Re: [PATCH RFC] KVM: async_pf: fix async_pf exception injection

[Wanpeng Li]

2017-06-08 19:52 GMT+08:00 Paolo Bonzini <pbonzini@xxxxxxxxxx>:
>
>
> On 08/06/2017 11:30, Wanpeng Li wrote:
>> However, I found that "nr == PF_VECTOR && vmx->apf_reason != 0" never be true
>> in nested_vmx_check_exception(). SVM depends on the similar stuff in
>> nested_svm_intercept() which makes me confusing how it can works. In addition,
>> vmx/svm->apf_reason should be got in L1 since apf_reason.reason will make sense
>> just in pv guest. So vmx/svm->apf_reason should always be 0 on L0.
>
> I agree.  My suggestion is a series like this:
>
> 1) remove all arguments except the first in
> kvm_x86_ops->queue_exception: they can extract the arguments from
> vcpu->arch.exception themselves
>
> 2) do the same in nested_{vmx,svm}_check_exception
>
> 3) add an async_page_fault member to vcpu->arch.exception

Agreed.

>
>> I change the
>> condition to "nr == PF_VECTOR && error_code == 0" to intercept async_pf, however,
>> the below bug will be splatted:
>
> Right, because error_code == 0 is a valid error code.
>
> For stable releases, this should be enough:

We should also take care PAGE_READY async_pfs injection. I will send a
patch to avoid async pf injection stuff in guest mode for stable
release.

>
> diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c
> index ac7810513d0e..c5f0023bac2d 100644
> --- a/arch/x86/kvm/mmu.c
> +++ b/arch/x86/kvm/mmu.c
> @@ -3689,6 +3689,9 @@ static bool can_do_async_pf(struct kvm_vcpu *vcpu)
>                      kvm_event_needs_reinjection(vcpu)))
>                 return false;
>
> +       if (is_guest_mode(vcpu))
> +               return false;
> +
>         return kvm_x86_ops->interrupt_allowed(vcpu);
>  }
>
>
> For proper nested async page fault support, you have to be careful,
> because old KVM as L1 you might crash due to the BUG(enable_ept); my
> suggestion is to add another flag bit to MSR_KVM_ASYNC_PF_EN.
>
> If bit 2 is 1, async page faults are delivered to L1 as #PF vmexits; if
> bit 2 is 0, can_do_async_pf returns 0 if in guest mode.  When bit 2 is
> 1, async page fault exceptions override the usual exception/intercept
> bitmaps, i.e. they always cause #PF vmexits.  What do you think?

Agreed.

>
> Thanks,
>
> Paolo
>
>>  BUG: unable to handle kernel paging request at ffffe305770a87e0
>>  IP: kfree+0x6f/0x300
>>  PGD 0
>>  P4D 0
>>
>>  Oops: 0000 [#1] PREEMPT SMP
>>  CPU: 3 PID: 2187 Comm: transhuge-stres Tainted: G           OE   4.12.0-rc4+ #9
>>  task: ffff8a9214b58000 task.stack: ffffb46bc34e4000
>>  RIP: 0010:kfree+0x6f/0x300
>>  RSP: 0000:ffffb46bc34e7b28 EFLAGS: 00010086
>>  RAX: ffffe305770a87c0 RBX: ffffb46bc2a1fe70 RCX: 0000000000000001
>>  RDX: 0000757180000000 RSI: 00000000ffffffff RDI: 0000000000000096
>>  RBP: ffffb46bc34e7b50 R08: 0000000000000000 R09: 0000000000000001
>>  R10: ffffb46bc34e7ac8 R11: 68b9962a00000000 R12: 000000a7770a87c0
>>  R13: ffffffff90059b75 R14: ffffffff913466c0 R15: ffffe25e06f18000
>>  FS:  00007f1904ae7700(0000) GS:ffff8a921a800000(0000) knlGS:0000000000000000
>>  CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
>>  CR2: ffffe305770a87e0 CR3: 000000040eb5c000 CR4: 00000000001426e0
>>  Call Trace:
>>   kvm_async_pf_task_wait+0xd5/0x280
>>   ? __this_cpu_preempt_check+0x13/0x20
>>   do_async_page_fault+0x77/0xb0
>>   ? do_async_page_fault+0x77/0xb0
>>   async_page_fault+0x28/0x30
>>
>> In additon, if svm->apf_reason doen't make sense on L0, then maybe it also will not
>> work in the function nested_svm_exit_special().
>>
>> The patch below is uncompleted, and your inputs to improve it is a great appreciated.
>
> The patch makes sense.  Let's try to make more code common to VMX and
> SVM too once the above is clarified.

Agreed.

Wanpeng Li

>
> Paolo
>
>> Cc: Paolo Bonzini <pbonzini@xxxxxxxxxx>
>> Cc: Radim KrÄmÃÅ <rkrcmar@xxxxxxxxxx>
>> Signed-off-by: Wanpeng Li <wanpeng.li@xxxxxxxxxxx>
>> ---
>>  arch/x86/kvm/vmx.c | 41 ++++++++++++++++++++++++++++++++---------
>>  1 file changed, 32 insertions(+), 9 deletions(-)
>>
>> diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
>> index ca5d2b9..21a1b44 100644
>> --- a/arch/x86/kvm/vmx.c
>> +++ b/arch/x86/kvm/vmx.c
>> @@ -616,6 +616,7 @@ struct vcpu_vmx {
>>       bool emulation_required;
>>
>>       u32 exit_reason;
>> +     u32 apf_reason;
>>
>>       /* Posted interrupt descriptor */
>>       struct pi_desc pi_desc;
>> @@ -2418,11 +2419,12 @@ static void skip_emulated_instruction(struct kvm_vcpu *vcpu)
>>   * KVM wants to inject page-faults which it got to the guest. This function
>>   * checks whether in a nested guest, we need to inject them to L1 or L2.
>>   */
>> -static int nested_vmx_check_exception(struct kvm_vcpu *vcpu, unsigned nr)
>> +static int nested_vmx_check_exception(struct kvm_vcpu *vcpu, unsigned nr, u32 error_code)
>>  {
>>       struct vmcs12 *vmcs12 = get_vmcs12(vcpu);
>>
>> -     if (!(vmcs12->exception_bitmap & (1u << nr)))
>> +     if (!((vmcs12->exception_bitmap & (1u << nr)) ||
>> +             (nr == PF_VECTOR && error_code == 0)))
>>               return 0;
>>
>>       nested_vmx_vmexit(vcpu, EXIT_REASON_EXCEPTION_NMI,
>> @@ -2439,7 +2441,7 @@ static void vmx_queue_exception(struct kvm_vcpu *vcpu, unsigned nr,
>>       u32 intr_info = nr | INTR_INFO_VALID_MASK;
>>
>>       if (!reinject && is_guest_mode(vcpu) &&
>> -         nested_vmx_check_exception(vcpu, nr))
>> +         nested_vmx_check_exception(vcpu, nr, error_code))
>>               return;
>>
>>       if (has_error_code) {
>> @@ -5646,14 +5648,31 @@ static int handle_exception(struct kvm_vcpu *vcpu)
>>       }
>>
>>       if (is_page_fault(intr_info)) {
>> -             /* EPT won't cause page fault directly */
>> -             BUG_ON(enable_ept);
>>               cr2 = vmcs_readl(EXIT_QUALIFICATION);
>> -             trace_kvm_page_fault(cr2, error_code);
>> +             switch (vmx->apf_reason) {
>> +             default:
>> +                     /* EPT won't cause page fault directly */
>> +                     BUG_ON(enable_ept);
>> +                     trace_kvm_page_fault(cr2, error_code);
>>
>> -             if (kvm_event_needs_reinjection(vcpu))
>> -                     kvm_mmu_unprotect_page_virt(vcpu, cr2);
>> -             return kvm_mmu_page_fault(vcpu, cr2, error_code, NULL, 0);
>> +                     if (kvm_event_needs_reinjection(vcpu))
>> +                             kvm_mmu_unprotect_page_virt(vcpu, cr2);
>> +                     return kvm_mmu_page_fault(vcpu, cr2, error_code, NULL, 0);
>> +                     break;
>> +             case KVM_PV_REASON_PAGE_NOT_PRESENT:
>> +                     vmx->apf_reason = 0;
>> +                     local_irq_disable();
>> +                     kvm_async_pf_task_wait(cr2);
>> +                     local_irq_enable();
>> +                     break;
>> +             case KVM_PV_REASON_PAGE_READY:
>> +                     vmx->apf_reason = 0;
>> +                     local_irq_disable();
>> +                     kvm_async_pf_task_wake(cr2);
>> +                     local_irq_enable();
>> +                     break;
>> +             }
>> +             return 0;
>>       }
>>
>>       ex_no = intr_info & INTR_INFO_VECTOR_MASK;
>> @@ -8600,6 +8619,10 @@ static void vmx_complete_atomic_exit(struct vcpu_vmx *vmx)
>>       vmx->exit_intr_info = vmcs_read32(VM_EXIT_INTR_INFO);
>>       exit_intr_info = vmx->exit_intr_info;
>>
>> +     /* if exit due to PF check for async PF */
>> +     if (is_page_fault(exit_intr_info))
>> +             vmx->apf_reason = kvm_read_and_reset_pf_reason();
>> +
>>       /* Handle machine checks before interrupts are enabled */
>>       if (is_machine_check(exit_intr_info))
>>               kvm_machine_check();
>>