guest connect to a Unix domain socket on the host rather than the
other way around. The former is a QEMU feature (guestfwd to a
socket) while the latter requires a patch (which we can now get rid
of).
svn path=/nixos/branches/boot-order/; revision=22331
machine can now declare an option `virtualisation.vlans' that causes
it to have network interfaces connected to each listed virtual
network. For instance,
virtualisation.vlans = [ 1 2 ];
causes the machine to have two interfaces (in addition to eth0, used
by the test driver to control the machine): eth1 connected to
network 1 with IP address 192.168.1.<i>, and eth2 connected to
network 2 with address 192.168.2.<i> (where <i> is the index of the
machine in the `nodes' attribute set). On the other hand,
virtualisation.vlans = [ 2 ];
causes the machine to only have an eth1 connected to network 2 with
address 192.168.2.<i>. So each virtual network <n> is assigned the
IP range 192.168.<n>.0/24.
Each virtual network is implemented using a separate multicast
address on the host, so guests really cannot talk to networks to
which they are not connected.
* Added a simple NAT test to demonstrate this.
* Added an option `virtualisation.qemu.options' to specify QEMU
command-line options. Used to factor out some commonality between
the test driver script and the interactive test script.
svn path=/nixos/trunk/; revision=21928
console. This uses the `sendkey' command in the QEMU monitor.
* For the block/unblock primitives, use the `set_link' command in the
QEMU monitor.
svn path=/nixos/trunk/; revision=19854
it special commands such as "screendump", "sendkey" and so on.
* Take screenshots using the "screendump" command. This has the
advantage over "scrot" that it also supports taking a picture of the
console, and is not affected by weird X visuals.
svn path=/nixos/trunk/; revision=19837
its default behaviour is to stop the emulator (i.e. suspend the VM).
For automated tests, this is bad, because is makes the VM appear to
hang without any error message. The "werror=report" flag causes
QEMU to report the problem to the VM. As a side effect QEMU exits
very elegantly:
[ 2.308668] end_request: I/O error, dev vda, sector 534400
[ 2.309611] Buffer I/O error on device vda, logical block 66800
...
*** glibc detected *** /nix/store/yhngqrww53j0aw7z7v4bv948x5g5fc3d-qemu-kvm-0.12.1.2/bin/qemu-system-x86_64: double free or corruption (!prev): 0x08e3e040 ***
Aborted
So I guess we now depend on a bug in QEMU :-)
svn path=/nixos/trunk/; revision=19703
write some magic string to ttyS0. This removes the dependency on
having a CIFS mount.
* Use a thread to process the stdout/stderr of each QEMU instance.
* Add a kernel command line parameter "stage1panic" to tell stage 1 to
panic if an error occurs. This is faster than waiting until
connect() times out.
svn path=/nixos/trunk/; revision=19212
lib/build-vms.nix contains a function `buildVirtualNetwork' that
takes a specification of a network of machines (as an attribute set
of NixOS machine configurations) and builds a script that starts
each configuration in a separate QEMU/KVM VM and connects them
together in a virtual network. This script can be run manually to
test the VMs interactively. There is also a function `runTests'
that starts and runs the virtual network in a derivation, and
then executes a test specification that tells the VMs to do certain
things (i.e., letting one VM send an HTTP request to a webserver on
another VM). The tests are written in Perl (for now).
tests/subversion.nix shows a simple example, namely a network of two
machines: a webserver that runs the Subversion subservice, and a
client. Apache, Subversion and a few other packages are built with
coverage analysis instrumentation. For instance,
$ nix-build tests/subversion.nix -A vms
$ ./result/bin/run-vms
starts two QEMU/KVM instances. When they have finished booting, the
webserver can be accessed from the host through
http://localhost:8081/.
It also has a small test suite:
$ nix-build tests/subversion.nix -A report
This runs the VMs in a derivation, runs the tests, and then produces
a distributed code coverage analysis report (i.e. it shows the
combined coverage on both machines).
The Perl test driver program is in lib/test-driver. It executes
commands on the guest machines by connecting to a root shell running
on port 514 (provided by modules/testing/test-instrumentation.nix).
The VMs are connected together in a virtual network using QEMU's
multicast feature. This isn't very secure. At the very least,
other processes on the same machine can listen to or send packets on
the virtual network. On the plus side, we don't need to be root to
set up a multicast virtual network, so we can do it from a
derivation. Maybe we can use VDE instead.
(Moved from the vario repository.)
svn path=/nixos/trunk/; revision=16899
