Re: [Fwd: [RFC PATCH 2/4] KVM: x86: Reduce retpoline performance impact in slot_handle_level_range()]

[Sironi, Filippo]

> On 8. Feb 2018, at 13:17, David Woodhouse <dwmw2@xxxxxxxxxxxxx> wrote:
> 
> 
> From: David Woodhouse <dwmw@xxxxxxxxxxxx>
> Subject: [RFC PATCH 2/4] KVM: x86: Reduce retpoline performance impact in slot_handle_level_range()
> Date: 7. February 2018 at 01:03:12 GMT+1
> To: tglx@xxxxxxxxxxxxx, torvalds@xxxxxxxxxxxxxxxxxxxx, x86@xxxxxxxxxx, linux-kernel@xxxxxxxxxxxxxxx, bp@xxxxxxxxx, peterz@xxxxxxxxxxxxx, tim.c.chen@xxxxxxxxxxxxxxx, dave.hansen@xxxxxxxxx, arjan.van.de.ven@xxxxxxxxx
> 
> 
> With retpoline, tight loops of "call this function for every XXX" are
> very much pessimised by taking a prediction miss *every* time. This one
> showed up very high in our early testing.
> 
> By marking the iterator slot_handle_â() functions always_inline, we can
> ensure that the indirect function call can be optimised away into a
> direct call and it actually generates slightly smaller code because
> some of the other conditionals can get optimised away too.
> 
> Suggested-by: Linus Torvalds <torvalds@xxxxxxxxxxxxxxxxxxxx>
> Signed-off-by: David Woodhouse <dwmw@xxxxxxxxxxxx>
> ---
> arch/x86/kvm/mmu.c | 10 +++++-----
> 1 file changed, 5 insertions(+), 5 deletions(-)
> 
> diff --git a/arch/x86/kvm/mmu.c b/arch/x86/kvm/mmu.c
> index 2b8eb4d..cc83bdc 100644
> --- a/arch/x86/kvm/mmu.c
> +++ b/arch/x86/kvm/mmu.c
> @@ -5058,7 +5058,7 @@ void kvm_mmu_uninit_vm(struct kvm *kvm)
> typedef bool (*slot_level_handler) (struct kvm *kvm, struct kvm_rmap_head *rmap_head);
> 
> /* The caller should hold mmu-lock before calling this function. */
> -static bool
> +static __always_inline bool
> slot_handle_level_range(struct kvm *kvm, struct kvm_memory_slot *memslot,
> 			slot_level_handler fn, int start_level, int end_level,
> 			gfn_t start_gfn, gfn_t end_gfn, bool lock_flush_tlb)
> @@ -5088,7 +5088,7 @@ slot_handle_level_range(struct kvm *kvm, struct kvm_memory_slot *memslot,
> 	return flush;
> }
> 
> -static bool
> +static __always_inline bool
> slot_handle_level(struct kvm *kvm, struct kvm_memory_slot *memslot,
> 		  slot_level_handler fn, int start_level, int end_level,
> 		  bool lock_flush_tlb)
> @@ -5099,7 +5099,7 @@ slot_handle_level(struct kvm *kvm, struct kvm_memory_slot *memslot,
> 			lock_flush_tlb);
> }
> 
> -static bool
> +static __always_inline bool
> slot_handle_all_level(struct kvm *kvm, struct kvm_memory_slot *memslot,
> 		      slot_level_handler fn, bool lock_flush_tlb)
> {
> @@ -5107,7 +5107,7 @@ slot_handle_all_level(struct kvm *kvm, struct kvm_memory_slot *memslot,
> 				 PT_MAX_HUGEPAGE_LEVEL, lock_flush_tlb);
> }
> 
> -static bool
> +static __always_inline bool
> slot_handle_large_level(struct kvm *kvm, struct kvm_memory_slot *memslot,
> 			slot_level_handler fn, bool lock_flush_tlb)
> {
> @@ -5115,7 +5115,7 @@ slot_handle_large_level(struct kvm *kvm, struct kvm_memory_slot *memslot,
> 				 PT_MAX_HUGEPAGE_LEVEL, lock_flush_tlb);
> }
> 
> -static bool
> +static __always_inline bool
> slot_handle_leaf(struct kvm *kvm, struct kvm_memory_slot *memslot,
> 		 slot_level_handler fn, bool lock_flush_tlb)
> {
> -- 
> 2.7.4

+kvm@xxxxxxxxxxxxxxx

With this patch, launches of "large instances" are pretty close to what we see with
nospectre_v2 (within tens of milliseconds).

Reviewed-by: Filippo Sironi <sironi@xxxxxxxxx>
Tested-by: Filippo Sironi <sironi@xxxxxxxxx>

Amazon Development Center Germany GmbH
Berlin - Dresden - Aachen
main office: Krausenstr. 38, 10117 Berlin
Geschaeftsfuehrer: Dr. Ralf Herbrich, Christian Schlaeger
Ust-ID: DE289237879
Eingetragen am Amtsgericht Charlottenburg HRB 149173 B