Re: [PATCH 4/5] add 2nd stage page fault handling during live migration

[Marc Zyngier <marc.zyngier@arm.com>]

On Thu, Apr 17 2014 at  2:34:39 am BST, Mario Smarduch <m.smarduch@samsung.com> wrote:
> Additional logic to handle second stage page faults during migration. Primarily
> page faults are prevented from creating huge pages.
>
>
> Signed-off-by: Mario Smarduch <m.smarduch@samsung.com>
> ---
>  arch/arm/kvm/mmu.c |   33 +++++++++++++++++++++++++++++++--
>  1 file changed, 31 insertions(+), 2 deletions(-)
>
> diff --git a/arch/arm/kvm/mmu.c b/arch/arm/kvm/mmu.c
> index 47bec1c..ebec33c 100644
> --- a/arch/arm/kvm/mmu.c
> +++ b/arch/arm/kvm/mmu.c
> @@ -839,6 +839,7 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
>  	struct kvm_mmu_memory_cache *memcache = &vcpu->arch.mmu_page_cache;
>  	struct vm_area_struct *vma;
>  	pfn_t pfn;
> +	bool migration_active;
>  
>  	write_fault = kvm_is_write_fault(kvm_vcpu_get_hsr(vcpu));
>  	if (fault_status == FSC_PERM && !write_fault) {
> @@ -890,12 +891,22 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
>  		return -EFAULT;
>  
>  	spin_lock(&kvm->mmu_lock);
> +	/* place inside lock to prevent race condition when whole VM is being
> +	 * write proteced initially, prevent pmd update if it's split up.
> +	 */
> +	migration_active = vcpu->kvm->arch.migration_in_progress;
> +
>  	if (mmu_notifier_retry(kvm, mmu_seq))
>  		goto out_unlock;
> -	if (!hugetlb && !force_pte)
> +
> +	/* During migration don't rebuild huge pages */
> +	if (!hugetlb && !force_pte && !migration_active)
>  		hugetlb = transparent_hugepage_adjust(&pfn, &fault_ipa);
>  
> -	if (hugetlb) {
> +	/* Steer away from installing PMDs if migrating, migration failed,
> +	 * or this an initial page fault. Migrating huge pages is too slow.
> +	 */
> +	if (!migration_active && hugetlb) {
>  		pmd_t new_pmd = pfn_pmd(pfn, PAGE_S2);
>  		new_pmd = pmd_mkhuge(new_pmd);
>  		if (writable) {
> @@ -907,6 +918,22 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
>  	} else {
>  		pte_t new_pte = pfn_pte(pfn, PAGE_S2);
>  		if (writable) {
> +			/* First convert huge page pfn to normal 4k page pfn,
> +			 * while  migration is in progress.
> +			 * Second in migration mode and rare case where
> +			 * splitting of huge pages fails check if pmd is
> +			 * mapping a huge page if it is then clear it so
> +			 * stage2_set_pte() can map in a small page.
> +			 */
> +			if (migration_active && hugetlb) {
> +				pmd_t *pmd;
> +				pfn += (fault_ipa >> PAGE_SHIFT) &
> +							(PTRS_PER_PMD-1);

Shouldn't that be "pfn += pte_index(fault_addr);"?

> +				new_pte = pfn_pte(pfn, PAGE_S2);
> +				pmd = stage2_get_pmd(kvm, NULL, fault_ipa);
> +				if (pmd && kvm_pmd_huge(*pmd))
> +					clear_pmd_entry(kvm, pmd, fault_ipa);
> +			}
>  			kvm_set_s2pte_writable(&new_pte);
>  			kvm_set_pfn_dirty(pfn);
>  		}
> @@ -914,6 +941,8 @@ static int user_mem_abort(struct kvm_vcpu *vcpu, phys_addr_t fault_ipa,
>  		ret = stage2_set_pte(kvm, memcache, fault_ipa, &new_pte, false);
>  	}
>  
> +	if (writable)

Shouldn't that be done only when migration is active?

> +		mark_page_dirty(kvm, gfn);
>  
>  out_unlock:
>  	spin_unlock(&kvm->mmu_lock);

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