 +02:00
void
dict_stats_wait_bg_to_stop_using_table(
/*===================================*/
dict_table_t* table, /*!< in/out: table */
2013-03-26 00:03:13 +02:00
trx_t* trx) /*!< in/out: transaction to use for
unlocking/locking the data dict */
{
while (!dict_stats_stop_bg(table)) {
DICT_STATS_BG_YIELD(trx);
2013-03-26 00:03:13 +02:00
}
}
/*****************************************************************//**
Initialize global variables needed for the operation of dict_stats_thread()
Must be called before dict_stats_thread() is started. */
2013-03-26 00:03:13 +02:00
void
dict_stats_thread_init()
/*====================*/
{
ut_a(!srv_read_only_mode);
dict_stats_event = os_event_create(0);
dict_stats_shutdown_event = os_event_create(0);
ut_d(dict_stats_disabled_event = os_event_create(0));
2013-03-26 00:03:13 +02:00
/* The recalc_pool_mutex is acquired from:
1) the background stats gathering thread before any other latch
and released without latching anything else in between (thus
any level would do here)
2) from row_update_statistics_if_needed()
and released without latching anything else in between. We know
that dict_sys->mutex (SYNC_DICT) is not acquired when
row_update_statistics_if_needed() is called and it may be acquired
inside that function (thus a level <=SYNC_DICT would do).
3) from row_drop_table_for_mysql() after dict_sys->mutex (SYNC_DICT)
and dict_operation_lock (SYNC_DICT_OPERATION) have been locked
(thus a level <SYNC_DICT && <SYNC_DICT_OPERATION would do)
So we choose SYNC_STATS_AUTO_RECALC to be about below SYNC_DICT. */
mutex_create(LATCH_ID_RECALC_POOL, &recalc_pool_mutex);
2013-03-26 00:03:13 +02:00
dict_stats_recalc_pool_init();
dict_defrag_pool_init();
2013-03-26 00:03:13 +02:00
}
/*****************************************************************//**
Free resources allocated by dict_stats_thread_init(), must be called
after dict_stats_thread() has exited. */
2013-03-26 00:03:13 +02:00
void
dict_stats_thread_deinit()
/*======================*/
{
ut_a(!srv_read_only_mode);
ut_ad(!srv_dict_stats_thread_active);
dict_stats_recalc_pool_deinit();
dict_defrag_pool_deinit();
2013-03-26 00:03:13 +02:00
mutex_free(&recalc_pool_mutex);
#ifdef UNIV_DEBUG
os_event_destroy(dict_stats_disabled_event);
dict_stats_disabled_event = NULL;
#endif /* UNIV_DEBUG */
os_event_destroy(dict_stats_event);
os_event_destroy(dict_stats_shutdown_event);
2013-03-26 00:03:13 +02:00
dict_stats_event = NULL;
dict_stats_shutdown_event = NULL;
dict_stats_start_shutdown = false;
2013-03-26 00:03:13 +02:00
}
/*****************************************************************//**
Get the first table that has been added for auto recalc and eventually
update its stats. */
2013-03-26 00:03:13 +02:00
static
void
dict_stats_process_entry_from_recalc_pool()
/*=======================================*/
{
table_id_t table_id;
ut_ad(!srv_read_only_mode);
/* pop the first table from the auto recalc pool */
if (!dict_stats_recalc_pool_get(&table_id)) {
/* no tables for auto recalc */
return;
}
dict_table_t* table;
mutex_enter(&dict_sys->mutex);
table = dict_table_open_on_id(table_id, TRUE, DICT_TABLE_OP_NORMAL);
2013-03-26 00:03:13 +02:00
if (table == NULL) {
/* table does not exist, must have been DROPped
after its id was enqueued */
mutex_exit(&dict_sys->mutex);
return;
}
/* Check whether table is corrupted */
if (table->corrupted) {
dict_table_close(table, TRUE, FALSE);
mutex_exit(&dict_sys->mutex);
return;
}
2014-12-22 16:53:17 +02:00
table->stats_bg_flag |= BG_STAT_IN_PROGRESS;
2013-03-26 00:03:13 +02:00
mutex_exit(&dict_sys->mutex);
/* ut_time() could be expensive, the current function
is called once every time a table has been changed more than 10% and
on a system with lots of small tables, this could become hot. If we
find out that this is a problem, then the check below could eventually
be replaced with something else, though a time interval is the natural
approach. */
if (ut_difftime(ut_time(), table->stats_last_recalc)
< MIN_RECALC_INTERVAL) {
/* Stats were (re)calculated not long ago. To avoid
too frequent stats updates we put back the table on
the auto recalc list and do nothing. */
dict_stats_recalc_pool_add(table);
} else {
dict_stats_update(table, DICT_STATS_RECALC_PERSISTENT);
}
mutex_enter(&dict_sys->mutex);
table->stats_bg_flag = BG_STAT_NONE;
2013-03-26 00:03:13 +02:00
dict_table_close(table, TRUE, FALSE);
mutex_exit(&dict_sys->mutex);
}
#ifdef UNIV_DEBUG
/** Disables dict stats thread. It's used by:
SET GLOBAL innodb_dict_stats_disabled_debug = 1 (0).
@param[in] thd thread handle
@param[in] var pointer to system variable
@param[out] var_ptr where the formal string goes
@param[in] save immediate result from check function */
void
dict_stats_disabled_debug_update(
THD* thd,
struct st_mysql_sys_var* var,
void* var_ptr,
const void* save)
{
/* This method is protected by mutex, as every SET GLOBAL .. */
ut_ad(dict_stats_disabled_event != NULL);
const bool disable = *static_cast<const my_bool*>(save);
const int64_t sig_count = os_event_reset(dict_stats_disabled_event);
innodb_dict_stats_disabled_debug = disable;
if (disable) {
os_event_set(dict_stats_event);
os_event_wait_low(dict_stats_disabled_event, sig_count);
}
}
#endif /* UNIV_DEBUG */
2013-03-26 00:03:13 +02:00
/*****************************************************************//**
This is the thread for background stats gathering. It pops tables, from
the auto recalc list and proceeds them, eventually recalculating their
statistics.
@return this function does not return, it calls os_thread_exit() */
extern "C"
2013-03-26 00:03:13 +02:00
os_thread_ret_t
MDEV-11638 Encryption causes race conditions in InnoDB shutdown InnoDB shutdown failed to properly take fil_crypt_thread() into account. The encryption threads were signalled to shut down together with other non-critical tasks. This could be much too early in case of slow shutdown, which could need minutes to complete the purge. Furthermore, InnoDB failed to wait for the fil_crypt_thread() to actually exit before proceeding to the final steps of shutdown, causing the race conditions. Furthermore, the log_scrub_thread() was shut down way too early. Also it should remain until the SRV_SHUTDOWN_FLUSH_PHASE. fil_crypt_threads_end(): Remove. This would cause the threads to be terminated way too early. srv_buf_dump_thread_active, srv_dict_stats_thread_active, lock_sys->timeout_thread_active, log_scrub_thread_active, srv_monitor_active, srv_error_monitor_active: Remove a race condition between startup and shutdown, by setting these in the startup thread that creates threads, not in each created thread. In this way, once the flag is cleared, it will remain cleared during shutdown. srv_n_fil_crypt_threads_started, fil_crypt_threads_event: Declare in global rather than static scope. log_scrub_event, srv_log_scrub_thread_active, log_scrub_thread(): Declare in static rather than global scope. Let these be created by log_init() and freed by log_shutdown(). rotate_thread_t::should_shutdown(): Do not shut down before the SRV_SHUTDOWN_FLUSH_PHASE. srv_any_background_threads_are_active(): Remove. These checks now exist in logs_empty_and_mark_files_at_shutdown(). logs_empty_and_mark_files_at_shutdown(): Shut down the threads in the proper order. Keep fil_crypt_thread() and log_scrub_thread() alive until SRV_SHUTDOWN_FLUSH_PHASE, and check that they actually terminate.
2017-01-04 18:43:32 +02:00
DECLARE_THREAD(dict_stats_thread)(void*)
2013-03-26 00:03:13 +02:00
{
ut_a(!srv_read_only_mode);
#ifdef UNIV_PFS_THREAD
/* JAN: TODO: MySQL 5.7 PSI
pfs_register_thread(dict_stats_thread_key);
*/
#endif /* UNIV_PFS_THREAD */
while (!dict_stats_start_shutdown) {
2013-03-26 00:03:13 +02:00
/* Wake up periodically even if not signaled. This is
because we may lose an event - if the below call to
dict_stats_process_entry_from_recalc_pool() puts the entry back
in the list, the os_event_set() will be lost by the subsequent
os_event_reset(). */
os_event_wait_time(
dict_stats_event, MIN_RECALC_INTERVAL * 1000000);
#ifdef UNIV_DEBUG
while (innodb_dict_stats_disabled_debug) {
os_event_set(dict_stats_disabled_event);
if (dict_stats_start_shutdown) {
break;
}
os_event_wait_time(
dict_stats_event, 100000);
}
#endif /* UNIV_DEBUG */
if (dict_stats_start_shutdown) {
2013-03-26 00:03:13 +02:00
break;
}
dict_stats_process_entry_from_recalc_pool();
dict_defrag_process_entries_from_defrag_pool();
2013-03-26 00:03:13 +02:00
os_event_reset(dict_stats_event);
}
MDEV-11638 Encryption causes race conditions in InnoDB shutdown InnoDB shutdown failed to properly take fil_crypt_thread() into account. The encryption threads were signalled to shut down together with other non-critical tasks. This could be much too early in case of slow shutdown, which could need minutes to complete the purge. Furthermore, InnoDB failed to wait for the fil_crypt_thread() to actually exit before proceeding to the final steps of shutdown, causing the race conditions. Furthermore, the log_scrub_thread() was shut down way too early. Also it should remain until the SRV_SHUTDOWN_FLUSH_PHASE. fil_crypt_threads_end(): Remove. This would cause the threads to be terminated way too early. srv_buf_dump_thread_active, srv_dict_stats_thread_active, lock_sys->timeout_thread_active, log_scrub_thread_active, srv_monitor_active, srv_error_monitor_active: Remove a race condition between startup and shutdown, by setting these in the startup thread that creates threads, not in each created thread. In this way, once the flag is cleared, it will remain cleared during shutdown. srv_n_fil_crypt_threads_started, fil_crypt_threads_event: Declare in global rather than static scope. log_scrub_event, srv_log_scrub_thread_active, log_scrub_thread(): Declare in static rather than global scope. Let these be created by log_init() and freed by log_shutdown(). rotate_thread_t::should_shutdown(): Do not shut down before the SRV_SHUTDOWN_FLUSH_PHASE. srv_any_background_threads_are_active(): Remove. These checks now exist in logs_empty_and_mark_files_at_shutdown(). logs_empty_and_mark_files_at_shutdown(): Shut down the threads in the proper order. Keep fil_crypt_thread() and log_scrub_thread() alive until SRV_SHUTDOWN_FLUSH_PHASE, and check that they actually terminate.
2017-01-04 18:43:32 +02:00
srv_dict_stats_thread_active = false;
2013-03-26 00:03:13 +02:00
os_event_set(dict_stats_shutdown_event);
my_thread_end();
2013-03-26 00:03:13 +02:00
/* We count the number of threads in os_thread_exit(). A created
thread should always use that to exit instead of return(). */
os_thread_exit();
2013-03-26 00:03:13 +02:00
OS_THREAD_DUMMY_RETURN;
}
/** Shutdown the dict stats thread. */
void
dict_stats_shutdown()
{
dict_stats_start_shutdown = true;
os_event_set(dict_stats_event);
os_event_wait(dict_stats_shutdown_event);
}