Comment 3 for bug 1632252

I am able to reproduce this locally by using the "ssh" runner on a manually started QEMU instance, instead of the "qemu" runner directly; so this is much easier to investigate.

First this needs a small new feature in autopkgtest's ssh runner: https://anonscm.debian.org/cgit/autopkgtest/autopkgtest.git/commit/?id=9aa6fdbef . Easiest to just run autopkgtest from git (see reproducer below).

 * Take a standard autopkgtest yakkety image (autopkgtest-buildvm-ubuntu-cloud) and run it in QEMU:

  qemu-system-x86_64 -enable-kvm -m 4096 -smp 2 -nographic -drive file=path/to/autopkgtest-yakkety-amd64.img,if=virtio -net nic,model=virtio -net user,hostfwd=tcp::22000-:22

 * Log in (ubuntu/ubuntu) and scp/install your host's ssh key info ~/.ssh/authorized_keys

 * Run the test:

   git clone https://anonscm.debian.org/git/autopkgtest/autopkgtest.git
   autopkgtest/runner/autopkgtest --testname ubuntu-regression-suite linux -- ssh -H localhost -l ubuntu -p 22000 --reboot --capability=isolation-machine --capability=revert --capability=revert-full-system

This takes an hour or so, then the test fails with "testbed failure: testbed auxverb failed with exit code 255" and the console (in qemu) is completely dead. So notably this fails later than on the production infra (where it OOM-kills sshd during AppArmor tests), here it fails right after

  12:39:40 DEBUG| [stdout] selftests: breakpoint_test [PASS]

but this could be variance due to the infra instances having more CPUs and memory, or it's just an artifact of truncating the log earlier.