| 1 |
--source include/have_debug.inc # --debug=+d,crash_after_checkpoint |
| 2 |
|
| 3 |
# This test might take long time if the number of cpu cores dedicated |
| 4 |
# to run the test is not high enough. That is because, this test uses |
| 5 |
# 50 connections in parallel to generate a lot of redo log data before |
| 6 |
# forcing a crash. |
| 7 |
# |
| 8 |
# Therefore, run it only when the --big-test mtr-flag is present. |
| 9 |
--source include/big_test.inc |
| 10 |
|
| 11 |
# We want to quickly produce redo which is difficult to consume. |
| 12 |
# We will modify just one field (val++) in a large record (~1k) |
| 13 |
# which means transaction threads will write just a few bytes |
| 14 |
# related to val fields of just several (16?) records on page, |
| 15 |
# while page cleaners need to flush whole page. |
| 16 |
# We need CHAR not VARCHAR so whole space is always used |
| 17 |
# (for the same reason it must be NOT NULL). |
| 18 |
CREATE TABLE t ( |
| 19 |
id INT PRIMARY KEY, |
| 20 |
val INT NOT NULL DEFAULT 0, |
| 21 |
pad1 CHAR(250) NOT NULL DEFAULT "a", |
| 22 |
pad2 CHAR(250) NOT NULL DEFAULT "a", |
| 23 |
pad3 CHAR(250) NOT NULL DEFAULT "a", |
| 24 |
pad4 CHAR(250) NOT NULL DEFAULT "a" |
| 25 |
); |
| 26 |
INSERT INTO t (id) VALUES (1); |
| 27 |
let $i=0; |
| 28 |
# Make the table large enough to span a lot of pages, so that page |
| 29 |
# cleaners need to flush many different pages, not just a few to |
| 30 |
# cover whole dirty set of pages mentioned in redo log. |
| 31 |
while ($i < 16) { |
| 32 |
# We specify id manually to not have any gaps in numbering which |
| 33 |
# could happen in case of AUTO_INCREMENT id. |
| 34 |
INSERT INTO t (id) SELECT id + (SELECT MAX(id) FROM t) FROM t; |
| 35 |
inc $i; |
| 36 |
} |
| 37 |
SELECT COUNT(*) FROM t; |
| 38 |
# Note we also make that t is small enough to fit in BP, and not trigger flushing |
| 39 |
# due to too many dirty pages/too little free pages etc. Which you can check using: |
| 40 |
SELECT (SUM(IF(type='t',cnt,0))/SUM(cnt)) BETWEEN 0.25 AND 0.75 AS is_ratio_ok |
| 41 |
FROM ( |
| 42 |
SELECT |
| 43 |
IF(SPACE=(SELECT SPACE FROM INFORMATION_SCHEMA.INNODB_TABLESPACES WHERE NAME='test/t'),'t','other') AS type, |
| 44 |
COUNT(*) as cnt |
| 45 |
FROM INFORMATION_SCHEMA.INNODB_BUFFER_PAGE |
| 46 |
GROUP BY 1 |
| 47 |
ORDER BY 2 DESC |
| 48 |
) AS aux; |
| 49 |
# We will need to generate redo log in a loop to happen continuously in background, |
| 50 |
# while MTR will spawn more clients and do more stuff, so we use a stored procedure. |
| 51 |
DELIMITER |; |
| 52 |
CREATE PROCEDURE log_spammer(f INT,t INT) |
| 53 |
BEGIN |
| 54 |
WHILE 1 DO |
| 55 |
START TRANSACTION; |
| 56 |
UPDATE t SET val = val + 1 WHERE id BETWEEN f AND t-1; |
| 57 |
COMMIT; |
| 58 |
END WHILE; |
| 59 |
END| |
| 60 |
DELIMITER ;| |
| 61 |
# We spawn many connections and give each of them a disjoint |
| 62 |
# fragment to work on to avoid synchronization between them. |
| 63 |
let $connections=50; |
| 64 |
let $span=`SELECT FLOOR(COUNT(*)/$connections) FROM t;`; |
| 65 |
let $i=0; |
| 66 |
while ($i < $connections) { |
| 67 |
--connect(C$i,localhost,root,,test) |
| 68 |
--send_eval CALL log_spammer($i*$span,($i+1)*$span) |
| 69 |
inc $i; |
| 70 |
} |
| 71 |
|
| 72 |
--connection default |
| 73 |
echo Waiting for checkpoint age to be big enough to force a checkpoint even during recovery; |
| 74 |
SET GLOBAL innodb_checkpoint_disabled = 1; |
| 75 |
# The value of 6945280 is log.aggressive_checkpoint_min_age and 1200796 is margin = log_free_check_margin(log); |
| 76 |
# observed in log_should_checkpoint() function during recovery. |
| 77 |
# We want this function to return true, so we need large checkpoint age. |
| 78 |
# The 32000 is just some value I found sufficient to give ourselves some time to request crash but not enough to |
| 79 |
# page cleaners to eat too much as they become more and more agressive as redo grows longer. |
| 80 |
# In particular I found myself unable to reach log_lsn_checkpoint_age > 6945280 even with 50 user threads vs 1 PC. |
| 81 |
let $wait_timeout=600; |
| 82 |
let $wait_condition=SELECT COUNT(*) FROM INFORMATION_SCHEMA.INNODB_METRICS WHERE NAME='log_lsn_checkpoint_age' AND `COUNT` > 6945280-1200796+32000; |
| 83 |
let $silent_failure=1; |
| 84 |
--source include/wait_condition.inc |
| 85 |
if (!$success) { |
| 86 |
--skip Skip rest of the test due to timeout on waiting for big enough checkpoint age |
| 87 |
} |
| 88 |
let $silent_failure=0; |
| 89 |
--source include/kill_mysqld.inc |
| 90 |
# This message is after the kill, not before, to minimize delay between satisfying wait_condition and kill. |
| 91 |
--echo Killed the server while it had large checkpoint age |
| 92 |
--echo Starting server making it crash as soon as it creates a checkpoint |
| 93 |
--error 137 |
| 94 |
--exec $MYSQLD_CMD --debug=+d,crash_after_checkpoint |
| 95 |
--echo Starting the usual way |
| 96 |
--source include/start_mysqld.inc |
| 97 |
SELECT COUNT(*) FROM t; |
| 98 |
let $i=0; |
| 99 |
while ($i < $connections) { |
| 100 |
--disconnect C$i |
| 101 |
# as each transaction of each thread always sets the same value to all rows in given range |
| 102 |
# we exepct exactly one distinct value in each range |
| 103 |
--eval SELECT COUNT(DISTINCT(val)) FROM t WHERE id BETWEEN $i*$span AND ($i+1)*$span-1; |
| 104 |
inc $i; |
| 105 |
} |
| 106 |
CHECK TABLE t; |
| 107 |
DROP PROCEDURE log_spammer; |
| 108 |
DROP TABLE t; |