Re: [PATCH kernel 3/9] KVM: PPC: Use preregistered memory API to access TCE list

[Alexey Kardashevskiy <aik@ozlabs.ru>]

On 03/08/2016 05:30 PM, David Gibson wrote:
> On Tue, Mar 08, 2016 at 04:47:20PM +1100, Alexey Kardashevskiy wrote:
>> On 03/07/2016 05:00 PM, David Gibson wrote:
>>> On Mon, Mar 07, 2016 at 02:41:11PM +1100, Alexey Kardashevskiy wrote:
>>>> VFIO on sPAPR already implements guest memory pre-registration
>>>> when the entire guest RAM gets pinned. This can be used to translate
>>>> the physical address of a guest page containing the TCE list
>>> >from H_PUT_TCE_INDIRECT.
>>>>
>>>> This makes use of the pre-registrered memory API to access TCE list
>>>> pages in order to avoid unnecessary locking on the KVM memory
>>>> reverse map.
>>>>
>>>> Signed-off-by: Alexey Kardashevskiy <aik@ozlabs.ru>
>>>
>>> Ok.. so, what's the benefit of not having to lock the rmap?
>>
>> Less locking -> less racing == good, no?
>
> Well.. maybe.  The increased difficulty in verifying that the code is
> correct isn't always a good price to pay.
 >
>>>> ---
>>>>   arch/powerpc/kvm/book3s_64_vio_hv.c | 86 ++++++++++++++++++++++++++++++-------
>>>>   1 file changed, 70 insertions(+), 16 deletions(-)
>>>>
>>>> diff --git a/arch/powerpc/kvm/book3s_64_vio_hv.c b/arch/powerpc/kvm/book3s_64_vio_hv.c
>>>> index 44be73e..af155f6 100644
>>>> --- a/arch/powerpc/kvm/book3s_64_vio_hv.c
>>>> +++ b/arch/powerpc/kvm/book3s_64_vio_hv.c
>>>> @@ -180,6 +180,38 @@ long kvmppc_gpa_to_ua(struct kvm *kvm, unsigned long gpa,
>>>>   EXPORT_SYMBOL_GPL(kvmppc_gpa_to_ua);
>>>>
>>>>   #ifdef CONFIG_KVM_BOOK3S_HV_POSSIBLE
>>>> +static mm_context_t *kvmppc_mm_context(struct kvm_vcpu *vcpu)
>>>> +{
>>>> +	struct task_struct *task;
>>>> +
>>>> +	task = vcpu->arch.run_task;
>>>> +	if (unlikely(!task || !task->mm))
>>>> +		return NULL;
>>>> +
>>>> +	return &task->mm->context;
>>>> +}
>>>> +
>>>> +static inline bool kvmppc_preregistered(struct kvm_vcpu *vcpu)
>>>> +{
>>>> +	mm_context_t *mm = kvmppc_mm_context(vcpu);
>>>> +
>>>> +	if (unlikely(!mm))
>>>> +		return false;
>>>> +
>>>> +	return mm_iommu_preregistered(mm);
>>>> +}
>>>> +
>>>> +static struct mm_iommu_table_group_mem_t *kvmppc_rm_iommu_lookup(
>>>> +		struct kvm_vcpu *vcpu, unsigned long ua, unsigned long size)
>>>> +{
>>>> +	mm_context_t *mm = kvmppc_mm_context(vcpu);
>>>> +
>>>> +	if (unlikely(!mm))
>>>> +		return NULL;
>>>> +
>>>> +	return mm_iommu_lookup_rm(mm, ua, size);
>>>> +}
>>>> +
>>>>   long kvmppc_h_put_tce(struct kvm_vcpu *vcpu, unsigned long liobn,
>>>>   		      unsigned long ioba, unsigned long tce)
>>>>   {
>>>> @@ -261,23 +293,44 @@ long kvmppc_rm_h_put_tce_indirect(struct kvm_vcpu *vcpu,
>>>>   	if (ret != H_SUCCESS)
>>>>   		return ret;
>>>>
>>>> -	if (kvmppc_gpa_to_ua(vcpu->kvm, tce_list, &ua, &rmap))
>>>> -		return H_TOO_HARD;
>>>> +	if (kvmppc_preregistered(vcpu)) {
>>>> +		/*
>>>> +		 * We get here if guest memory was pre-registered which
>>>> +		 * is normally VFIO case and gpa->hpa translation does not
>>>> +		 * depend on hpt.
>>>> +		 */
>>>> +		struct mm_iommu_table_group_mem_t *mem;
>>>>
>>>> -	rmap = (void *) vmalloc_to_phys(rmap);
>>>> +		if (kvmppc_gpa_to_ua(vcpu->kvm, tce_list, &ua, NULL))
>>>> +			return H_TOO_HARD;
>>>>
>>>> -	/*
>>>> -	 * Synchronize with the MMU notifier callbacks in
>>>> -	 * book3s_64_mmu_hv.c (kvm_unmap_hva_hv etc.).
>>>> -	 * While we have the rmap lock, code running on other CPUs
>>>> -	 * cannot finish unmapping the host real page that backs
>>>> -	 * this guest real page, so we are OK to access the host
>>>> -	 * real page.
>>>> -	 */
>>>> -	lock_rmap(rmap);
>>>> -	if (kvmppc_rm_ua_to_hpa(vcpu, ua, &tces)) {
>>>> -		ret = H_TOO_HARD;
>>>> -		goto unlock_exit;
>>>> +		mem = kvmppc_rm_iommu_lookup(vcpu, ua, IOMMU_PAGE_SIZE_4K);
>>>> +		if (!mem || mm_iommu_rm_ua_to_hpa(mem, ua, &tces))
>>>> +			return H_TOO_HARD;
>>>> +	} else {
>>>> +		/*
>>>> +		 * This is emulated devices case.
>>>> +		 * We do not require memory to be preregistered in this case
>>>> +		 * so lock rmap and do __find_linux_pte_or_hugepte().
>>>> +		 */
>>>> +		if (kvmppc_gpa_to_ua(vcpu->kvm, tce_list, &ua, &rmap))
>>>> +			return H_TOO_HARD;
>>>> +
>>>> +		rmap = (void *) vmalloc_to_phys(rmap);
>>>> +
>>>> +		/*
>>>> +		 * Synchronize with the MMU notifier callbacks in
>>>> +		 * book3s_64_mmu_hv.c (kvm_unmap_hva_hv etc.).
>>>> +		 * While we have the rmap lock, code running on other CPUs
>>>> +		 * cannot finish unmapping the host real page that backs
>>>> +		 * this guest real page, so we are OK to access the host
>>>> +		 * real page.
>>>> +		 */
>>>> +		lock_rmap(rmap);
>>>> +		if (kvmppc_rm_ua_to_hpa(vcpu, ua, &tces)) {
>>>> +			ret = H_TOO_HARD;
>>>> +			goto unlock_exit;
>>>> +		}
>>>>   	}
>>>>
>>>>   	for (i = 0; i < npages; ++i) {
>>>> @@ -291,7 +344,8 @@ long kvmppc_rm_h_put_tce_indirect(struct kvm_vcpu *vcpu,
>>>>   	}
>>>>
>>>>   unlock_exit:
>>>> -	unlock_rmap(rmap);
>>>> +	if (rmap)
>>>
>>> I don't see where rmap is initialized to NULL in the case where it's
>>> not being used.
>>
>> @rmap is not new to this function, and it has always been initialized to
>> NULL as it was returned via a pointer from kvmppc_gpa_to_ua().
>
> This comment confuses me.  Looking closer at the code I see you're
> right, and it's initialized to NULL where defined, which I missed.
>
> But that has nothing to do with being returned by pointer from
> kvmppc_gpa_to_ua(), since one of your branches in the new code no
> longer passes &rmap to that function.


So? The code is still correct - the "preregistered branch" does not touch 
NULL pointer, it remains NULL and unlock_rmap() is not called. I agree the 
patch is not the easiest to read but how can I improve it to get your "rb"? 
Replace "if(rmap)" with "if (kvmppc_preregistered(vcpu))"? Move that loop 
between lock_rmap/unlock_rmap to a helper? kvmppc_rm_h_put_tce_indirect() 
is not big enough to justify splitting, the comments inside it are though...



-- 
Alexey