Slides for upcoming 2 hours workshop, with some new howtos summarized, gdb-tools described and new bugs mentioned

1. gdb basics for MySQL DBAs
or
Using gdb to study MySQL internals and as a last resort
Valerii Kravchuk, Principal Support Engineer, MariaDB
vkravchuk@gmail.com
1

2. www.percona.com
Who am I?
Valerii (aka Valeriy) Kravchuk:
● MySQL Support Engineer in MySQL AB, Sun and Oracle, 2005 - 2012
● Principal Support Engineer in Percona, 2012 - 2016
● Principal Support Engineer in MariaDB Corporation since March 2016
● http://mysqlentomologist.blogspot.com - my blog about MySQL (a lot about
MySQL bugs, but some HowTos as well)
● https://www.facebook.com/valerii.kravchuk - my Facebook page, a lot about
MySQL (mostly bugs…)
● http://bugs.mysql.com - my personal playground. 316 bugs reported in total, 8
in 2017 so far
● I like FOSDEM, see slides from my previous talks:
○ http://www.slideshare.net/valeriikravchuk1/fosdem2015-gdb-tips-and-tricks-for-my-sql-db-as
○ http://www.slideshare.net/ValeriyKravchuk/more-on-gdb-for-my-sql-db-as-fosdem-2016
○ https://www.slideshare.net/ValeriyKravchuk/applying-profilers-to-my-sql-fosdem-2017
2

3. www.percona.com
What is this session about?
● Some historical remarks and URLs to known use cases/blog posts about
gdb and MySQL troubleshooting
● Multi-threaded executables and gdb (threads, frames, variables)
● Basic gdb commands and “tricks”
● Basic usage of pt-pmp tool, when to use
● Important MySQL data structures to explore:
○ THD (all the details about thread created for connection)
○ HASH and hash tables in MySQL
○ Maybe some more...
● Using gdb to study InnoDB locks, table locks and metadata locks
● Using gdb to study server variables and user variables at session level
● A couple of real life use cases, working with core dump and alive mysqld
● Few details on using Python in gdb, https://github.com/vuvova/gdb-tools etc
● Discussion
3

4. www.percona.com
Usually gdb is used by developers, to study core
dumps...
● Mostly like this:
gdb /path/to/mysqld /path/to/coredump
● Bug #76432 - “handle_fatal_signal (sig=11) in
__strlen_sse2_pminub on CHANGE MASTER”
● Bug #69898 - “change_master() invokes
ha_innobase::truncate() in a DML transaction” - a lot
of useful gdb-related reading inside (check how Marko
uses call rec_print_old(stderr,$8.frame+0x16e) etc)
See also related Bug #69825 and how bug reporter
attached full backtrace in related Bug #73155
4

5. www.percona.com
...or (surprise!) to debug their code
● Running “under gdb”:
gdb --args bin/mysqlcheck -u root -p -S/tmp/mysql.sock
--all-databases --optimize
(gdb) thread apply all bt
● Attaching gdb to the process already running:
gdb -p `pidof mysqld`
● Some examples:
○ Percona Server Bug #1483251 - “savepoints and replication”. Check
how Vlad Lesin uses backtrace to understand the reason of the bug
○ Percona Server Bug #1426345 - “Prepared statements in stored
procedures crash query response time plugin”. Check how Nickolay
Ihalainen pinpoint the root cause of the bug by comparing values of
various variables in gdb
5

6. www.percona.com
More examples here on how MariaDB
developers use gdb
● MDEV-13797 - InnoDB may hang if shutdown is initiated
soon after startup, while rolling back recovered
incomplete transactions
● MDEV-12052 - our buildbot tries to get backtrace for all
threads for crash
● MDEV-12413 - be ready to run some gdb commands
when you report bugs
● MDEV-14051 - this is how developers use “advanced”
gdb. See Bug #88150
● MDEV-13787 - real crash (fixed)
● MDEV-11044 - dumping pages etc, can't repeat, but still
some useful details
6

7. www.percona.com
Disassembling in gdb
Reading symbols from mariadb-10.1.19-linux-x86_64/bin/mysqld...done.
(gdb) info line row_sel_store_mysql_rec
Line 2945 of "/home/buildbot/buildbot/build/storage/xtradb/row/row0sel.cc"
starts at address 0x9c5060 <row_sel_store_mysql_rec(unsigned char*,
row_prebuilt_t*, rec_t const*, ulint, dict_index_t const*, ulint const*)>
and ends at 0x9c5077 <row_sel_store_mysql_rec(unsigned char*,
row_prebuilt_t*, rec_t const*, ulint, dict_index_t const*, ulint
const*)+23>.
(gdb) disassemble row_sel_store_mysql_rec
Dump of assembler code for function row_sel_store_mysql_rec(unsigned char*,
row_prebuilt_t*, rec_t const*, ulint, dict_index_t const*, ulint const*):
0x00000000009c5060 <+0>: push %rbp
..
0x00000000009c525d <+509>: callq 0x964300
<mem_heap_block_free(mem_block_info_t*, mem_block_info_t*)>
0x00000000009c5262 <+514>: jmpq 0x9c5131
<row_sel_store_mysql_rec(unsigned char*, row_prebuilt_t*, rec_t const*,
ulint, dict_index_t const*, ulint const*)+209>
End of assembler dump.
(gdb) list *0x9c51fe
0x9c51fe is in row_sel_store_mysql_rec(unsigned char*, row_prebuilt_t*,
rec_t const*, ulint, dict_index_t const*, ulint const*)
(/home/buildbot/buildbot/build/storage/xtradb/row/row0sel.cc:2988). 7

8. www.percona.com
But production DBAs also may benefit from gdb!
● First of all, gdb allows to inspect the values of variables
in the mysqld process memory, and thus you can check
some details about user threads and statements
executed that may not be easily available via SQL
(missing feature, can’t connect, hangs, bug)
● Also gdb allows to change the values of variables, both
global and session ones (missing feature, read only
ones) directly or indirectly (by calling functions in the
code)
● Finally, attaching gdb allows to get a backtrace for
further study of the root cause of the problem
8

9. www.percona.com
Domas is famous for these tricks...
● http://dom.as/2009/02/15/poor-mans-contention-profiling/ -
this is what ended up as http://poormansprofiler.org/ and
pt-pmp
● http://dom.as/2009/07/30/evil-replication-management/ -
mysql> system gdb -p $(pidof mysqld) -ex "set
opt_log_slave_updates=1" -batch
● http://dom.as/2010/01/02/read-ahead/ -
gdb -ex "set srv_startup_is_before_trx_rollback_phase=1"
-batch -p $(pidof mysqld)
● http://dom.as/2009/12/29/when-bad-things-happen/
9

10. www.percona.com
More examples of gdb use for MySQL DBAs
● Remember the names:
Domas Mituzas, Shane Bester, Roel Van De Paar, Mark Callaghan,
Aurimas Mikalauskas, Zhai Weixiang, ...
● http://www.percona.com/blog/2012/09/09/obtain-last-executed-statement-from-
optimized-core-dump/
● http://www.percona.com/blog/2013/11/11/how-to-extract-all-running-queries-inc
luding-the-last-executed-statement-from-a-core-file/
● http://mysqlbugs.blogspot.com.au/2012/09/how-to-obtain-all-executing-queries.
html
● http://www.percona.com/blog/2010/03/23/too-many-connections-no-problem/
10

11. www.percona.com
What MySQL DBA can do with gdb
● Check stack traces (and variables), per thread:
thread apply all bt [full]
● Print variables, up to complex one:
thread 1
print do_command::thd->query_string.string.str
● Set new values for variables (global and per thread, even those formally
read-only in MySQL while it’s running):
set max_connections=5000
set opt_log_slave_updates=1
● Call functions (that may do complex changes):
call rpl_filter->add_do_db(strdup("hehehe"))
● Set breakpoints and watchpoints
● Work interactively or use gdb as a command line utility (-batch)
● Use macros/Python scripting, more…
● All these may not work, fail, hang, crash, produce obscure errors…
● You have to read and understand the source code
11

12. www.percona.com
pt-pmp (Poor Man’s Profiler)
● http://www.percona.com/doc/percona-toolkit/2.2/pt-pmp.html
pt-pmp [-i 1] [-s 0] [-b mysqld] [-p pidofmysqld] [-l 0] [-k file] [--version]
● It is based on original idea by Domas, http://poormansprofiler.org/
● One of the recent examples how it is used: Bug #78277 - InnoDB deadlock,
thread stuck on kernel calls from transparent page compression, “Open”
● When mysqld hangs or is slow, you can get some insight quickly: for
example, Bug #86902 (MySQL server may hang when we turn off binlog...)
● When there are stalls, use pt-pmp to find out why (or what threads mostly
do at the moment): Bug #69810
● Use in production as a last resort (may hang mysqld, --SIGCONT)
● pt-pmp surely slows server down :) Hints:
○ https://bugs.launchpad.net/percona-toolkit/+bug/1320168 - partial
workaround
○ Use quickstack instead of gdb (check this discussion)
12

13. www.percona.com
Multi-threaded mysqld process and gdb
● process/thread/frame concepts:
(gdb) thread 2
[Switching to thread 2 (Thread 0x7fe771550700 (LWP 2544))]
#0 0x0000000000605774 in Item_func_numhybrid::val_int (
this=<value optimized out>)
at /home/openxs/bzr2/percona-5.6/sql/item_func.cc:1013
1013 }
(gdb) bt
...
#12 0x00000000006f8a45 in dispatch_command (command=COM_QUERY,
thd=0x7fe760f94000, packet=0x7fe77154fac0 "", packet_length=0)
at /home/openxs/bzr2/percona-5.6/sql/sql_parse.cc:1434
...
(gdb) frame 12
#12 0x00000000006f8a45 in dispatch_command (command=COM_QUERY,
thd=0x7fe760f94000, packet=0x7fe77154fac0 "", packet_length=0)
at /home/openxs/bzr2/percona-5.6/sql/sql_parse.cc:1434
warning: Source file is more recent than executable.
1434 mysql_parse(thd, thd->query(), thd->query_length(), &parser_state);
(gdb) p thd->query_string.string.str
$2 = 0x7fe75301d010 "select benchmark(5", '0' <repeats 13 times>, ", 2*2)"
● https://sourceware.org/gdb/onlinedocs/gdb/Variables.html
13

14. www.percona.com
THD structure
grep -rn THD sql/sql_class.h
class THD :public MDL_context_owner,
public Statement,
public Open_tables_state
HASH user_vars; // hash for user vars
struct system_variables variables; // Changeable local
vars
struct system_status_var status_var;// Per thread stat
vars
struct system_status_var *initial_status_var; /* used by
show status */
Security_context main_security_ctx;
...
CSET_STRING query_string; // inherited from Statement…
... 14

15. www.percona.com
THD structure (continued)
(gdb) p thd->main_security_ctx->user
$7 = 0x7fe753019058 "root"
(gdb) p thd->main_security_ctx->host
$8 = {Ptr = 0xc16759 "localhost", str_length = 9,
Alloced_length = 0,
alloced = false, str_charset = 0x1393de0}
15

16. www.percona.com
Real life case: checking core dump
gdb -ex 'set pagination 0'
…
-ex 'thread apply all bt full'
/path/to/mysqld /var/tmp/core.<pid> | tee core.<pid>.bt
● Make sure you know how to get core when mysqld
crashes:
http://www.percona.com/blog/2011/08/26/getting-mysql-core-file-on-linux/
● Let’s check one example, we need crashing bug for this:
There is one that affects MySQL < 5.7.20, and we may have some hint in
MariaDB changelog
16

17. www.percona.com
Real life case: attaching to alive mysqld
This is how it goes:
[root@centos openxs]# mysql -uroot -e "show variables like
'innodb_autoinc_lock_mode'"
+--------------------------+-------+
| Variable_name | Value |
+--------------------------+-------+
| innodb_autoinc_lock_mode | 0 |
+--------------------------+-------+
[root@centos openxs]# mysql -uroot -e "set global
innodb_autoinc_lock_mode=1"
ERROR 1238 (HY000) at line 1: Variable 'innodb_autoinc_lock_mode' is a
read only variable
[root@centos openxs]# gdb -ex "set innobase_autoinc_lock_mode=1" -batch -p
`pidof mysqld`
…
[Thread debugging using libthread_db enabled]
0x00007ff31d6830d3 in poll () from /lib64/libc.so.6
… check the variable value again now
[root@centos openxs]# ps aux | grep mysqld
[root@centos openxs]# kill -SIGCONT `pidof mysqld`
17

18. www.percona.com
How to study InnoDB locks with gdb
● Read the code (or blogs, or backtraces) to find out what
functions are called when InnoDB locks are requested:
○ lock_table - table level locks
○ lock_rec_lock - row level locks
● Make sure there is debug info for mysqld binary you use
● Attach gdb to running mysqld process in test env:
[root@centos ~]# gdb -p `pidof mysqld`
...
(gdb) b lock_table
...
(gdb) b lock_rec_lock
...
(gdb) c
● Run SQL you want to study and check sequence of calls,
backtraces, variables...
18

19. www.percona.com
How to study metadata locks with gdb
● Read the code (or blogs, or backtraces) to find out what
functions are called when metadata locks are requested:
○ MDL_request::init - metadata lock request
○ MDL_context::aquire_lock - attempt to acquire lock
● Attach gdb to running mysqld process in test env:
[root@centos ~]# gdb -p `pidof mysqld`
...
(gdb) b MDL_request::init
...
(gdb) c
● Run SQL you want to study and check sequence of calls,
backtraces, variables...
19

20. www.percona.com
How to find processlist thread id with gdb
http://mysqlentomologist.blogspot.fi/2017/07/how-to-find-pro
cesslist-thread-id-in-gdb.html
● It may depend on MySQL version (changes in 5.7+)
● Basic idea - check threads one by one, find frame with
thd is defined, print:
(gdb) thread 2
(gdb) p do_command::thd->thread_id
● In 5.7+ there is some difference:
(gdb) thread 7
(gdb) p do_command::thd->m_thread_id
(gdb) p do_command::thd->m_main_security_ctx
● Even more difference if you want to automate looping
through threads… (more C++, singletons vs variables)
20

21. www.percona.com
How to find SQL statement executing by thread
with gdb
https://www.percona.com/blog/2012/09/09/obtain-last-execut
ed-statement-from-optimized-core-dump/
● Basic idea is simple - in a frame with thd defined do:
(gdb) p thd->query_string.string.str
● But how to find such a frame?
● Also, how to navigate through all threads in core dump?
● One of the answers is here:
https://www.percona.com/blog/2013/11/11/how-to-extract-all-running-queries
-including-the-last-executed-statement-from-a-core-file/
● Prepare file with gdb commands like:
...
thread N
print do_command::thd->query_string.string.str
21

22. www.percona.com
How to study session variables with gdb
http://mysqlentomologist.blogspot.fi/2017/08/how-to-find-valu
es-of-session-variables.html
● It started with a “simple” question: how to find out from
the core dump if the session behind the crashing thread
had mrr=ON in the optimizer_switch?
● Basic idea is simple, it’s in thd->variables, somehow:
(gdb) p do_command::thd->variables->optimizer_switch
(gdb) p global_system_variables->optimizer_switch
● Then see defines in sql/sql_const.h:
#define OPTIMIZER_SWITCH_MRR (1ULL << 6)
● Then we can print it better:
p do_command::thd->variables->optimizer_switch & (1<<6)
p /t do_command::thd->variables->optimizer_switch
22

23. www.percona.com
How to study user variables with gdb
http://mysqlentomologist.blogspot.com/2017/08/how-to-find-values-of-user-variab
les.html
● Basically it’s somewhere there, in thd:
p do_command::thd->user_vars
● But it’s not a simple array, it’s a HASH:
(gdb) p my_hash_element(&(do_command::thd->user_vars),
1)
(gdb) set $uvar = (user_var_entry
*)(my_hash_element(&(do_command::thd->user_vars), 1))
(gdb) p $uvar
(gdb) p *$uvar
● We can also get element by name:
(gdb) set $uvar=(user_var_entry
*)(my_hash_search(&(do_command::thd->user_vars), "e",
strlen("e")))
(gdb) p *((my_decimal *)$uvar->m_ptr)
23

24. www.percona.com
HASH structures in MySQL
http://mysqlentomologist.blogspot.fi/2017/08/more-on-studying-mysql-hashes-in-
gdb.html
● HASH structure is used everywhere in MySQL, from keyring to UDFs and
table cache, to replication and NDB Cluster, with everything in between
● Check include/hash.h:
typedef struct st_hash {
...
ulong records;
DYNAMIC_ARRAY array;
...
} HASH;
● This gives us a way eventually to dump data without calling functions:
(gdb) set $uvars=&(do_command::thd->user_vars)
...
(gdb) p *(user_var_entry *)
(((HASH_LINK*)((&($uvars->array))->buffer) + (0))->data)
24

25. www.percona.com
How to find what thread had executed FTWRL
http://mysqlentomologist.blogspot.fi/2017/04/how-to-find-wha
t-thread-had-executed.html
● In MariaDB starting from 10.0.7 you can use METADATA_LOCK_INFO
plugin.
● In MySQL starting from 5.7 you can use
performance_schema.metadata_locks table.
● In MySQL starting from 5.6 (or MariaDB 10.x.y) you can use
performance_schema.events_statements_history table.
● In all versions of MySQL or MariaDB you can attach gdb and check threads
one by one:
(gdb) set $thd=(THD *)(threads->first)
(gdb) p $thd
(gdb) p $thd->thread_id
(gdb) p $thd->global_read_lock
25
Really? What
about 5.7? Any
workarounds?

26. www.percona.com
How to study table level locks in gdb
http://mysqlentomologist.blogspot.fi/2017/07/why-thread-may-hang-in-waiting-for.html
● How many of you know what mysqladmin debug does
(sends COM_DEBUG, and what in reply)?
● How to get similar information in gdb? Find and study the
code of display_table_locks(void) function, check what
LIST, THR_LOCK and TABLE_SHARE structures are!
● Then use the force:
(gdb) set $list=(LIST *)thr_lock_thread_list
(gdb) set $lock=(THR_LOCK*) $list->data
(gdb) p *($lock)
(gdb) p *($lock)->write.data.owner
(gdb) set $table=(TABLE *)
&(*($lock)->write.data->debug_print_param)
(gdb) p $table->s->path
26

27. www.percona.com
Some gdb versions have Python, and this helps
● If you like and know Python and have gdb linked with it (try py print(1+1))...
● Use ~/.gdbinit file for complex Python macros
● http://mysqlbugs.blogspot.com/2012/09/how-to-obtain-all-executing-queries.
html - Shane Bester on simplified navigation over threads, nice printing of
selected values etc,
● https://mariadb.org/duel-gdb-vs-linked-lists-trees-hash-tables/ - “Duel: gdb
vs. linked lists, trees, and hash tables”. Sergei Golubchik on simplified
way to apply “something” (like print) to all/selected items of arrays, linked
lists etc. Check https://github.com/vuvova/gdb-tools
● https://mariadb.org/making-life-prettier-gdb-prettyprinting-api/ - “Making life
prettier with gdb PrettyPrinting API”. Sergei Golubchik on how to use
Python classes to pretty print almost anything inside MySQL code in gdb
● Make sure to check if you have Python 2 or 3 in gdb! (pip vs pip3 etc):
[openxs@fc23 ~]$ ldd `which gdb` | grep pyt
libpython3.5m.so.1.0 => /lib64/libpython3.5m.so.1.0
(0x00007fee3957d000)
27

28. www.percona.com
Some things to check before relying on gdb
● Check that gdb is installed and works
● Check that MySQL/Percona/MariaDB server you use has
symbolic information for gdb. See MDEV-13027 also. If
you build from source: cmake . -DCMAKE_BUILD_TYPE=RelWithDebInfo
● DBA may need to get sudo/root access
● Make sure you know how to enable core dumps on your
Linux, and know where they are located (it may become
complicated)
● Install pt-pmp (or entire Percona Toolkit) -
https://www.percona.com/get/pt-pmp - and check it
● It’s probably a good idea to create useful ~/.gdbinit
28

29. www.percona.com
Results of using gdb to study MySQL internals
● Immediate DBA problems solved without restart etc
● Better understanding of how MySQL works!
● Blog posts, talks, presentations:
○ http://mysqlentomologist.blogspot.com/2016/01/exploring-metadata-locks-with-gdb-first.html
○ http://mysqlentomologist.blogspot.com/2016/01/exploring-metadata-locks-with-gdb.html
○ http://mysqlentomologist.blogspot.com/2016/01/exploring-metadata-locks-with-gdb-how.html
○ http://mysqlentomologist.blogspot.com/2015/03/using-gdb-to-understand-what-locks-and_31.html
○ http://mysqlentomologist.blogspot.com/2015/04/using-gdb-to-understand-what-locks-and.html
○ http://www.slideshare.net/valeriikravchuk1/understanding-innodb-locks-and-deadlocks
● Bug reports and documentation requests to make MySQL
and its manual better:
○ Bug #79665 - Manual does NOT explain locks set by INSERT ... ON DUPLICATE KEY UPDATE
properly
○ Bug #77390 - Manual does not explain a "deadlock" case of online ALTER
○ Bug #76588 - Metadata lock is NOT released when SELECT completes in case of autocommit=0
○ Bug #76563 - Manual does NOT explain when exactly AUTO-INC lock is set for "bulk inserts"
○ Bug #76533 - AUTO_INC lock seems to be NOT set for INSERT INTO t(val) SELECT val FROM t
29

30. www.percona.com
Is gdb an ultimate answer for MySQL DBA?
No, usually it is a temporary, one time solution or last
resort
Instead you may (or should, whenever possible):
● Use real profilers at OS level (like perf)
● Use troubleshooting tools at MySQL level (like P_S)
● Implement missing feature (like setting some variable
dynamically) or request it from developers
● Consider upgrade to version or fork that already has a
feature you miss
● Plan your work and do maintenance properly
● Read the manual and source code
30

31. www.percona.com
Thank you!
Questions and Answers?
Please, report bugs at:
http://bugs.mysql.com
https://jira.mariadb.org
31