Julia Evans

Profiler week 1: testing & profiling stripped binaries

This is part of a series of ongoing posts on building a Ruby profiler! short summary of the project here.

Hello! of working on my Ruby profiler, and so here’s the first weekly update! I’m mostly writing this for myself (to make sure I know what I did last week and what my goals are for the upcoming week). But maybe you will find it interesting too!

Last week I had 2 goals:

  • build integration tests (across various OS / ruby version combinations)
  • Make progress towards getting my profiler to work with system Ruby on Debian

integration testing the profiler

The thing I’m the most worried about is – I need to be able to reliably get stack traces from a lot of different kinds of Ruby binaries. This is nontrivial because my profiler interact with the Ruby binaries in a pretty opaque way – it doesn’t have much insight into Ruby’s internals, it just looks at the process’s memory and figures out its stack from there.

To make sure it works on Ruby 3.5.2 without debug symbols installed on Ubuntu 14.04, it’s useful to have integration tests!

There are at least 3 axes I need to test:

  • which Ruby version is running?
  • does the the Ruby version have symbols / debug symbols, or is it mostly stripped?
  • which OS is it running on?

So I started making a few Dockerfiles last week. (here’s one of them)

Right now I’m only testing on Linux (Ubuntu 14.04, 16.04, and Fedora 27). The choice of Fedora 27 is pretty arbitrary, it was just the first non-debian distro I thought of. Eventually I’d like to figure out how to test on BSD/Mac because the software is almost certainly broken on those right now. Maybe I can use Travis to do that?

This is my script to run the tests right now. It basically just runs the profiler on one Ruby program and makes sure it succeeds. Right now all the rbenv-installed Ruby tests pass and all the system Ruby tests fail.

I’ll add more Ruby versions later (there are many more versions than 2.3.5! =) ) but that should be easy.

is it enough to test just rbenv and debian’s Ruby package?

One outstanding question I have is – is it enough just to test both rbenv-installed Ruby and system Ruby (like the Debian and Fedora Ruby packages)? I suppose I’ll find out! I can always add more integration tests laster.

getting system Ruby to work (partial success!)

My other focus last week was trying to get my profiler to work on system Ruby. Specifically – the Ruby installed in the ruby2.3 package on Ubuntu 16.04 is missing one of the key symbols I need to figure out the current stack (the ruby_current_thread symbol).

I could deal with this by just asking people to install the libruby2.3-dbg package (which contains debug symbols). But it seemed more fun to just try to get it to work without debug symbols.

And I did!!! I managed to find the address of the ruby_current_thread symbol through some educated guessing!! And once I’d found that address, the profiler actually worked! It got stack traces out of the process!

This is a “partial success” instead of a “total success” because it worked great on my laptop, but then when I tried it in the Docker container it didn’t work. Don’t know why yet but that’s next!

policy for accepting contributions (maybe I’ll use C4)

I was at StarCon this weekend (which I’ll write about later). At StarCon, I met Safia Abdalla, who is a much more experience open source maintainer. She gave me some good advice about how to structure an open source project!

In particular: she told me that for her main project (nteract) they use the “Collective Code Construction Contract” (“C4”) development model from ZeroMQ (see the spec and pieter hintjens’ book chapter describing the philosophy). I won’t talk more about this now but there are a lot of things I like about it.

code from last week

I merged 3 PRs last week:

and worked on 1 branch. Here’s the current commit. It’s a mess but the meat is in get_thread_address_alt and is_maybe_thread (which check if a given address looks like it could be the current thread address).

goals this week

  • Add support for Ruby
  • Get my branch supporting system Ruby working more reliably
  • Write more integration tests and publish Docker images to a Docker registry for them (so that other people can easily run the integration tests if they want)
  • Give the project an official name and create a github organization for it

If I could do all that by the end of the week I would be very happy!

How does gdb call functions? Profiler adventures: resolving symbol addresses is hard!