KVM: x86: count number of zapped pages for tdp_mmu
PATCH 1/1 KVM: x86: count number of zapped pages for tdp_mmu
Count the number of zapped pages of tdp_mmu for vm stat.
--- arch/x86/kvm/mmu/tdp_mmu.c | 1 + 1 file changed, 1 insertion(+)
arch/x86/kvm/mmu/tdp_mmu.c +1 -0
@@ -325,6 +325,7 @@ static void handle_removed_pt(struct kvm *kvm, tdp_ptep_t pt, bool shared)int i;trace_kvm_mmu_prepare_zap_page (sp); + ++kvm->stat.mmu_shadow_zapped;tdp_mmu_unlink_sp (kvm, sp, shared);
Signature · 1 line
2.40.1
Sean Christopherson
maintainer+David
Quoted 1 line · Liang Chen
Count the number of zapped pages of
tdp_mmufor vm stat.
Why? I don't necessarily disagree with the change, but it's also not obvious that this information is all that useful for the TDP MMU, e.g. the pf_fixed/taken stats largely capture the same information.
Quoted 10 lines
Signed-off-by Liang Chen <[email protected]>--- arch/x86/kvm/mmu/tdp_mmu.c | 1 + 1 file changed, 1 insertion(+)arch/x86/kvm/mmu/tdp_mmu.c +0
-0
arch/x86/kvm/mmu/tdp_mmu.c static void handle_removed_pt(struct kvm *kvm, tdp_ptep_t pt, bool shared) · line 328
Quoted 4 lines
arch/x86/kvm/mmu/tdp_mmu.c +1
-0@@ -325,6 +325,7 @@ static void handle_removed_pt(struct kvm *kvm, tdp_ptep_t pt, bool shared)int i;trace_kvm_mmu_prepare_zap_page (sp); + ++kvm->stat.mmu_shadow_zapped;
This isn't thread safe. The TDP MMU can zap PTEs with mmu_lock held for read, i.e. this needs to be an atomic access. And tdp_mmu_unlink_sp() or even tdp_unaccount_mmu_page() seems like a better fit for the stats update.
Quoted 5 lines
arch/x86/kvm/mmu/tdp_mmu.c +0
-0tdp_mmu_unlink_sp (kvm, sp, shared);Signature · 1 line
2.40.1
Quoted 10 lines · Sean Christopherson
+David
Quoted 1 line · Liang Chen
Count the number of zapped pages of
tdp_mmufor vm stat.Why? I don't necessarily disagree with the change, but it's also not obvious that this information is all that useful for the TDP MMU, e.g. the pf_fixed/taken stats largely capture the same information.
We are attempting to make zapping specific to a particular memory slot, something like below.
void kvm_tdp_zap_pages_in_memslot(struct kvm *kvm, struct kvm_memory_slot *slot)
{
struct kvm_mmu_page *root;
bool shared = false;
struct tdp_iter iter;
gfn_t end = slot->base_gfn + slot->npages;
gfn_t start = slot->base_gfn;
write_lock(&kvm->mmu_lock);
rcu_read_lock();
for_each_tdp_mmu_root_yield_safe(kvm, root, false) {
for_each_tdp_pte_min_level(iter, root,
root->role.level, start, end) {
if (tdp_mmu_iter_cond_resched(kvm, &iter, false, false))
continue;
if (!is_shadow_present_pte(iter.old_spte))
continue;
tdp_mmu_set_spte(kvm, &iter, 0);
}
}
kvm_flush_remote_tlbs(kvm);
rcu_read_unlock();
write_unlock(&kvm->mmu_lock);
}I noticed that it was previously done to the legacy MMU, but encountered some subtle issues with VFIO. I'm not sure if the issue is still there with TDP_MMU. So we are trying to do more tests and analyses before submitting a patch. This provides me a convenient way to observe the number of pages being zapped.
Quoted 12 lines
Quoted 12 lines
Signed-off-by Liang Chen <[email protected]>--- arch/x86/kvm/mmu/tdp_mmu.c | 1 + 1 file changed, 1 insertion(+)arch/x86/kvm/mmu/tdp_mmu.c +0
-0@@ -325,6 +325,7 @@ static void handle_removed_pt(struct kvm *kvm, tdp_ptep_t pt, bool shared)int i;
Quoted 7 lines
arch/x86/kvm/mmu/tdp_mmu.c static void handle_removed_pt(struct kvm *kvm, tdp_ptep_t pt, bool shared) · line 328
Quoted 2 lines
arch/x86/kvm/mmu/tdp_mmu.c +1
-0trace_kvm_mmu_prepare_zap_page (sp); + ++kvm->stat.mmu_shadow_zapped;This isn't thread safe. The TDP MMU can zap PTEs with
mmu_lockheld for read, i.e. this needs to be an atomic access. Andtdp_mmu_unlink_sp()or eventdp_unaccount_mmu_page()seems like a better fit for the stats update.
Yeah, that's an issue. Perhaps it isn't worth the effort to convert mmu_shadow_zapped to atomic type or use any concurrency control means. I will just add an attribute in debugfs to keep track of the number.
Quoted 5 lines
Quoted 5 lines
arch/x86/kvm/mmu/tdp_mmu.c +0
-1tdp_mmu_unlink_sp (kvm, sp, shared); --2.40.1
Quoted 48 lines · Liang Chen
Quoted 10 lines · Sean Christopherson
+David
Quoted 1 line · Liang Chen
Count the number of zapped pages of
tdp_mmufor vm stat.Why? I don't necessarily disagree with the change, but it's also not obvious that this information is all that useful for the TDP MMU, e.g. the pf_fixed/taken stats largely capture the same information.
We are attempting to make zapping specific to a particular memory slot, something like below.
void
kvm_tdp_zap_pages_in_memslot(struct kvm *kvm, struct kvm_memory_slot *slot){ struct kvm_mmu_page *root; bool shared = false; struct tdp_iter iter; gfn_t end = slot->base_gfn + slot->npages; gfn_t start = slot->base_gfn; write_lock(&kvm->mmu_lock); rcu_read_lock(); for_each_tdp_mmu_root_yield_safe(kvm, root, false) { for_each_tdp_pte_min_level(iter, root, root->role.level, start, end) { if (tdp_mmu_iter_cond_resched(kvm, &iter, false, false)) continue; if (!is_shadow_present_pte(iter.old_spte)) continue; tdp_mmu_set_spte(kvm, &iter, 0); } } kvm_flush_remote_tlbs(kvm); rcu_read_unlock(); write_unlock(&kvm->mmu_lock); }
Quoted 5 lines
I noticed that it was previously done to the legacy MMU, but encountered some subtle issues with VFIO. I'm not sure if the issue is still there with
TDP_MMU. So we are trying to do more tests and analyses before submitting a patch. This provides me a convenient way to observe the number of pages being zapped.
Note you could also use the existing tracepoint to observe the number of pages being zapped in a given test run. e.g.
perf stat -e kvmmmu:kvm_mmu_prepare_zap_page -- <cmd>
Quoted 56 lines · David Matlack
Quoted 48 lines · Liang Chen
Quoted 10 lines · Sean Christopherson
+David
Quoted 1 line · Liang Chen
Count the number of zapped pages of
tdp_mmufor vm stat.Why? I don't necessarily disagree with the change, but it's also not obvious that this information is all that useful for the TDP MMU, e.g. the pf_fixed/taken stats largely capture the same information.
We are attempting to make zapping specific to a particular memory slot, something like below.
void
kvm_tdp_zap_pages_in_memslot(struct kvm *kvm, struct kvm_memory_slot *slot){ struct kvm_mmu_page *root; bool shared = false; struct tdp_iter iter; gfn_t end = slot->base_gfn + slot->npages; gfn_t start = slot->base_gfn; write_lock(&kvm->mmu_lock); rcu_read_lock(); for_each_tdp_mmu_root_yield_safe(kvm, root, false) { for_each_tdp_pte_min_level(iter, root, root->role.level, start, end) { if (tdp_mmu_iter_cond_resched(kvm, &iter, false, false)) continue; if (!is_shadow_present_pte(iter.old_spte)) continue; tdp_mmu_set_spte(kvm, &iter, 0); } } kvm_flush_remote_tlbs(kvm); rcu_read_unlock(); write_unlock(&kvm->mmu_lock); }Quoted 5 lines
I noticed that it was previously done to the legacy MMU, but encountered some subtle issues with VFIO. I'm not sure if the issue is still there with
TDP_MMU. So we are trying to do more tests and analyses before submitting a patch. This provides me a convenient way to observe the number of pages being zapped.
Quoted 4 lines
Note you could also use the existing tracepoint to observe the number of pages being zapped in a given test run. e.g.
perf stat -e kvmmmu:kvm_mmu_prepare_zap_page -- <cmd>
Sure. Thank you!