While some increase was necessary anyway because the idle system requires
about 5-10 MiB more memory, the main issue is resolving the code line and
function name in case of a memory leak. Calling addr2line requires a lot
more memory than before. Using backtraces via libbfd doesn't help either
because the trigger is the bfd_find_nearest_line() call we use as well.
And because we'd try to resolve all symbols that way (for whitelisting),
the memory overhead would be even higher and affect every shutdown, even
if no leak occurred. It also causes a significant time overhead (running
all tests took 75m instead of 48m).
I also tested switching to ASAN/LSAN. The peak memory usage is slightly
higher than when using libbfd, but enabling it also increased the runtime
overhead a lot (the daemon and swanctl both required about 10-20 MiB more
memory, not just during the shutdown).
The services running on alice seem to require a bit more memory with
Debian bookworm, so increase the memory allocation. But at the same
time reduce winnetou's allocation by the same amount as it really doesn't
require that much memory.
The unit change makes it easier to read.
It might not exist on all platforms and according to the man page:
The kvm wrapper script is used to provide compatibility with old
qemu-kvm package which has been merged into qemu as of version 1.3.
The script executes
qemu-system-x86_64 -enable-kvm
passing all other command-line arguments to the qemu binary.
Closesstrongswan/strongswan#385.
This should give us the best performance and feature set on modern
hardware (in particular when compared to code2duo, which e.g. does not allow
nested virtualization).
Closesstrongswan/strongswan#340.
It's ever so close with strongTNC, sometimes the OOM killer got triggered
and the tests failed, or even worse, the whole guest system got stuck.
This might just be enough for now.
Newer versions of systemd etc. seem to require quite a lot of entropy
from /dev/random while booting, which can block and therefore delay the
start of other services (in particular sshd) by more than a minute.
Using the host's /dev/urandom via VirtIO RNG, we can avoid blocking the
guests.
The required kernel options are added for kernel versions 5.4+.
This avoids having to copy testresults, makes results of cancelled runs
browsable (runs may actually be followed live) and preserves old results
when rebuilding guest images (e.g. when using the build-strongswan script).
The number of consecutive test runs without any intermittent rebuild of the
guest images is also not limited by the image size anymore.
This allows accessing the guests with `virsh console <name>`.
Using a serial console would also be possible but our kernel configs
have no serial drivers enabled, CONFIG_VIRTIO_CONSOLE is enabled though.
So to avoid having to recompile the kernels let's do it this way, only
requires rebuilding the guest images.
References #729.
This simplifies capturing traffic with Wireshark on the host as each of
the guest's interfaces is clearly identified.
The three bridges were previously numbered starting from 0, this scheme
is restored here.
Passthrough mode only works as expected when running as root. On
Debian/Ubuntu systems qemu runs as user 'libvirt-qemu' and group 'kvm'
so all shared files must be chowned to grant access from guests.
Symlinks created on the host are still problematic because the Plan 9
filesystem has no direct notion of symbolic links, see [1].
[1] - http://ericvh.github.com/9p-rfc/rfc9p2000.u.html
Guest and network configuration is setup using the libvirt
virtualization API. The [start|stop]_testing scripts have been updated
accordingly.
qemu/KVM does not currently support a hostfs, so the shared build tree
mount has been dropped for now.
