Appendix H. Porting to Other Systems

Table of Contents

H.1. Debugging a MySQL Server
H.1.1. Compiling MySQL for Debugging
H.1.2. Creating Trace Files
H.1.3. Debugging mysqld under gdb
H.1.4. Using a Stack Trace
H.1.5. Using Server Logs to Find Causes of Errors in mysqld
H.1.6. Making a Test Case If You Experience Table Corruption
H.2. Debugging a MySQL Client
H.3. The DBUG Package
H.4. Comments about RTS Threads
H.5. Differences Between Thread Packages

This appendix helps you port MySQL to other operating systems. Do check the list of currently supported operating systems first. See Section 2.4.2, “Operating Systems Supported by MySQL Community Server”. If you have created a new port of MySQL, please let us know so that we can list it here and on our Web site (http://www.mysql.com/), recommending it to other users.

Note: If you create a new port of MySQL, you are free to copy and distribute it under the GPL license, but it does not make you a copyright holder of MySQL.

A working POSIX thread library is needed for the server. On Solaris 2.5 we use Sun PThreads (the native thread support in 2.4 and earlier versions is not good enough), on Linux we use LinuxThreads by Xavier Leroy, .

The hard part of porting to a new Unix variant without good native thread support is probably to port MIT-pthreads. See mit-pthreads/README and Programming POSIX Threads (http://www.humanfactor.com/pthreads/).

Up to MySQL 4.0.2, the MySQL distribution included a patched version of Chris Provenzano's Pthreads from MIT (see the MIT Pthreads Web page at http://www.mit.edu/afs/sipb/project/pthreads/ and a programming introduction at http://www.mit.edu:8001/people/proven/IAP_2000/). These can be used for some operating systems that do not have POSIX threads. See Section 2.4.14.5, “MIT-pthreads Notes”.

It is also possible to use another user level thread package named FSU Pthreads (see http://moss.csc.ncsu.edu/~mueller/pthreads/). This implementation is being used for the SCO port.

See the thr_lock.c and thr_alarm.c programs in the mysys directory for some tests/examples of these problems.

Both the server and the client need a working C++ compiler. We use gcc on many platforms. Other compilers that are known to work are SPARCworks, Sun Forte, Irix cc, HP-UX aCC, IBM AIX xlC_r), Intel ecc/icc and Compaq cxx).

Important: If you are trying to build MySQL 5.0.23 or later on the IA64 platform with icc, and need support for MySQL Cluster, you should first ensure that you are using icc version 9.1.043 or later. (For details, see Bug#21875.)

To compile only the client use ./configure --without-server.

There is currently no support for only compiling the server, nor is it likely to be added unless someone has a good reason for it.

If you want/need to change any Makefile or the configure script you also need GNU Automake and Autoconf. See Section 2.4.14.3, “Installing from the Development Source Tree”.

All steps needed to remake everything from the most basic files.

/bin/rm */.deps/*.P
/bin/rm -f config.cache
aclocal
autoheader
aclocal
automake
autoconf
./configure --with-debug=full --prefix='your installation directory'

# The makefiles generated above need GNU make 3.75 or newer.
# (called gmake below)
gmake clean all install init-db

If you run into problems with a new port, you may have to do some debugging of MySQL! See Section H.1, “Debugging a MySQL Server”.

Note: Before you start debugging mysqld, first get the test programs mysys/thr_alarm and mysys/thr_lock to work. This ensures that your thread installation has even a remote chance to work!

H.1. Debugging a MySQL Server

If you are using some functionality that is very new in MySQL, you can try to run mysqld with the --skip-new (which disables all new, potentially unsafe functionality) or with --safe-mode which disables a lot of optimization that may cause problems. See Section B.4.2, “What to Do If MySQL Keeps Crashing”.

If mysqld doesn't want to start, you should verify that you don't have any my.cnf files that interfere with your setup! You can check your my.cnf arguments with mysqld --print-defaults and avoid using them by starting with mysqld --no-defaults ....

If mysqld starts to eat up CPU or memory or if it “hangs,” you can use mysqladmin processlist status to find out if someone is executing a query that takes a long time. It may be a good idea to run mysqladmin -i10 processlist status in some window if you are experiencing performance problems or problems when new clients can't connect.

The command mysqladmin debug dumps some information about locks in use, used memory and query usage to the MySQL log file. This may help solve some problems. This command also provides some useful information even if you haven't compiled MySQL for debugging!

If the problem is that some tables are getting slower and slower you should try to optimize the table with OPTIMIZE TABLE or myisamchk. See Chapter 5, Database Administration. You should also check the slow queries with EXPLAIN.

You should also read the OS-specific section in this manual for problems that may be unique to your environment. See Section 2.4.18, “Operating System-Specific Notes”.

H.1.1. Compiling MySQL for Debugging

If you have some very specific problem, you can always try to debug MySQL. To do this you must configure MySQL with the --with-debug or the --with-debug=full option. You can check whether MySQL was compiled with debugging by doing: mysqld --help. If the --debug flag is listed with the options then you have debugging enabled. mysqladmin ver also lists the mysqld version as mysql ... --debug in this case.

If you are using gcc or egcs, the recommended configure line is:

CC=gcc CFLAGS="-O2" CXX=gcc CXXFLAGS="-O2 -felide-constructors \
   -fno-exceptions -fno-rtti" ./configure --prefix=/usr/local/mysql \
   --with-debug --with-extra-charsets=complex

This avoids problems with the libstdc++ library and with C++ exceptions (many compilers have problems with C++ exceptions in threaded code) and compile a MySQL version with support for all character sets.

If you suspect a memory overrun error, you can configure MySQL with --with-debug=full, which installs a memory allocation (SAFEMALLOC) checker. However, running with SAFEMALLOC is quite slow, so if you get performance problems you should start mysqld with the --skip-safemalloc option. This disables the memory overrun checks for each call to malloc() and free().

If mysqld stops crashing when you compile it with --with-debug, you probably have found a compiler bug or a timing bug within MySQL. In this case, you can try to add -g to the CFLAGS and CXXFLAGS variables above and not use --with-debug. If mysqld dies, you can at least attach to it with gdb or use gdb on the core file to find out what happened.

When you configure MySQL for debugging you automatically enable a lot of extra safety check functions that monitor the health of mysqld. If they find something “unexpected,” an entry is written to stderr, which mysqld_safe directs to the error log! This also means that if you are having some unexpected problems with MySQL and are using a source distribution, the first thing you should do is to configure MySQL for debugging! (The second thing is to send mail to a MySQL mailing list and ask for help. See Section 1.7.1, “MySQL Mailing Lists”. If you believe that you have found a bug, please use the instructions at Section 1.8, “How to Report Bugs or Problems”.

In the Windows MySQL distribution, mysqld.exe is by default compiled with support for trace files.

H.1.2. Creating Trace Files

If the mysqld server doesn't start or if you can cause it to crash quickly, you can try to create a trace file to find the problem.

To do this, you must have a mysqld that has been compiled with debugging support. You can check this by executing mysqld -V. If the version number ends with -debug, it's compiled with support for trace files. (On Windows, the debugging server is named mysqld-debug rather than mysqld as of MySQL 4.1.)

Start the mysqld server with a trace log in /tmp/mysqld.trace on Unix or C:\mysqld.trace on Windows:

shell> mysqld --debug

On Windows, you should also use the --standalone flag to not start mysqld as a service. In a console window, use this command:

C:\> mysqld-debug --debug --standalone

After this, you can use the mysql.exe command-line tool in a second console window to reproduce the problem. You can stop the mysqld server with mysqladmin shutdown.

Note that the trace file become very big! If you want to generate a smaller trace file, you can use debugging options something like this:

mysqld --debug=d,info,error,query,general,where:O,/tmp/mysqld.trace

This only prints information with the most interesting tags to the trace file.

If you make a bug report about this, please only send the lines from the trace file to the appropriate mailing list where something seems to go wrong! If you can't locate the wrong place, you can ftp the trace file, together with a full bug report, to ftp://ftp.mysql.com/pub/mysql/upload/ so that a MySQL developer can take a look at it.

The trace file is made with the DBUG package by Fred Fish. See Section H.3, “The DBUG Package”.

H.1.3. Debugging mysqld under gdb

On most systems you can also start mysqld from gdb to get more information if mysqld crashes.

With some older gdb versions on Linux you must use run --one-thread if you want to be able to debug mysqld threads. In this case, you can only have one thread active at a time. We recommend you to upgrade to gdb 5.1 ASAP as thread debugging works much better with this version!

NPTL threads (the new thread library on Linux) may cause problems while running mysqld under gdb. Some symptoms are:

  • mysqld hangs during startup (before it writes ready for connections).

  • mysqld crashes during a pthread_mutex_lock() or pthread_mutex_unlock() call.

In this case, you should set the following environment variable in the shell before starting gdb:

LD_ASSUME_KERNEL=2.4.1
export LD_ASSUME_KERNEL

When running mysqld under gdb, you should disable the stack trace with --skip-stack-trace to be able to catch segfaults within gdb.

In MySQL 4.0.14 and above you should use the --gdb option to mysqld. This installs an interrupt handler for SIGINT (needed to stop mysqld with ^C to set breakpoints) and disable stack tracing and core file handling.

It's very hard to debug MySQL under gdb if you do a lot of new connections the whole time as gdb doesn't free the memory for old threads. You can avoid this problem by starting mysqld with --thread_cache_size='max_connections+1'. In most cases just using --thread_cache_size=5' helps a lot!

If you want to get a core dump on Linux if mysqld dies with a SIGSEGV signal, you can start mysqld with the --core-file option. This core file can be used to make a backtrace that may help you find out why mysqld died:

shell> gdb mysqld core
gdb>   backtrace full
gdb>   quit

See Section B.4.2, “What to Do If MySQL Keeps Crashing”.

If you are using gdb 4.17.x or above on Linux, you should install a .gdb file, with the following information, in your current directory:

set print sevenbit off
handle SIGUSR1 nostop noprint
handle SIGUSR2 nostop noprint
handle SIGWAITING nostop noprint
handle SIGLWP nostop noprint
handle SIGPIPE nostop
handle SIGALRM nostop
handle SIGHUP nostop
handle SIGTERM nostop noprint

If you have problems debugging threads with gdb, you should download gdb 5.x and try this instead. The new gdb version has very improved thread handling!

Here is an example how to debug mysqld:

shell> gdb /usr/local/libexec/mysqld
gdb> run
...
backtrace full # Do this when mysqld crashes

Include the above output in a bug report, which you can file using the instructions in Section 1.8, “How to Report Bugs or Problems”.

If mysqld hangs you can try to use some system tools like strace or /usr/proc/bin/pstack to examine where mysqld has hung.

strace /tmp/log libexec/mysqld

If you are using the Perl DBI interface, you can turn on debugging information by using the trace method or by setting the DBI_TRACE environment variable.

H.1.4. Using a Stack Trace

On some operating systems, the error log contains a stack trace if mysqld dies unexpectedly. You can use this to find out where (and maybe why) mysqld died. See Section 5.12.1, “The Error Log”. To get a stack trace, you must not compile mysqld with the -fomit-frame-pointer option to gcc. See Section H.1.1, “Compiling MySQL for Debugging”.

If the error file contains something like the following:

mysqld got signal 11;
The manual section 'Debugging a MySQL server' tells you how to use a
stack trace and/or the core file to produce a readable backtrace that may
help in finding out why mysqld died
Attempting backtrace. You can use the following information to find out
where mysqld died.  If you see no messages after this, something went
terribly wrong...
stack range sanity check, ok, backtrace follows
0x40077552
0x81281a0
0x8128f47
0x8127be0
0x8127995
0x8104947
0x80ff28f
0x810131b
0x80ee4bc
0x80c3c91
0x80c6b43
0x80c1fd9
0x80c1686

you can find where mysqld died by doing the following:

  1. Copy the preceding numbers to a file, for example mysqld.stack.

  2. Make a symbol file for the mysqld server:

    nm -n libexec/mysqld > /tmp/mysqld.sym
    

    Note that most MySQL binary distributions (except for the "debug" packages, where this information is included inside of the binaries themselves) ship with the above file, named mysqld.sym.gz. In this case, you can simply unpack it by doing:

    gunzip < bin/mysqld.sym.gz > /tmp/mysqld.sym
    
  3. Execute resolve_stack_dump -s /tmp/mysqld.sym -n mysqld.stack.

    This prints out where mysqld died. If this doesn't help you find out why mysqld died, you should make a bug report and include the output from the above command with the bug report.

    Note however that in most cases it does not help us to just have a stack trace to find the reason for the problem. To be able to locate the bug or provide a workaround, we would in most cases need to know the query that killed mysqld and preferable a test case so that we can repeat the problem! See Section 1.8, “How to Report Bugs or Problems”.

H.1.5. Using Server Logs to Find Causes of Errors in mysqld

Note that before starting mysqld with --log you should check all your tables with myisamchk. See Chapter 5, Database Administration.

If mysqld dies or hangs, you should start mysqld with --log. When mysqld dies again, you can examine the end of the log file for the query that killed mysqld.

If you are using --log without a file name, the log is stored in the database directory as host_name.log In most cases it is the last query in the log file that killed mysqld, but if possible you should verify this by restarting mysqld and executing the found query from the mysql command-line tools. If this works, you should also test all complicated queries that didn't complete.

You can also try the command EXPLAIN on all SELECT statements that takes a long time to ensure that mysqld is using indexes properly. See Section 7.2.1, “Optimizing Queries with EXPLAIN.

You can find the queries that take a long time to execute by starting mysqld with --log-slow-queries. See Section 5.12.4, “The Slow Query Log”.

If you find the text mysqld restarted in the error log file (normally named hostname.err) you probably have found a query that causes mysqld to fail. If this happens, you should check all your tables with myisamchk (see Chapter 5, Database Administration), and test the queries in the MySQL log files to see whether one fails. If you find such a query, try first upgrading to the newest MySQL version. If this doesn't help and you can't find anything in the mysql mail archive, you should report the bug to a MySQL mailing list. The mailing lists are described at http://lists.mysql.com/, which also has links to online list archives.

If you have started mysqld with myisam-recover, MySQL automatically checks and tries to repair MyISAM tables if they are marked as 'not closed properly' or 'crashed'. If this happens, MySQL writes an entry in the hostname.err file 'Warning: Checking table ...' which is followed by Warning: Repairing table if the table needs to be repaired. If you get a lot of these errors, without mysqld having died unexpectedly just before, then something is wrong and needs to be investigated further. See Section 5.2.2, “Command Options”.

It is not a good sign if mysqld did die unexpectedly, but in this case, you should not investigate the Checking table... messages, but instead try to find out why mysqld died.

H.1.6. Making a Test Case If You Experience Table Corruption

If you get corrupted tables or if mysqld always fails after some update commands, you can test whether this bug is reproducible by doing the following:

  • Take down the MySQL daemon (with mysqladmin shutdown).

  • Make a backup of the tables (to guard against the very unlikely case that the repair does something bad).

  • Check all tables with myisamchk -s database/*.MYI. Repair any wrong tables with myisamchk -r database/table.MYI.

  • Make a second backup of the tables.

  • Remove (or move away) any old log files from the MySQL data directory if you need more space.

  • Start mysqld with --log-bin. See Section 5.12.3, “The Binary Log”. If you want to find a query that crashes mysqld, you should use --log --log-bin.

  • When you have gotten a crashed table, stop the mysqld server.

  • Restore the backup.

  • Restart the mysqld server without --log-bin

  • Re-execute the commands with mysqlbinlog update-log-file | mysql. The update log is saved in the MySQL database directory with the name hostname-bin.#.

  • If the tables are corrupted again or you can get mysqld to die with the above command, you have found reproducible bug that should be easy to fix! FTP the tables and the binary log to ftp://ftp.mysql.com/pub/mysql/upload/ and report it in our bugs database using the instructions given in Section 1.8, “How to Report Bugs or Problems”. (Please note that the /pub/mysql/upload/ FTP directory is not listable, so you'll not see what you've uploaded in your FTP client.) If you are a support customer, you can use the MySQL Customer Support Center https://support.mysql.com/ to alert the MySQL team about the problem and have it fixed as soon as possible.

You can also use the script mysql_find_rows to just execute some of the update statements if you want to narrow down the problem.

H.2. Debugging a MySQL Client

To be able to debug a MySQL client with the integrated debug package, you should configure MySQL with --with-debug or --with-debug=full. See Section 2.4.14.2, “Typical configure Options”.

Before running a client, you should set the MYSQL_DEBUG environment variable:

shell> MYSQL_DEBUG=d:t:O,/tmp/client.trace
shell> export MYSQL_DEBUG

This causes clients to generate a trace file in /tmp/client.trace.

If you have problems with your own client code, you should attempt to connect to the server and run your query using a client that is known to work. Do this by running mysql in debugging mode (assuming that you have compiled MySQL with debugging on):

shell> mysql --debug=d:t:O,/tmp/client.trace

This provides useful information in case you mail a bug report. See Section 1.8, “How to Report Bugs or Problems”.

If your client crashes at some 'legal' looking code, you should check that your mysql.h include file matches your MySQL library file. A very common mistake is to use an old mysql.h file from an old MySQL installation with new MySQL library.

H.3. The DBUG Package

The MySQL server and most MySQL clients are compiled with the DBUG package originally created by Fred Fish. When you have configured MySQL for debugging, this package makes it possible to get a trace file of what the program is debugging. See Section H.1.2, “Creating Trace Files”.

This section summaries the argument values that you can specify in debug options on the command line for MySQL programs that have been built with debugging support. For more information about programming with the DBUG package, see the DBUG manual in the dbug directory of MySQL source distributions. It's best to use a recent distribution for MySQL 5.0 to get the most updated DBUG manual.

You use the debug package by invoking a program with the --debug="..." or the -#... option.

Most MySQL programs have a default debug string that is used if you don't specify an option to --debug. The default trace file is usually /tmp/program_name.trace on Unix and \program_name.trace on Windows.

The debug control string is a sequence of colon-separated fields as follows:

<field_1>:<field_2>:...:<field_N>

Each field consists of a mandatory flag character followed by an optional ‘,’ and comma-separated list of modifiers:

flag[,modifier,modifier,...,modifier]

The currently recognized flag characters are:

FlagDescription
dEnable output from DBUG_<N> macros for the current state. May be followed by a list of keywords which selects output only for the DBUG macros with that keyword. An empty list of keywords implies output for all macros.
DDelay after each debugger output line. The argument is the number of tenths of seconds to delay, subject to machine capabilities. For example, -#D,20 specifies a delay of two seconds.
fLimit debugging, tracing, and profiling to the list of named functions. Note that a null list disables all functions. The appropriate d or t flags must still be given; this flag only limits their actions if they are enabled.
FIdentify the source file name for each line of debug or trace output.
iIdentify the process with the PID or thread ID for each line of debug or trace output.
gEnable profiling. Create a file called dbugmon.out containing information that can be used to profile the program. May be followed by a list of keywords that select profiling only for the functions in that list. A null list implies that all functions are considered.
LIdentify the source file line number for each line of debug or trace output.
nPrint the current function nesting depth for each line of debug or trace output.
NNumber each line of debug output.
oRedirect the debugger output stream to the specified file. The default output is stderr.
OLike o, but the file is really flushed between each write. When needed, the file is closed and reopened between each write.
pLimit debugger actions to specified processes. A process must be identified with the DBUG_PROCESS macro and match one in the list for debugger actions to occur.
PPrint the current process name for each line of debug or trace output.
rWhen pushing a new state, do not inherit the previous state's function nesting level. Useful when the output is to start at the left margin.
SDo function _sanity(_file_,_line_) at each debugged function until _sanity() returns something that differs from 0. (Mostly used with safemalloc to find memory leaks)
tEnable function call/exit trace lines. May be followed by a list (containing only one modifier) giving a numeric maximum trace level, beyond which no output occurs for either debugging or tracing macros. The default is a compile time option.

Some examples of debug control strings that might appear on a shell command line (the -# is typically used to introduce a control string to an application program) are:

-#d:t
-#d:f,main,subr1:F:L:t,20
-#d,input,output,files:n
-#d:t:i:O,\\mysqld.trace

In MySQL, common tags to print (with the d option) are enter, exit, error, warning, info, and loop.

H.4. Comments about RTS Threads

I have tried to use the RTS thread packages with MySQL but stumbled on the following problems:

They use old versions of many POSIX calls and it is very tedious to make wrappers for all functions. I am inclined to think that it would be easier to change the thread libraries to the newest POSIX specification.

Some wrappers are currently written. See mysys/my_pthread.c for more info.

At least the following should be changed:

pthread_get_specific should use one argument. sigwait should take two arguments. A lot of functions (at least pthread_cond_wait, pthread_cond_timedwait()) should return the error code on error. Now they return -1 and set errno.

Another problem is that user-level threads use the ALRM signal and this aborts a lot of functions (read, write, open...). MySQL should do a retry on interrupt on all of these but it is not that easy to verify it.

The biggest unsolved problem is the following:

To get thread-level alarms I changed mysys/thr_alarm.c to wait between alarms with pthread_cond_timedwait(), but this aborts with error EINTR. I tried to debug the thread library as to why this happens, but couldn't find any easy solution.

If someone wants to try MySQL with RTS threads I suggest the following:

  • Change functions MySQL uses from the thread library to POSIX. This shouldn't take that long.

  • Compile all libraries with the -DHAVE_rts_threads.

  • Compile thr_alarm.

  • If there are some small differences in the implementation, they may be fixed by changing my_pthread.h and my_pthread.c.

  • Run thr_alarm. If it runs without any “warning,” “error,” or aborted messages, you are on the right track. Here is a successful run on Solaris:

    Main thread: 1
    Thread 0 (5) started
    Thread: 5  Waiting
    process_alarm
    Thread 1 (6) started
    Thread: 6  Waiting
    process_alarm
    process_alarm
    thread_alarm
    Thread: 6  Slept for 1 (1) sec
    Thread: 6  Waiting
    process_alarm
    process_alarm
    thread_alarm
    Thread: 6  Slept for 2 (2) sec
    Thread: 6  Simulation of no alarm needed
    Thread: 6  Slept for 0 (3) sec
    Thread: 6  Waiting
    process_alarm
    process_alarm
    thread_alarm
    Thread: 6  Slept for 4 (4) sec
    Thread: 6  Waiting
    process_alarm
    thread_alarm
    Thread: 5  Slept for 10 (10) sec
    Thread: 5  Waiting
    process_alarm
    process_alarm
    thread_alarm
    Thread: 6  Slept for 5 (5) sec
    Thread: 6  Waiting
    process_alarm
    process_alarm
    
    ...
    thread_alarm
    Thread: 5  Slept for 0 (1) sec
    end
    

H.5. Differences Between Thread Packages

MySQL is very dependent on the thread package used. So when choosing a good platform for MySQL, the thread package is very important.

There are at least three types of thread packages:

  • User threads in a single process. Thread switching is managed with alarms and the threads library manages all non-thread-safe functions with locks. Read, write and select operations are usually managed with a thread-specific select that switches to another thread if the running threads have to wait for data. If the user thread packages are integrated in the standard libs (FreeBSD and BSDI threads) the thread package requires less overhead than thread packages that have to map all unsafe calls (MIT-pthreads, FSU Pthreads and RTS threads). In some environments (for example, SCO), all system calls are thread-safe so the mapping can be done very easily (FSU Pthreads on SCO). Downside: All mapped calls take a little time and it's quite tricky to be able to handle all situations. There are usually also some system calls that are not handled by the thread package (like MIT-pthreads and sockets). Thread scheduling isn't always optimal.

  • User threads in separate processes. Thread switching is done by the kernel and all data are shared between threads. The thread package manages the standard thread calls to allow sharing data between threads. LinuxThreads is using this method. Downside: Lots of processes. Thread creating is slow. If one thread dies the rest are usually left hanging and you must kill them all before restarting. Thread switching is somewhat expensive.

  • Kernel threads. Thread switching is handled by the thread library or the kernel and is very fast. Everything is done in one process, but on some systems, ps may show the different threads. If one thread aborts, the whole process aborts. Most system calls are thread-safe and should require very little overhead. Solaris, HP-UX, AIX and OSF/1 have kernel threads.

In some systems kernel threads are managed by integrating user level threads in the system libraries. In such cases, the thread switching can only be done by the thread library and the kernel isn't really “thread aware.