KVM: VMX: dynamise PLE window

Window is increased on every PLE exit and decreased on every sched_in.
The idea is that we don't want to PLE exit if there is no preemption
going on.
We do this with sched_in() because it does not hold rq lock.

There are two new kernel parameters for changing the window:
 ple_window_grow and ple_window_shrink
ple_window_grow affects the window on PLE exit and ple_window_shrink
does it on sched_in;  depending on their value, the window is modifier
like this: (ple_window is kvm_intel's global)

  ple_window_shrink/ |
  ple_window_grow    | PLE exit           | sched_in
  -------------------+--------------------+---------------------
  < 1                |  = ple_window      |  = ple_window
  < ple_window       | *= ple_window_grow | /= ple_window_shrink
  otherwise          | += ple_window_grow | -= ple_window_shrink

A third new parameter, ple_window_max, controls the maximal ple_window;
it is internally rounded down to a closest multiple of ple_window_grow.

VCPU's PLE window is never allowed below ple_window.

Signed-off-by: Radim Krčmář <rkrcmar@redhat.com>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>

diff --git a/arch/x86/kvm/vmx.c b/arch/x86/kvm/vmx.c
index 70175990536be68b62745852717172618c0c8539..baeac7f580a73c3a5032d8c54209390dae4b7d08 100644 (file)
--- a/arch/x86/kvm/vmx.c
+++ b/arch/x86/kvm/vmx.c
@@ -125,14 +125,32 @@ module_param(nested, bool, S_IRUGO);
  * Time is measured based on a counter that runs at the same rate as the TSC,
  * refer SDM volume 3b section 21.6.13 & 22.1.3.
  */
-#define KVM_VMX_DEFAULT_PLE_GAP    128
-#define KVM_VMX_DEFAULT_PLE_WINDOW 4096
+#define KVM_VMX_DEFAULT_PLE_GAP           128
+#define KVM_VMX_DEFAULT_PLE_WINDOW        4096
+#define KVM_VMX_DEFAULT_PLE_WINDOW_GROW   2
+#define KVM_VMX_DEFAULT_PLE_WINDOW_SHRINK 0
+#define KVM_VMX_DEFAULT_PLE_WINDOW_MAX    \
+               INT_MAX / KVM_VMX_DEFAULT_PLE_WINDOW_GROW
+
 static int ple_gap = KVM_VMX_DEFAULT_PLE_GAP;
 module_param(ple_gap, int, S_IRUGO);
 
 static int ple_window = KVM_VMX_DEFAULT_PLE_WINDOW;
 module_param(ple_window, int, S_IRUGO);
 
+/* Default doubles per-vcpu window every exit. */
+static int ple_window_grow = KVM_VMX_DEFAULT_PLE_WINDOW_GROW;
+module_param(ple_window_grow, int, S_IRUGO);
+
+/* Default resets per-vcpu window every exit to ple_window. */
+static int ple_window_shrink = KVM_VMX_DEFAULT_PLE_WINDOW_SHRINK;
+module_param(ple_window_shrink, int, S_IRUGO);
+
+/* Default is to compute the maximum so we can never overflow. */
+static int ple_window_actual_max = KVM_VMX_DEFAULT_PLE_WINDOW_MAX;
+static int ple_window_max        = KVM_VMX_DEFAULT_PLE_WINDOW_MAX;
+module_param(ple_window_max, int, S_IRUGO);
+
 extern const ulong vmx_return;
 
 #define NR_AUTOLOAD_MSRS 8
@@ -5683,12 +5701,81 @@ out:
        return ret;
 }
 
+static int __grow_ple_window(int val)
+{
+       if (ple_window_grow < 1)
+               return ple_window;
+
+       val = min(val, ple_window_actual_max);
+
+       if (ple_window_grow < ple_window)
+               val *= ple_window_grow;
+       else
+               val += ple_window_grow;
+
+       return val;
+}
+
+static int __shrink_ple_window(int val, int modifier, int minimum)
+{
+       if (modifier < 1)
+               return ple_window;
+
+       if (modifier < ple_window)
+               val /= modifier;
+       else
+               val -= modifier;
+
+       return max(val, minimum);
+}
+
+static void grow_ple_window(struct kvm_vcpu *vcpu)
+{
+       struct vcpu_vmx *vmx = to_vmx(vcpu);
+       int old = vmx->ple_window;
+
+       vmx->ple_window = __grow_ple_window(old);
+
+       if (vmx->ple_window != old)
+               vmx->ple_window_dirty = true;
+}
+
+static void shrink_ple_window(struct kvm_vcpu *vcpu)
+{
+       struct vcpu_vmx *vmx = to_vmx(vcpu);
+       int old = vmx->ple_window;
+
+       vmx->ple_window = __shrink_ple_window(old,
+                                             ple_window_shrink, ple_window);
+
+       if (vmx->ple_window != old)
+               vmx->ple_window_dirty = true;
+}
+
+/*
+ * ple_window_actual_max is computed to be one grow_ple_window() below
+ * ple_window_max. (See __grow_ple_window for the reason.)
+ * This prevents overflows, because ple_window_max is int.
+ * ple_window_max effectively rounded down to a multiple of ple_window_grow in
+ * this process.
+ * ple_window_max is also prevented from setting vmx->ple_window < ple_window.
+ */
+static void update_ple_window_actual_max(void)
+{
+       ple_window_actual_max =
+                       __shrink_ple_window(max(ple_window_max, ple_window),
+                                           ple_window_grow, INT_MIN);
+}
+
 /*
  * Indicate a busy-waiting vcpu in spinlock. We do not enable the PAUSE
  * exiting, so only get here on cpu with PAUSE-Loop-Exiting.
  */
 static int handle_pause(struct kvm_vcpu *vcpu)
 {
+       if (ple_gap)
+               grow_ple_window(vcpu);
+
        skip_emulated_instruction(vcpu);
        kvm_vcpu_on_spin(vcpu);
 
@@ -8860,6 +8947,8 @@ static int vmx_check_intercept(struct kvm_vcpu *vcpu,
 
 void vmx_sched_in(struct kvm_vcpu *vcpu, int cpu)
 {
+       if (ple_gap)
+               shrink_ple_window(vcpu);
 }
 
 static struct kvm_x86_ops vmx_x86_ops = {
@@ -9082,6 +9171,8 @@ static int __init vmx_init(void)
        } else
                kvm_disable_tdp();
 
+       update_ple_window_actual_max();
+
        return 0;
 
 out7: