Message-ID: <20200925222220.GA977797@google.com>
Date:   Fri, 25 Sep 2020 15:22:20 -0700
From:   Vipin Sharma <vipinsh@...gle.com>
To:     Tom Lendacky <thomas.lendacky@....com>,
        Sean Christopherson <sean.j.christopherson@...el.com>,
        pbonzini@...hat.com
Cc:     tj@...nel.org, lizefan@...wei.com, joro@...tes.org, corbet@....net,
        brijesh.singh@....com, jon.grimm@....com, eric.vantassell@....com,
        gingell@...gle.com, rientjes@...gle.com, kvm@...r.kernel.org,
        x86@...nel.org, cgroups@...r.kernel.org, linux-doc@...r.kernel.org,
        linux-kernel@...r.kernel.org
Subject: Re: [RFC Patch 0/2] KVM: SVM: Cgroup support for SVM SEV ASIDs

On Thu, Sep 24, 2020 at 02:55:18PM -0500, Tom Lendacky wrote:
> On 9/24/20 2:21 PM, Sean Christopherson wrote:
> > On Tue, Sep 22, 2020 at 02:14:04PM -0700, Vipin Sharma wrote:
> > > On Mon, Sep 21, 2020 at 06:48:38PM -0700, Sean Christopherson wrote:
> > > > On Mon, Sep 21, 2020 at 05:40:22PM -0700, Vipin Sharma wrote:
> > > > > Hello,
> > > > > 
> > > > > This patch series adds a new SEV controller for tracking and limiting
> > > > > the usage of SEV ASIDs on the AMD SVM platform.
> > > > > 
> > > > > SEV ASIDs are used in creating encrypted VM and lightweight sandboxes
> > > > > but this resource is in very limited quantity on a host.
> > > > > 
> > > > > This limited quantity creates issues like SEV ASID starvation and
> > > > > unoptimized scheduling in the cloud infrastructure.
> > > > > 
> > > > > SEV controller provides SEV ASID tracking and resource control
> > > > > mechanisms.
> > > > 
> > > > This should be genericized to not be SEV specific.  TDX has a similar
> > > > scarcity issue in the form of key IDs, which IIUC are analogous to SEV ASIDs
> > > > (gave myself a quick crash course on SEV ASIDs).  Functionally, I doubt it
> > > > would change anything, I think it'd just be a bunch of renaming.  The hardest
> > > > part would probably be figuring out a name :-).
> > > > 
> > > > Another idea would be to go even more generic and implement a KVM cgroup
> > > > that accounts the number of VMs of a particular type, e.g. legacy, SEV,
> > > > SEV-ES?, and TDX.  That has potential future problems though as it falls
> > > > apart if hardware every supports 1:MANY VMs:KEYS, or if there is a need to
> > > > account keys outside of KVM, e.g. if MKTME for non-KVM cases ever sees the
> > > > light of day.
> > > 
> > > I read about the TDX and its use of the KeyID for encrypting VMs. TDX
> > > has two kinds of KeyIDs private and shared.
> > 
> > To clarify, "shared" KeyIDs are simply legacy MKTME KeyIDs.  This is relevant
> > because those KeyIDs can be used without TDX or KVM in the picture.
> > 
> > > On AMD platform there are two types of ASIDs for encryption.
> > > 1. SEV ASID - Normal runtime guest memory encryption.
> > > 2. SEV-ES ASID - Extends SEV ASID by adding register state encryption with
> > > 		 integrity.
> > > 
> > > Both types of ASIDs have their own maximum value which is provisioned in
> > > the firmware
> > 
> > Ugh, I missed that detail in the SEV-ES RFC.  Does SNP add another ASID type,
> > or does it reuse SEV-ES ASIDs?  If it does add another type, is that trend
> > expected to continue, i.e. will SEV end up with SEV, SEV-ES, SEV-ES-SNP,
> > SEV-ES-SNP-X, SEV-ES-SNP-X-Y, etc...?
> 
> SEV-SNP and SEV-ES share the same ASID range.
> 
> Thanks,
> Tom
> 
> > 
> > > So, we are talking about 4 different types of resources:
> > > 1. AMD SEV ASID (implemented in this patch as sev.* files in SEV cgroup)
> > > 2. AMD SEV-ES ASID (in future, adding files like sev_es.*)
> > > 3. Intel TDX private KeyID
> > > 4. Intel TDX shared KeyID
> > > 
> > > TDX private KeyID is similar to SEV and SEV-ES ASID. I think coming up
> > > with the same name which can be used by both platforms will not be easy,
> > > and extensible with the future enhancements. This will get even more
> > > difficult if Arm also comes up with something similar but with different
> > > nuances.
> > 
> > Honest question, what's easier for userspace/orchestration layers?  Having an
> > abstract but common name, or conrete but different names?  My gut reaction is
> > to provide a common interface, but I can see how that could do more harm than
> > good, e.g. some amount of hardware capabilitiy discovery is possible with
> > concrete names.  And I'm guessing there's already a fair amount of vendor
> > specific knowledge bleeding into userspace for these features...

I agree with you that the abstract name is better than the concrete
name, I also feel that we must provide HW extensions. Here is one
approach:

Cgroup name: cpu_encryption, encryption_slots, or memcrypt (open to
suggestions)

Control files: slots.{max, current, events}

Contents of the slot.max:
default max
	default: Corresponds to all kinds of encryption capabilities on
		 a platform. For AMD, it will be SEV and SEV-ES.  For
		 Intel, it will be TDX and MKTME. This can also be used
		 by other devices not just CPU.

	max: max or any number to denote limit on the cgroup.

A user who wants the finer control, then they need to know about the
capabilities a platform provides and use them, e.g. on AMD:

$ echo "sev-es 1000" > slot.max
$ cat slots.max
default max sev-es 1000

This means that in the cgroup maximum SEV-ES ASIDs which can be used is
1000 and SEV ASIDs is max (default, no limit).  Each platform can
provide their own extensions which can be overwritten by a user,
otherwise extensions will have the default limit.

This is kind of similar to the IO and the rdma controller.

I think it is keeping abstraction for userspace and also providing finer
control for HW specific features.

What do you think about the above approach?  

> > 
> > And if SNP is adding another ASID namespace, trying to abstract the types is
> > probably a lost cause.
> > 
> >  From a code perspective, I doubt it will matter all that much, e.g. it should
> > be easy enough to provide helpers for exposing a new asid/key type.
> > 
> > > I like the idea of the KVM cgroup and when it is mounted it will have
> > > different files based on the hardware platform.
> > 
> > I don't think a KVM cgroup is the correct approach, e.g. there are potential
> > use cases for "legacy" MKTME without KVM.  Maybe something like Encryption
> > Keys cgroup?

Added some suggestion in the above approach. I think we should not add
key in the name because here limitation is on the number of keys that
can be used simultaneously. 

> > 
> > > 1. KVM cgroup on AMD will have:
> > > sev.max & sev.current.
> > > sev_es.max & sev_es.current.
> > > 
> > > 2. KVM cgroup mounted on Intel:
> > > tdx_private_keys.max
> > > tdx_shared_keys.max
> > > 
> > > The KVM cgroup can be used to have control files which are generic (no
> > > use case in my mind right now) and hardware platform specific files
> > > also.
> > 
> > My "generic KVM cgroup" suggestion was probably a pretty bad suggestion.
> > Except for ASIDs/KeyIDs, KVM itself doesn't manage any constrained resources,
> > e.g. memory, logical CPUs, time slices, etc... are all generic resources that
> > are consumed by KVM but managed elsewhere.  We definitely don't want to change
> > that, nor do I think we want to do anything, such as creating a KVM cgroup,
> > that would imply that having KVM manage resources is a good idea.
> > 

Powered by blists - more mailing lists