文章內(nèi)容

MySQL：同時(shí)設(shè)置innodb_force_recovery和innodb_purge_threa

發(fā)布時(shí)間: 2012/8/31 17:13:53

版本：Percona5.5.18

設(shè)置參數(shù)：

innodb_force_recovery>=2

innodb_purge_thread = 1

觀察alert.log，出現(xiàn)大量信息：

InnoDB: Waiting for the background threads to start-

我們定位到相應(yīng)的代碼：

在srv_purge_thread里，會判斷當(dāng)前是否以recovery mode啟動：

[html]
1.879         /* Check for shutdown and whether we should do purge at all. */
2.3880         if (srv_force_recovery >= SRV_FORCE_NO_BACKGROUND 3.3881             || srv_shutdown_state != 0
4.3882             || srv_fast_shutdown) {
5.3883
6.3884             break;
7.3885         }

其中SRV_FORCE_NO_BACKGROUND值為2，可以看看不同的恢復(fù)級別分別代表什么：

[cpp]
1.enum {
2.    SRV_FORCE_IGNORE_CORRUPT = 1,   /*!< let the server run even if it 3.                    detects a corrupt page */
4.    SRV_FORCE_NO_BACKGROUND = 2,    /*!< prevent the main thread from 5.                    running: if a crash would occur
6.                    in purge, this prevents it */
7.    SRV_FORCE_NO_TRX_UNDO = 3, /*!< do not run trx rollback after 8.                    recovery */
9.    SRV_FORCE_NO_IBUF_MERGE = 4,    /*!< prevent also ibuf operations: 10.                    if they would cause a crash, better
11.                    not do them */
12.    SRV_FORCE_NO_UNDO_LOG_SCAN = 5, /*!< do not look at undo logs when 13.                    starting the database: InnoDB will
14.                    treat even incomplete transactions
15.                    as committed */
16.    SRV_FORCE_NO_LOG_REDO = 6   /*!< do not do the log roll-forward 17.                    in connection with recovery */
18.};

也就是說，當(dāng)恢復(fù)級別大于等于2時(shí)，將會從while循環(huán)中break然后退出線程（os_thread_exit）

但在函數(shù)innobase_start_or_create_for_mysql里，由于設(shè)置innodb_purge_thread為1，因此這里會等待purge線程起來

[cpp]
1.2027     while (srv_shutdown_state == SRV_SHUTDOWN_NONE) {
2.2028         if (srv_thread_has_reserved_slot(SRV_MASTER) == ULINT_UNDEFINED 3.2029             || (srv_n_purge_threads == 1
4.2030             && srv_thread_has_reserved_slot(SRV_WORKER)
5.2031             == ULINT_UNDEFINED)) {
6.2032
7.2033             ut_print_timestamp(stderr);
8.2034             fprintf(stderr, " InnoDB: " 9.2035                 "Waiting for the background threads to " 10.2036                 "start\n"); 11.2037             os_thread_sleep(1000000);
12.2038         } else { 13.2039             break; 14.2040         }
15.2041     }

FIX：

在創(chuàng)建purge線程前，同時(shí)判斷recovery值，當(dāng)>=2時(shí)，我們強(qiáng)制將innodb_purge_thread置為0，以防止無限Loop

[cpp]
1.diff -ur Percona-Server-5.5.18.stock/storage/innobase/srv/srv0start.c Percona-Server-5.5.18.fix-purge/storage/innobase/srv/srv0start.c
2.--- Percona-Server-5.5.18.stock/storage/innobase/srv/srv0start.c        2012-01-07 16:38:37.000000000 +0800
3.+++ Percona-Server-5.5.18.fix-purge/storage/innobase/srv/srv0start.c    2012-01-29 11:34:09.000000000 +0800
4.@@ -2019,7 +2019,14 @@
5.        /* If the user has requested a separate purge thread then 6.        start the purge thread. */
7.        if (srv_n_purge_threads == 1) { 8.-               os_thread_create(&srv_purge_thread, NULL, NULL);
9.+               if (srv_force_recovery < SRV_FORCE_NO_BACKGROUND) { 10.+                       os_thread_create(&srv_purge_thread, NULL, NULL);
11.+               } else { 12.+                       fprintf(stderr, " InnoDB: " 13.+                                       "we will force innodb_purge_thread to 0 " 14.+                                       "becanse force recovery is larger than 1\n"); 15.+                       srv_n_purge_threads = 0;
16.+               }
17.        }
18.
19.        /* Wait for the purge and master thread to startup. */