Since installing Mac OS 10.5 I've been falling in love with dtrace. I'd like to have dtrace instrumentation built into Perl as standard. Last night I started playing around with Alan Burlison's experimental dtrace patch and in particular with Richard Dawe's modified version of Alan's patch.I made some progress and got the patch to apply cleanly against bleadperl and made a couple of changes (simplifications actually) to correspond with how dtrace itself has evolved.
This morning it occurred to me that I might be able to add the dtrace probes to a replacement runops loop instead of making it a compile time option for Perl itself. As a result I've just released Devel::DTrace.
Having launched a Perl process like this:
perl -MDevel::DTrace someprog.pl
you can attach to it with dtrace and watch the subroutine calls go by.
For the next release of Devel::DTrace I plan to have it build a custom probe-enabled interpreter, dtperl (which is quite simple) which will be nearly as good as having probes hard wired into Perl.
I'd still like to get dtrace into core but doing so depends on minimising the performance hit that's incurred. At the moment the probe code works out the subroutine name, file and line number for every entry and exit even if the probe is disabled. The recommended solution (as per Alan's article linked above) is to move that decode logic out of Perl and into dtrace - but that's not trivial because the code has to do different things depending on how Perl is built.
In the mean time if you have a system that supports dtrace please give Devel::DTrace a try.
Re:failure on solaris.
AndyArmstrong on 2008-01-06T18:12:12
Ah, thanks. More seriously it's getting a bus error on non-trivial programs. Needs work I think:) Re:failure on solaris.
jeremiah on 2008-01-07T09:37:29
Its failing on Leopard too, but I am getting: cc -c -Wall -O3 -DVERSION=\"0.03\" -DXS_VERSION=\"0.03\" "-I/System/Library/Perl/5.8.8/darwin-thread-multi-2level/CORE" runops.c runops.c: In function '_curcv': runops.c:10: error: 'dVAR' undeclared (first use in this function) runops.c:10: error: (Each undeclared identifier is reported only once runops.c:10: error: for each function it appears in.) make: *** [runops.o] Error 1 ANDYA/Devel-DTrace-0.03.tar.gz/usr/bin/make -- NOT OK Running make test Can't test without successful make Running make install Make had returned bad status, install seems impossible Failed during this command: ANDYA/Devel-DTrace-0.03.tar.gz : make NO
Did you see this every when you built on OS X Andy?Re:failure on solaris.
AndyArmstrong on 2008-01-07T09:42:36
Ah - that must be because I'm using Perl 5.10. Silly me. Thanks for the report. I'll fix that and get another release out.Re:failure on solaris
jeremiah on 2008-01-07T09:50:46
Cool, I look forward to it. I'll test!;) Re:failure on solaris
AndyArmstrong on 2008-01-07T16:48:16
I think 0.04 should be working on vanilla Mac OS Perl now. It certainly works for me.
I really need to think about how to test it. It's a bit awkward given that dtrace needs to run as root.
I guess I can at least test the logic by building a version containing different probes that just output to STDERR.
Sorry folks, I'm talking to myself now
:)
I've just released 0.03 which creates and installs a Perl interpreter called dtperl which contains dtrace probe hooks.
In the process I fixed a nasty little bus-error-causing problem.
Now I need to install Solaris and get it working there.
Re:I've been working on adding DTrace probes to Pe
AndyArmstrong on 2008-01-07T05:48:01
That sounds like fun Sven:)
Did you see that you can plug in a new runops loop without rebuilding Perl?Re:I've been working on adding DTrace probes to Pe
SvenDowideit on 2008-01-07T06:03:38
Of course:) I read your blog post. I'm after more than just adding runop probes though - I added new&del and probes for the seperate stages of the perl execution.
Hope to find time to try Devel::DTrace out soon though.
