On 06/08/2014 05:05 AM, Christoffer Dall wrote:
> On Tue, Jun 03, 2014 at 04:19:27PM -0700, Mario Smarduch wrote:
>> This patch adds support for handling 2nd stage page faults during migration,
>> it disables faulting in huge pages, and disolves huge pages to page tables.
>
> s/disolves/dissolves/g
Will do.
>
>> In case migration is canceled huge pages will be used again.
>>
>> Signed-off-by: Mario Smarduch <m.smard...@samsung.com>
>> ---
>> arch/arm/kvm/mmu.c | 36 ++++++++++++++++++++++++++++++++++--
>> 1 file changed, 34 insertions(+), 2 deletions(-)
>>
>> diff --git a/arch/arm/kvm/mmu.c b/arch/arm/kvm/mmu.c
>> index 1c546c9..aca4fbf 100644
>> --- a/arch/arm/kvm/mmu.c
>> +++ b/arch/arm/kvm/mmu.c
>> @@ -966,6 +966,8 @@ 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;
>> + /* Get logging status, if dirty_bitmap is not NULL then logging is on */
>> + bool logging_active = !!memslot->dirty_bitmap;
>
>>
>> write_fault = kvm_is_write_fault(kvm_vcpu_get_hsr(vcpu));
>> if (fault_status == FSC_PERM && !write_fault) {
>> @@ -1019,10 +1021,16 @@ 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)
>> +
>> + /* When logging don't spend cycles to check for huge pages */
>
> drop the comment: either explain the entire clause (which would be too
> long) or don't explain anything.
>
Ok.
>> + if (!hugetlb && !force_pte && !logging_active)
>
> instead of having all this, can't you just change
>
> if (is_vm_hugetlb_page(vma)) to
> if (is_vm_hugetlb_page(vma) && !logging_active)
>
> then you're also not mucking around with the gfn etc.
I didn't want to modify this function too much, but if that's ok that
simplifies things a lot.
>
>> hugetlb = transparent_hugepage_adjust(&pfn, &fault_ipa);
>>
>> - if (hugetlb) {
>> + /*
>> + * Force all not present/perm faults to PTE handling, address both
>> + * PMD and PTE faults
>> + */
>
> I don't understand this comment? In which case does this apply?
>
The cases I see here -
- huge page permission fault is forced into page table code while logging
- pte permission/not present handled by page table code as before.
>> + if (hugetlb && !logging_active) {
>> pmd_t new_pmd = pfn_pmd(pfn, PAGE_S2);
>> new_pmd = pmd_mkhuge(new_pmd);
>> if (writable) {
>> @@ -1034,6 +1042,22 @@ 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) {
>> + /*
>> + * If pmd is mapping a huge page then clear it and let
>> + * stage2_set_pte() create a pte table. At the sametime
>> + * you write protect the pte (PAGE_S2 pgprot_t).
>> + */
>> + if (logging_active) {
>> + pmd_t *pmd;
>> + if (hugetlb) {
>> + pfn += pte_index(fault_ipa);
>> + gfn = fault_ipa >> PAGE_SHIFT;
>> + 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);
>> + }
>
> now instead of all this, you just need to check for kvm_pmd_huge() in
> stage2_set_pte() and if that's true, you clear it, and then then install
> your new pte.
Yes this really simplifies things!
>
>> kvm_set_s2pte_writable(&new_pte);
>> kvm_set_pfn_dirty(pfn);
>> }
>> @@ -1041,6 +1065,14 @@ 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);
>> }
>>
>> + /*
>> + * Log the dirty page in dirty_bitmap[], call regardless if logging is
>> + * disabled or enabled both cases handled safely.
>> + * TODO: for larger page size mark mulitple dirty page bits for each
>> + * 4k page.
>> + */
>> + if (writable)
>> + mark_page_dirty(kvm, gfn);
>
> what if you just faulted in a page on a read which wasn't present
> before but it happens to belong to a writeable memslot, is that page
> then dirty? hmmm.
>
A bug, must also check if it was a write fault not just that we're dealing with
a writable region. This one could be pretty bad on performance, not to mention
in accurate. It will be interesting to see new test results, glad you caught
that.
Thanks,
Mario.
>
>>
>> out_unlock:
>> spin_unlock(&kvm->mmu_lock);
>> --
>> 1.7.9.5
>>
>
> Thanks,
> -Christoffer
>
--
To unsubscribe from this list: send the line "unsubscribe kvm" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
