Message-ID: <CALMp9eSJgGJxSjej85yYvTav-n=KHNEPo4m2hEqsET+bHrXLew@mail.gmail.com>
Date:   Tue, 1 Mar 2022 13:57:52 -0800
From:   Jim Mattson <jmattson@...gle.com>
To:     Xiaoyao Li <xiaoyao.li@...el.com>
Cc:     Paolo Bonzini <pbonzini@...hat.com>,
        Chenyi Qiang <chenyi.qiang@...el.com>,
        Sean Christopherson <seanjc@...gle.com>,
        Vitaly Kuznetsov <vkuznets@...hat.com>,
        Wanpeng Li <wanpengli@...cent.com>,
        Joerg Roedel <joro@...tes.org>, kvm@...r.kernel.org,
        linux-kernel@...r.kernel.org
Subject: Re: [PATCH v3] KVM: VMX: Enable Notify VM exit

On Mon, Feb 28, 2022 at 9:30 PM Xiaoyao Li <xiaoyao.li@...el.com> wrote:
>
> On 3/1/2022 12:32 PM, Jim Mattson wrote:
> > On Mon, Feb 28, 2022 at 5:41 PM Xiaoyao Li <xiaoyao.li@...el.com> wrote:
> >>
> >> On 2/28/2022 10:30 PM, Jim Mattson wrote:
> >>> On Sun, Feb 27, 2022 at 11:10 PM Xiaoyao Li <xiaoyao.li@...el.com> wrote:
> >>>>
> >>>> On 2/26/2022 10:24 PM, Jim Mattson wrote:
> >>>>> On Fri, Feb 25, 2022 at 10:24 PM Xiaoyao Li <xiaoyao.li@...el.com> wrote:
> >>>>>>
> >>>>>> On 2/26/2022 12:53 PM, Jim Mattson wrote:
> >>>>>>> On Fri, Feb 25, 2022 at 8:25 PM Jim Mattson <jmattson@...gle.com> wrote:
> >>>>>>>>
> >>>>>>>> On Fri, Feb 25, 2022 at 8:07 PM Xiaoyao Li <xiaoyao.li@...el.com> wrote:
> >>>>>>>>>
> >>>>>>>>> On 2/25/2022 11:13 PM, Paolo Bonzini wrote:
> >>>>>>>>>> On 2/25/22 16:12, Xiaoyao Li wrote:
> >>>>>>>>>>>>>>
> >>>>>>>>>>>>>
> >>>>>>>>>>>>> I don't like the idea of making things up without notifying userspace
> >>>>>>>>>>>>> that this is fictional. How is my customer running nested VMs supposed
> >>>>>>>>>>>>> to know that L2 didn't actually shutdown, but L0 killed it because the
> >>>>>>>>>>>>> notify window was exceeded? If this information isn't reported to
> >>>>>>>>>>>>> userspace, I have no way of getting the information to the customer.
> >>>>>>>>>>>>
> >>>>>>>>>>>> Then, maybe a dedicated software define VM exit for it instead of
> >>>>>>>>>>>> reusing triple fault?
> >>>>>>>>>>>>
> >>>>>>>>>>>
> >>>>>>>>>>> Second thought, we can even just return Notify VM exit to L1 to tell
> >>>>>>>>>>> L2 causes Notify VM exit, even thought Notify VM exit is not exposed
> >>>>>>>>>>> to L1.
> >>>>>>>>>>
> >>>>>>>>>> That might cause NULL pointer dereferences or other nasty occurrences.
> >>>>>>>>>
> >>>>>>>>> IMO, a well written VMM (in L1) should handle it correctly.
> >>>>>>>>>
> >>>>>>>>> L0 KVM reports no Notify VM Exit support to L1, so L1 runs without
> >>>>>>>>> setting Notify VM exit. If a L2 causes notify_vm_exit with
> >>>>>>>>> invalid_vm_context, L0 just reflects it to L1. In L1's view, there is no
> >>>>>>>>> support of Notify VM Exit from VMX MSR capability. Following L1 handler
> >>>>>>>>> is possible:
> >>>>>>>>>
> >>>>>>>>> a)      if (notify_vm_exit available & notify_vm_exit enabled) {
> >>>>>>>>>                     handle in b)
> >>>>>>>>>             } else {
> >>>>>>>>>                     report unexpected vm exit reason to userspace;
> >>>>>>>>>             }
> >>>>>>>>>
> >>>>>>>>> b)      similar handler like we implement in KVM:
> >>>>>>>>>             if (!vm_context_invalid)
> >>>>>>>>>                     re-enter guest;
> >>>>>>>>>             else
> >>>>>>>>>                     report to userspace;
> >>>>>>>>>
> >>>>>>>>> c)      no Notify VM Exit related code (e.g. old KVM), it's treated as
> >>>>>>>>> unsupported exit reason
> >>>>>>>>>
> >>>>>>>>> As long as it belongs to any case above, I think L1 can handle it
> >>>>>>>>> correctly. Any nasty occurrence should be caused by incorrect handler in
> >>>>>>>>> L1 VMM, in my opinion.
> >>>>>>>>
> >>>>>>>> Please test some common hypervisors (e.g. ESXi and Hyper-V).
> >>>>>>>
> >>>>>>> I took a look at KVM in Linux v4.9 (one of our more popular guests),
> >>>>>>> and it will not handle this case well:
> >>>>>>>
> >>>>>>>             if (exit_reason < kvm_vmx_max_exit_handlers
> >>>>>>>                 && kvm_vmx_exit_handlers[exit_reason])
> >>>>>>>                     return kvm_vmx_exit_handlers[exit_reason](vcpu);
> >>>>>>>             else {
> >>>>>>>                     WARN_ONCE(1, "vmx: unexpected exit reason 0x%x\n", exit_reason);
> >>>>>>>                     kvm_queue_exception(vcpu, UD_VECTOR);
> >>>>>>>                     return 1;
> >>>>>>>             }
> >>>>>>>
> >>>>>>> At least there's an L1 kernel log message for the first unexpected
> >>>>>>> NOTIFY VM-exit, but after that, there is silence. Just a completely
> >>>>>>> inexplicable #UD in L2, assuming that L2 is resumable at this point.
> >>>>>>
> >>>>>> At least there is a message to tell L1 a notify VM exit is triggered in
> >>>>>> L2. Yes, the inexplicable #UD won't be hit unless L2 triggers Notify VM
> >>>>>> exit with invalid_context, which is malicious to L0 and L1.
> >>>>>
> >>>>> There is only an L1 kernel log message *the first time*. That's not
> >>>>> good enough. And this is just one of the myriad of possible L1
> >>>>> hypervisors.
> >>>>>
> >>>>>> If we use triple_fault (i.e., shutdown), then no info to tell L1 that
> >>>>>> it's caused by Notify VM exit with invalid context. Triple fault needs
> >>>>>> to be extended and L1 kernel needs to be enlightened. It doesn't help
> >>>>>> old guest kernel.
> >>>>>>
> >>>>>> If we use Machine Check, it's somewhat same inexplicable to L2 unless
> >>>>>> it's enlightened. But it doesn't help old guest kernel.
> >>>>>>
> >>>>>> Anyway, for Notify VM exit with invalid context from L2, I don't see a
> >>>>>> good solution to tell L1 VMM it's a "Notify VM exit with invalid context
> >>>>>> from L2" and keep all kinds of L1 VMM happy, especially for those with
> >>>>>> old kernel versions.
> >>>>>
> >>>>> I agree that there is no way to make every conceivable L1 happy.
> >>>>> That's why the information needs to be surfaced to the L0 userspace. I
> >>>>> contend that any time L0 kvm violates the architectural specification
> >>>>> in its emulation of L1 or L2, the L0 userspace *must* be informed.
> >>>>
> >>>> We can make the design to exit to userspace on notify vm exit
> >>>> unconditionally with exit_qualification passed, then userspace can take
> >>>> the same action like what this patch does in KVM that
> >>>>
> >>>>     - re-enter guest when context_invalid is false;
> >>>>     - stop running the guest if context_invalid is true; (userspace can
> >>>> definitely re-enter the guest in this case, but it needs to take the
> >>>> fall on this)
> >>>>
> >>>> Then, for nested case, L0 needs to enable it transparently for L2 if
> >>>> this feature is enabled for L1 guest (the reason as we all agreed that
> >>>> cannot allow L1 to escape just by creating a L2). Then what should KVM
> >>>> do when notify vm exit from L2?
> >>>>
> >>>>     - Exit to L0 userspace on L2's notify vm exit. L0 userspace takes the
> >>>> same action:
> >>>>           - re-enter if context-invalid is false;
> >>>>           - kill L1 if context-invalid is true; (I don't know if there is any
> >>>> interface for L0 userspace to kill L2). Then it opens the potential door
> >>>> for malicious user to kill L1 by creating a L2 to trigger fatal notify
> >>>> vm exit. If you guys accept it, we can implement in this way.
> >>>>
> >>>>
> >>>> in conclusion, we have below solution:
> >>>>
> >>>> 1. Take this patch as is. The drawback is L1 VMM receives a triple_fault
> >>>> from L2 when L2 triggers notify vm exit with invalid context. Neither of
> >>>> L1 VMM, L1 userspace, nor L2 kernel know it's caused due to notify vm
> >>>> exit. There is only kernel log in L0, which seems not accessible for L1
> >>>> user or L2 guest.
> >>>
> >>> You are correct on that last point, and I feel that I cannot stress it
> >>> enough. In a typical environment, the L0 kernel log is only available
> >>> to the administrator of the L0 host.
> >>>
> >>>> 2. a) Inject notify vm exit back to L1 if L2 triggers notify vm exit
> >>>> with invalid context. The drawback is, old L1 hypervisor is not
> >>>> enlightened of it and maybe misbehave on it.
> >>>>
> >>>>       b) Inject a synthesized SHUTDOWN exit to L1, with additional info to
> >>>> tell it's caused by fatal notify vm exit from L2. It has the same
> >>>> drawback that old hypervisor has no idea of it and maybe misbehave on it.
> >>>>
> >>>> 3. Exit to L0 usersapce unconditionally no matter it's caused from L1 or
> >>>> L2. Then it may open the door for L1 user to kill L1.
> >>>>
> >>>> Do you have any better solution other than above? If no, we need to pick> >> one from above though it cannot make everyone happy.
> >>>
> >>> Yes, I believe I have a better solution. We obviously need an API for
> >>> userspace to synthesize a SHUTDOWN event for a vCPU.
> >>
> >> Can you elaborate on it? Do you mean userspace to inject a synthesized
> >> SHUTDOWN to guest? If so, I have no idea how it will work.
> >
> > It can probably be implemented as an extension of KVM_SET_VCPU_EVENTS
> > that invokes kvm_make_request(KVM_REQ_TRIPLE_FAULT).
>
> Then, you mean
>
> 1. notify vm exit from guest;
> 2. exit to userspace on notify vm exit;
> 3. a. if context_invalid, inject SHUTDOWN to vcpu from userspace to
> request KVM_REQ_TRIPLE_FAULT; goto step 4;
>     b. if !context_invalid, re-run vcpu; no step 4 and 5;
> 4. exit to userspace again with KVM_EXIT_SHUTDOWN due to triple fault;
> 5. userspace stop running the vcpu/VM
>
> Then why not handle it as KVM_EXIT_SHUTDOWN directly in 3.a ? I don't
> get the point of userspace to inject TRIPLE_FAULT to KVM.

Sure, that should work, as long as L0 userspace is notified of the
emulation error.

Going back to something you said previously:

>> In addition, to avoid breaking legacy userspace, the NOTIFY VM-exit should be opt-in.

> Yes, it's designed as opt-in already that the feature is off by default.

I meant that userspace should opt-in, per VM. I believe your design is
opt-in by system administrator, host-wide.

Powered by blists - more mailing lists