Re: [RFC PATCH] KVM: PPC: Book3S: MMIO emulation support for little endian guests

[Cedric Le Goater <clg@fr.ibm.com>]

Hi Alex,

On 10/04/2013 02:50 PM, Alexander Graf wrote:
> 
> On 03.10.2013, at 13:03, Cédric Le Goater wrote:
> 
>> MMIO emulation reads the last instruction executed by the guest
>> and then emulates. If the guest is running in Little Endian mode,
>> the instruction needs to be byte-swapped before being emulated.
>>
>> This patch stores the last instruction in the endian order of the
>> host, primarily doing a byte-swap if needed. The common code
>> which fetches last_inst uses a helper routine kvmppc_need_byteswap().
>> and the exit paths for the Book3S PV and HR guests use their own
>> version in assembly.
>>
>> kvmppc_emulate_instruction() also uses kvmppc_need_byteswap() to
>> define in which endian order the mmio needs to be done.
>>
>> The patch is based on Alex Graf's kvm-ppc-queue branch and it
>> has been tested on Big Endian and Little Endian HV guests and
>> Big Endian PR guests.
>>
>> Signed-off-by: Cédric Le Goater <clg@fr.ibm.com>
>> ---
>>
>> Here are some comments/questions : 
>>
>> * the host is assumed to be running in Big Endian. when Little Endian
>>   hosts are supported in the future, we will use the cpu features to
>>   fix kvmppc_need_byteswap()
>>
>> * the 'is_bigendian' parameter of the routines kvmppc_handle_load()
>>   and kvmppc_handle_store() seems redundant but the *BRX opcodes 
>>   make the improvements unclear. We could eventually rename the
>>   parameter to byteswap and the attribute vcpu->arch.mmio_is_bigendian
>>   to vcpu->arch.mmio_need_byteswap. Anyhow, the current naming sucks
>>   and I would happy to have some directions to fix it.
>>
>>
>>
>> arch/powerpc/include/asm/kvm_book3s.h   |   15 ++++++-
>> arch/powerpc/kvm/book3s_64_mmu_hv.c     |    4 ++
>> arch/powerpc/kvm/book3s_hv_rmhandlers.S |   14 +++++-
>> arch/powerpc/kvm/book3s_segment.S       |   14 +++++-
>> arch/powerpc/kvm/emulate.c              |   71 +++++++++++++++++--------------
>> 5 files changed, 83 insertions(+), 35 deletions(-)
>>
>> diff --git a/arch/powerpc/include/asm/kvm_book3s.h b/arch/powerpc/include/asm/kvm_book3s.h
>> index 0ec00f4..36c5573 100644
>> --- a/arch/powerpc/include/asm/kvm_book3s.h
>> +++ b/arch/powerpc/include/asm/kvm_book3s.h
>> @@ -270,14 +270,22 @@ static inline ulong kvmppc_get_pc(struct kvm_vcpu *vcpu)
>> 	return vcpu->arch.pc;
>> }
>>
>> +static inline bool kvmppc_need_byteswap(struct kvm_vcpu *vcpu)
>> +{
>> +	return vcpu->arch.shared->msr & MSR_LE;
>> +}
>> +
>> static inline u32 kvmppc_get_last_inst(struct kvm_vcpu *vcpu)
>> {
>> 	ulong pc = kvmppc_get_pc(vcpu);
>>
>> 	/* Load the instruction manually if it failed to do so in the
>> 	 * exit path */
>> -	if (vcpu->arch.last_inst == KVM_INST_FETCH_FAILED)
>> +	if (vcpu->arch.last_inst == KVM_INST_FETCH_FAILED) {
>> 		kvmppc_ld(vcpu, &pc, sizeof(u32), &vcpu->arch.last_inst, false);
>> +		if (kvmppc_need_byteswap(vcpu))
>> +			vcpu->arch.last_inst = swab32(vcpu->arch.last_inst);
> 
> Could you please introduce a new helper to load 32bit numbers? Something like kvmppc_ldl or kvmppc_ld32. That'll be easier to read here then :).

ok. I did something in that spirit in the next patchset I am about to send. I will
respin if needed but there is one fuzzy area though : kvmppc_read_inst().  

It calls kvmppc_get_last_inst() and then again kvmppc_ld(). Is that actually useful ? 

>> +	}
>>
>> 	return vcpu->arch.last_inst;
>> }
>> @@ -293,8 +301,11 @@ static inline u32 kvmppc_get_last_sc(struct kvm_vcpu *vcpu)
>>
>> 	/* Load the instruction manually if it failed to do so in the
>> 	 * exit path */
>> -	if (vcpu->arch.last_inst == KVM_INST_FETCH_FAILED)
>> +	if (vcpu->arch.last_inst == KVM_INST_FETCH_FAILED) {
>> 		kvmppc_ld(vcpu, &pc, sizeof(u32), &vcpu->arch.last_inst, false);
>> +		if (kvmppc_need_byteswap(vcpu))
>> +			vcpu->arch.last_inst = swab32(vcpu->arch.last_inst);
>> +	}
>>
>> 	return vcpu->arch.last_inst;
>> }
>> diff --git a/arch/powerpc/kvm/book3s_64_mmu_hv.c b/arch/powerpc/kvm/book3s_64_mmu_hv.c
>> index 3a89b85..28130c7 100644
>> --- a/arch/powerpc/kvm/book3s_64_mmu_hv.c
>> +++ b/arch/powerpc/kvm/book3s_64_mmu_hv.c
>> @@ -547,6 +547,10 @@ static int kvmppc_hv_emulate_mmio(struct kvm_run *run, struct kvm_vcpu *vcpu,
>> 		ret = kvmppc_ld(vcpu, &srr0, sizeof(u32), &last_inst, false);
>> 		if (ret != EMULATE_DONE || last_inst == KVM_INST_FETCH_FAILED)
>> 			return RESUME_GUEST;
>> +
>> +		if (kvmppc_need_byteswap(vcpu))
>> +			last_inst = swab32(last_inst);
>> +
>> 		vcpu->arch.last_inst = last_inst;
>> 	}
>>
>> diff --git a/arch/powerpc/kvm/book3s_hv_rmhandlers.S b/arch/powerpc/kvm/book3s_hv_rmhandlers.S
>> index dd80953..1d3ee40 100644
>> --- a/arch/powerpc/kvm/book3s_hv_rmhandlers.S
>> +++ b/arch/powerpc/kvm/book3s_hv_rmhandlers.S
>> @@ -1393,14 +1393,26 @@ fast_interrupt_c_return:
>> 	lwz	r8, 0(r10)
>> 	mtmsrd	r3
>>
>> +	ld	r0, VCPU_MSR(r9)
>> +
>> +	/* r10 = vcpu->arch.msr & MSR_LE */
>> +	rldicl	r10, r0, 0, 63
> 
> rldicl.?

sure.

>> +	cmpdi	r10, 0
>> +	bne	2f
> 
> I think it makes sense to inline that branch in here instead. Just make this
> 
>   stw r8, VCPU_LAST_INST(r9)
>   beq after_inst_store
>   /* Little endian instruction, swap for big endian hosts */
>   addi ...
>   stwbrx ...
> 
> after_inst_store:
> 
> The duplicate store shouldn't really hurt too badly, but in our "fast path" we're only doing one store anyway :). And the code becomes more readable.

It is indeed more readable. I have changed that.

>> +
>> 	/* Store the result */
>> 	stw	r8, VCPU_LAST_INST(r9)
>>
>> 	/* Unset guest mode. */
>> -	li	r0, KVM_GUEST_MODE_NONE
>> +1:	li	r0, KVM_GUEST_MODE_NONE
>> 	stb	r0, HSTATE_IN_GUEST(r13)
>> 	b	guest_exit_cont
>>
>> +	/* Swap and store the result */
>> +2:	addi	r11, r9, VCPU_LAST_INST
>> +	stwbrx	r8, 0, r11
>> +	b	1b
>> +
>> /*
>>  * Similarly for an HISI, reflect it to the guest as an ISI unless
>>  * it is an HPTE not found fault for a page that we have paged out.
>> diff --git a/arch/powerpc/kvm/book3s_segment.S b/arch/powerpc/kvm/book3s_segment.S
>> index 1abe478..bf20b45 100644
>> --- a/arch/powerpc/kvm/book3s_segment.S
>> +++ b/arch/powerpc/kvm/book3s_segment.S
>> @@ -287,7 +287,19 @@ ld_last_inst:
>> 	sync
>>
>> #endif
>> -	stw	r0, SVCPU_LAST_INST(r13)
>> +	ld	r8, SVCPU_SHADOW_SRR1(r13)
>> +
>> +	/* r10 = vcpu->arch.msr & MSR_LE */
>> +	rldicl	r10, r0, 0, 63
>> +	cmpdi	r10, 0
>> +	beq	1f
>> +
>> +	/* swap and store the result */
>> +	addi	r11, r13, SVCPU_LAST_INST
>> +	stwbrx	r0, 0, r11
>> +	b	no_ld_last_inst
>> +
>> +1:	stw	r0, SVCPU_LAST_INST(r13)
>>
>> no_ld_last_inst:
>>
>> diff --git a/arch/powerpc/kvm/emulate.c b/arch/powerpc/kvm/emulate.c
>> index 751cd45..20529ca 100644
>> --- a/arch/powerpc/kvm/emulate.c
>> +++ b/arch/powerpc/kvm/emulate.c
>> @@ -232,6 +232,7 @@ int kvmppc_emulate_instruction(struct kvm_run *run, struct kvm_vcpu *vcpu)
>> 	int sprn = get_sprn(inst);
>> 	enum emulation_result emulated = EMULATE_DONE;
>> 	int advance = 1;
>> +	int dont_byteswap = !kvmppc_need_byteswap(vcpu);
> 
> The parameter to kvmppc_handle_load is "is_bigendian" which is also the flag that we interpret 
> for our byte swaps later. I think we should preserve that semantic. Please call your variable 
> "is_bigendian" and create a separate helper for that one.
> 
> When little endian host kernels come, we only need to change the way kvmppc_complete_mmio_load 
> and kvmppc_handle_store swap things - probably according to user space endianness even.

ok.

Thanks for the review Alex, I will be sending a new patchset shortly.

C.