[PATCH v2] KVM: x86/mmu: Make HVA handler retpoline-friendly

From: Maciej S. Szmigiero
Date: Mon Feb 08 2021 - 15:18:29 EST


From: "Maciej S. Szmigiero" <maciej.szmigiero@xxxxxxxxxx>

When retpolines are enabled they have high overhead in the inner loop
inside kvm_handle_hva_range() that iterates over the provided memory area.

Let's mark this function and its TDP MMU equivalent __always_inline so
compiler will be able to change the call to the actual handler function
inside each of them into a direct one.

This significantly improves performance on the unmap test on the existing
kernel memslot code (tested on a Xeon 8167M machine):
30 slots in use:
Test Before After Improvement
Unmap 0.0353s 0.0334s 5%
Unmap 2M 0.00104s 0.000407s 61%

509 slots in use:
Test Before After Improvement
Unmap 0.0742s 0.0740s None
Unmap 2M 0.00221s 0.00159s 28%

Looks like having an indirect call in these functions (and, so, a
retpoline) might have interfered with unrolling of the whole loop in the
CPU.

Signed-off-by: Maciej S. Szmigiero <maciej.szmigiero@xxxxxxxxxx>
---
Changes from v1:
* Switch from static dispatch to __always_inline annotation.

* Separate this patch from the rest of log(n) memslot code changes.

* Redo benchmarks.

arch/x86/kvm/mmu/mmu.c | 21 +++++++++++----------
arch/x86/kvm/mmu/tdp_mmu.c | 16 +++++++++++-----
2 files changed, 22 insertions(+), 15 deletions(-)

diff --git a/arch/x86/kvm/mmu/mmu.c b/arch/x86/kvm/mmu/mmu.c
index 6d16481aa29d..38d7a38609d4 100644
--- a/arch/x86/kvm/mmu/mmu.c
+++ b/arch/x86/kvm/mmu/mmu.c
@@ -1456,16 +1456,17 @@ static void slot_rmap_walk_next(struct slot_rmap_walk_iterator *iterator)
slot_rmap_walk_okay(_iter_); \
slot_rmap_walk_next(_iter_))

-static int kvm_handle_hva_range(struct kvm *kvm,
- unsigned long start,
- unsigned long end,
- unsigned long data,
- int (*handler)(struct kvm *kvm,
- struct kvm_rmap_head *rmap_head,
- struct kvm_memory_slot *slot,
- gfn_t gfn,
- int level,
- unsigned long data))
+static __always_inline int
+kvm_handle_hva_range(struct kvm *kvm,
+ unsigned long start,
+ unsigned long end,
+ unsigned long data,
+ int (*handler)(struct kvm *kvm,
+ struct kvm_rmap_head *rmap_head,
+ struct kvm_memory_slot *slot,
+ gfn_t gfn,
+ int level,
+ unsigned long data))
{
struct kvm_memslots *slots;
struct kvm_memory_slot *memslot;
diff --git a/arch/x86/kvm/mmu/tdp_mmu.c b/arch/x86/kvm/mmu/tdp_mmu.c
index b56d604809b8..f26c2269291f 100644
--- a/arch/x86/kvm/mmu/tdp_mmu.c
+++ b/arch/x86/kvm/mmu/tdp_mmu.c
@@ -639,11 +639,17 @@ int kvm_tdp_mmu_map(struct kvm_vcpu *vcpu, gpa_t gpa, u32 error_code,
return ret;
}

-static int kvm_tdp_mmu_handle_hva_range(struct kvm *kvm, unsigned long start,
- unsigned long end, unsigned long data,
- int (*handler)(struct kvm *kvm, struct kvm_memory_slot *slot,
- struct kvm_mmu_page *root, gfn_t start,
- gfn_t end, unsigned long data))
+static __always_inline int
+kvm_tdp_mmu_handle_hva_range(struct kvm *kvm,
+ unsigned long start,
+ unsigned long end,
+ unsigned long data,
+ int (*handler)(struct kvm *kvm,
+ struct kvm_memory_slot *slot,
+ struct kvm_mmu_page *root,
+ gfn_t start,
+ gfn_t end,
+ unsigned long data))
{
struct kvm_memslots *slots;
struct kvm_memory_slot *memslot;