Re: [PATCH v2] KVM: arm64: Skip more of the SError vaxorcism

[James Morse <james.morse@arm.com>]

Hi Marc,

On 10/06/2019 17:58, Marc Zyngier wrote:
> On 10/06/2019 17:30, James Morse wrote:
>> During __guest_exit() we need to consume any SError left pending by the
>> guest so it doesn't contaminate the host. With v8.2 we use the
>> ESB-instruction. For systems without v8.2, we use dsb+isb and unmask
>> SError. We do this on every guest exit.
>>
>> Use the same dsb+isr_el1 trick, this lets us know if an SError is pending
>> after the dsb, allowing us to skip the isb and self-synchronising PSTATE
>> write if its not.
>>
>> This means SError remains masked during KVM's world-switch, so any SError
>> that occurs during this time is reported by the host, instead of causing
>> a hyp-panic.
> 
> Ah, that'd be pretty good.

I'll add a patch to re-mask it so this intent is clear, and the behaviour/performance
stuff is done in separate patches.


>> If you give gcc likely()/unlikely() hints in an if() condition, it
>> shuffles the generated assembly so that the likely case is immediately
>> after the branch. Lets do the same here.
>>
>> Signed-off-by: James Morse <james.morse@arm.com>
>> ---
>> This patch was previously posted as part of:
>> [v1] https://lore.kernel.org/linux-arm-kernel/20190604144551.188107-1-james.morse@arm.com/
>>
>>  arch/arm64/kvm/hyp/entry.S | 14 ++++++++++----
>>  1 file changed, 10 insertions(+), 4 deletions(-)
>>
>> diff --git a/arch/arm64/kvm/hyp/entry.S b/arch/arm64/kvm/hyp/entry.S
>> index a5a4254314a1..c2de1a1faaf4 100644
>> --- a/arch/arm64/kvm/hyp/entry.S
>> +++ b/arch/arm64/kvm/hyp/entry.S
>> @@ -161,18 +161,24 @@ alternative_if ARM64_HAS_RAS_EXTN
>>  	orr	x0, x0, #(1<<ARM_EXIT_WITH_SERROR_BIT)
>>  1:	ret
>>  alternative_else
>> -	// If we have a pending asynchronous abort, now is the
>> -	// time to find out. From your VAXorcist book, page 666:
>> +	dsb	sy		// Synchronize against in-flight ld/st
>> +	mrs	x2, isr_el1
> 
> The CPU is allowed to perform a system register access before the DSB
> completes if it doesn't have a side effect. Reading ISR_EL1 doesn't have
> such side effect, so you could end-up missing the abort. An ISB after
> DSB should cure that,

... bother ...


> but you'll need to verify that it doesn't make
> things much worse than what we already have.

Retested with isb in both patches, and Robin's better assembly.

This still saves the self-synchronising pstate modification, (of which we'd need two if we
want to keep SError masked over the rest of world-switch)

On Xgene:
| 5.2.0-rc2-00006-g9b94314 mean:3215 stddev:45
| 5.2.0-rc2-00007-g5d37b0b mean:3176 stddev:30
| with this patch 1.23% faster

On Seattle:
| 5.2.0-rc2-00006-g9b9431445730 mean:4474 stddev:10
| 5.2.0-rc2-00007-g5d37b0b5dd65 mean:4410 stddev:27
| with this patch: 1.44% faster


Thanks,

James

_______________________________________________
linux-arm-kernel mailing list
linux-arm-kernel@lists.infradead.org
http://lists.infradead.org/mailman/listinfo/linux-arm-kernel