KVM: Move IO APIC to its own lock
Gleb Natapov [Mon, 24 Aug 2009 08:54:25 +0000 (11:54 +0300)]
The allows removal of irq_lock from the injection path.

Signed-off-by: Gleb Natapov <gleb@redhat.com>
Signed-off-by: Avi Kivity <avi@redhat.com>

arch/ia64/kvm/kvm-ia64.c
arch/x86/kvm/i8259.c
arch/x86/kvm/lapic.c
arch/x86/kvm/x86.c
virt/kvm/ioapic.c
virt/kvm/ioapic.h
virt/kvm/irq_comm.c

index 0ad09f0..4a98314 100644 (file)
@@ -851,8 +851,7 @@ static int kvm_vm_ioctl_get_irqchip(struct kvm *kvm,
        r = 0;
        switch (chip->chip_id) {
        case KVM_IRQCHIP_IOAPIC:
-               memcpy(&chip->chip.ioapic, ioapic_irqchip(kvm),
-                               sizeof(struct kvm_ioapic_state));
+               r = kvm_get_ioapic(kvm, &chip->chip.ioapic);
                break;
        default:
                r = -EINVAL;
@@ -868,9 +867,7 @@ static int kvm_vm_ioctl_set_irqchip(struct kvm *kvm, struct kvm_irqchip *chip)
        r = 0;
        switch (chip->chip_id) {
        case KVM_IRQCHIP_IOAPIC:
-               memcpy(ioapic_irqchip(kvm),
-                               &chip->chip.ioapic,
-                               sizeof(struct kvm_ioapic_state));
+               r = kvm_set_ioapic(kvm, &chip->chip.ioapic);
                break;
        default:
                r = -EINVAL;
index ccc941a..d057c0c 100644 (file)
@@ -38,7 +38,15 @@ static void pic_clear_isr(struct kvm_kpic_state *s, int irq)
        s->isr_ack |= (1 << irq);
        if (s != &s->pics_state->pics[0])
                irq += 8;
+       /*
+        * We are dropping lock while calling ack notifiers since ack
+        * notifier callbacks for assigned devices call into PIC recursively.
+        * Other interrupt may be delivered to PIC while lock is dropped but
+        * it should be safe since PIC state is already updated at this stage.
+        */
+       spin_unlock(&s->pics_state->lock);
        kvm_notify_acked_irq(s->pics_state->kvm, SELECT_PIC(irq), irq);
+       spin_lock(&s->pics_state->lock);
 }
 
 void kvm_pic_clear_isr_ack(struct kvm *kvm)
@@ -176,16 +184,18 @@ int kvm_pic_set_irq(void *opaque, int irq, int level)
 static inline void pic_intack(struct kvm_kpic_state *s, int irq)
 {
        s->isr |= 1 << irq;
-       if (s->auto_eoi) {
-               if (s->rotate_on_auto_eoi)
-                       s->priority_add = (irq + 1) & 7;
-               pic_clear_isr(s, irq);
-       }
        /*
         * We don't clear a level sensitive interrupt here
         */
        if (!(s->elcr & (1 << irq)))
                s->irr &= ~(1 << irq);
+
+       if (s->auto_eoi) {
+               if (s->rotate_on_auto_eoi)
+                       s->priority_add = (irq + 1) & 7;
+               pic_clear_isr(s, irq);
+       }
+
 }
 
 int kvm_pic_read_irq(struct kvm *kvm)
@@ -294,9 +304,9 @@ static void pic_ioport_write(void *opaque, u32 addr, u32 val)
                                priority = get_priority(s, s->isr);
                                if (priority != 8) {
                                        irq = (priority + s->priority_add) & 7;
-                                       pic_clear_isr(s, irq);
                                        if (cmd == 5)
                                                s->priority_add = (irq + 1) & 7;
+                                       pic_clear_isr(s, irq);
                                        pic_update_irq(s->pics_state);
                                }
                                break;
index 23c2176..df8bcb0 100644 (file)
@@ -471,11 +471,8 @@ static void apic_set_eoi(struct kvm_lapic *apic)
                trigger_mode = IOAPIC_LEVEL_TRIG;
        else
                trigger_mode = IOAPIC_EDGE_TRIG;
-       if (!(apic_get_reg(apic, APIC_SPIV) & APIC_SPIV_DIRECTED_EOI)) {
-               mutex_lock(&apic->vcpu->kvm->irq_lock);
+       if (!(apic_get_reg(apic, APIC_SPIV) & APIC_SPIV_DIRECTED_EOI))
                kvm_ioapic_update_eoi(apic->vcpu->kvm, vector, trigger_mode);
-               mutex_unlock(&apic->vcpu->kvm->irq_lock);
-       }
 }
 
 static void apic_send_ipi(struct kvm_lapic *apic)
index 1687d12..fdf989f 100644 (file)
@@ -2038,9 +2038,7 @@ static int kvm_vm_ioctl_get_irqchip(struct kvm *kvm, struct kvm_irqchip *chip)
                        sizeof(struct kvm_pic_state));
                break;
        case KVM_IRQCHIP_IOAPIC:
-               memcpy(&chip->chip.ioapic,
-                       ioapic_irqchip(kvm),
-                       sizeof(struct kvm_ioapic_state));
+               r = kvm_get_ioapic(kvm, &chip->chip.ioapic);
                break;
        default:
                r = -EINVAL;
@@ -2070,11 +2068,7 @@ static int kvm_vm_ioctl_set_irqchip(struct kvm *kvm, struct kvm_irqchip *chip)
                spin_unlock(&pic_irqchip(kvm)->lock);
                break;
        case KVM_IRQCHIP_IOAPIC:
-               mutex_lock(&kvm->irq_lock);
-               memcpy(ioapic_irqchip(kvm),
-                       &chip->chip.ioapic,
-                       sizeof(struct kvm_ioapic_state));
-               mutex_unlock(&kvm->irq_lock);
+               r = kvm_set_ioapic(kvm, &chip->chip.ioapic);
                break;
        default:
                r = -EINVAL;
index 9fe140b..38a2d20 100644 (file)
@@ -182,6 +182,7 @@ int kvm_ioapic_set_irq(struct kvm_ioapic *ioapic, int irq, int level)
        union kvm_ioapic_redirect_entry entry;
        int ret = 1;
 
+       mutex_lock(&ioapic->lock);
        if (irq >= 0 && irq < IOAPIC_NUM_PINS) {
                entry = ioapic->redirtbl[irq];
                level ^= entry.fields.polarity;
@@ -198,34 +199,51 @@ int kvm_ioapic_set_irq(struct kvm_ioapic *ioapic, int irq, int level)
                }
                trace_kvm_ioapic_set_irq(entry.bits, irq, ret == 0);
        }
+       mutex_unlock(&ioapic->lock);
+
        return ret;
 }
 
-static void __kvm_ioapic_update_eoi(struct kvm_ioapic *ioapic, int pin,
-                                   int trigger_mode)
+static void __kvm_ioapic_update_eoi(struct kvm_ioapic *ioapic, int vector,
+                                    int trigger_mode)
 {
-       union kvm_ioapic_redirect_entry *ent;
+       int i;
+
+       for (i = 0; i < IOAPIC_NUM_PINS; i++) {
+               union kvm_ioapic_redirect_entry *ent = &ioapic->redirtbl[i];
 
-       ent = &ioapic->redirtbl[pin];
+               if (ent->fields.vector != vector)
+                       continue;
 
-       kvm_notify_acked_irq(ioapic->kvm, KVM_IRQCHIP_IOAPIC, pin);
+               /*
+                * We are dropping lock while calling ack notifiers because ack
+                * notifier callbacks for assigned devices call into IOAPIC
+                * recursively. Since remote_irr is cleared only after call
+                * to notifiers if the same vector will be delivered while lock
+                * is dropped it will be put into irr and will be delivered
+                * after ack notifier returns.
+                */
+               mutex_unlock(&ioapic->lock);
+               kvm_notify_acked_irq(ioapic->kvm, KVM_IRQCHIP_IOAPIC, i);
+               mutex_lock(&ioapic->lock);
+
+               if (trigger_mode != IOAPIC_LEVEL_TRIG)
+                       continue;
 
-       if (trigger_mode == IOAPIC_LEVEL_TRIG) {
                ASSERT(ent->fields.trig_mode == IOAPIC_LEVEL_TRIG);
                ent->fields.remote_irr = 0;
-               if (!ent->fields.mask && (ioapic->irr & (1 << pin)))
-                       ioapic_service(ioapic, pin);
+               if (!ent->fields.mask && (ioapic->irr & (1 << i)))
+                       ioapic_service(ioapic, i);
        }
 }
 
 void kvm_ioapic_update_eoi(struct kvm *kvm, int vector, int trigger_mode)
 {
        struct kvm_ioapic *ioapic = kvm->arch.vioapic;
-       int i;
 
-       for (i = 0; i < IOAPIC_NUM_PINS; i++)
-               if (ioapic->redirtbl[i].fields.vector == vector)
-                       __kvm_ioapic_update_eoi(ioapic, i, trigger_mode);
+       mutex_lock(&ioapic->lock);
+       __kvm_ioapic_update_eoi(ioapic, vector, trigger_mode);
+       mutex_unlock(&ioapic->lock);
 }
 
 static inline struct kvm_ioapic *to_ioapic(struct kvm_io_device *dev)
@@ -250,8 +268,8 @@ static int ioapic_mmio_read(struct kvm_io_device *this, gpa_t addr, int len,
        ioapic_debug("addr %lx\n", (unsigned long)addr);
        ASSERT(!(addr & 0xf));  /* check alignment */
 
-       mutex_lock(&ioapic->kvm->irq_lock);
        addr &= 0xff;
+       mutex_lock(&ioapic->lock);
        switch (addr) {
        case IOAPIC_REG_SELECT:
                result = ioapic->ioregsel;
@@ -265,6 +283,8 @@ static int ioapic_mmio_read(struct kvm_io_device *this, gpa_t addr, int len,
                result = 0;
                break;
        }
+       mutex_unlock(&ioapic->lock);
+
        switch (len) {
        case 8:
                *(u64 *) val = result;
@@ -277,7 +297,6 @@ static int ioapic_mmio_read(struct kvm_io_device *this, gpa_t addr, int len,
        default:
                printk(KERN_WARNING "ioapic: wrong length %d\n", len);
        }
-       mutex_unlock(&ioapic->kvm->irq_lock);
        return 0;
 }
 
@@ -293,15 +312,15 @@ static int ioapic_mmio_write(struct kvm_io_device *this, gpa_t addr, int len,
                     (void*)addr, len, val);
        ASSERT(!(addr & 0xf));  /* check alignment */
 
-       mutex_lock(&ioapic->kvm->irq_lock);
        if (len == 4 || len == 8)
                data = *(u32 *) val;
        else {
                printk(KERN_WARNING "ioapic: Unsupported size %d\n", len);
-               goto unlock;
+               return 0;
        }
 
        addr &= 0xff;
+       mutex_lock(&ioapic->lock);
        switch (addr) {
        case IOAPIC_REG_SELECT:
                ioapic->ioregsel = data;
@@ -312,15 +331,14 @@ static int ioapic_mmio_write(struct kvm_io_device *this, gpa_t addr, int len,
                break;
 #ifdef CONFIG_IA64
        case IOAPIC_REG_EOI:
-               kvm_ioapic_update_eoi(ioapic->kvm, data, IOAPIC_LEVEL_TRIG);
+               __kvm_ioapic_update_eoi(ioapic, data, IOAPIC_LEVEL_TRIG);
                break;
 #endif
 
        default:
                break;
        }
-unlock:
-       mutex_unlock(&ioapic->kvm->irq_lock);
+       mutex_unlock(&ioapic->lock);
        return 0;
 }
 
@@ -349,6 +367,7 @@ int kvm_ioapic_init(struct kvm *kvm)
        ioapic = kzalloc(sizeof(struct kvm_ioapic), GFP_KERNEL);
        if (!ioapic)
                return -ENOMEM;
+       mutex_init(&ioapic->lock);
        kvm->arch.vioapic = ioapic;
        kvm_ioapic_reset(ioapic);
        kvm_iodevice_init(&ioapic->dev, &ioapic_mmio_ops);
@@ -360,3 +379,26 @@ int kvm_ioapic_init(struct kvm *kvm)
        return ret;
 }
 
+int kvm_get_ioapic(struct kvm *kvm, struct kvm_ioapic_state *state)
+{
+       struct kvm_ioapic *ioapic = ioapic_irqchip(kvm);
+       if (!ioapic)
+               return -EINVAL;
+
+       mutex_lock(&ioapic->lock);
+       memcpy(state, ioapic, sizeof(struct kvm_ioapic_state));
+       mutex_unlock(&ioapic->lock);
+       return 0;
+}
+
+int kvm_set_ioapic(struct kvm *kvm, struct kvm_ioapic_state *state)
+{
+       struct kvm_ioapic *ioapic = ioapic_irqchip(kvm);
+       if (!ioapic)
+               return -EINVAL;
+
+       mutex_lock(&ioapic->lock);
+       memcpy(ioapic, state, sizeof(struct kvm_ioapic_state));
+       mutex_unlock(&ioapic->lock);
+       return 0;
+}
index 6e461ad..419c43b 100644 (file)
@@ -45,6 +45,7 @@ struct kvm_ioapic {
        struct kvm_io_device dev;
        struct kvm *kvm;
        void (*ack_notifier)(void *opaque, int irq);
+       struct mutex lock;
 };
 
 #ifdef DEBUG
@@ -74,4 +75,7 @@ int kvm_ioapic_set_irq(struct kvm_ioapic *ioapic, int irq, int level);
 void kvm_ioapic_reset(struct kvm_ioapic *ioapic);
 int kvm_irq_delivery_to_apic(struct kvm *kvm, struct kvm_lapic *src,
                struct kvm_lapic_irq *irq);
+int kvm_get_ioapic(struct kvm *kvm, struct kvm_ioapic_state *state);
+int kvm_set_ioapic(struct kvm *kvm, struct kvm_ioapic_state *state);
+
 #endif
index 6c94614..fadf440 100644 (file)
@@ -146,8 +146,8 @@ static int kvm_set_msi(struct kvm_kernel_irq_routing_entry *e,
  */
 int kvm_set_irq(struct kvm *kvm, int irq_source_id, u32 irq, int level)
 {
-       struct kvm_kernel_irq_routing_entry *e;
-       int ret = -1;
+       struct kvm_kernel_irq_routing_entry *e, irq_set[KVM_NR_IRQCHIPS];
+       int ret = -1, i = 0;
        struct kvm_irq_routing_table *irq_rt;
        struct hlist_node *n;
 
@@ -162,14 +162,19 @@ int kvm_set_irq(struct kvm *kvm, int irq_source_id, u32 irq, int level)
        rcu_read_lock();
        irq_rt = rcu_dereference(kvm->irq_routing);
        if (irq < irq_rt->nr_rt_entries)
-               hlist_for_each_entry(e, n, &irq_rt->map[irq], link) {
-                       int r = e->set(e, kvm, irq_source_id, level);
-                       if (r < 0)
-                               continue;
-
-                       ret = r + ((ret < 0) ? 0 : ret);
-               }
+               hlist_for_each_entry(e, n, &irq_rt->map[irq], link)
+                       irq_set[i++] = *e;
        rcu_read_unlock();
+
+       while(i--) {
+               int r;
+               r = irq_set[i].set(&irq_set[i], kvm, irq_source_id, level);
+               if (r < 0)
+                       continue;
+
+               ret = r + ((ret < 0) ? 0 : ret);
+       }
+
        return ret;
 }