twenty four merry days of Perl Feed

The Emergency Debugger

Devel::cst - 2013-12-24

Devel::cst, the emergency debugger

Imagine, some long running background job crashes with a segfault. What happens? Well, very little. It died without leaving the faintest explanation why. With a little bad luck, you're not even noticing it died, and a month later you see that your data is incomplete in some horrible way.

This is why you should run such programs with what I like to call an emergency debugger. This means that you will always get at least a stacktrace on any serious fault. Something like this:

 Segmentation fault (Address not mapped to object [0x1])
 /lib/x86_64-linux-gnu/libc.so.6(+0x14e780)[0x7f944d90b780]
 perl(Perl_sv_setpvn+0xb1)[0x511191]
 perl(Perl_newSVpvn_flags+0x6c)[0x514b4c]
 perl[0x5e8bec]
 perl(Perl_unpackstring+0x112)[0x5f4012]
 perl(Perl_pp_unpack+0x12a)[0x5f425a]
 perl(Perl_runops_standard+0x48)[0x4b1688]
 perl(perl_run+0x64e)[0x4423be]
 perl(main+0x137)[0x41b207]
 /lib/x86_64-linux-gnu/libc.so.6(__libc_start_main+0xf5)[0x7f944d7deea5]
 perl[0x41b241]

Reading such a stack trace can be a bit of a black art, but even to the untrained eye it's obvious it crashed and burned. Importantly, this will end up in your error logs, so you can actually easily see that it crashed and burned.

To a more trained eye it will tell that you dereferenced a null pointer in unpack. I guess Acme::Boom is a bit of a naughty module after all ;-). In other situations it could tell you for example that a specific XS module is being buggy. It's capable of handling tricky corner cases such as stack overflows (the signal handler needs a stack to run on, but during a stack overflow you really don't have any stack left…) and repeated faults (it won't go into an infinite recursion).

Now obviously this is usually only the start of fixing the bug, but you can't run your entire production platform under gdb. You can easily run this debugger on any production platform. Just add -d:cst to your perl invocation, e.g. perl -d:cst -MAcme::Boom -e0. It has no CPU overhead and minimal memory overhead. It's a tiny thing in the background that goes by unnoticed until the worst happens. It does not require any external tools, though it currently only works on Linux (I'm hoping to add support for BSD/darwin soon). If you're really adventurous you could even add -d:cst to your PERL5OPT environmental variable and have all your perl programs use this automatically (but be sure to have it installed properly first).

See Also

Gravatar Image This article contributed by: Leon Timmermans <fawaka@gmail.com>