x86/irq: do not release irq until all cleanup is done

Current code in _clear_irq_vector() will mark the irq as unused before
doing the cleanup required when move_in_progress is true.

This can lead to races in create_irq() if the function picks an irq
desc that's been marked as unused but has move_in_progress set, as the
call to assign_irq_vector() in that function can then fail with
-EAGAIN.

Prevent that by only marking irq descs as unused when all the cleanup
has been done.  While there also use write_atomic() when setting
IRQ_UNUSED in _clear_irq_vector() and add a barrier in order to
prevent the setting of IRQ_UNUSED getting reordered by the compiler.

The check for move_in_progress cannot be removed from
_assign_irq_vector(), as other users (io_apic_set_pci_routing() and
ioapic_guest_write()) can still pass active irq descs to
assign_irq_vector().

Note the trace point is not moved and is now set before the irq is
marked as unused.  This is done so that the CPU mask provided in the
trace point is the one belonging to the current vector, not the old
one.

Signed-off-by: Roger Pau Monné <roger.pau@citrix.com>
Reviewed-by: Jan Beulich <jbeulich@suse.com>

xen/arch/x86/irq.c

@@ -220,27 +220,28 @@ static void _clear_irq_vector(struct irq_desc *desc)
         clear_bit(vector, desc->arch.used_vectors);
     }
 
-    desc->arch.used = IRQ_UNUSED;
-
     trace_irq_mask(TRC_HW_IRQ_CLEAR_VECTOR, irq, vector, tmp_mask);
 
-    if ( likely(!desc->arch.move_in_progress) )
-        return;
-
-    /* If we were in motion, also clear desc->arch.old_vector */
-    old_vector = desc->arch.old_vector;
-    cpumask_and(tmp_mask, desc->arch.old_cpu_mask, &cpu_online_map);
-
-    for_each_cpu(cpu, tmp_mask)
+    if ( unlikely(desc->arch.move_in_progress) )
     {
-        ASSERT(per_cpu(vector_irq, cpu)[old_vector] == irq);
-        TRACE_3D(TRC_HW_IRQ_MOVE_FINISH, irq, old_vector, cpu);
-        per_cpu(vector_irq, cpu)[old_vector] = ~irq;
+        /* If we were in motion, also clear desc->arch.old_vector */
+        old_vector = desc->arch.old_vector;
+        cpumask_and(tmp_mask, desc->arch.old_cpu_mask, &cpu_online_map);
+
+        for_each_cpu(cpu, tmp_mask)
+        {
+            ASSERT(per_cpu(vector_irq, cpu)[old_vector] == irq);
+            TRACE_3D(TRC_HW_IRQ_MOVE_FINISH, irq, old_vector, cpu);
+            per_cpu(vector_irq, cpu)[old_vector] = ~irq;
+        }
+
+        release_old_vec(desc);
+
+        desc->arch.move_in_progress = 0;
     }
 
-    release_old_vec(desc);
-
-    desc->arch.move_in_progress = 0;
+    smp_wmb();
+    write_atomic(&desc->arch.used, IRQ_UNUSED);
 }
 
 void __init clear_irq_vector(int irq)