Dynamically allocate the alternate signal stack on 4.09 and earlier

[dra27]

Glibc 2.34 is starting to trickle into mainstream distros (see repology for more detail). In particular, Ubuntu 21.10 now has it, as will Fedora 35.

This affects opam installations of older compilers on those distros, so as we did with the GCC 10 -fcommon change, we'll expect to carry patches in opam-repository.

In previous dev meetings, we've discussed (and IIRC agreed) to committing these changes to the maintenance branches. The 4.10-4.12 backport of #10266 is trivial. Here is a candidate for 4.09 and earlier, as these versions don't include #8670.

cc @xavierleroy, @stedolan and all other signals experts 🙂

[dra27]

I'm attempting to run this through precheck#631 - I'm not sure if AppVeyor still works on the 4.09 branch and obviously we turned off Travis ages ago.

[dra27]

No, precheck doesn't work - I have however tested on Linux locally.

[dra27]

Aha: it's running less unsuccessfully in precheck#632. macOS 10.15 and OpenBSD are failing because those workers are no longer configured to have libbfd and OmniOS isn't supported.

[dra27]

FreeBSD failure is #9981 (also probably a candidate for an opam-repository, but not right now)

[xavierleroy]

I think it makes sense to backport #10266, but I was hoping to fix the leak reported in #10698 before, and I just submitted a PR that does this: #10726.

[dra27]

For 4.09 and earlier, we have to back-port #8670 as well - is that what you intend? The full back-port of 10266 for 4.10 and 4.12 are presently on my own fork (e.g. https://github.com/dra27/ocaml/commits/4.10) ready to push - I also agree with adding #10698 to it before pushing that here and to opam-repository.

[xavierleroy]

For 4.09 and earlier, we have to back-port #8670 as well - is that what you intend?

I don't know. If all we want for 4.09 and earlier is to compile with Glibc 2.34, we don't have to back-port #8670: we can elect to keep stack overflow handling in threads as broken as before. Likewise, avoiding the memory leak like #10698 does may not be absolutely necessary.

[Zimmi48]

Hello! Is there any hope to make progress on this backport? It looks like we might have an issue related to this making it impossible to install the Coq platform with opam on Ubuntu 21.10 (because some included packages are not yet compatible with OCaml 4.13 and earlier versions of OCaml do not compile on Ubuntu 21.10).

[dra27]

At the moment, I've been waiting until #10726 is merged to decide whether to back-port that part of the machinery to 4.09 and earlier.

[xavierleroy]

Hello! Is there any hope to make progress on this backport?

Should make progress pretty soon. Please review #10726 ASAP and we'll make a backport, with or without #10726 depending on the outcome of the review.

[cheese]

Great job! Tested 4.08.1 build on CentOS Stream 9.

[dra27]

This will be pushing pushed directly, rather than merged through a PR and this version is out of date. See dra27#56 for the actual proposed commits for the 4.09 branch.