KVM: Fix racy in kvm_free_assigned_irq

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>

2 files changed, 20 insertions, 9 deletions

diff --git a/arch/x86/kvm/x86.c b/arch/x86/kvm/x86.c
index b0fc079f1bee..fc3e329f6ade 100644
--- a/arch/x86/kvm/x86.c
+++ b/arch/x86/kvm/x86.c
@@ -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);
diff --git a/virt/kvm/kvm_main.c b/virt/kvm/kvm_main.c
index 68e3f1ec1674..277ea7f39fc8 100644
--- a/virt/kvm/kvm_main.c
+++ b/virt/kvm/kvm_main.c
@@ -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);