Re: KVM vs AMD: Re: [PATCH v3 48/59] x86/retbleed: Add SKL return thunk

[Andrew Cooper]

On 04/11/2022 12:44, Peter Zijlstra wrote:
> On Thu, Nov 03, 2022 at 10:53:54PM +0000, Andrew Cooper wrote:
>> On 21/10/2022 16:21, Nathan Chancellor wrote:
>>> On Fri, Oct 21, 2022 at 11:53:09AM +0200, Peter Zijlstra wrote:
>>>> On Thu, Oct 20, 2022 at 04:10:28PM -0700, Nathan Chancellor wrote:
>>>>> This commit is now in -next as commit 5d8213864ade ("x86/retbleed: Add
>>>>> SKL return thunk"). I just bisected an immediate reboot on my AMD test
>>>>> system when starting a virtual machine with QEMU + KVM to it (see the
>>>>> bisect log below). My Intel test systems do not show this.
>>>>> Unfortunately, I do not have much more information, as there are no logs
>>>>> in journalctl, which makes sense as the reboot occurs immediately after
>>>>> I hit the enter key for the QEMU command.
>>>>>
>>>>> If there is any further information I can provide or patches I can test
>>>>> for further debugging, I am more than happy to do so.
>>>> Moo :-(
>>>>
>>>> you happen to have a .config for me?
>>> Sure thing, sorry I did not provide it in the first place! Attached. It
>>> has been run through localmodconfig for the particular machine but I
>>> assume the core pieces should still be present.
>> Following up from some debugging on IRC.
>>
>> The problem is that FILL_RETURN_BUFFER now has a per-cpu variable
>> access, and AMD SVM has a fun optimisation where the VMRUN instruction
>> doesn't swap, amongst other things, %gs.
>>
>> per-cpu variables only become safe following
>> vmload(__sme_page_pa(sd->save_area)); in svm_vcpu_enter_exit().
>>
>> Given that retbleed=force ought to work on non-skylake hardware, the
>> appropriate fix is to move the VMLOAD/VMSAVE's down into asm and put
>> them adjacent to VMRUN.
>>
>> This also addresses an undocumented dependency where its only the memory
>> clobber in vmload() which stops the compiler moving
>> svm_vcpu_enter_exit()'s calculation of sd into an unsafe position.
> So, aside from wasting the entire morning on resuscitating my AMD
> Interlagos, I ended up with the below patch which seems to work.
>
> Not being a virt person, I'm sure I've messed up something, please
> advise.
>
> ---
> diff --git a/arch/x86/kvm/svm/svm.c b/arch/x86/kvm/svm/svm.c
> index 58f0077d9357..f7ee1eedacfe 100644
> --- a/arch/x86/kvm/svm/svm.c
> +++ b/arch/x86/kvm/svm/svm.c
> @@ -3929,11 +3929,8 @@ static noinstr void svm_vcpu_enter_exit(struct kvm_vcpu *vcpu)
>  		 * the state doesn't need to be copied between vmcb01 and
>  		 * vmcb02 when switching vmcbs for nested virtualization.
>  		 */
> -		vmload(svm->vmcb01.pa);
> -		__svm_vcpu_run(vmcb_pa, (unsigned long *)&vcpu->arch.regs);
> -		vmsave(svm->vmcb01.pa);
> -
> -		vmload(__sme_page_pa(sd->save_area));
> +		__svm_vcpu_run(vmcb_pa, (unsigned long *)&vcpu->arch.regs,
> +			       svm->vmcb01.pa, __sme_page_pa(sd->save_area));
>  	}
>  
>  	guest_state_exit_irqoff();
> diff --git a/arch/x86/kvm/svm/svm.h b/arch/x86/kvm/svm/svm.h
> index 6a7686bf6900..2a038def7ac7 100644
> --- a/arch/x86/kvm/svm/svm.h
> +++ b/arch/x86/kvm/svm/svm.h
> @@ -684,6 +684,7 @@ void sev_es_unmap_ghcb(struct vcpu_svm *svm);
>  /* vmenter.S */
>  
>  void __svm_sev_es_vcpu_run(unsigned long vmcb_pa);
> -void __svm_vcpu_run(unsigned long vmcb_pa, unsigned long *regs);
> +void __svm_vcpu_run(unsigned long vmcb_pa, unsigned long *regs,
> +		    unsigned long guest_vmcb_pa, unsigned long host_vmcb_pa);
>  
>  #endif
> diff --git a/arch/x86/kvm/svm/vmenter.S b/arch/x86/kvm/svm/vmenter.S
> index 09eacf19d718..50f200f7b773 100644
> --- a/arch/x86/kvm/svm/vmenter.S
> +++ b/arch/x86/kvm/svm/vmenter.S
> @@ -32,8 +32,10 @@
>  
>  /**
>   * __svm_vcpu_run - Run a vCPU via a transition to SVM guest mode
> - * @vmcb_pa:	unsigned long
> - * @regs:	unsigned long * (to guest registers)
> + * @vmcb_pa:		unsigned long
> + * @regs:		unsigned long * (to guest registers)
> + * @guest_vmcb_pa:	unsigned long
> + * @host_vmcb_pa:	unsigned long
>   */
>  SYM_FUNC_START(__svm_vcpu_run)
>  	push %_ASM_BP
> @@ -51,9 +53,18 @@ SYM_FUNC_START(__svm_vcpu_run)
>  	/* Save @regs. */
>  	push %_ASM_ARG2
>  
> +	/* Save host_vmcb_pa */
> +	push %_ASM_ARG4
> +
> +	/* Save guest_vmcb_pa */
> +	push %_ASM_ARG3
> +
>  	/* Save @vmcb. */
>  	push %_ASM_ARG1
>  
> +	/* Save guest_vmcb_pa */
> +	push %_ASM_ARG3
> +
>  	/* Move @regs to RAX. */
>  	mov %_ASM_ARG2, %_ASM_AX
>  
> @@ -75,15 +86,29 @@ SYM_FUNC_START(__svm_vcpu_run)
>  	mov VCPU_R15(%_ASM_AX), %r15
>  #endif
>  
> +	/* POP and VMLOAD @guest_vmcb01_pa */
> +	pop %_ASM_AX
> +1:	vmload %_ASM_AX
> +2:
>  	/* "POP" @vmcb to RAX. */
>  	pop %_ASM_AX
>  
>  	/* Enter guest mode */
>  	sti
>  
> -1:	vmrun %_ASM_AX
> +3:	vmrun %_ASM_AX
> +4:
> +	cli
>  
> -2:	cli
> +	/* POP and VMSAVE @guest_vmcb01_pa */
> +	pop %_ASM_AX
> +5:	vmsave %_ASM_AX
> +6:
> +	/* POP and VMLOAD @host_vmcb01_pa */
> +	pop %_ASM_AX
> +7:	vmload %_ASM_AX
> +8:
> +	/* Now host %GS is live */
>  
>  #ifdef CONFIG_RETPOLINE
>  	/* IMPORTANT: Stuff the RSB immediately after VM-Exit, before RET! */
> @@ -160,11 +185,26 @@ SYM_FUNC_START(__svm_vcpu_run)
>  	pop %_ASM_BP
>  	RET
>  
> -3:	cmpb $0, kvm_rebooting
> +10:	cmpb $0, kvm_rebooting
>  	jne 2b
>  	ud2
>  
> -	_ASM_EXTABLE(1b, 3b)
> +30:	cmpb $0, kvm_rebooting
> +	jne 4b
> +	ud2
> +
> +50:	cmpb $0, kvm_rebooting
> +	jne 6b
> +	ud2
> +
> +70:	cmpb $0, kvm_rebooting
> +	jne 8b
> +	ud2
> +
> +	_ASM_EXTABLE(1b, 10b)
> +	_ASM_EXTABLE(3b, 30b)
> +	_ASM_EXTABLE(5b, 50b)
> +	_ASM_EXTABLE(7b, 70b)

Following up again from IRC.  This #UD handling is pre-existingly broken.

Between the two VMLOADs, no exceptions are safe.  paranoid_entry can
cope with %gs.base being wrong (if #UD was paranoid, but it isn't), but
nothing can deal with %tr being wrong for IST vectors.

But that's fine because we're in a GIF=0 region anyway and not even NMIs
can hit.  (Whether SMIs can hit is model specific, but if RSM corrupts
EFER.SVME, then its game over anyway).

~Andrew