VMX: properly handle pi when all the assigned devices are removed

This patch handles some corner cases when the last assigned device
is removed from the domain. In this case we should carefully handle
pi descriptor and the per-cpu blocking list, to make sure:
- all the PI descriptor are in the right state when next time a
devices is assigned to the domain again.
- No remaining vcpus of the domain in the per-cpu blocking list.

Here we call vmx_pi_unblock_vcpu() to remove the vCPU from the blocking list
if it is on the list. However, this could happen when vmx_vcpu_block() is
being called, hence we might incorrectly add the vCPU to the blocking list
while the last devcie is detached from the domain. Consider that the situation
can only occur when detaching the last device from the domain and it is not
a frequent operation, so we use domain_pause before that, which is considered
as an clean and maintainable solution for the situation.

Signed-off-by: Feng Wu <feng.wu@intel.com>
Signed-off-by: Chao Gao <chao.gao@intel.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>
Acked-by: Kevin Tian <kevin.tian@intel.com>

diff --git a/xen/arch/x86/hvm/vmx/vmx.c b/xen/arch/x86/hvm/vmx/vmx.c
index fe2799727c79a06eeb56b27bc93b1dd9477d5e32..894d7d419f8b6b04cdfe4191d1959b745c16c905 100644 (file)
--- a/xen/arch/x86/hvm/vmx/vmx.c
+++ b/xen/arch/x86/hvm/vmx/vmx.c
@@ -154,14 +154,12 @@ static void vmx_pi_switch_to(struct vcpu *v)
     pi_clear_sn(pi_desc);
 }
 
-static void vmx_pi_do_resume(struct vcpu *v)
+static void vmx_pi_unblock_vcpu(struct vcpu *v)
 {
     unsigned long flags;
     spinlock_t *pi_blocking_list_lock;
     struct pi_desc *pi_desc = &v->arch.hvm_vmx.pi_desc;
 
-    ASSERT(!test_bit(_VPF_blocked, &v->pause_flags));
-
     /*
      * Set 'NV' field back to posted_intr_vector, so the
      * Posted-Interrupts can be delivered to the vCPU when
@@ -169,12 +167,12 @@ static void vmx_pi_do_resume(struct vcpu *v)
      */
     write_atomic(&pi_desc->nv, posted_intr_vector);
 
-    /* The vCPU is not on any blocking list. */
     pi_blocking_list_lock = v->arch.hvm_vmx.pi_blocking.lock;
 
     /* Prevent the compiler from eliminating the local variable.*/
     smp_rmb();
 
+    /* The vCPU is not on any blocking list. */
     if ( pi_blocking_list_lock == NULL )
         return;
 
@@ -194,6 +192,13 @@ static void vmx_pi_do_resume(struct vcpu *v)
     spin_unlock_irqrestore(pi_blocking_list_lock, flags);
 }
 
+static void vmx_pi_do_resume(struct vcpu *v)
+{
+    ASSERT(!test_bit(_VPF_blocked, &v->pause_flags));
+
+    vmx_pi_unblock_vcpu(v);
+}
+
 /*
  * To handle posted interrupts correctly, we need to set the following
  * state:
@@ -279,11 +284,22 @@ void vmx_pi_hooks_assign(struct domain *d)
 /* This function is called when pcidevs_lock is held */
 void vmx_pi_hooks_deassign(struct domain *d)
 {
+    struct vcpu *v;
+
     if ( !iommu_intpost || !has_hvm_container_domain(d) )
         return;
 
     ASSERT(d->arch.hvm_domain.pi_ops.vcpu_block);
 
+    /*
+     * Pausing the domain can make sure the vCPUs are not
+     * running and hence not calling the hooks simultaneously
+     * when deassigning the PI hooks and removing the vCPU
+     * from the blocking list.
+     */
+    ASSERT(current->domain != d);
+    domain_pause(d);
+
     /*
      * Note that we don't set 'd->arch.hvm_domain.pi_ops.switch_to' to NULL
      * here. If we deassign the hooks while the vCPU is runnable in the
@@ -301,6 +317,11 @@ void vmx_pi_hooks_deassign(struct domain *d)
     d->arch.hvm_domain.pi_ops.vcpu_block = NULL;
     d->arch.hvm_domain.pi_ops.switch_from = NULL;
     d->arch.hvm_domain.pi_ops.do_resume = NULL;
+
+    for_each_vcpu ( d, v )
+        vmx_pi_unblock_vcpu(v);
+
+    domain_unpause(d);
 }
 
 static int vmx_domain_initialise(struct domain *d)