4:12 a.m.
On 28.03.20 08:47, PRAKHAR BANSAL wrote:
Hi Jan,
Thanks for the reply!
I was only considering the command-line tool "code" for reference to the
Jailhouse kernel API(ioctl calls) because I didn't find a documentation
of the Jailhouse kernel APIs.
Right, the IOCTL API is not documented so far. It is currently only used
inside the Jailhouse project. This needs to be formalized when there
shall be external users like a libvirt driver.
That might be a nice small contribution task: Create some
Documentation/driver-interfaces.md that describes the IOCTLs along with
their parameter structures and that also includes the current
sysfs-entries.txt as a section. Then send this as patch here. I'll help
out when details are not clear from reading the code.
For the second part as you mentioned that Jailhouse can only create
cells with the constraints defined in the root cell configuration. I
have a few questions regarding that.
1. Is there a way to know if Jailhouse is enabled on the host and get
the root cell configuration(s) from Jailhouse through an API? This can
be used while binding the libvirt to the Jailhouse hypervisor.
Look at
https://github.com/siemens/jailhouse/blob/master/Documentation/sysfs-entr...
for what is reported as runtime information. Full configurations can't
be read back at this point. This might be reconsidered in the light of
[1], but I wouldn't plat for that yet.
2. If Jailhouse is not enabled(again can we know this using some API)
then, can libvirt enable/disable Jailhouse during the libvirt binding of
the Jailhouse driver with a given set of Jailhouse cell configurations
describing a complete partitioned system?
With the API above and a given configuration set, yes. The config set
would have to be provided to the libvirt driver in some to-be-defined
way (e.g. /etc/libvirt/jailhouse.conf -> /etc/libvirt/jailhouse/*.cell).
3. I was wondering, as you mentioned that libvirt driver should check
for mismatch of the cell configuration with the root cell configuration,
the question is, isn't that done by Jailhouse itself? If yes, then
libvirt can just pass on the cell creation requests to Jailhouse and
return the response to the user as it is, rather than driver doing any
such mismatch check.
With matching I'm referring to a libvirt user request like "create a
domain with 2 CPUs", while there are no cells left that have more than
one CPU. Or "give the domain 1G RAM", and you need to find an available
cell with that much memory. Those are simple examples. A request that
states "connect the domain to the host network A" implies that a cell
has a shared-memory link to, say, the root cell that can be configured
to bridge this. But let's keep that for later and start as simple as
possible.
Jan
[1]
https://groups.google.com/d/msgid/jailhouse-dev/CADiTV-1QiRhSWZnw%2BkHhJM...
-Prakhar
On Thu, Mar 26, 2020 at 1:49 AM Jan Kiszka <jan.kiszka(a)web.de
<mailto:jan.kiszka@web.de>> wrote:
Hi Prakhar,
On 25.03.20 05:36, PRAKHAR BANSAL wrote:
> Hi Jan,
>
> Thanks for the reply. I looked deeper into the libvirt and Jailhouse
> source code and found following two things that seem relevant to the
> project I am interested in.
>
> - Libvirt driver interface at [libvirt.git]
> <https://libvirt.org/git/?p=libvirt.git;a=tree;hb=HEAD> / src
> <https://libvirt.org/git/?p=libvirt.git;a=tree;f=src;hb=HEAD> /
driver.h
>
<https://libvirt.org/git/?p=libvirt.git;a=blob_plain;f=src/driver.h;hb=HEA...
> - Jailhouse tool, which is using the ioctl API of the Jailhouse,
> available at
> https://github.com/siemens/jailhouse/blob/master/tools/jailhouse.c.
>
> With the help of the above two, it looks like, a libvirt driver
for the
> Jailhouse can be implemented. Let me know if I am moving in the right
> direction so far.
From the Jailhouse perspective, it is important to not consider the
command line tool an interface anymore (like in the first prototype) but
build on top of the Linux driver API (ioctls, sysfs). There is already a
Python library which started to abstract this interface for
Jailhouse-internal use cases. However, I strongly suspect libvirt will
rather want a native binding.
>
> I have been looking at the other libvirt driver implementations for
> hypervisors like HyperV and VMware to understand their implementation
> and learn from there.
As Jailhouse is a static partitioning hypervisor without abstraction of
the underlying hardware, your starting point for the libvirt binding
should be a given set of Jailhouse cell configurations describing a
complete partitioned system. So rather than instantiating on demand a
domain (Jailhouse cell) with, say, a network adapter, the driver should
match a user request for a domain against the configuration set and use
what is there - or report the mismatch. What it could organize, though,
is interconnecting cells that have a (preconfigured) virtual network
link to the root cell.
Due to this different concept, there will be no 1:1 mapping for
commodity hypervisor drivers to the Jailhouse scenario. Still, studying
what they do is useful and needed in order to understand what "normally"
happens and find a reasonable translation. This is probably the most
challenging part of the project.
Jan