[PATCH v13 7/7] arm: KVM: ARMv7 dirty page logging 2nd stage page fault

[Marc Zyngier]

On 07/11/14 00:40, Mario Smarduch wrote:
> This patch adds support for handling 2nd stage page faults during migration,
> it disables faulting in huge pages, and dissolves huge pages to page tables.
> In case migration is canceled huge pages are used again.
> 
> Reviewed-by: Christoffer Dall <christoffer.dall at linaro.org>
> Signed-off-by: Mario Smarduch <m.smarduch at samsung.com>
> ---
>  arch/arm/kvm/mmu.c |   47 ++++++++++++++++++++++++++++++++++++++++-------
>  1 file changed, 40 insertions(+), 7 deletions(-)
> 
> diff --git a/arch/arm/kvm/mmu.c b/arch/arm/kvm/mmu.c
> index 2f5131e..d511fc0 100644
> --- a/arch/arm/kvm/mmu.c
> +++ b/arch/arm/kvm/mmu.c
> @@ -47,6 +47,15 @@ static phys_addr_t hyp_idmap_vector;
>  #define kvm_pmd_huge(_x)	(pmd_huge(_x) || pmd_trans_huge(_x))
>  #define kvm_pud_huge(_x)	pud_huge(_x)
>  
> +static bool kvm_get_logging_state(struct kvm_memory_slot *memslot)
> +{
> +#ifdef CONFIG_ARM
> +	return !!memslot->dirty_bitmap;
> +#else
> +	return false;
> +#endif
> +}
> +
>  static void kvm_tlb_flush_vmid_ipa(struct kvm *kvm, phys_addr_t ipa)
>  {
>  	/*
> @@ -626,7 +635,8 @@ static int stage2_set_pmd_huge(struct kvm *kvm, struct kvm_mmu_memory_cache
>  }
>  
>  static int stage2_set_pte(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
> -			  phys_addr_t addr, const pte_t *new_pte, bool iomap)
> +			  phys_addr_t addr, const pte_t *new_pte, bool iomap,
> +			  bool logging_active)

Yuk. Yet another parameter. Can't we have a set of flags instead,
indicating both iomap and logging in one go? That would make things more
readable (at least for me).

>  {
>  	pmd_t *pmd;
>  	pte_t *pte, old_pte;
> @@ -641,6 +651,18 @@ static int stage2_set_pte(struct kvm *kvm, struct kvm_mmu_memory_cache *cache,
>  		return 0;
>  	}
>  
> +	/*
> +	 * While dirty memory logging, clear PMD entry for huge page and split
> +	 * into smaller pages, to track dirty memory at page granularity.
> +	 */
> +	if (logging_active && kvm_pmd_huge(*pmd)) {
> +		phys_addr_t ipa = pmd_pfn(*pmd) << PAGE_SHIFT;
> +
> +		pmd_clear(pmd);
> +		kvm_tlb_flush_vmid_ipa(kvm, ipa);
> +		put_page(virt_to_page(pmd));
> +	}
> +

If we have huge PUDs (like on arn64 with 4k Pages and 4 levels of
translation), would we need something similar? If that's the case, a
comment would be very welcome.

>  	/* Create stage-2 page mappings - Level 2 */
>  	if (pmd_none(*pmd)) {
>  		if (!cache)
> @@ -693,7 +715,7 @@ int kvm_phys_addr_ioremap(struct kvm *kvm, phys_addr_t guest_ipa,
>  		if (ret)
>  			goto out;
>  		spin_lock(&kvm->mmu_lock);
> -		ret = stage2_set_pte(kvm, &cache, addr, &pte, true);
> +		ret = stage2_set_pte(kvm, &cache, addr, &pte, true, false);
>  		spin_unlock(&kvm->mmu_lock);
>  		if (ret)
>  			goto out;
> @@ -910,6 +932,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
>  	struct vm_area_struct *vma;
>  	pfn_t pfn;
>  	pgprot_t mem_type = PAGE_S2;
> +	bool logging_active = kvm_get_logging_state(memslot);
>  
>  	write_fault = kvm_is_write_fault(kvm_vcpu_get_hsr(vcpu));
>  	if (fault_status == FSC_PERM && !write_fault) {
> @@ -920,7 +943,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
>  	/* Let's check if we will get back a huge page backed by hugetlbfs */
>  	down_read(&current->mm->mmap_sem);
>  	vma = find_vma_intersection(current->mm, hva, hva + 1);
> -	if (is_vm_hugetlb_page(vma)) {
> +	if (is_vm_hugetlb_page(vma) && !logging_active) {
>  		hugetlb = true;
>  		gfn = (fault_ipa & PMD_MASK) >> PAGE_SHIFT;
>  	} else {
> @@ -966,7 +989,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
>  	spin_lock(&kvm->mmu_lock);
>  	if (mmu_notifier_retry(kvm, mmu_seq))
>  		goto out_unlock;
> -	if (!hugetlb && !force_pte)
> +	if (!hugetlb && !force_pte && !logging_active)
>  		hugetlb = transparent_hugepage_adjust(&pfn, &fault_ipa);
>  
>  	if (hugetlb) {
> @@ -986,10 +1009,12 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
>  		}
>  		coherent_cache_guest_page(vcpu, hva, PAGE_SIZE);
>  		ret = stage2_set_pte(kvm, memcache, fault_ipa, &new_pte,
> -				     mem_type == PAGE_S2_DEVICE);
> +					mem_type == PAGE_S2_DEVICE,
> +					logging_active);
>  	}
>  
> -
> +	if (write_fault)
> +		mark_page_dirty(kvm, gfn);
>  out_unlock:
>  	spin_unlock(&kvm->mmu_lock);
>  	kvm_release_pfn_clean(pfn);
> @@ -1139,7 +1164,15 @@ static void kvm_set_spte_handler(struct kvm *kvm, gpa_t gpa, void *data)
>  {
>  	pte_t *pte = (pte_t *)data;
>  
> -	stage2_set_pte(kvm, NULL, gpa, pte, false);
> +	/*
> +	 * We can always call stage2_set_pte with logging_active == false,
> +	 * because MMU notifiers will have unmapped a huge PMD before calling
> +	 * ->change_pte() (which in turn calls kvm_set_spte_hva()) and therefore
> +	 * stage2_set_pte() never needs to clear out a huge PMD through this
> +	 * calling path.
> +	 */
> +
> +	stage2_set_pte(kvm, NULL, gpa, pte, false, false);
>  }
>  
>  
> 

Thanks,

	M.
-- 
Jazz is not dead. It just smells funny...