Message-ID: <CA+EHjTwTgEVtea7wgef5G6EEgFa0so_GbNXTMZNKyFE=ucyV0g@mail.gmail.com>
Date:   Thu, 5 Oct 2023 15:42:42 +0100
From:   Fuad Tabba <tabba@...gle.com>
To:     Sean Christopherson <seanjc@...gle.com>
Cc:     Paolo Bonzini <pbonzini@...hat.com>, Marc Zyngier <maz@...nel.org>,
        Oliver Upton <oliver.upton@...ux.dev>,
        Huacai Chen <chenhuacai@...nel.org>,
        Michael Ellerman <mpe@...erman.id.au>,
        Anup Patel <anup@...infault.org>,
        Paul Walmsley <paul.walmsley@...ive.com>,
        Palmer Dabbelt <palmer@...belt.com>,
        Albert Ou <aou@...s.berkeley.edu>,
        "Matthew Wilcox (Oracle)" <willy@...radead.org>,
        Andrew Morton <akpm@...ux-foundation.org>,
        Paul Moore <paul@...l-moore.com>,
        James Morris <jmorris@...ei.org>,
        "Serge E. Hallyn" <serge@...lyn.com>, KVM <kvm@...r.kernel.org>,
        "moderated list:ARM64 PORT (AARCH64 ARCHITECTURE)" 
        <linux-arm-kernel@...ts.infradead.org>,
        KVMARM <kvmarm@...ts.linux.dev>,
        LinuxMIPS <linux-mips@...r.kernel.org>,
        linuxppc-dev@...ts.ozlabs.org, kvm-riscv@...ts.infradead.org,
        linux-riscv@...ts.infradead.org, linux-fsdevel@...r.kernel.org,
        linux-mm@...ck.org, linux-security-module@...r.kernel.org,
        open list <linux-kernel@...r.kernel.org>,
        Chao Peng <chao.p.peng@...ux.intel.com>,
        Jarkko Sakkinen <jarkko@...nel.org>,
        Anish Moorthy <amoorthy@...gle.com>,
        Yu Zhang <yu.c.zhang@...ux.intel.com>,
        Isaku Yamahata <isaku.yamahata@...el.com>,
        Xu Yilun <yilun.xu@...el.com>,
        Vlastimil Babka <vbabka@...e.cz>,
        Vishal Annapurve <vannapurve@...gle.com>,
        Ackerley Tng <ackerleytng@...gle.com>,
        Maciej Szmigiero <mail@...iej.szmigiero.name>,
        David Hildenbrand <david@...hat.com>,
        Quentin Perret <qperret@...gle.com>,
        Michael Roth <michael.roth@....com>,
        Wang <wei.w.wang@...el.com>,
        Liam Merwick <liam.merwick@...cle.com>,
        Isaku Yamahata <isaku.yamahata@...il.com>,
        "Kirill A . Shutemov" <kirill.shutemov@...ux.intel.com>
Subject: Re: [RFC PATCH v12 11/33] KVM: Introduce per-page memory attributes

Hi Sean,

On Tue, Oct 3, 2023 at 9:51 PM Sean Christopherson <seanjc@...gle.com> wrote:
>
> On Tue, Oct 03, 2023, Fuad Tabba wrote:
> > On Tue, Oct 3, 2023 at 4:59 PM Sean Christopherson <seanjc@...gle.com> wrote:
> > > On Tue, Oct 03, 2023, Fuad Tabba wrote:
> > > > > +#define KVM_MEMORY_ATTRIBUTE_PRIVATE           (1ULL << 3)
> > > > > +
> > > >
> > > > In pKVM, we don't want to allow setting (or clearing) of PRIVATE/SHARED
> > > > attributes from userspace.
> > >
> > > Why not?  The whole thing falls apart if userspace doesn't *know* the state of a
> > > page, and the only way for userspace to know the state of a page at a given moment
> > > in time is if userspace controls the attributes.  E.g. even if KVM were to provide
> > > a way for userspace to query attributes, the attributes exposed to usrspace would
> > > become stale the instant KVM drops slots_lock (or whatever lock protects the attributes)
> > > since userspace couldn't prevent future changes.
> >
> > I think I might not quite understand the purpose of the
> > KVM_SET_MEMORY_ATTRIBUTES ABI. In pKVM, all of a protected guest's memory is
> > private by default, until the guest shares it with the host (via a
> > hypercall), or another guest (future work). When the guest shares it,
> > userspace is notified via KVM_EXIT_HYPERCALL. In many use cases, userspace
> > doesn't need to keep track directly of all of this, but can reactively un/map
> > the memory being un/shared.
>
> Yes, and then userspace needs to tell KVM, via KVM_SET_MEMORY_ATTRIBUTES, that
> userspace has agreed to change the state of the page.  Userspace may not need/want
> to explicitly track the state of pages, but userspace still needs to tell KVM what
> userspace wants.
>
> KVM is primarily an accelerator, e.g. KVM's role is to make things go fast (relative
> to doing things in userspace) and provide access to resources/instructions that
> require elevated privileges.  As a general rule, we try to avoid defining the vCPU
> model, security policies, etc. in KVM, because hardcoding policy into KVM (and the
> kernel as a whole) eventually limits the utility of KVM.
>
> As it pertains to PRIVATE vs. SHARED, KVM's role is to define and enforce the basic
> rules, but KVM shouldn't do things like define when it is (il)legal to convert
> memory to/from SHARED, what pages can be converted, what happens if the guest and
> userspace disagree, etc.

Thanks for clarifying that. My initial understanding of the purpose of
KVM_SET_MEMORY_ATTRIBUTES wasn't clear. Now I see how having the
userspace view in KVM would avoid the need to hardcore many policies,
and I can see how this could come in handy in the future when we start
going into multi-sharing, for example.

> > > Why does pKVM need to prevent userspace from stating *its* view of attributes?
> > >
> > > If the goal is to reduce memory overhead, that can be solved by using an internal,
> > > non-ABI attributes flag to track pKVM's view of SHARED vs. PRIVATE.  If the guest
> > > attempts to access memory where pKVM and userspace don't agree on the state,
> > > generate an exit to userspace.  Or kill the guest.  Or do something else entirely.
> >
> > For the pKVM hypervisor the guest's view of the attributes doesn't
> > matter. The hypervisor at the end of the day is the ultimate arbiter
> > for what is shared and with how. For pKVM (at least in my port of
> > guestmem), we use the memory attributes from guestmem essentially to
> > control which memory can be mapped by the host.
>
> The guest's view absolutely matters.  The guest's view may not be expressed at
> access time, e.g. as you note below, pKVM and other software-protected VMs don't
> have a dedicated shared vs. private bit like TDX and SNP.  But the view is still
> there, e.g. in the pKVM model, the guest expresses its desire for shared vs.
> private via hypercall, and IIRC, the guest's view is tracked by the hypervisor
> in the stage-2 PTEs.  pKVM itself may track the guest's view on things, but the
> view is still the guest's.

This was poorly worded on my part. You're right that in practice the
pKVM hypervisor is the one tracking the guest's view, based on the
hypercalls from the guest.

> E.g. if the guest thinks a page is private, but in reality KVM and host userspace
> have it as shared, then the guest may unintentionally leak data to the untrusted
> world.
>
> IIUC, you have implemented guest_memfd support in pKVM by changing the attributes
> when the guest makes the hypercall.  This can work, but only so long as the guest
> and userspace are well-behaved, and it will likely paint pKVM into a corner in
> the long run.
>
> E.g. if the guest makes a hypercall to convert memory to PRIVATE, but there is
> no memslot or the memslot doesn't support private memory, then unless there is
> policy baked into KVM, or an ABI for the guest<=>host hypercall interface that
> allows unwinding the program counter, you're stuck.  Returning an error for the
> hypercall straight from KVM is undesirable as that would put policy into KVM that
> doesn't need to be there, e.g. that would prevent userspace from manipulating
> memslots in response to (un)share requests from the guest.  It's a similar story
> if KVM marks the page as PRIVATE, as that would prevent userspace from returning
> an error for the hypercall, i.e. would prevent usersepace from denying the request
> to convert to PRIVATE.

Ack.

> > One difference between pKVM and TDX (as I understand it), is that TDX
> > uses the msb of the guest's IPA to indicate whether memory is shared
> > or private, and that can generate a mismatch on guest memory access
> > between what it thinks the state is, and what the sharing state in
> > reality is. pKVM doesn't have that. Memory is private by default, and
> > can be shared in-place, both in the guest's IPA space as well as the
> > underlying physical page.
>
> TDX's shared bit and SNP's encryption bit are just a means of hardware enforcement.
> pKVM does have a hardware bit because hardware doesn't provide any enforcement.
> But as above, pKVM does have an equivalent *somewhere*.
>
> > > > The other thing, which we need for pKVM anyway, is to make
> > > > kvm_vm_set_mem_attributes() global, so that it can be called from outside of
> > > > kvm_main.c (already have a local patch for this that declares it in
> > > > kvm_host.h),
> > >
> > > That's no problem, but I am definitely opposed to KVM modifying attributes that
> > > are owned by userspace.
> > >
> > > > and not gate this function by KVM_GENERIC_MEMORY_ATTRIBUTES.
> > >
> > > As above, I am opposed to pKVM having a completely different ABI for managing
> > > PRIVATE vs. SHARED.  I have no objection to pKVM using unclaimed flags in the
> > > attributes to store extra metadata, but if KVM_SET_MEMORY_ATTRIBUTES doesn't work
> > > for pKVM, then we've failed miserably and should revist the uAPI.
> >
> > Like I said, pKVM doesn't need a userspace ABI for managing PRIVATE/SHARED,
> > just a way of tracking in the host kernel of what is shared (as opposed to
> > the hypervisor, which already has the knowledge). The solution could simply
> > be that pKVM does not enable KVM_GENERIC_MEMORY_ATTRIBUTES, has its own
> > tracking of the status of the guest pages, and only selects KVM_PRIVATE_MEM.
>
> At the risk of overstepping my bounds, I think that effectively giving the guest
> full control over what is shared vs. private is a mistake.  It more or less locks
> pKVM into a single model, and even within that model, dealing with errors and/or
> misbehaving guests becomes unnecessarily problematic.
>
> Using KVM_SET_MEMORY_ATTRIBUTES may not provide value *today*, e.g. the userspace
> side of pKVM could simply "reflect" all conversion hypercalls, and terminate the
> VM on errors.  But the cost is very minimal, e.g. a single extra ioctl() per
> converion, and the upside is that pKVM won't be stuck if a use case comes along
> that wants to go beyond "all conversion requests either immediately succeed or
> terminate the guest".

Now that I understand the purpose of KVM_SET_MEMORY_ATTRIBUTES, I
agree. However, pKVM needs to track at the host kernel (i.e., EL1)
whether guest memory is shared or private.

One approach would be to add another flag to the attributes that
tracks the host kernel view. The way KVM_SET_MEMORY_ATTRIBUTES is
implemented now, userspace can zero it, so in that case, that
operation would need to be masked to avoid that.

Another approach would be to have a pKVM-specific xarray (or similar)
to do the tracking, but since there is a structure that's already
doing something similar (i.e.,the attributes array), it seems like it
would be unnecessary overhead.

Do you have any ideas or preferences?

Cheers,
/fuad

Powered by blists - more mailing lists