Segfault-on-exit due to lazy LibGit2 initialization

[timholy]

#28113 reduced loading times (yay!) by lazy-initialization of libgit2, however, now certain Revise tests lead to a segfault when Julia is exited. Here's a small reproducer:

julia> import LibGit2

julia> path = dirname(dirname(Sys.BINDIR))
"/home/tim/src/julia-1.0"

julia> repo = LibGit2.GitRepo(path)
LibGit2.GitRepo("/home/tim/src/julia-1.0")

julia> status = LibGit2.GitStatus(repo;
                                  status_opts=LibGit2.StatusOptions(flags=LibGit2.Consts.STATUS_OPT_INCLUDE_UNMODIFIED))
LibGit2.GitStatus(LibGit2.GitRepo("/home/tim/src/julia-1.0"), Ptr{Nothing} @0x0000555b2d20e0f0)

julia> exit()

signal (11): Segmentation fault
in expression starting at no file:0
git_strmap_lookup_index at /home/tim/src/julia-1.0/usr/bin/../lib/libgit2.so (unknown line)
unknown function (ip: 0x1)
Allocations: 11218380 (Pool: 11216399; Big: 1981); GC: 24
Segmentation fault (core dumped)

You can avoid the segfault by adding

julia> status = repo = 0

julia> GC.gc()

before exit(), so it seems likely to be a resource-cleanup issue.

This segfault does not happen if you revert c670f1a.

[bjarthur]

thanks for the hack around @timholy !

[timholy]

@StefanKarpinski, since this was caused by #28113, any chance this is on your radar?

[StefanKarpinski]

No idea, but is it possible that this uncovers some use case whereby the libgit2 initialization doesn't happen before usage? It's possible I missed something when making the initialization lazy.

[timholy]

From the symptoms, I'm guessing it's more about the order in which various resources get shut down or something. But I haven't poked at this, so take it for what it's worth::Nothing.