Julia: Intermittent libgit2 test segfault

[tkelman]

Can be reproduced on current master on 64 bit linux by running make test-libgit2 a bunch of times in a row. Looks like

signal (11): Segmentation fault
while loading no file, in expression starting on line 0
git_config_free at /home/tkelman2/Julia/julia-0.6/usr/bin/../lib/libgit2.so (unknown line)
close at ./libgit2/types.jl:547
unknown function (ip: 0x2b3d1e01712a)
jl_call_fptr_internal at /home/tkelman2/Julia/julia-0.6/src/julia_internal.h:337 [inlined]
jl_call_method_internal at /home/tkelman2/Julia/julia-0.6/src/julia_internal.h:356 [inlined]
jl_apply_generic at /home/tkelman2/Julia/julia-0.6/src/gf.c:1930
jl_apply at /home/tkelman2/Julia/julia-0.6/src/julia.h:1422 [inlined]
run_finalizer at /home/tkelman2/Julia/julia-0.6/src/gc.c:111
jl_gc_run_finalizers_in_list at /home/tkelman2/Julia/julia-0.6/src/gc.c:200
run_finalizers at /home/tkelman2/Julia/julia-0.6/src/gc.c:234
jl_atexit_hook at /home/tkelman2/Julia/julia-0.6/src/init.c:265
main at /home/tkelman2/Julia/julia-0.6/ui/repl.c:265
__libc_start_main at /build/eglibc-oGUzwX/eglibc-2.19/csu/libc-start.c:287
unknown function (ip: 0x4015fa)
Allocations: 9425217 (Pool: 9416101; Big: 9116); GC: 13
Segmentation fault
make[1]: *** [libgit2] Error 139
make: *** [test-libgit2] Error 2

Not sure yet whether this was introduced by #21603 or #21597. 36b056e547344ad3f4828067b174a240ecbb7475 can reproduce the problem, and at 991b581c2723c57fc0d5005c1c7b61a1b0458308 I can run make test-libgit2 40 times in a row with no problems. Still testing at the commits between there.

[yuyichao]

Caught in rr locally, will have a breif look....

[yuyichao]

Currently reduced to

mktempdir() do dir
    # test parameters
    cache_repo = joinpath(dir, "Example")

    repo = LibGit2.init(cache_repo)
    close(repo)

    up_path = joinpath(dir, "Example.PushUp")
    up_repo = LibGit2.clone(cache_repo, up_path)
    our_path = joinpath(dir, "Example.Push")
    our_repo = LibGit2.clone(cache_repo, our_path) # 0->1
    # need to set this for merges to succeed
    global our_cfg = LibGit2.GitConfig(our_repo) # 1->2
    LibGit2.set!(our_cfg, "user.name", "AAAA")
    LibGit2.set!(our_cfg, "user.email", "[email protected]")
    global up_cfg = LibGit2.GitConfig(up_repo)
    LibGit2.set!(up_cfg, "user.name", "AAAA")
    LibGit2.set!(up_cfg, "user.email", "[email protected]")
    try
        open(joinpath(LibGit2.path(our_repo),"file1"),"w") do f
            write(f, "111\n")
        end
        LibGit2.add!(our_repo, "file1")
        LibGit2.commit(our_repo, "add file1")
        # we cannot yet locally push to non-bare repos
        try
            LibGit2.push(our_repo, remoteurl=up_path) # 2->1
        end
    finally
        close(our_repo) # 1->0
        close(up_repo)
    end
end
ccall(:jl_, Void, (Any,), our_cfg)
close(our_cfg)

Note the comment about the refcount on our_cfg. Somehow the LibGit2.push(our_repo, remoteurl=up_path) decreases the refcount on the config

[yuyichao]

This is an upstream bug. Ref https://github.com/libgit2/libgit2/pull/4224. I'll push the fix to https://github.com/JuliaLang/julia/pull/21636 to reduce conflicts. Also, this get's triggered on travis on that PR .... https://travis-ci.org/JuliaLang/julia/jobs/227153347

[yuyichao]

This might also explain some of the hangs on the CI since the double free corrupts libc internal datastructure causing random error when printing error messages....