KVM: PPC: Book3S: PR: Rework irq disabling

Today, we disable preemption while inside guest context, because we need
to expose to the world that we are not in a preemptible context. However,
during that time we already have interrupts disabled, which would indicate
that we are in a non-preemptible context.

The reason the checks for irqs_disabled() fail for us though is that we
manually control hard IRQs and ignore all the lazy EE framework. Let's
stop doing that. Instead, let's always use lazy EE to indicate when we
want to disable IRQs, but do a special final switch that gets us into
EE disabled, but soft enabled state. That way when we get back out of
guest state, we are immediately ready to process interrupts.

This simplifies the code drastically and reduces the time that we appear
as preempt disabled.

Signed-off-by: Alexander Graf <agraf@suse.de>

arch/powerpc/include/asm/kvm_ppc.h

@@ -234,5 +234,15 @@ static inline void kvmppc_mmu_flush_icache(pfn_t pfn)
 	}
 }
 
+/* Please call after prepare_to_enter. This function puts the lazy ee state
+   back to normal mode, without actually enabling interrupts. */
+static inline void kvmppc_lazy_ee_enable(void)
+{
+#ifdef CONFIG_PPC64
+	/* Only need to enable IRQs by hard enabling them after this */
+	local_paca->irq_happened = 0;
+	local_paca->soft_enabled = 1;
+#endif
+}
 
 #endif /* __POWERPC_KVM_PPC_H__ */

arch/powerpc/kvm/book3s_pr.c

@@ -52,8 +52,6 @@ static int kvmppc_handle_ext(struct kvm_vcpu *vcpu, unsigned int exit_nr,
 #define MSR_USER32 MSR_USER
 #define MSR_USER64 MSR_USER
 #define HW_PAGE_SIZE PAGE_SIZE
-#define __hard_irq_disable local_irq_disable
-#define __hard_irq_enable local_irq_enable
 #endif
 
 void kvmppc_core_vcpu_load(struct kvm_vcpu *vcpu, int cpu)
@@ -597,12 +595,10 @@ int kvmppc_handle_exit(struct kvm_run *run, struct kvm_vcpu *vcpu,
 	run->exit_reason = KVM_EXIT_UNKNOWN;
 	run->ready_for_interrupt_injection = 1;
 
-	/* We get here with MSR.EE=0, so enable it to be a nice citizen */
+	/* We get here with MSR.EE=1 */
-	__hard_irq_enable();
 
 	trace_kvm_exit(exit_nr, vcpu);
 	kvm_guest_exit();
-	preempt_enable();
 
 	switch (exit_nr) {
 	case BOOK3S_INTERRUPT_INST_STORAGE:
@@ -854,7 +850,6 @@ program_interrupt:
 	}
 	}
 
-	preempt_disable();
 	if (!(r & RESUME_HOST)) {
 		/* To avoid clobbering exit_reason, only check for signals if
 		 * we aren't already exiting to userspace for some other
@@ -866,14 +861,15 @@ program_interrupt:
 		 * and if we really did time things so badly, then we just exit
 		 * again due to a host external interrupt.
 		 */
-		__hard_irq_disable();
+		local_irq_disable();
 		if (kvmppc_prepare_to_enter(vcpu)) {
-			/* local_irq_enable(); */
+			local_irq_enable();
 			run->exit_reason = KVM_EXIT_INTR;
 			r = -EINTR;
 		} else {
 			/* Going back to guest */
 			kvm_guest_enter();
+			kvmppc_lazy_ee_enable();
 		}
 	}
 
@@ -1066,8 +1062,6 @@ int kvmppc_vcpu_run(struct kvm_run *kvm_run, struct kvm_vcpu *vcpu)
 #endif
 	ulong ext_msr;
 
-	preempt_disable();
-
 	/* Check if we can run the vcpu at all */
 	if (!vcpu->arch.sane) {
 		kvm_run->exit_reason = KVM_EXIT_INTERNAL_ERROR;
@@ -1081,9 +1075,9 @@ int kvmppc_vcpu_run(struct kvm_run *kvm_run, struct kvm_vcpu *vcpu)
 	 * really did time things so badly, then we just exit again due to
 	 * a host external interrupt.
 	 */
-	__hard_irq_disable();
+	local_irq_disable();
 	if (kvmppc_prepare_to_enter(vcpu)) {
-		__hard_irq_enable();
+		local_irq_enable();
 		kvm_run->exit_reason = KVM_EXIT_INTR;
 		ret = -EINTR;
 		goto out;
@@ -1122,7 +1116,7 @@ int kvmppc_vcpu_run(struct kvm_run *kvm_run, struct kvm_vcpu *vcpu)
 	if (vcpu->arch.shared->msr & MSR_FP)
 		kvmppc_handle_ext(vcpu, BOOK3S_INTERRUPT_FP_UNAVAIL, MSR_FP);
 
-	kvm_guest_enter();
+	kvmppc_lazy_ee_enable();
 
 	ret = __kvmppc_vcpu_run(kvm_run, vcpu);
 
@@ -1157,7 +1151,6 @@ int kvmppc_vcpu_run(struct kvm_run *kvm_run, struct kvm_vcpu *vcpu)
 
 out:
 	vcpu->mode = OUTSIDE_GUEST_MODE;
-	preempt_enable();
 	return ret;
 }
 

arch/powerpc/kvm/book3s_rmhandlers.S

@@ -170,20 +170,21 @@ kvmppc_handler_skip_ins:
  * Call kvmppc_handler_trampoline_enter in real mode
  *
  * On entry, r4 contains the guest shadow MSR
+ * MSR.EE has to be 0 when calling this function
  */
 _GLOBAL(kvmppc_entry_trampoline)
 	mfmsr	r5
 	LOAD_REG_ADDR(r7, kvmppc_handler_trampoline_enter)
 	toreal(r7)
 
-	li	r9, MSR_RI
-	ori	r9, r9, MSR_EE
-	andc	r9, r5, r9	/* Clear EE and RI in MSR value */
 	li	r6, MSR_IR | MSR_DR
-	ori	r6, r6, MSR_EE
+	andc	r6, r5, r6	/* Clear DR and IR in MSR value */
-	andc	r6, r5, r6	/* Clear EE, DR and IR in MSR value */
+	/*
-	MTMSR_EERI(r9)		/* Clear EE and RI in MSR */
+	 * Set EE in HOST_MSR so that it's enabled when we get into our
-	mtsrr0	r7		/* before we set srr0/1 */
+	 * C exit handler function
+	 */
+	ori	r5, r5, MSR_EE
+	mtsrr0	r7
 	mtsrr1	r6
 	RFI
 

arch/powerpc/kvm/booke.c

@@ -486,6 +486,7 @@ int kvmppc_vcpu_run(struct kvm_run *kvm_run, struct kvm_vcpu *vcpu)
 		ret = -EINTR;
 		goto out;
 	}
+	kvmppc_lazy_ee_enable();
 
 	kvm_guest_enter();
 
@@ -955,6 +956,7 @@ int kvmppc_handle_exit(struct kvm_run *run, struct kvm_vcpu *vcpu,
 		} else {
 			/* Going back to guest */
 			kvm_guest_enter();
+			kvmppc_lazy_ee_enable();
 		}
 	}
 

arch/powerpc/kvm/powerpc.c

@@ -30,6 +30,7 @@
 #include <asm/kvm_ppc.h>
 #include <asm/tlbflush.h>
 #include <asm/cputhreads.h>
+#include <asm/irqflags.h>
 #include "timing.h"
 #include "../mm/mmu_decl.h"
 
@@ -93,6 +94,19 @@ int kvmppc_prepare_to_enter(struct kvm_vcpu *vcpu)
 			break;
 		}
 
+#ifdef CONFIG_PPC64
+		/* lazy EE magic */
+		hard_irq_disable();
+		if (lazy_irq_pending()) {
+			/* Got an interrupt in between, try again */
+			local_irq_enable();
+			local_irq_disable();
+			continue;
+		}
+
+		trace_hardirqs_on();
+#endif
+
 		/* Going into guest context! Yay! */
 		vcpu->mode = IN_GUEST_MODE;
 		smp_wmb();