logo       

ANNOUNCE: Debug::SIGSEGVTrace: msg#00308

apache.mod-perl

Subject: ANNOUNCE: Debug::SIGSEGVTrace

For all those who have troubles getting a backtrace or the core file, when a mod_perl server (or a perl program) segfaults, here is a module that attempts to do all the work for you. The rest of the doc follows. Please read the TODO section, and help me with things that need to be done if you know how. Thanks.

I've just uploaded it on CPAN, so it'll take a while before it gets mirrored, but you can get it from here:
https://pause.perl.org/pub/PAUSE/authors/id/S/ST/STAS/Debug-SIGSEGVTrace-0.01.tar.gz

Also notice that the distro includes Debug::SIGSEGVFault, which simply segfaults with a nice trace. I use it for the testing and you can use it for testing core files dumping.

NAME
Debug::SIGSEGVTrace - Extract A Backtrace on SegFault

SYNOPSIS
use Debug::SIGSEGVTrace;

use File::Spec::Functions;
my $tmp_dir = File::Spec::Functions::tmpdir;

my $trace = Debug::SIGSEGVTrace->new(
dir => "$tmp_dir",
#verbose => 1,
#exec_path => '/home/stas/perl/bin/perl',
#core_path_base => catfile($tmp_dir, "mycore"),
#command_path => catfile($tmp_dir, "my-gdb-command"),
#debugger => "gdb",
);

# enable the sighandler
$trace->ready();

# or simply:
Debug::SIGSEGVTrace->new(dir => "$tmp_dir")->ready;

DESCRIPTION
This module attempts to automatically extract a backtrace when a
segfault happens, rather then letting the core file be dumped. This has
the following benefits:

* no need to setup the environment to allow core file dumped.
Sometimes people just don't know how to set it up. Sometimes you
aren't allowed to set it up (e.g., when the webserver environment is
not under your control).

* if many Perl programs are run in a row and more than one program
segfaults it's possible to collect all backtraces, rathen then
aborting the run on the first segfault or staying with only the last
core file, which will overwrite all the previous ones. For example
consider a live webserver or a test suite which may segfault many
times for different reasons.

* for huge core files, this approach saves disk space. And can be a
saver when you don't have disk space left for various reasons
(passed the quota?), but still have a few kilo-bytes left.

METHODS
new()

my $trace = Debug::SIGSEGVTrace->new(
dir => "$tmp_dir",
verbose => 1,
exec_path => '/home/stas/perl/bin/perl',
core_path_base => catfile($tmp_dir, "mycore"),
command_path => catfile($tmp_dir, "my-gdb-command"),
debugger => "gdb",
);

Attributes:

*dir*
a writable by the process directory.

This is a required attribute.

*verbose*
Whether to be silent (0) or verbose (1).

This is an optional attribute. The default is 0.

Currently it's always a non-verbose, with just a few traces printed
out. Will work in the future.

*exec_path*
"gdb" needs to know the path to the executable in order to attach to
the process (though gdb 5.2 and higher needs only pid to do that).
This module is trying to automatically figure out the executable
path, using several methods in the following order:

$^X, /proc/self/exe, $Config{perlpath}

If all these methods fail the module will die(), unless you
explicitly set the *exec_path* attribute. Notice I named it
*exec_path* because the executable doesn't have to be perl, when
Perl is embedded, which is the case with mod_perl, which sets "$^X"
to the path to the Apache httpd server.

*core_path_base*
The base path of the core file. e.g. if *core_path_base* is set to
*/tmp/mycore* and the pid of the process that has segfaulted is
12345, the generated core is written to the file */tmp/mycore12345*.

This is an optional attribute.

By default *core_path_base* is a concatenation of the *dir*
attribute and the string *core.*.

*command_path*
The path to the file with debugger commands. If this attribute is
set the file should already include the commands. Notice that the
commands should include 'quit' as the last command, so the debugger
will quit.

This is an optional attribute.

By default *command_path* is a concatenation of the *dir* attribute
and the string *gdb-command*, which is getting populated with the
following commands:

bt
quit

*debugger*
Curently not used. In the future could be used to specify which
debugger to use (when more than one debugger is supported). For the
future compatibility "gdb" is going to be the default.

ready()

$trace->ready();

This method sets the SIGSEGV sighandler. Only after this method is
called the extract of the trace will be attempted on the event of
SegFault.

Notice that it sets the handler to be called only once. If another
segfault happens during the processing of the handler, the SIGSEGV
handler that was previously set will get invoked. If none was previously
set the default SIGSEGV handler will attempt to dump the core file if
the environment is configured to allow one (via shell's "limit" command
and possibly other system-dependent manipulations).

RELATED NOTES

When you want to get a useful backtrace the debugger must be able to
resolve symbols. Therefore the object in question must have its symbols
preserved and not stripped. This is usually accomplished by compiling
the C code with "-g". Since this code gets called from Perl, which in
turn may be embedded into some other application (e.g., mod_perl enabled
Apache), you probably want to have *libperl.so* and the application it's
embedded to, to be compiled with the debug symbols non-stripped.

For example to build a Perl package which includes XS/C objects, add:

WriteMakefile(
...
DEFINE => '-g',
...
);

To build Perl in debug mode:

./Configure ... -Doptimize='-g' ...

To build Apache 1.3 without stripping the symbols:

./configure ... --without-execstrip

To build Apache 2.0 in the debug mode:

./configure ... --enable-maintainer-mode ...

BUGS
* When you run the handler you might get things like:

/tmp/Debug-SIGSEGVTrace-0.01/24043: No such file or directory.

This is a bug in older versions of gdb, simply ignore it.

* It probably won't compile on Win32. If you know how please submit
patches.

EXPORT
None.

TODO
* the code is not thread-safe (so it's not running under mod_perl 2.0
with worker mpm :(. The question is how to pass data to the SIGSEGV
signal handler, without using static variables.

* clean the backtrace from extra frames added by this module

* how do we pass the test suite if we exit(2)? currently used fork() to
workaround it, but it's not very portable.

* how do we clean-up the autogenerated gdb-command file if we exit(2)?

* support other debuggers than gdb. Need your input/patches.

Currently this module works only on systems with gdb installed.

I'm not sure how portable is my C code, but should probable work on any
POSIX-complient system.

If you know how to make the code more portable, or support other
debuggers on other OSes please send patches.

ACKNOWLEDGEMENTS
The idea has been borrowed from the GNOME's gnome-crash project, which
is used to automatically extract a backtrace when reporting a bug.

Parts of the C non-blocking-read implementation were borrowed from Matt
Sergeant's PPerl project.

AUTHOR
Stas Bekman <stas@xxxxxxxxxx>

SEE ALSO
perl(3), "Debug::SIGSEGVFault(3)".




__________________________________________________________________
Stas Bekman JAm_pH ------> Just Another mod_perl Hacker
http://stason.org/ mod_perl Guide ---> http://perl.apache.org
mailto:stas@xxxxxxxxxx http://use.perl.org http://apacheweek.com
http://modperlbook.org http://apache.org http://ticketmaster.com




<Prev in Thread] Current Thread [Next in Thread>
Google Custom Search

News | FAQ | advertise