[Intel-gfx] [PATCH v7 20/20] drm/i915/vm_bind: Async vm_unbind support

[Niranjana Vishwanathapura]

Asynchronously unbind the vma upon vm_unbind call.
Fall back to synchronous unbind if backend doesn't support
async unbind or if async unbind fails.

No need for vm_unbind out fence support as i915 will internally
handle all sequencing and user need not try to sequence any
operation with the unbind completion.

v2: use i915_vma_destroy_async in vm_unbind ioctl

Signed-off-by: Niranjana Vishwanathapura <niranjana.vishwanathapura at intel.com>
---
 .../drm/i915/gem/i915_gem_vm_bind_object.c    |  2 +-
 drivers/gpu/drm/i915/i915_vma.c               | 51 +++++++++++++++++--
 drivers/gpu/drm/i915/i915_vma.h               |  1 +
 include/uapi/drm/i915_drm.h                   |  3 +-
 4 files changed, 51 insertions(+), 6 deletions(-)

diff --git a/drivers/gpu/drm/i915/gem/i915_gem_vm_bind_object.c b/drivers/gpu/drm/i915/gem/i915_gem_vm_bind_object.c
index d87d1210365b..36651b447966 100644
--- a/drivers/gpu/drm/i915/gem/i915_gem_vm_bind_object.c
+++ b/drivers/gpu/drm/i915/gem/i915_gem_vm_bind_object.c
@@ -210,7 +210,7 @@ static int i915_gem_vm_unbind_vma(struct i915_address_space *vm,
 	 */
 	obj = vma->obj;
 	i915_gem_object_lock(obj, NULL);
-	i915_vma_destroy(vma);
+	i915_vma_destroy_async(vma);
 	i915_gem_object_unlock(obj);
 
 	i915_gem_object_put(obj);
diff --git a/drivers/gpu/drm/i915/i915_vma.c b/drivers/gpu/drm/i915/i915_vma.c
index 7cf77c67d755..483d25f2425c 100644
--- a/drivers/gpu/drm/i915/i915_vma.c
+++ b/drivers/gpu/drm/i915/i915_vma.c
@@ -42,6 +42,8 @@
 #include "i915_vma.h"
 #include "i915_vma_resource.h"
 
+static struct dma_fence *__i915_vma_unbind_async(struct i915_vma *vma);
+
 static inline void assert_vma_held_evict(const struct i915_vma *vma)
 {
 	/*
@@ -1713,7 +1715,7 @@ void i915_vma_reopen(struct i915_vma *vma)
 	spin_unlock_irq(&gt->closed_lock);
 }
 
-static void force_unbind(struct i915_vma *vma)
+static void force_unbind(struct i915_vma *vma, bool async)
 {
 	if (!drm_mm_node_allocated(&vma->node))
 		return;
@@ -1727,7 +1729,21 @@ static void force_unbind(struct i915_vma *vma)
 		i915_vma_set_purged(vma);
 
 	atomic_and(~I915_VMA_PIN_MASK, &vma->flags);
-	WARN_ON(__i915_vma_unbind(vma));
+	if (async) {
+		struct dma_fence *fence;
+
+		fence = __i915_vma_unbind_async(vma);
+		if (IS_ERR_OR_NULL(fence)) {
+			async = false;
+		} else {
+			dma_resv_add_fence(vma->obj->base.resv, fence,
+					   DMA_RESV_USAGE_READ);
+			dma_fence_put(fence);
+		}
+	}
+
+	if (!async)
+		WARN_ON(__i915_vma_unbind(vma));
 	GEM_BUG_ON(drm_mm_node_allocated(&vma->node));
 }
 
@@ -1787,7 +1803,7 @@ void i915_vma_destroy_locked(struct i915_vma *vma)
 {
 	lockdep_assert_held(&vma->vm->mutex);
 
-	force_unbind(vma);
+	force_unbind(vma, false);
 	list_del_init(&vma->vm_link);
 	release_references(vma, vma->vm->gt, false);
 }
@@ -1798,7 +1814,34 @@ void i915_vma_destroy(struct i915_vma *vma)
 	bool vm_ddestroy;
 
 	mutex_lock(&vma->vm->mutex);
-	force_unbind(vma);
+	force_unbind(vma, false);
+	list_del_init(&vma->vm_link);
+	vm_ddestroy = vma->vm_ddestroy;
+	vma->vm_ddestroy = false;
+
+	/* vma->vm may be freed when releasing vma->vm->mutex. */
+	gt = vma->vm->gt;
+	mutex_unlock(&vma->vm->mutex);
+	release_references(vma, gt, vm_ddestroy);
+}
+
+void i915_vma_destroy_async(struct i915_vma *vma)
+{
+	bool vm_ddestroy, async = vma->obj->mm.rsgt;
+	struct intel_gt *gt;
+
+	if (dma_resv_reserve_fences(vma->obj->base.resv, 1))
+		async = false;
+
+	mutex_lock(&vma->vm->mutex);
+	/*
+	 * Ensure any asynchronous binding is complete while using
+	 * async unbind as we will be releasing the vma here.
+	 */
+	if (async && i915_active_wait(&vma->active))
+		async = false;
+
+	force_unbind(vma, async);
 	list_del_init(&vma->vm_link);
 	vm_ddestroy = vma->vm_ddestroy;
 	vma->vm_ddestroy = false;
diff --git a/drivers/gpu/drm/i915/i915_vma.h b/drivers/gpu/drm/i915/i915_vma.h
index 737ef310d046..25f15965dab8 100644
--- a/drivers/gpu/drm/i915/i915_vma.h
+++ b/drivers/gpu/drm/i915/i915_vma.h
@@ -272,6 +272,7 @@ void i915_vma_reopen(struct i915_vma *vma);
 
 void i915_vma_destroy_locked(struct i915_vma *vma);
 void i915_vma_destroy(struct i915_vma *vma);
+void i915_vma_destroy_async(struct i915_vma *vma);
 
 #define assert_vma_held(vma) dma_resv_assert_held((vma)->obj->base.resv)
 
diff --git a/include/uapi/drm/i915_drm.h b/include/uapi/drm/i915_drm.h
index e5600f358a15..431d40bb1dee 100644
--- a/include/uapi/drm/i915_drm.h
+++ b/include/uapi/drm/i915_drm.h
@@ -3969,7 +3969,8 @@ struct drm_i915_gem_vm_bind {
  * any error.
  *
  * VM_BIND/UNBIND ioctl calls executed on different CPU threads concurrently
- * are not ordered.
+ * are not ordered. Furthermore, parts of the VM_UNBIND operation can be done
+ * asynchronously.
  */
 struct drm_i915_gem_vm_unbind {
 	/** @vm_id: VM (address space) id to bind */
-- 
2.21.0.rc0.32.g243a4c7e27