On 31/05/18 13:38, Marc Zyngier wrote:
> On 31/05/18 12:49, Mark Rutland wrote:
>> On Wed, May 30, 2018 at 01:47:01PM +0100, Marc Zyngier wrote:
>>> Up to ARMv8.3, the combinaison of Stage-1 and Stage-2 attributes
>>> results in the strongest attribute of the two stages. This means
>>> that the hypervisor has to perform quite a lot of cache maintenance
>>> just in case the guest has some non-cacheable mappings around.
>>>
>>> ARMv8.4 solves this problem by offering a different mode (FWB) where
>>> Stage-2 has total control over the memory attribute (this is limited
>>> to systems where both I/O and instruction caches are coherent with
>>
>> s/caches/fetches/ -- the I-caches themselves aren't coherent with the
>> D-caches (or we could omit I-cache maintenance).
>>
>> i.e. this implies IDC, but not DIC.
>
> It may imply IDC behaviour, but not quite IDC itself. I agree, this
> looks dodgy. I've asked for clarification on the spec.
I've now received confirmation that FWB implies the IDC behaviour. It
doesn't guarantee that CTR_EL0.IDC will be set though, only that
CLIDR_EL1.LOU{U,IS} are both 0.
>
>>
>>> the dcache). This is achieved by having a different set of memory
>>> attributes in the page tables, and a new bit set in HCR_EL2.
>>>
>>> On such a system, we can then safely sidestep any form of dcache
>>> management.
>>>
>>> Acked-by: Catalin Marinas <catalin.marinas@xxxxxxx>
>>> Signed-off-by: Marc Zyngier <marc.zyngier@xxxxxxx>
>>> ---
>>
>>> static inline bool vcpu_has_cache_enabled(struct kvm_vcpu *vcpu)
>>> {
>>> @@ -268,7 +269,10 @@ static inline void __clean_dcache_guest_page(kvm_pfn_t pfn, unsigned long size)
>>> {
>>> void *va = page_address(pfn_to_page(pfn));
>>>
>>> - kvm_flush_dcache_to_poc(va, size);
>>> + if (!cpus_have_const_cap(ARM64_HAS_STAGE2_FWB))
>>> + kvm_flush_dcache_to_poc(va, size);
>>> + else
>>> + kvm_flush_dcache_to_pou(va, size);
>>> }
>>
>> Te commit message said instruction fetches were coherent, and that no
>> D-cache maintenance was necessary, so why do we need maintenance to the
>> PoU?
>
> That maintenance will be elided if we actually have IDC set. I'm happy
> to drop it once I have confirmation that we have an identical behaviour.
Given the above, I'll drop the clean to PoU.
>
>>
>>> +static void cpu_has_fwb(const struct arm64_cpu_capabilities *__unused)
>>> +{
>>> + u64 val = read_sysreg_s(SYS_CLIDR_EL1);
>>> +
>>> + /* Check that CLIDR_EL1.LOU{U,IS} are both 0 */
>>> + WARN_ON(val & (7 << 27 | 7 << 21));
>>> +}
>>
>> What about CTR_EL0.IDC?
>
> Again, that depends on whether FWB implies IDC or not.
The spec doesn't seem to guarantee IDC, but does mandate that
CLIDR_EL1.LOU{U,IS} are set 0.
Thanks,
M.
--
Jazz is not dead. It just smells funny...
