Re: [PATCH 3/8] KVM: arm64: Map AArch32 cp15 register to AArch64 sysregs

[Marc Zyngier <maz@kernel.org>]

On 2020-11-03 18:29, James Morse wrote:
> Hi Marc,
> 
> On 02/11/2020 19:16, Marc Zyngier wrote:
>> Move all the cp15 registers over to their AArch64 counterpart.
>> This requires the annotation of a few of them (such as the usual
>> DFAR/IFAR vs FAR_EL1), and a new helper that generates mask/shift
>> pairs for the various configurations.
> 
>> diff --git a/arch/arm64/kvm/sys_regs.c b/arch/arm64/kvm/sys_regs.c
>> index 26c7c25f8a6d..137818793a4a 100644
>> --- a/arch/arm64/kvm/sys_regs.c
>> +++ b/arch/arm64/kvm/sys_regs.c
>> @@ -138,26 +156,16 @@ static bool access_vm_reg(struct kvm_vcpu *vcpu,
>>  			  const struct sys_reg_desc *r)
>>  {
>>  	bool was_enabled = vcpu_has_cache_enabled(vcpu);
>> -	u64 val;
>> -	int reg = r->reg;
>> +	u64 val, mask, shift;
>> 
>>  	BUG_ON(!p->is_write);
>> 
>> -	/* See the 32bit mapping in kvm_host.h */
>> -	if (p->is_aarch32)
>> -		reg = r->reg / 2;
>> +	get_access_mask(r, &mask, &shift);
>> 
>> -	if (!p->is_aarch32 || !p->is_32bit) {
>> -		val = p->regval;
>> -	} else {
>> -		val = vcpu_read_sys_reg(vcpu, reg);
>> -		if (r->reg % 2)
>> -			val = (p->regval << 32) | (u64)lower_32_bits(val);
>> -		else
>> -			val = ((u64)upper_32_bits(val) << 32) |
>> -				lower_32_bits(p->regval);
>> -	}
>> -	vcpu_write_sys_reg(vcpu, val, reg);
> 
>> +	val = vcpu_read_sys_reg(vcpu, r->reg);
>> +	val &= ~mask;
>> +	val |= (p->regval & (mask >> shift)) << shift;
>> +	vcpu_write_sys_reg(vcpu, val, r->reg);
> 
> I can't tell if the compiler has worked out ithat it can skip the
> sys_read most of the
> time... Won't some of these trap for a nested VHE hypervisor?
> 
> | if (~mask) {
> |	val = vcpu_read_sys_reg(vcpu, r->reg);
> |	val &= ~mask;
> | }

Seems like a good call. I'm happy to fold that in.

> 
> 
> But, as the sys_reg arrays already have indirection for this function, 
> why does
> access_vm_reg() have to deal with this at all? Its not even needed for
> all the aarch32 registers...
> 
> 
> | { AA32(LO), Op1( 0), CRn(10), CRm( 2), Op2( 0),
> access_aarch32_alias, NULL, MAIR_EL1 },
> 
> Where access_aarch32_alias() does the shift+mask and read_modify-write
> on the sys_reg?


I don't really like the idea of separate handlers. The whole point is to
unify the two view, and I feel like having yet another indirection makes
it harder to maintain.

> 
>>  	kvm_toggle_cache(vcpu, was_enabled);
>>  	return true;
> 
>> @@ -1919,19 +1919,24 @@ static const struct sys_reg_desc 
>> cp14_64_regs[] = {
>>   */
>>  static const struct sys_reg_desc cp15_regs[] = {
> 
>> -	{ Op1( 0), CRn( 2), CRm( 0), Op2( 2), access_vm_reg, NULL, c2_TTBCR 
>> },
> 
>> +	{ Op1( 0), CRn( 2), CRm( 0), Op2( 2), access_vm_reg, NULL, TCR_EL1 
>> },
> 
> Don't look now ... TTBRCR2 means this one is a LO/HI job.

That's the problem with AArch32. You stop looking for a couple of years,
and it decides to throw a new sysreg at you without any notice. WTF?

> (I'm guessing that should be fixed before this series to make the 
> backport easy)

I'll work something out as a prologue to this series.

Thanks,

         M.
-- 
Jazz is not dead. It just smells funny...