drm/i915: Fix a race between vma / object destruction and unbinding

The vma destruction code was using an unlocked advisory check for
drm_mm_node_allocated() to avoid racing with eviction code unbinding
the vma.

This is very fragile and prohibits the dereference of non-refcounted
pointers of dying vmas after a call to __i915_vma_unbind(). It also
prohibits the dereference of vma->obj of refcounted pointers of
dying vmas after a call to __i915_vma_unbind(), since even if a
refcount is held on the vma, that won't guarantee that its backing
object doesn't get destroyed.

So introduce an unbind under the vm mutex at object destroy time,
removing all weak references of the vma and its object from the
object vma list and from the vm bound list.

Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220127115622.302970-1-thomas.hellstrom@linux.intel.com

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_object.c b/drivers/gpu/drm/i915/gem/i915_gem_object.c
index 1a9e1f9..e03e362 100644 (file)
--- a/drivers/gpu/drm/i915/gem/i915_gem_object.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_object.c
@@ -280,6 +280,12 @@ void __i915_gem_object_pages_fini(struct drm_i915_gem_object *obj)
                        GEM_BUG_ON(vma->obj != obj);
                        spin_unlock(&obj->vma.lock);
 
+                       /* Verify that the vma is unbound under the vm mutex. */
+                       mutex_lock(&vma->vm->mutex);
+                       atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
+                       __i915_vma_unbind(vma);
+                       mutex_unlock(&vma->vm->mutex);
+
                        __i915_vma_put(vma);
 
                        spin_lock(&obj->vma.lock);