Re: [PATCH v3] kvm: mmu: lazy collapse small sptes into large sptes

[Andres Lagar-Cavilla <andreslc@google.com>]

On Fri, Apr 3, 2015 at 12:40 AM, Wanpeng Li <wanpeng.li@linux.intel.com> wrote:
>
> There are two scenarios for the requirement of collapsing small sptes
> into large sptes.
> - dirty logging tracks sptes in 4k granularity, so large sptes are split,
>   the large sptes will be reallocated in the destination machine and the
>   guest in the source machine will be destroyed when live migration successfully.
>   However, the guest in the source machine will continue to run if live migration
>   fail due to some reasons, the sptes still keep small which lead to bad
>   performance.
> - our customers write tools to track the dirty speed of guests by EPT D bit/PML
>   in order to determine the most appropriate one to be live migrated, however
>   sptes will still keep small after tracking dirty speed.
>
> This patch introduce lazy collapse small sptes into large sptes, the memory region
> will be scanned on the ioctl context when dirty log is stopped, the ones which can
> be collapsed into large pages will be dropped during the scan, it depends the on
> later #PF to reallocate all large sptes.
>
> Reviewed-by: Xiao Guangrong <guangrong.xiao@linux.intel.com>
> Signed-off-by: Wanpeng Li <wanpeng.li@linux.intel.com>

Hi, apologies for late review (vacation), but wanted to bring
attention to a few matters:

>
> ---
> v2 -> v3:
>  * update comments
>  * fix infinite for loop
> v1 -> v2:
>  * use 'bool' instead of 'int'
>  * add more comments
>  * fix can not get the next spte after drop the current spte
>
>  arch/x86/include/asm/kvm_host.h |  2 ++
>  arch/x86/kvm/mmu.c              | 73 +++++++++++++++++++++++++++++++++++++++++
>  arch/x86/kvm/x86.c              | 19 +++++++++++
>  3 files changed, 94 insertions(+)
>
> diff --git a/arch/x86/include/asm/kvm_host.h b/arch/x86/include/asm/kvm_host.h
> index 30b28dc..91b5bdb 100644
> --- a/arch/x86/include/asm/kvm_host.h
> +++ b/arch/x86/include/asm/kvm_host.h
> @@ -854,6 +854,8 @@ void kvm_mmu_set_mask_ptes(u64 user_mask, u64 accessed_mask,
>  void kvm_mmu_reset_context(struct kvm_vcpu *vcpu);
>  void kvm_mmu_slot_remove_write_access(struct kvm *kvm,
>                                       struct kvm_memory_slot *memslot);
> +void kvm_mmu_zap_collapsible_sptes(struct kvm *kvm,
> +                                       struct kvm_memory_slot *memslot);
>  void kvm_mmu_slot_leaf_clear_dirty(struct kvm *kvm,
>                                    struct kvm_memory_slot *memslot);
>  void kvm_mmu_slot_largepage_remove_write_access(struct kvm *kvm,
> diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c
> index cee7592..ba002a0 100644
> --- a/arch/x86/kvm/mmu.c
> +++ b/arch/x86/kvm/mmu.c
> @@ -4465,6 +4465,79 @@ void kvm_mmu_slot_remove_write_access(struct kvm *kvm,
>                 kvm_flush_remote_tlbs(kvm);
>  }
>
> +static bool kvm_mmu_zap_collapsible_spte(struct kvm *kvm,
> +               unsigned long *rmapp)
> +{
> +       u64 *sptep;
> +       struct rmap_iterator iter;
> +       int need_tlb_flush = 0;
> +       pfn_t pfn;
> +       struct kvm_mmu_page *sp;
> +
> +       for (sptep = rmap_get_first(*rmapp, &iter); sptep;) {
> +               BUG_ON(!(*sptep & PT_PRESENT_MASK));
> +
> +               sp = page_header(__pa(sptep));
> +               pfn = spte_to_pfn(*sptep);
> +
> +               /*
> +                * Lets support EPT only for now, there still needs to figure
> +                * out an efficient way to let these codes be aware what mapping
> +                * level used in guest.
> +                */
> +               if (sp->role.direct &&
> +                       !kvm_is_reserved_pfn(pfn) &&
> +                       PageTransCompound(pfn_to_page(pfn))) {

Not your fault, but PageTransCompound is very unhappy naming, as it
also yields true for PageHuge. Suggestion: document this check covers
static hugetlbfs, or switch to PageCompound() check.

A slightly bolder approach would be to refactor and reuse the nearly
identical check done in transparent_hugepage_adjust, instead of
open-coding here. In essence this code is asking for the same check,
plus the out-of-band check for static hugepages.


> +                       drop_spte(kvm, sptep);
> +                       sptep = rmap_get_first(*rmapp, &iter);
> +                       need_tlb_flush = 1;
> +               } else
> +                       sptep = rmap_get_next(&iter);
> +       }
> +
> +       return need_tlb_flush;
> +}
> +
> +void kvm_mmu_zap_collapsible_sptes(struct kvm *kvm,
> +                       struct kvm_memory_slot *memslot)
> +{
> +       bool flush = false;
> +       unsigned long *rmapp;
> +       unsigned long last_index, index;
> +       gfn_t gfn_start, gfn_end;
> +
> +       spin_lock(&kvm->mmu_lock);
> +
> +       gfn_start = memslot->base_gfn;
> +       gfn_end = memslot->base_gfn + memslot->npages - 1;
> +
> +       if (gfn_start >= gfn_end)
> +               goto out;

I don't understand the value of this check here. Are we looking for a
broken memslot? Shouldn't this be a BUG_ON? Is this the place to care
about these things? npages is capped to KVM_MEM_MAX_NR_PAGES, i.e.
2^31. A 64 bit overflow would be caused by a gigantic gfn_start which
would be trouble in many other ways.

All this to say: please remove the above 5 lines and make code simpler.

> +
> +       rmapp = memslot->arch.rmap[0];
> +       last_index = gfn_to_index(gfn_end, memslot->base_gfn,
> +                                       PT_PAGE_TABLE_LEVEL);
> +
> +       for (index = 0; index <= last_index; ++index, ++rmapp) {

One could argue that the cleaner iteration should be over the gfn
space covered by the memslot, thus leaving the gfn <--> rmap <--> spte
interactions hidden under the hood of __gfn_to_rmap. That yields much
cleaner (IMHO) code:

    for (gfn = memslot->base_gfn; gfn <= memslot->base_gfn +
memslot->npages; gfn++) {
        flush |= kvm_mmu_zap_collapsible_spte(kvm, __gfn_to_rmap(gfn,
1, memslot));
        ....

Now you can also get rid of index, last_index and rmapp. And more
importantly, the code is more understandable, and follows pattern as
established in x86/kvm/mmu.

> +               if (*rmapp)
> +                       flush |= kvm_mmu_zap_collapsible_spte(kvm, rmapp);
> +
> +               if (need_resched() || spin_needbreak(&kvm->mmu_lock)) {
> +                       if (flush) {
> +                               kvm_flush_remote_tlbs(kvm);
> +                               flush = false;
> +                       }
> +                       cond_resched_lock(&kvm->mmu_lock);

Relinquishing this spinlock is problematic, because
commit_memory_region has not gotten around to removing write
protection. Are you certain no new write-protected PTEs will be
inserted by a racing fault that sneaks in while the spinlock is
relinquished?

If that can happen, you can decide that does not really matter, or the
collapsing is best effort, but either of those decisions should be
documented.

If you absolutely want to go for complete removal of PTEs, then you
need to restart the loop (costly!), or retain you structure that
iterates on the rmapp array, but restart from the beginning after each
spinlock removal (less costly, but not happy).

Thanks
Andres

> +               }
> +       }
> +
> +       if (flush)
> +               kvm_flush_remote_tlbs(kvm);
> +
> +out:
> +       spin_unlock(&kvm->mmu_lock);
> +}
> +
>  void kvm_mmu_slot_leaf_clear_dirty(struct kvm *kvm,
>                                    struct kvm_memory_slot *memslot)
>  {
> diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
> index 50861dd..a6cd10b 100644
> --- a/arch/x86/kvm/x86.c
> +++ b/arch/x86/kvm/x86.c
> @@ -7647,6 +7647,25 @@ void kvm_arch_commit_memory_region(struct kvm *kvm,
>         new = id_to_memslot(kvm->memslots, mem->slot);
>
>         /*
> +        * Dirty logging tracks sptes in 4k granularity, so large sptes are
> +        * split, the large sptes will be reallocated in the destination
> +        * machine and the guest in the source machine will be destroyed
> +        * when live migration successfully. However, the guest in the source
> +        * machine will continue to run if live migration fail due to some
> +        * reasons, the sptes still keep small which lead to bad performance.
> +        *
> +        * Lazy collapse small sptes into large sptes is intended to handle
> +        * this, the memory region will be scanned on the ioctl context when
> +        * dirty log is stopped, the ones which can be collapsed into large
> +        * pages will be dropped during the scan, it depends the on later #PF
> +        * to reallocate all large sptes.
> +        */
> +       if ((change != KVM_MR_DELETE) &&
> +               (old->flags & KVM_MEM_LOG_DIRTY_PAGES) &&
> +               !(new->flags & KVM_MEM_LOG_DIRTY_PAGES))
> +               kvm_mmu_zap_collapsible_sptes(kvm, new);
> +
> +       /*
>          * Set up write protection and/or dirty logging for the new slot.
>          *
>          * For KVM_MR_DELETE and KVM_MR_MOVE, the shadow pages of old slot have
> --
> 1.9.1
>
> --
> To unsubscribe from this list: send the line "unsubscribe kvm" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html




-- 
Andres Lagar-Cavilla | Google Kernel Team | andreslc@google.com