Re: [PATCH 00/13] kvm: selftests: add aarch64 framework and dirty

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

 



On Thu, Nov 01, 2018 at 10:08:25AM +0100, Christoffer Dall wrote:
> On Tue, Oct 30, 2018 at 06:38:20PM +0100, Andrew Jones wrote:
> > 
> > Hi Christoffer,
> > 
> > Thanks for your interest in these tests. There isn't any documentation
> > that I know of, but it's a good idea to have some. I'll write something
> > up soon. I'll also try to answer your questions now.
> > 
> 
> That sounds great, thanks!
> 
> > On Mon, Oct 29, 2018 at 06:40:02PM +0100, Christoffer Dall wrote:
> > > Hi Drew,
> > > 
> > > On Tue, Sep 18, 2018 at 07:54:23PM +0200, Andrew Jones wrote:
> > > > This series provides KVM selftests that test dirty log tracking on
> > > > AArch64 for both 4K and 64K guest page sizes. Additionally the
> > > > framework provides an easy way to test dirty log tracking with the
> > > > recently posted dynamic IPA and 52bit IPA series[1].
> > > 
> > > I was trying to parse the commit text of patch 2, and I realized that I
> > > don't understand the 'hypercall to userspace' thing at all, which
> > > probably means I have no idea how the selftests work overall.
> > 
> > There are three parts to a kvm selftest: 1) the test code which runs in
> > host userspace and _is_ the kvm userspace used with kvm, 2) the vcpu
> > thread code which executes KVM_RUN for the guest code and possibly also
> > some host userspace test code, and 3) the guest code, which is naturally
> > run in the vcpu thread, but in guest mode.
> > 
> > The need for a "ucall" arises for 2's "possibly also some host userspace
> > test code". In that case the guest code needs to invoke an exit from guest
> > mode, not just to kvm, but all the way to kvm userspace. For AArch64, as
> > you know, this can be done with an MMIO access. The reason patch 2
> > generalizes the concept is because for x86 this can and is done with a
> > PIO access.
> > 
> 
> So in the world of normal KVM userspace, (2) would be a thread in the
> same process as (1), sharing its mm.  Is this a different setup somehow,
> why?

It's the same setup. Actually the only difference is what Paolo pointed
out in his reply. There's no need for an independent vcpu thread to be
spawned when only one vcpu thread is needed and no additional main thread
is needed. I.e. the main test / kvm userspace code can call KVM_RUN
itself.

> 
> > > 
> > > I then spent a while reading various bits of documentation in the kernel
> > > tree, LWN, etc., only to realize that I don't understand how this test
> > > framework actually works.
> > > 
> > > Are the selftests modules, userspace programs, or code that is compiled
> > > with the kernel, and (somehow?) run from userspace.  I thought the
> > > latter, partially based on your explanation at ELC, but then I don't
> > > understand how the "compile and run" make target works.
> > 
> > The tests are standalone userspace programs which are compiled separately,
> > but have dependencies on kernel headers. As stated above, for kvm, each
> > selftest is a kvm userspace (including its vcpu thread code) and guest
> > code combined. While there's a lot of complexity in the framework,
> > particularly for memory management, and a bit for vcpu setup, most of that
> > can be shared among tests using the kvm_util.h and test_util.h APIs,
> > allowing a given test to only have a relatively simple main(), vcpu thread
> > "vcpu_worker()" function, and "guest_code()" function. Guest mode code can
> > easily share code with the kvm userspace test code (assuming the guest
> > page tables are set up in the default way) and even data can be shared as
> > long the accesses are done with the appropriate mappings (gva vs. hva).
> > There's a small API to help with that as well.
> > 
> 
> Sounds cool.  Beware of the attributes of the mappings such that both
> the guest and host have mapped the memory cacheable etc., but I'm sure
> you've thought of that already.

Right. If you look at virt_pg_map(), then you'll see that I have the
default set to NORMAL memory. It can be overridden by calling
_virt_pg_map() directly - which might be nice to do to specifically
test stage1/stage2 mapping combinations.

> 
> > > 
> > > Can you help me paint the overall picture, or point me to the piece of
> > > documentation/presentation that explains the high-level picture, which I
> > > must have obviously missed somehow?
> > 
> > We definitely need the documentation and, in hindsight, it looks like it
> > would have been a good BoF topic last week too.
> 
> An overview of the different testing approaches would be a good KVM
> Forum talk for next year, IMHO.  When should you use kvm-unit-tests, and
> when should you use kselftests, some examples, etc.  Just saying ;)

:-)

> 
> > 
> > I think this framework has a lot of potential for KVM API testing and
> > even for quick & dirty guest code instruction sequence tests (although
> > instruction sequences would also fit kvm-unit-tests). I hope I can help
> > get you and anyone else interested started.
> > 
> 
> I'll have a look at this series and glance at the code some more, it
> would be interesting to consider if using some of this for nested virt
> tests makes sense.

Yes. x86 has many nested tests. I think the framework was originally
created with that in mind.

Thanks,
drew



[Index of Archives]     [KVM ARM]     [KVM ia64]     [KVM ppc]     [Virtualization Tools]     [Spice Development]     [Libvirt]     [Libvirt Users]     [Linux USB Devel]     [Linux Audio Users]     [Yosemite Questions]     [Linux Kernel]     [Linux SCSI]     [XFree86]

  Powered by Linux