[PATCH] arm64: Add KRYO4XX gold CPU core to spectre-v2 safe list

[Marc Zyngier]

On 2020-08-13 13:33, Sai Prakash Ranjan wrote:
> On 2020-08-13 16:09, Marc Zyngier wrote:
>> On 2020-08-13 10:40, Will Deacon wrote:
>>> On Thu, Aug 13, 2020 at 02:49:37PM +0530, Sai Prakash Ranjan wrote:
>>>> On 2020-08-13 14:33, Will Deacon wrote:
>>>> > On Thu, Aug 13, 2020 at 01:48:34PM +0530, Sai Prakash Ranjan wrote:
>>>> > > KRYO4XX gold/big CPU cores are based on Cortex-A76 which has CSV2
>>>> > > bits set and are spectre-v2 safe. But on big.LITTLE systems where
>>>> > > they are coupled with other CPU cores such as the KRYO4XX silver
>>>> > > based on Cortex-A55 which are spectre-v2 safe but do not have CSV2
>>>> > > bits set, the system wide safe value will be set to the lowest value
>>>> > > of CSV2 bits as per FTR_LOWER_SAFE defined for CSV2 bits of register
>>>> > > ID_AA64PFR0_EL1.
>>>> > >
>>>> > > This is a problem when booting a guest kernel on gold CPU cores
>>>> > > where it will incorrectly report ARM_SMCCC_ARCH_WORKAROUND_1 warning
>>>> > > and consider them as vulnerable for Spectre variant 2 due to system
>>>> > > wide safe value which is used in kvm emulation code when reading id
>>>> > > registers. One wrong way of fixing this is to set the FTR_HIGHER_SAFE
>>>> > > for CSV2 bits, so instead add the KRYO4XX gold CPU core to the safe
>>>> > > list which will be consulted even when the sanitised read reports
>>>> > > that CSV2 bits are not set for KRYO4XX gold cores.
>>>> > >
>>>> > > Reported-by: Stephen Boyd <swboyd at chromium.org>
>>>> > > Signed-off-by: Sai Prakash Ranjan <saiprakash.ranjan at codeaurora.org>
>>>> > > ---
>>>> > >  arch/arm64/kernel/cpu_errata.c | 1 +
>>>> > >  1 file changed, 1 insertion(+)
>>>> > >
>>>> > > diff --git a/arch/arm64/kernel/cpu_errata.c
>>>> > > b/arch/arm64/kernel/cpu_errata.c
>>>> > > index 6bd1d3ad037a..6cbdd2d98a2a 100644
>>>> > > --- a/arch/arm64/kernel/cpu_errata.c
>>>> > > +++ b/arch/arm64/kernel/cpu_errata.c
>>>> > > @@ -545,6 +545,7 @@ static const struct midr_range
>>>> > > spectre_v2_safe_list[] = {
>>>> > >  	MIDR_ALL_VERSIONS(MIDR_HISI_TSV110),
>>>> > >  	MIDR_ALL_VERSIONS(MIDR_QCOM_KRYO_3XX_SILVER),
>>>> > >  	MIDR_ALL_VERSIONS(MIDR_QCOM_KRYO_4XX_SILVER),
>>>> > > +	MIDR_ALL_VERSIONS(MIDR_QCOM_KRYO_4XX_GOLD),
>>>> >
>>>> > We shouldn't be putting CPUs in the safe list when they have CSV2
>>>> > reporting
>>>> > that they are mitigated in hardware, so I don't think this is the right
>>>> > approach.
>>>> >
>>>> 
>>>> Ok but the only thing I find wrong in this approach is that it is a
>>>> redundant
>>>> information because CSV2 is already advertising the mitigation, but 
>>>> again
>>>> CSV2 check is done first so it doesn't really hurt to add it to the 
>>>> safe
>>>> list because we already know that it is safe.
>>> 
>>> It simply doesn't scale. That's why CSV2 exists in the first place, 
>>> so we
>>> don't have to modify the kernel everytime a new CPU is invented.
>>> 
>>>> > Sounds more like KVM should advertise CSV2 for the vCPUs if all of the
>>>> > physical CPUs without CSV2 set are on the safe list. But then again, KVM
>>>> > has always been slightly in denial about big.LITTLE because you can't
>>>> > sensibly expose it to a guest if there are detectable differences...
>>>> >
>>>> 
>>>> Sorry but I don't see how the guest kernel will see the CSV2 bits 
>>>> set for
>>>> gold CPU cores without actually adding them to the safe list or 
>>>> reading the
>>>> not sanitised value of ID_AA64PFR0_EL1 ?
>>> 
>>> Well that's for somebody to figure out in the patch. I'm just saying 
>>> that
>>> adding cores to the safe list when they already have a CSV2 field 
>>> conveying
>>> the same information is the wrong approach. The right appproach is 
>>> for KVM
>>> to expose CSV2 as set when the system is not affected by the erratum.
>> 
>> A sensible way to fix this would be with something like that:
>> 
>> diff --git a/arch/arm64/kvm/sys_regs.c b/arch/arm64/kvm/sys_regs.c
>> index 077293b5115f..2735db21ff0d 100644
>> --- a/arch/arm64/kvm/sys_regs.c
>> +++ b/arch/arm64/kvm/sys_regs.c
>> @@ -1131,6 +1131,9 @@ static u64 read_id_reg(const struct kvm_vcpu 
>> *vcpu,
>>  		if (!vcpu_has_sve(vcpu))
>>  			val &= ~(0xfUL << ID_AA64PFR0_SVE_SHIFT);
>>  		val &= ~(0xfUL << ID_AA64PFR0_AMU_SHIFT);
>> +		if (!(val & (0xfUL << ID_AA64PFR0_CSV2_SHIFT)) &&
>> +		    get_spectre_v2_workaround_state() == 
>> ARM64_BP_HARDEN_NOT_REQUIRED)
>> +			val |= (1UL << ID_AA64PFR0_CSV2_SHIFT);
>>  	} else if (id == SYS_ID_AA64ISAR1_EL1 && !vcpu_has_ptrauth(vcpu)) {
>>  		val &= ~((0xfUL << ID_AA64ISAR1_APA_SHIFT) |
>>  			 (0xfUL << ID_AA64ISAR1_API_SHIFT) |
>> 
> 
> Thanks Marc, I gave this a go on SC7180 where the issue was seen and
> this fix is good.
> 
> Tested-by: Sai Prakash Ranjan <saiprakash.ranjan at codeaurora.org>

There is still a problem with this approach. A late CPU could
come up after a guest has been started. If that CPU identified
as vulnerable by the host kernel, get_spectre_v2_workaround_state()
will return a different value, breaking the guest (or more
likely, leaving it exposed to Spectre-v2 attacks).

We'd need to disable the late onlining of CPUs that would change
the mitigation state, and this is... ugly.

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