KVM: host side for eoi optimization
authorMichael S. Tsirkin <mst@redhat.com>
Sun, 24 Jun 2012 16:25:07 +0000 (19:25 +0300)
committerAvi Kivity <avi@redhat.com>
Mon, 25 Jun 2012 09:40:55 +0000 (12:40 +0300)
Implementation of PV EOI using shared memory.
This reduces the number of exits an interrupt
causes as much as by half.

The idea is simple: there's a bit, per APIC, in guest memory,
that tells the guest that it does not need EOI.
We set it before injecting an interrupt and clear
before injecting a nested one. Guest tests it using
a test and clear operation - this is necessary
so that host can detect interrupt nesting -
and if set, it can skip the EOI MSR.

There's a new MSR to set the address of said register
in guest memory. Otherwise not much changed:
- Guest EOI is not required
- Register is tested & ISR is automatically cleared on exit

For testing results see description of previous patch
'kvm_para: guest side for eoi avoidance'.

Signed-off-by: Michael S. Tsirkin <mst@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
arch/x86/include/asm/kvm_host.h
arch/x86/kvm/cpuid.c
arch/x86/kvm/lapic.c
arch/x86/kvm/lapic.h
arch/x86/kvm/trace.h
arch/x86/kvm/x86.c

index db7c1f2..24b7647 100644 (file)
@@ -175,6 +175,13 @@ enum {
 
 /* apic attention bits */
 #define KVM_APIC_CHECK_VAPIC   0
+/*
+ * The following bit is set with PV-EOI, unset on EOI.
+ * We detect PV-EOI changes by guest by comparing
+ * this bit with PV-EOI in guest memory.
+ * See the implementation in apic_update_pv_eoi.
+ */
+#define KVM_APIC_PV_EOI_PENDING        1
 
 /*
  * We don't want allocation failures within the mmu code, so we preallocate
@@ -484,6 +491,11 @@ struct kvm_vcpu_arch {
                u64 length;
                u64 status;
        } osvw;
+
+       struct {
+               u64 msr_val;
+               struct gfn_to_hva_cache data;
+       } pv_eoi;
 };
 
 struct kvm_lpage_info {
index 7df1c6d..61ccbdf 100644 (file)
@@ -409,6 +409,7 @@ static int do_cpuid_ent(struct kvm_cpuid_entry2 *entry, u32 function,
                             (1 << KVM_FEATURE_NOP_IO_DELAY) |
                             (1 << KVM_FEATURE_CLOCKSOURCE2) |
                             (1 << KVM_FEATURE_ASYNC_PF) |
+                            (1 << KVM_FEATURE_PV_EOI) |
                             (1 << KVM_FEATURE_CLOCKSOURCE_STABLE_BIT);
 
                if (sched_info_on())
index 805d887..ce87878 100644 (file)
@@ -311,6 +311,54 @@ int kvm_apic_set_irq(struct kvm_vcpu *vcpu, struct kvm_lapic_irq *irq)
                        irq->level, irq->trig_mode);
 }
 
+static int pv_eoi_put_user(struct kvm_vcpu *vcpu, u8 val)
+{
+
+       return kvm_write_guest_cached(vcpu->kvm, &vcpu->arch.pv_eoi.data, &val,
+                                     sizeof(val));
+}
+
+static int pv_eoi_get_user(struct kvm_vcpu *vcpu, u8 *val)
+{
+
+       return kvm_read_guest_cached(vcpu->kvm, &vcpu->arch.pv_eoi.data, val,
+                                     sizeof(*val));
+}
+
+static inline bool pv_eoi_enabled(struct kvm_vcpu *vcpu)
+{
+       return vcpu->arch.pv_eoi.msr_val & KVM_MSR_ENABLED;
+}
+
+static bool pv_eoi_get_pending(struct kvm_vcpu *vcpu)
+{
+       u8 val;
+       if (pv_eoi_get_user(vcpu, &val) < 0)
+               apic_debug("Can't read EOI MSR value: 0x%llx\n",
+                          (unsigned long long)vcpi->arch.pv_eoi.msr_val);
+       return val & 0x1;
+}
+
+static void pv_eoi_set_pending(struct kvm_vcpu *vcpu)
+{
+       if (pv_eoi_put_user(vcpu, KVM_PV_EOI_ENABLED) < 0) {
+               apic_debug("Can't set EOI MSR value: 0x%llx\n",
+                          (unsigned long long)vcpi->arch.pv_eoi.msr_val);
+               return;
+       }
+       __set_bit(KVM_APIC_PV_EOI_PENDING, &vcpu->arch.apic_attention);
+}
+
+static void pv_eoi_clr_pending(struct kvm_vcpu *vcpu)
+{
+       if (pv_eoi_put_user(vcpu, KVM_PV_EOI_DISABLED) < 0) {
+               apic_debug("Can't clear EOI MSR value: 0x%llx\n",
+                          (unsigned long long)vcpi->arch.pv_eoi.msr_val);
+               return;
+       }
+       __clear_bit(KVM_APIC_PV_EOI_PENDING, &vcpu->arch.apic_attention);
+}
+
 static inline int apic_find_highest_isr(struct kvm_lapic *apic)
 {
        int result;
@@ -527,15 +575,18 @@ int kvm_apic_compare_prio(struct kvm_vcpu *vcpu1, struct kvm_vcpu *vcpu2)
        return vcpu1->arch.apic_arb_prio - vcpu2->arch.apic_arb_prio;
 }
 
-static void apic_set_eoi(struct kvm_lapic *apic)
+static int apic_set_eoi(struct kvm_lapic *apic)
 {
        int vector = apic_find_highest_isr(apic);
+
+       trace_kvm_eoi(apic, vector);
+
        /*
         * Not every write EOI will has corresponding ISR,
         * one example is when Kernel check timer on setup_IO_APIC
         */
        if (vector == -1)
-               return;
+               return vector;
 
        apic_clear_isr(vector, apic);
        apic_update_ppr(apic);
@@ -550,6 +601,7 @@ static void apic_set_eoi(struct kvm_lapic *apic)
                kvm_ioapic_update_eoi(apic->vcpu->kvm, vector, trigger_mode);
        }
        kvm_make_request(KVM_REQ_EVENT, apic->vcpu);
+       return vector;
 }
 
 static void apic_send_ipi(struct kvm_lapic *apic)
@@ -1132,6 +1184,7 @@ void kvm_lapic_reset(struct kvm_vcpu *vcpu)
        atomic_set(&apic->lapic_timer.pending, 0);
        if (kvm_vcpu_is_bsp(vcpu))
                vcpu->arch.apic_base |= MSR_IA32_APICBASE_BSP;
+       vcpu->arch.pv_eoi.msr_val = 0;
        apic_update_ppr(apic);
 
        vcpu->arch.apic_arb_prio = 0;
@@ -1332,11 +1385,51 @@ void __kvm_migrate_apic_timer(struct kvm_vcpu *vcpu)
                hrtimer_start_expires(timer, HRTIMER_MODE_ABS);
 }
 
+/*
+ * apic_sync_pv_eoi_from_guest - called on vmexit or cancel interrupt
+ *
+ * Detect whether guest triggered PV EOI since the
+ * last entry. If yes, set EOI on guests's behalf.
+ * Clear PV EOI in guest memory in any case.
+ */
+static void apic_sync_pv_eoi_from_guest(struct kvm_vcpu *vcpu,
+                                       struct kvm_lapic *apic)
+{
+       bool pending;
+       int vector;
+       /*
+        * PV EOI state is derived from KVM_APIC_PV_EOI_PENDING in host
+        * and KVM_PV_EOI_ENABLED in guest memory as follows:
+        *
+        * KVM_APIC_PV_EOI_PENDING is unset:
+        *      -> host disabled PV EOI.
+        * KVM_APIC_PV_EOI_PENDING is set, KVM_PV_EOI_ENABLED is set:
+        *      -> host enabled PV EOI, guest did not execute EOI yet.
+        * KVM_APIC_PV_EOI_PENDING is set, KVM_PV_EOI_ENABLED is unset:
+        *      -> host enabled PV EOI, guest executed EOI.
+        */
+       BUG_ON(!pv_eoi_enabled(vcpu));
+       pending = pv_eoi_get_pending(vcpu);
+       /*
+        * Clear pending bit in any case: it will be set again on vmentry.
+        * While this might not be ideal from performance point of view,
+        * this makes sure pv eoi is only enabled when we know it's safe.
+        */
+       pv_eoi_clr_pending(vcpu);
+       if (pending)
+               return;
+       vector = apic_set_eoi(apic);
+       trace_kvm_pv_eoi(apic, vector);
+}
+
 void kvm_lapic_sync_from_vapic(struct kvm_vcpu *vcpu)
 {
        u32 data;
        void *vapic;
 
+       if (test_bit(KVM_APIC_PV_EOI_PENDING, &vcpu->arch.apic_attention))
+               apic_sync_pv_eoi_from_guest(vcpu, vcpu->arch.apic);
+
        if (!test_bit(KVM_APIC_CHECK_VAPIC, &vcpu->arch.apic_attention))
                return;
 
@@ -1347,17 +1440,44 @@ void kvm_lapic_sync_from_vapic(struct kvm_vcpu *vcpu)
        apic_set_tpr(vcpu->arch.apic, data & 0xff);
 }
 
+/*
+ * apic_sync_pv_eoi_to_guest - called before vmentry
+ *
+ * Detect whether it's safe to enable PV EOI and
+ * if yes do so.
+ */
+static void apic_sync_pv_eoi_to_guest(struct kvm_vcpu *vcpu,
+                                       struct kvm_lapic *apic)
+{
+       if (!pv_eoi_enabled(vcpu) ||
+           /* IRR set or many bits in ISR: could be nested. */
+           apic->irr_pending ||
+           /* Cache not set: could be safe but we don't bother. */
+           apic->highest_isr_cache == -1 ||
+           /* Need EOI to update ioapic. */
+           kvm_ioapic_handles_vector(vcpu->kvm, apic->highest_isr_cache)) {
+               /*
+                * PV EOI was disabled by apic_sync_pv_eoi_from_guest
+                * so we need not do anything here.
+                */
+               return;
+       }
+
+       pv_eoi_set_pending(apic->vcpu);
+}
+
 void kvm_lapic_sync_to_vapic(struct kvm_vcpu *vcpu)
 {
        u32 data, tpr;
        int max_irr, max_isr;
-       struct kvm_lapic *apic;
+       struct kvm_lapic *apic = vcpu->arch.apic;
        void *vapic;
 
+       apic_sync_pv_eoi_to_guest(vcpu, apic);
+
        if (!test_bit(KVM_APIC_CHECK_VAPIC, &vcpu->arch.apic_attention))
                return;
 
-       apic = vcpu->arch.apic;
        tpr = apic_get_reg(apic, APIC_TASKPRI) & 0xff;
        max_irr = apic_find_highest_irr(apic);
        if (max_irr < 0)
@@ -1443,3 +1563,16 @@ int kvm_hv_vapic_msr_read(struct kvm_vcpu *vcpu, u32 reg, u64 *data)
 
        return 0;
 }
+
+int kvm_lapic_enable_pv_eoi(struct kvm_vcpu *vcpu, u64 data)
+{
+       u64 addr = data & ~KVM_MSR_ENABLED;
+       if (!IS_ALIGNED(addr, 4))
+               return 1;
+
+       vcpu->arch.pv_eoi.msr_val = data;
+       if (!pv_eoi_enabled(vcpu))
+               return 0;
+       return kvm_gfn_to_hva_cache_init(vcpu->kvm, &vcpu->arch.pv_eoi.data,
+                                        addr);
+}
index 5ac9e5e..4af5405 100644 (file)
@@ -69,4 +69,6 @@ static inline bool kvm_hv_vapic_assist_page_enabled(struct kvm_vcpu *vcpu)
 {
        return vcpu->arch.hv_vapic & HV_X64_MSR_APIC_ASSIST_PAGE_ENABLE;
 }
+
+int kvm_lapic_enable_pv_eoi(struct kvm_vcpu *vcpu, u64 data);
 #endif
index 911d264..851914e 100644 (file)
@@ -517,6 +517,40 @@ TRACE_EVENT(kvm_apic_accept_irq,
                  __entry->coalesced ? " (coalesced)" : "")
 );
 
+TRACE_EVENT(kvm_eoi,
+           TP_PROTO(struct kvm_lapic *apic, int vector),
+           TP_ARGS(apic, vector),
+
+       TP_STRUCT__entry(
+               __field(        __u32,          apicid          )
+               __field(        int,            vector          )
+       ),
+
+       TP_fast_assign(
+               __entry->apicid         = apic->vcpu->vcpu_id;
+               __entry->vector         = vector;
+       ),
+
+       TP_printk("apicid %x vector %d", __entry->apicid, __entry->vector)
+);
+
+TRACE_EVENT(kvm_pv_eoi,
+           TP_PROTO(struct kvm_lapic *apic, int vector),
+           TP_ARGS(apic, vector),
+
+       TP_STRUCT__entry(
+               __field(        __u32,          apicid          )
+               __field(        int,            vector          )
+       ),
+
+       TP_fast_assign(
+               __entry->apicid         = apic->vcpu->vcpu_id;
+               __entry->vector         = vector;
+       ),
+
+       TP_printk("apicid %x vector %d", __entry->apicid, __entry->vector)
+);
+
 /*
  * Tracepoint for nested VMRUN
  */
index 7ea0f61..8eacb2e 100644 (file)
@@ -795,6 +795,7 @@ static u32 msrs_to_save[] = {
        MSR_KVM_SYSTEM_TIME_NEW, MSR_KVM_WALL_CLOCK_NEW,
        HV_X64_MSR_GUEST_OS_ID, HV_X64_MSR_HYPERCALL,
        HV_X64_MSR_APIC_ASSIST_PAGE, MSR_KVM_ASYNC_PF_EN, MSR_KVM_STEAL_TIME,
+       MSR_KVM_PV_EOI_EN,
        MSR_IA32_SYSENTER_CS, MSR_IA32_SYSENTER_ESP, MSR_IA32_SYSENTER_EIP,
        MSR_STAR,
 #ifdef CONFIG_X86_64
@@ -1653,6 +1654,10 @@ int kvm_set_msr_common(struct kvm_vcpu *vcpu, u32 msr, u64 data)
                kvm_make_request(KVM_REQ_STEAL_UPDATE, vcpu);
 
                break;
+       case MSR_KVM_PV_EOI_EN:
+               if (kvm_lapic_enable_pv_eoi(vcpu, data))
+                       return 1;
+               break;
 
        case MSR_IA32_MCG_CTL:
        case MSR_IA32_MCG_STATUS:
@@ -5394,6 +5399,8 @@ static int vcpu_enter_guest(struct kvm_vcpu *vcpu)
 
 cancel_injection:
        kvm_x86_ops->cancel_injection(vcpu);
+       if (unlikely(vcpu->arch.apic_attention))
+               kvm_lapic_sync_from_vapic(vcpu);
 out:
        return r;
 }