KVM: Fix racy in kvm_free_assigned_irq
authorSheng Yang <sheng@linux.intel.com>
Tue, 6 Jan 2009 02:03:03 +0000 (10:03 +0800)
committerAvi Kivity <avi@redhat.com>
Sun, 15 Feb 2009 00:47:36 +0000 (02:47 +0200)
In the past, kvm_get_kvm() and kvm_put_kvm() was called in assigned device irq
handler and interrupt_work, in order to prevent cancel_work_sync() in
kvm_free_assigned_irq got a illegal state when waiting for interrupt_work done.
But it's tricky and still got two problems:

1. A bug ignored two conditions that cancel_work_sync() would return true result
in a additional kvm_put_kvm().

2. If interrupt type is MSI, we would got a window between cancel_work_sync()
and free_irq(), which interrupt would be injected again...

This patch discard the reference count used for irq handler and interrupt_work,
and ensure the legal state by moving the free function at the very beginning of
kvm_destroy_vm(). And the patch fix the second bug by disable irq before
cancel_work_sync(), which may result in nested disable of irq but OK for we are
going to free it.

Signed-off-by: Sheng Yang <sheng@linux.intel.com>
Signed-off-by: Avi Kivity <avi@redhat.com>
arch/x86/kvm/x86.c
virt/kvm/kvm_main.c

index b0fc079f1bee54f97a4c21affa4ef380e10a1cbe..fc3e329f6ade807e720adbd21aa737a5d5932e6a 100644 (file)
@@ -4129,11 +4129,11 @@ static void kvm_free_vcpus(struct kvm *kvm)
 
 void kvm_arch_sync_events(struct kvm *kvm)
 {
+       kvm_free_all_assigned_devices(kvm);
 }
 
 void kvm_arch_destroy_vm(struct kvm *kvm)
 {
-       kvm_free_all_assigned_devices(kvm);
        kvm_iommu_unmap_guest(kvm);
        kvm_free_pit(kvm);
        kfree(kvm->arch.vpic);
index 68e3f1ec1674200f151b75abaf4fdfa7b4da04de..277ea7f39fc8e360e9e552b01b3a18c88559ca12 100644 (file)
@@ -173,7 +173,6 @@ static void kvm_assigned_dev_interrupt_work_handler(struct work_struct *work)
                assigned_dev->host_irq_disabled = false;
        }
        mutex_unlock(&assigned_dev->kvm->lock);
-       kvm_put_kvm(assigned_dev->kvm);
 }
 
 static irqreturn_t kvm_assigned_dev_intr(int irq, void *dev_id)
@@ -181,8 +180,6 @@ static irqreturn_t kvm_assigned_dev_intr(int irq, void *dev_id)
        struct kvm_assigned_dev_kernel *assigned_dev =
                (struct kvm_assigned_dev_kernel *) dev_id;
 
-       kvm_get_kvm(assigned_dev->kvm);
-
        schedule_work(&assigned_dev->interrupt_work);
 
        disable_irq_nosync(irq);
@@ -213,6 +210,7 @@ static void kvm_assigned_dev_ack_irq(struct kvm_irq_ack_notifier *kian)
        }
 }
 
+/* The function implicit hold kvm->lock mutex due to cancel_work_sync() */
 static void kvm_free_assigned_irq(struct kvm *kvm,
                                  struct kvm_assigned_dev_kernel *assigned_dev)
 {
@@ -228,11 +226,24 @@ static void kvm_free_assigned_irq(struct kvm *kvm,
        if (!assigned_dev->irq_requested_type)
                return;
 
-       if (cancel_work_sync(&assigned_dev->interrupt_work))
-               /* We had pending work. That means we will have to take
-                * care of kvm_put_kvm.
-                */
-               kvm_put_kvm(kvm);
+       /*
+        * In kvm_free_device_irq, cancel_work_sync return true if:
+        * 1. work is scheduled, and then cancelled.
+        * 2. work callback is executed.
+        *
+        * The first one ensured that the irq is disabled and no more events
+        * would happen. But for the second one, the irq may be enabled (e.g.
+        * for MSI). So we disable irq here to prevent further events.
+        *
+        * Notice this maybe result in nested disable if the interrupt type is
+        * INTx, but it's OK for we are going to free it.
+        *
+        * If this function is a part of VM destroy, please ensure that till
+        * now, the kvm state is still legal for probably we also have to wait
+        * interrupt_work done.
+        */
+       disable_irq_nosync(assigned_dev->host_irq);
+       cancel_work_sync(&assigned_dev->interrupt_work);
 
        free_irq(assigned_dev->host_irq, (void *)assigned_dev);