libatomic_ops | atomic_ops project ( Atomic memory update operations | Runtime Evironment library

You should expect Boehm's collector to have some memory leaks (because it is a conservative GC). Since Boehm GC is a conservative GC, it does not (and cannot) provide strong guarantees. But you hope that it won't have too much leak or memory waste (some papers mention a 20% leak rate as typical with Boehm's GC on Linux/x86-64). Boehm's GC has a page about Advantages and Disadvantages of Conservative Garbage Collection that you should absolutely read. And there is also a detailed description of it, and finally it is free software, so you can (and perhaps should) study its source code.

Both Boehm and valgrind use similar technology, so they cannot play well together. Obviously valgrind would detect a lot of memory leaks in any code using Boehm GC. Using valgrind on code linking Boehm's GC is useless. You might clear explicitly every memory zone obtained with GC_MALLOC.

If you need a precise GC (in particular if you need more guarantees about GC), choose something else, or code your own one (a naive precise mark&sweep stop-the-world GC is easy to code, at least in a single-thread program; the boring part is to maintain the GC roots and give access to your local "variables" containing pointers. You'll put them in some struct in every call frame, and link these struct-s together). Maybe look into Ravenbrook's MPS, or my old, unmaintained and buggy, Qish (perhaps it could inspire you). Look also into Ocaml GC and how you should interface C with Ocaml.

Read also the GC handbook.

BTW, your question is surprising: valgrind (its memcheck tool) is for hunting missing free-s, and the whole point of Boehm GC is to render free "useless" by providing a GC_MALLOC (replacement of malloc) which does not need any kind of freeing operation (so there is no point on using valgrind on a program which do GC_MALLOC-s).

Here is a Dockerfile based on your example which have installed nokogiri from Gemfile.

Dockerfile: