From: "Maciej S. Szmigiero" <[email protected]> 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 <[email protected]> --- 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;
