Re-enable rr for Pkg tests

[Keno]

@DilumAluthge notes that rr is currently disabled when running the Pkg tests on CI because libgit2 blows up rr trace file size due to excessive syscall traffic. This is an upstream libgit2 issue: libgit2/libgit2#6038. Opening this hear to record the knowledge and track the action item of updating the CI config once upstream libgit2 is fixed.

Our configuration:

julia/.buildkite/pipelines/main/platforms/tester_linux.arches

Line 13 in f728038

linux 64_g3_st g3 . 64 x86_64 . . . 3 . yes . v4.8 2a058481b567f0e91b9aa3ce4ad4f09e6419355a

[vtjnash]

Can we build libgit2 with mmap disabled?

[Keno]

Potentially. Reading the upstream discussion, it appears that this used to be done with read/write but was changed to work around coherence issues mmaping network drives on Windows. So hopefully changing this back should be easy, though I don't think there's a build flag for it or anything.

[Keno]

Actually, libgit2/libgit2#6039, may have already done exactly that. So we should probably try building with that and seeing if the issue is resolved.

[giordano]

That seems to be in 1.3.0: JuliaPackaging/Yggdrasil#3962

[vtjnash]

Yeah, my suggestion was they should completely eliminate use of mmap everywhere, since read/write are usually much more optimal and reliable, particularly once adding threads. Let's see how it goes with that PR to start.

[vtjnash]

We could still build with NO_MMAP, but it seems that this is already tracked upstream in libgit2 (libgit2/libgit2#6038), so we don't need to track it here

[Keno]

We've had to re-disable Pkg tests on CI. A brief investigation shows that the primary source of syscall traffic is a 20 byte read done using a full mmap/munmap inside git_packfile_unpack_header. I think this could be fixed in libgit2, but for now just writing down the observation.

[KristofferC]

This doesn't feel like it has to be on the milestone. Can be reenabled whenever libgit2 has better support for rr but I don't see why we should tie that to our release.