| Message ID | 8baa40dad8496abb2adb1096e0cf50dcc5f66802.1661331396.git.houwenlong.hwl@antgroup.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | KVM: x86/mmu: Fix wrong usages of range-based tlb flushing | expand |
On Wed, Aug 24, 2022 at 05:29:23PM +0800, Hou Wenlong wrote: > Only SP with PG_LEVLE_4K level could be unsync, so the size of gfn range > must be 1. > > Signed-off-by: Hou Wenlong <houwenlong.hwl@antgroup.com> > --- > arch/x86/kvm/mmu/paging_tmpl.h | 3 ++- > 1 file changed, 2 insertions(+), 1 deletion(-) > > diff --git a/arch/x86/kvm/mmu/paging_tmpl.h b/arch/x86/kvm/mmu/paging_tmpl.h > index 04149c704d5b..486a3163b1e4 100644 > --- a/arch/x86/kvm/mmu/paging_tmpl.h > +++ b/arch/x86/kvm/mmu/paging_tmpl.h > @@ -937,7 +937,8 @@ static void FNAME(invlpg)(struct kvm_vcpu *vcpu, gva_t gva, hpa_t root_hpa) > > mmu_page_zap_pte(vcpu->kvm, sp, sptep, NULL); > if (is_shadow_present_pte(old_spte)) > - kvm_flush_remote_tlbs_sptep(vcpu->kvm, sptep); > + kvm_flush_remote_tlbs_gfn(vcpu->kvm, > + kvm_mmu_page_get_gfn(sp, sptep - sp->spt), 1); The third argument to kvm_flush_remote_tlbs_gfn() is the level, not the number of pages. But that aside, I don't understand why this patch is necessary. kvm_flush_remote_tlbs_sptep() should already do the right thing. > > if (!rmap_can_add(vcpu)) > break; > -- > 2.31.1 >
On Thu, Sep 08, 2022 at 01:40:16AM +0800, David Matlack wrote: > On Wed, Aug 24, 2022 at 05:29:23PM +0800, Hou Wenlong wrote: > > Only SP with PG_LEVLE_4K level could be unsync, so the size of gfn range > > must be 1. > > > > Signed-off-by: Hou Wenlong <houwenlong.hwl@antgroup.com> > > --- > > arch/x86/kvm/mmu/paging_tmpl.h | 3 ++- > > 1 file changed, 2 insertions(+), 1 deletion(-) > > > > diff --git a/arch/x86/kvm/mmu/paging_tmpl.h b/arch/x86/kvm/mmu/paging_tmpl.h > > index 04149c704d5b..486a3163b1e4 100644 > > --- a/arch/x86/kvm/mmu/paging_tmpl.h > > +++ b/arch/x86/kvm/mmu/paging_tmpl.h > > @@ -937,7 +937,8 @@ static void FNAME(invlpg)(struct kvm_vcpu *vcpu, gva_t gva, hpa_t root_hpa) > > > > mmu_page_zap_pte(vcpu->kvm, sp, sptep, NULL); > > if (is_shadow_present_pte(old_spte)) > > - kvm_flush_remote_tlbs_sptep(vcpu->kvm, sptep); > > + kvm_flush_remote_tlbs_gfn(vcpu->kvm, > > + kvm_mmu_page_get_gfn(sp, sptep - sp->spt), 1); > > The third argument to kvm_flush_remote_tlbs_gfn() is the level, not the > number of pages. But that aside, I don't understand why this patch is > necessary. kvm_flush_remote_tlbs_sptep() should already do the right > thing. > Since only SP with PG_LEVEL_4K level could be unsync, so the level must be PG_LEVEL_4K, then sp->role.level access could be dropped. However, I'm not sure whether it is useful. I can drop it if it is useless. > > > > if (!rmap_can_add(vcpu)) > > break; > > -- > > 2.31.1 > >
On Tue, Sep 13, 2022 at 5:58 AM Hou Wenlong <houwenlong.hwl@antgroup.com> wrote: > > On Thu, Sep 08, 2022 at 01:40:16AM +0800, David Matlack wrote: > > On Wed, Aug 24, 2022 at 05:29:23PM +0800, Hou Wenlong wrote: > > > Only SP with PG_LEVLE_4K level could be unsync, so the size of gfn range > > > must be 1. > > > > > > Signed-off-by: Hou Wenlong <houwenlong.hwl@antgroup.com> > > > --- > > > arch/x86/kvm/mmu/paging_tmpl.h | 3 ++- > > > 1 file changed, 2 insertions(+), 1 deletion(-) > > > > > > diff --git a/arch/x86/kvm/mmu/paging_tmpl.h b/arch/x86/kvm/mmu/paging_tmpl.h > > > index 04149c704d5b..486a3163b1e4 100644 > > > --- a/arch/x86/kvm/mmu/paging_tmpl.h > > > +++ b/arch/x86/kvm/mmu/paging_tmpl.h > > > @@ -937,7 +937,8 @@ static void FNAME(invlpg)(struct kvm_vcpu *vcpu, gva_t gva, hpa_t root_hpa) > > > > > > mmu_page_zap_pte(vcpu->kvm, sp, sptep, NULL); > > > if (is_shadow_present_pte(old_spte)) > > > - kvm_flush_remote_tlbs_sptep(vcpu->kvm, sptep); > > > + kvm_flush_remote_tlbs_gfn(vcpu->kvm, > > > + kvm_mmu_page_get_gfn(sp, sptep - sp->spt), 1); > > > > The third argument to kvm_flush_remote_tlbs_gfn() is the level, not the > > number of pages. But that aside, I don't understand why this patch is > > necessary. kvm_flush_remote_tlbs_sptep() should already do the right > > thing. > > > Since only SP with PG_LEVEL_4K level could be unsync, so the level must > be PG_LEVEL_4K, then sp->role.level access could be dropped. However, > I'm not sure whether it is useful. I can drop it if it is useless. Ah, I see. I would be surprised if avoiding the read of sp->role.level has any noticeable impact on VM performance so I vote to drop this patch. > > > > > > > if (!rmap_can_add(vcpu)) > > > break; > > > -- > > > 2.31.1 > > >
On Tue, Sep 13, 2022, David Matlack wrote: > On Tue, Sep 13, 2022 at 5:58 AM Hou Wenlong <houwenlong.hwl@antgroup.com> wrote: > > > > On Thu, Sep 08, 2022 at 01:40:16AM +0800, David Matlack wrote: > > > On Wed, Aug 24, 2022 at 05:29:23PM +0800, Hou Wenlong wrote: > > > > Only SP with PG_LEVLE_4K level could be unsync, so the size of gfn range > > > > must be 1. > > > > > > > > Signed-off-by: Hou Wenlong <houwenlong.hwl@antgroup.com> > > > > --- > > > > arch/x86/kvm/mmu/paging_tmpl.h | 3 ++- > > > > 1 file changed, 2 insertions(+), 1 deletion(-) > > > > > > > > diff --git a/arch/x86/kvm/mmu/paging_tmpl.h b/arch/x86/kvm/mmu/paging_tmpl.h > > > > index 04149c704d5b..486a3163b1e4 100644 > > > > --- a/arch/x86/kvm/mmu/paging_tmpl.h > > > > +++ b/arch/x86/kvm/mmu/paging_tmpl.h > > > > @@ -937,7 +937,8 @@ static void FNAME(invlpg)(struct kvm_vcpu *vcpu, gva_t gva, hpa_t root_hpa) > > > > > > > > mmu_page_zap_pte(vcpu->kvm, sp, sptep, NULL); > > > > if (is_shadow_present_pte(old_spte)) > > > > - kvm_flush_remote_tlbs_sptep(vcpu->kvm, sptep); > > > > + kvm_flush_remote_tlbs_gfn(vcpu->kvm, > > > > + kvm_mmu_page_get_gfn(sp, sptep - sp->spt), 1); > > > > > > The third argument to kvm_flush_remote_tlbs_gfn() is the level, not the > > > number of pages. But that aside, I don't understand why this patch is > > > necessary. kvm_flush_remote_tlbs_sptep() should already do the right > > > thing. > > > > > Since only SP with PG_LEVEL_4K level could be unsync, so the level must > > be PG_LEVEL_4K, then sp->role.level access could be dropped. However, > > I'm not sure whether it is useful. I can drop it if it is useless. > > Ah, I see. I would be surprised if avoiding the read of sp->role.level > has any noticeable impact on VM performance so I vote to drop this patch. Agreed, the cost of the sp->role.level lookup is negligible in this case, and IMO using kvm_flush_remote_tlbs_sptep() is more intuitive. If kvm_flush_remote_tlbs_sptep() didn't exist and this was open coding the use of kvm_flush_remote_tlbs_with_address() + KVM_PAGES_PER_HPAGE(), then I would be in favor of hardcoding '1', because at that point the use of KVM_PAGES_PER_HPAGE() is misleading in its own way.
diff --git a/arch/x86/kvm/mmu/paging_tmpl.h b/arch/x86/kvm/mmu/paging_tmpl.h index 04149c704d5b..486a3163b1e4 100644 --- a/arch/x86/kvm/mmu/paging_tmpl.h +++ b/arch/x86/kvm/mmu/paging_tmpl.h @@ -937,7 +937,8 @@ static void FNAME(invlpg)(struct kvm_vcpu *vcpu, gva_t gva, hpa_t root_hpa) mmu_page_zap_pte(vcpu->kvm, sp, sptep, NULL); if (is_shadow_present_pte(old_spte)) - kvm_flush_remote_tlbs_sptep(vcpu->kvm, sptep); + kvm_flush_remote_tlbs_gfn(vcpu->kvm, + kvm_mmu_page_get_gfn(sp, sptep - sp->spt), 1); if (!rmap_can_add(vcpu)) break;
Only SP with PG_LEVLE_4K level could be unsync, so the size of gfn range must be 1. Signed-off-by: Hou Wenlong <houwenlong.hwl@antgroup.com> --- arch/x86/kvm/mmu/paging_tmpl.h | 3 ++- 1 file changed, 2 insertions(+), 1 deletion(-)