Re: [Qemu-devel] [PATCH] KVM: MMU: fast cleanup D bit based on fast write protect
On 12/01/19 09:20, Zhuangyanying wrote:
> From: Zhuang Yanying
>
> Recently I tested live-migration with large-memory guests, find vcpu may hang
> for a long time while starting migration, such as 9s for
> 2048G(linux-4.20.1+qemu-3.1.0).
> The reason is memory_global_dirty_log_start() taking too long, and the vcpu
> is waiting for BQL. The page-by-page D bit clearup is the main time
> consumption.
> I think that the idea of "KVM: MMU: fast write protect" by xiaoguangrong,
> especially the function kvm_mmu_write_protect_all_pages(), is very helpful.
> After a little modifcation, on his patch, can solve this problem, 9s to 0.5s.
>
> At the begining of live migration, write protection is only applied to the
> top-level SPTE. Then the write from vm trigger the EPT violation, with
> for_each_shadow_entry write protection is performed at dirct_map.
> Finally the Dirty bit of the target page(at level 1 page table) is cleared,
> and the dirty page tracking is started. Of coure, the page where GPA is
> located is marked dirty when mmu_set_spte.
> A similar implementation on xen, just emt instead of write protection.
>
> What do you think about this solution?
What tree does this patch apply to?
Paolo
> ---
> mmu.c | 5 -
> vmx.c | 3 +--
> 2 files changed, 5 insertions(+), 3 deletions(-)
>
> diff --git a/mmu.c b/mmu.c
> index b079d74..f49d316 100755
> --- a/mmu.c
> +++ b/mmu.c
> @@ -3210,7 +3210,10 @@ static bool mmu_load_shadow_page(struct kvm *kvm,
> struct kvm_mmu_page *sp)
> break;
>
> if (is_last_spte(spte, sp->role.level)) {
> - flush |= spte_write_protect(sptep, false);
> + if (sp->role.level == PT_PAGE_TABLE_LEVEL)
> + flush |= spte_clear_dirty(sptep);
> + else
> + flush |= spte_write_protect(sptep, false);
> continue;
> }
>
> diff --git a/vmx.c b/vmx.c
> index 95784bc..7ec717f 100755
> --- a/vmx.c
> +++ b/vmx.c
> @@ -14421,8 +14421,7 @@ static void vmx_sched_in(struct kvm_vcpu *vcpu, int
> cpu)
> static void vmx_slot_enable_log_dirty(struct kvm *kvm,
>struct kvm_memory_slot *slot)
> {
> - kvm_mmu_slot_leaf_clear_dirty(kvm, slot);
> - kvm_mmu_slot_largepage_remove_write_access(kvm, slot);
> + kvm_mmu_write_protect_all_pages(kvm, true);
> }
>
> static void vmx_slot_disable_log_dirty(struct kvm *kvm,
>
[Qemu-devel] [PATCH] KVM: MMU: fast cleanup D bit based on fast write protect
From: Zhuang Yanying
Recently I tested live-migration with large-memory guests, find vcpu may hang
for a long time while starting migration, such as 9s for
2048G(linux-4.20.1+qemu-3.1.0).
The reason is memory_global_dirty_log_start() taking too long, and the vcpu is
waiting for BQL. The page-by-page D bit clearup is the main time consumption.
I think that the idea of "KVM: MMU: fast write protect" by xiaoguangrong,
especially the function kvm_mmu_write_protect_all_pages(), is very helpful.
After a little modifcation, on his patch, can solve this problem, 9s to 0.5s.
At the begining of live migration, write protection is only applied to the
top-level SPTE. Then the write from vm trigger the EPT violation, with
for_each_shadow_entry write protection is performed at dirct_map.
Finally the Dirty bit of the target page(at level 1 page table) is cleared, and
the dirty page tracking is started. Of coure, the page where GPA is located is
marked dirty when mmu_set_spte.
A similar implementation on xen, just emt instead of write protection.
What do you think about this solution?
---
mmu.c | 5 -
vmx.c | 3 +--
2 files changed, 5 insertions(+), 3 deletions(-)
diff --git a/mmu.c b/mmu.c
index b079d74..f49d316 100755
--- a/mmu.c
+++ b/mmu.c
@@ -3210,7 +3210,10 @@ static bool mmu_load_shadow_page(struct kvm *kvm, struct
kvm_mmu_page *sp)
break;
if (is_last_spte(spte, sp->role.level)) {
- flush |= spte_write_protect(sptep, false);
+ if (sp->role.level == PT_PAGE_TABLE_LEVEL)
+ flush |= spte_clear_dirty(sptep);
+ else
+ flush |= spte_write_protect(sptep, false);
continue;
}
diff --git a/vmx.c b/vmx.c
index 95784bc..7ec717f 100755
--- a/vmx.c
+++ b/vmx.c
@@ -14421,8 +14421,7 @@ static void vmx_sched_in(struct kvm_vcpu *vcpu, int cpu)
static void vmx_slot_enable_log_dirty(struct kvm *kvm,
struct kvm_memory_slot *slot)
{
- kvm_mmu_slot_leaf_clear_dirty(kvm, slot);
- kvm_mmu_slot_largepage_remove_write_access(kvm, slot);
+ kvm_mmu_write_protect_all_pages(kvm, true);
}
static void vmx_slot_disable_log_dirty(struct kvm *kvm,
--
1.8.3.1
