Re: [PATCH v2 2/9] KVM: arm64: Add a range to __pkvm_host_share_guest()

[Vincent Donnefort]

On Thu, Apr 03, 2025 at 03:27:15PM +0000, Quentin Perret wrote:
> On Thursday 06 Mar 2025 at 11:00:31 (+0000), Vincent Donnefort wrote:
> > +int __pkvm_host_share_guest(u64 pfn, u64 gfn, u64 nr_pages, struct pkvm_hyp_vcpu *vcpu,
> >  			    enum kvm_pgtable_prot prot)
> >  {
> >  	struct pkvm_hyp_vm *vm = pkvm_hyp_vcpu_to_hyp_vm(vcpu);
> >  	u64 phys = hyp_pfn_to_phys(pfn);
> >  	u64 ipa = hyp_pfn_to_phys(gfn);
> > +	enum pkvm_page_state state;
> >  	struct hyp_page *page;
> > +	u64 size;
> >  	int ret;
> >  
> >  	if (prot & ~KVM_PGTABLE_PROT_RWX)
> >  		return -EINVAL;
> >  
> > -	ret = check_range_allowed_memory(phys, phys + PAGE_SIZE);
> > +	ret = __guest_check_transition_size(phys, ipa, nr_pages, &size);
> > +	if (ret)
> > +		return ret;
> > +
> > +	ret = check_range_allowed_memory(phys, phys + size);
> >  	if (ret)
> >  		return ret;
> >  
> >  	host_lock_component();
> >  	guest_lock_component(vm);
> >  
> > -	ret = __guest_check_page_state_range(vcpu, ipa, PAGE_SIZE, PKVM_NOPAGE);
> > +	ret = __guest_check_page_state_range(vm, ipa, size, PKVM_NOPAGE);
> >  	if (ret)
> >  		goto unlock;
> >  
> > -	page = hyp_phys_to_page(phys);
> > -	switch (page->host_state) {
> > +	state = hyp_phys_to_page(phys)->host_state;
> > +	for_each_hyp_page(phys, size, page) {
> > +		if (page->host_state != state) {
> > +			ret = -EPERM;
> > +			goto unlock;
> > +		}
> > +	}
> > +
> > +	switch (state) {
> >  	case PKVM_PAGE_OWNED:
> > -		WARN_ON(__host_set_page_state_range(phys, PAGE_SIZE, PKVM_PAGE_SHARED_OWNED));
> > +		WARN_ON(__host_set_page_state_range(phys, size, PKVM_PAGE_SHARED_OWNED));
> >  		break;
> >  	case PKVM_PAGE_SHARED_OWNED:
> > -		if (page->host_share_guest_count)
> > -			break;
> > -		/* Only host to np-guest multi-sharing is tolerated */
> > -		WARN_ON(1);
> > -		fallthrough;
> > +		for_each_hyp_page(phys, size, page) {
> > +			/* Only host to np-guest multi-sharing is tolerated */
> > +			if (WARN_ON(!page->host_share_guest_count)) {
> > +				ret = -EPERM;
> > +				goto unlock;
> > +			}
> > +		}
> > +		break;
> >  	default:
> >  		ret = -EPERM;
> >  		goto unlock;
> >  	}
> >  
> > -	WARN_ON(kvm_pgtable_stage2_map(&vm->pgt, ipa, PAGE_SIZE, phys,
> > +	WARN_ON(kvm_pgtable_stage2_map(&vm->pgt, ipa, size, phys,
> >  				       pkvm_mkstate(prot, PKVM_PAGE_SHARED_BORROWED),
> >  				       &vcpu->vcpu.arch.pkvm_memcache, 0));
> > -	page->host_share_guest_count++;
> > +	__host_update_share_guest_count(phys, size, true);
> 
> So we're walking the entire phys range 3 times;
> 
> 	1. to check the host_state is consistent with that of the first
> 	page;
> 
> 	2. to set the state to SHARED_OWNED or to check the
> 	host_share_guest_count;
> 
> 	3. and then again here to update the host share guest count
> 
> I feel like we could probably remove at least one loop with a pattern
> like so:
> 
> 	for_each_hyp_page(phys, size, page) {
> 		switch (page->state) {
> 		case PKVM_PAGE_OWNED:
> 			continue;
> 		case PKVM_PAGE_SHARED_BORROWED:
> 			if (page->host_shared_guest_count)
> 				continue;
> 			fallthrough;
> 		default;
> 			ret = -EPERM;
> 			goto unlock;
> 		}
> 	}
> 
> 	for_each_hyp_page(phys, size, page) {
> 		page->host_state = PKVM_PAGE_SHARED_OWNED;
> 		page->host_share_guest_count++;
> 	}
> 
> That would also tolerate a mix of OWNED and SHARED_OWNED page in the
> range, which I'm not sure is needed but it doesn't cost us anything to
> support so ... :-)
> 
> Wdyt?

That sounds good, I'll drop __host_update_share_guest_count at the same
time to fold it directly into the share/unshare functions.

> 
> >  unlock:
> >  	guest_unlock_component(vm);
> > diff --git a/arch/arm64/kvm/pkvm.c b/arch/arm64/kvm/pkvm.c
> > index 930b677eb9b0..00fd9a524bf7 100644
> > --- a/arch/arm64/kvm/pkvm.c
> > +++ b/arch/arm64/kvm/pkvm.c
> > @@ -361,7 +361,7 @@ int pkvm_pgtable_stage2_map(struct kvm_pgtable *pgt, u64 addr, u64 size,
> >  		return -EINVAL;
> >  
> >  	lockdep_assert_held_write(&kvm->mmu_lock);
> > -	ret = kvm_call_hyp_nvhe(__pkvm_host_share_guest, pfn, gfn, prot);
> > +	ret = kvm_call_hyp_nvhe(__pkvm_host_share_guest, pfn, gfn, 1, prot);
> >  	if (ret) {
> >  		/* Is the gfn already mapped due to a racing vCPU? */
> >  		if (ret == -EPERM)
> > -- 
> > 2.48.1.711.g2feabab25a-goog
> >