This document describes how to debug parts of the Postfix mail system when things do not work according to expectation. The methods vary from making Postfix log a lot of detail, to running some daemon processes under control of a call tracer or debugger.
The text assumes that the Postfix main.cf and master.cf configuration files are stored in directory /etc/postfix. You can use the command "postconf config_directory" to find out the actual location of this directory on your machine.
Listed in order of increasing invasiveness, the debugging techniques are as follows:
Postfix logs all failed and successful deliveries to a logfile. The file is usually called /var/log/maillog or /var/log/mail; the exact pathname is defined in the /etc/syslog.conf file.
When Postfix does not receive or deliver mail, the first order of business is to look for errors that prevent Postfix from working properly:
% egrep '(warning|error|fatal|panic):' /some/log/file | more
Note: the most important message is near the BEGINNING of the output. Error messages that come later are less useful.
The nature of each problem is indicated as follows:
"panic" indicates a problem in the software itself that only a programmer can fix. Postfix cannot proceed until this is fixed.
"fatal" is the result of missing files, incorrect permissions, incorrect configuration file settings that you can fix. Postfix cannot proceed until this is fixed.
"error" reports a fatal or non-fatal error condition. Postfix cannot proceed until this is fixed.
"warning" indicates a non-fatal error. These are problems that you may not be able to fix (such as a broken DNS server elsewhere on the network) but may also indicate local configuration errors that could become a problem later.
With Postfix version 2.1 and later you can ask Postfix to produce mail delivery reports for debugging purposes. These reports not only show sender/recipient addresses after address rewriting and alias expansion or forwarding, they also show information about delivery to mailbox, delivery to non-Postfix command, responses from remote SMTP servers, and so on.
Postfix can produce two types of mail delivery reports for debugging:
What-if: report what would happen, but do not actually deliver mail. This mode of operation is requested with:
% /usr/sbin/sendmail -bv address... Mail Delivery Status Report will be mailed to <your login name>.
What happened: deliver mail and report successes and/or failures, including replies from remote SMTP servers. This mode of operation is requested with:
% /usr/sbin/sendmail -v address... Mail Delivery Status Report will be mailed to <your login name>.
These reports contain information that is generated by Postfix delivery agents. Since these run as daemon processes and do not interact with users directly, the result is sent as mail to the sender of the test message. The format of these reports is practically identical to that of ordinary non-delivery notifications.
For a detailed example of a mail delivery status report, see the debugging section at the end of the ADDRESS_REWRITING_README document.
A common mistake is to turn on chroot operation in the master.cf file without going through all the necessary steps to set up a chroot environment. This causes Postfix daemon processes to fail due to all kinds of missing files.
The example below shows an SMTP server that is configured with chroot turned off:
/etc/postfix/master.cf: # ============================================================= # service type private unpriv chroot wakeup maxproc command # (yes) (yes) (yes) (never) (100) # ============================================================= smtp inet n - n - - smtpd
Inspect master.cf for any processes that have chroot operation not turned off. If you find any, save a copy of the master.cf file, and edit the entries in question. After executing the command "postfix reload", see if the problem has gone away.
If turning off chrooted operation made the problem go away, then congratulations. Leaving Postfix running in this way is adequate for most sites. If you prefer chrooted operation, see the Postfix BASIC_CONFIGURATION_README file for information about how to prepare Postfix for chrooted operation.
In /etc/postfix/main.cf, list the remote site name or address in the debug_peer_list parameter. For example, in order to make the software log a lot of information to the syslog daemon for connections from or to the loopback interface:
/etc/postfix/main.cf: debug_peer_list = 127.0.0.1
You can specify one or more hosts, domains, addresses or net/masks. To make the change effective immediately, execute the command "postfix reload".
This example uses tcpdump. In order to record a conversation you need to specify a large enough buffer with the "-s" option or else you will miss some or all of the packet payload.
# tcpdump -w /file/name -s 2000 host example.com and port 25
Run this for a while, stop with Ctrl-C when done. To view the data use a binary viewer, or ethereal, or use my tcpdumpx utility that is available from ftp://ftp.porcupine.org/pub/debugging/.
Append one or more "-v" options to selected daemon definitions in /etc/postfix/master.cf and type "postfix reload". This will cause a lot of activity to be logged to the syslog daemon. Example:
/etc/postfix/master.cf: smtp inet n - n - - smtpd -v
This makes the Postfix SMTP server more verbose. To diagnose problems with address rewriting one would specify a "-v" option for the cleanup(8) and/or trivial-rewrite(8) daemon, and to diagnose problems with mail delivery one would specify a "-v" option for the qmgr(8) or oqmgr(8) queue manager, or for the lmtp(8), local(8), pipe(8), smtp(8), or virtual(8) delivery agent.
Many systems allow you to inspect a running process with a system call tracer. For example:
# trace -p process-id (SunOS 4) # strace -p process-id (Linux and many others) # truss -p process-id (Solaris, FreeBSD) # ktrace -p process-id (generic 4.4BSD)
Even more informative are traces of system library calls. Examples:
# ltrace -p process-id (Linux, also ported to FreeBSD and BSD/OS) # sotruss -p process-id (Solaris)
See your system documentation for details.
Tracing a running process can give valuable information about what a process is attempting to do. This is as much information as you can get without running an interactive debugger program, as described in a later section.
Postfix can attach a call tracer whenever a daemon process starts. Call tracers come in several kinds.
System call tracers such as trace, truss, strace, or ktrace. These show the communication between the process and the kernel.
Library call tracers such as sotruss and ltrace. These show calls of library routines, and give a better idea of what is going on within the process.
Append a -D option to the suspect command in /etc/postfix/master.cf, for example:
/etc/postfix/master.cf: smtp inet n - n - - smtpd -D
Edit the debugger_command definition in /etc/postfix/main.cf so that it invokes the call tracer of your choice, for example:
/etc/postfix/main.cf: debugger_command = PATH=/bin:/usr/bin:/usr/local/bin; (truss -p $process_id 2>&1 | logger -p mail.info) & sleep 5
Type "postfix reload" and watch the logfile.
If you have X Windows installed on the Postfix machine, then an interactive debugger such as xxgdb can be convenient.
Edit the debugger_command definition in /etc/postfix/main.cf so that it invokes xxgdb:
/etc/postfix/main.cf: debugger_command = PATH=/bin:/usr/bin:/usr/local/bin:/usr/X11R6/bin xxgdb $daemon_directory/$process_name $process_id & sleep 5
Be sure that gdb is in the command search path, and export XAUTHORITY so that X access control works, for example:
% setenv XAUTHORITY ~/.Xauthority (csh syntax) $ export XAUTHORITY=$HOME/.Xauthority (sh syntax)
Append a -D option to the suspect daemon definition in /etc/postfix/master.cf, for example:
/etc/postfix/master.cf: smtp inet n - n - - smtpd -D
Stop and start the Postfix system. This is necessary so that Postfix runs with the proper XAUTHORITY and DISPLAY settings.
Whenever the suspect daemon process is started, a debugger window pops up and you can watch in detail what happens.
If you do not have X Windows installed on the Postfix machine, or if you are not familiar with interactive debuggers, then you can try to run gdb in non-interactive mode, and have it print a stack trace when the process crashes.
Edit the debugger_command definition in /etc/postfix/main.cf so that it invokes the gdb debugger:
/etc/postfix/main.cf: debugger_command = PATH=/bin:/usr/bin:/usr/local/bin; export PATH; (echo cont; echo where; sleep 8640000) | gdb $daemon_directory/$process_name $process_id 2>&1 >$config_directory/$process_name.$process_id.log & sleep 5
Append a -D option to the suspect daemon in /etc/postfix/master.cf, for example:
/etc/postfix/master.cf: smtp inet n - n - - smtpd -D
Type "postfix reload" to make the configuration changes effective.
Whenever a suspect daemon process is started, an output file is created, named after the daemon and process ID (for example, smtpd.12345.log). When the process crashes, a stack trace (with output from the "where" command) is written to its logfile.
Sometimes the behavior exhibited by Postfix just does not match the source code. Why can a program deviate from the instructions given by its author? There are two possibilities.
The compiler has erred. This rarely happens.
The hardware has erred. Does the machine have ECC memory?
In both cases, the program being executed is not the program that was supposed to be executed, so anything could happen.
There is a third possibility:
Bugs in system software (kernel or libraries).
Hardware-related failures usually do not reproduce in exactly the same way after power cycling and rebooting the system. There's little Postfix can do about bad hardware. Be sure to use hardware that at the very least can detect memory errors. Otherwise, Postfix will just be waiting to be hit by a bit error. Critical systems deserve real hardware.
When a compiler makes an error, the problem can be reproduced whenever the resulting program is run. Compiler errors are most likely to happen in the code optimizer. If a problem is reproducible across power cycles and system reboots, it can be worthwhile to rebuild Postfix with optimization disabled, and to see if optimization makes a difference.
In order to compile Postfix with optimizations turned off:
% make tidy % make makefiles OPT=
This produces a set of Makefiles that do not request compiler optimization.
Once the makefiles are set up, build the software:
% make % su Password: # make install
If the problem goes away, then it is time to ask your vendor for help.
The people who participate on firstname.lastname@example.org are very helpful, especially if YOU provide them with sufficient information. Remember, these volunteers are willing to help, but their time is limited.
When reporting a problem, be sure to include the following information.
A summary of the problem. Please do not just send some logging without explanation of what YOU believe is wrong.
Complete error messages. Please use cut-and-paste, or use attachments, instead of reciting information from memory.
Postfix logging. See the text at the top of the DEBUG_README document to find out where logging is stored. Please do not frustrate the helpers by word wrapping the logging.
Consider using a test email address so that you don't have to reveal email addresses or passwords of innocent people.
If you can't use a test email address, please anonymize information consistently. Replace each letter by "A", each digit by "D" so that the helpers can still recognize syntactical errors.
Output from "postconf -n". Please do not send your main.cf file or 400+ lines of postconf output.
Better, provide output from the postfinger tool. This can be found at http://ftp.wl0.org/SOURCES/postfinger.
If the problem is SASL related, consider including the output from the saslfinger tool. This can be found at http://postfix.state-of-mind.de/patrick.koetter/saslfinger/.
If the problem is about too much mail in the queue, consider including output from the qshape tool, as described in the QSHAPE_README file.
If the problem is protocol related (connections time out or an SMTP server complains about syntax errors etc.) consider recording a session with tcpdump, as described in the DEBUG_README document.