drm/msm: Defer VMA unmap for fb unpins

With the conversion to drm_gpuvm, we lost the lazy VMA cleanup, which
means that fb cleanup/unpin when pageflipping to new scanout buffers
immediately unmaps the scanout buffer.  This is costly (with tlbinv,
it can be 4-6ms for a 1080p scanout buffer, and more for higher
resolutions)!

To avoid this, introduce a vma_ref, which is incremented whenever
userspace has a GEM handle or dma-buf fd.  When unpinning if the
vm is the kms->vm we defer tearing down the VMA until the vma_ref
drops to zero.  If the buffer is still part of a flip-chain then
userspace will be holding some sort of reference to the BO, either
via a GEM handle and/or dma-buf fd.  So this avoids unmapping the VMA
when there is a strong possibility that it will be needed again.

Signed-off-by: Rob Clark <robin.clark@oss.qualcomm.com>
Tested-by: Antonino Maniscalco <antomani103@gmail.com>
Reviewed-by: Antonino Maniscalco <antomani103@gmail.com>
Patchwork: https://patchwork.freedesktop.org/patch/661538/

1 files changed, 37 insertions, 23 deletions

diff --git a/drivers/gpu/drm/msm/msm_gem.c b/drivers/gpu/drm/msm/msm_gem.c
index 3e87d27dfcb6..33d3354c6102 100644
--- a/drivers/gpu/drm/msm/msm_gem.c
+++ b/drivers/gpu/drm/msm/msm_gem.c
@@ -19,6 +19,7 @@
 #include "msm_drv.h"
 #include "msm_gem.h"
 #include "msm_gpu.h"
+#include "msm_kms.h"
 
 static void update_device_mem(struct msm_drm_private *priv, ssize_t size)
 {
@@ -39,6 +40,7 @@ static void update_ctx_mem(struct drm_file *file, ssize_t size)
 
 static int msm_gem_open(struct drm_gem_object *obj, struct drm_file *file)
 {
+	msm_gem_vma_get(obj);
 	update_ctx_mem(file, obj->size);
 	return 0;
 }
@@ -46,33 +48,13 @@ static int msm_gem_open(struct drm_gem_object *obj, struct drm_file *file)
 static void put_iova_spaces(struct drm_gem_object *obj, struct drm_gpuvm *vm,
 			    bool close, const char *reason);
 
-static void detach_vm(struct drm_gem_object *obj, struct drm_gpuvm *vm)
-{
-	msm_gem_assert_locked(obj);
-	drm_gpuvm_resv_assert_held(vm);
-
-	struct drm_gpuvm_bo *vm_bo = drm_gpuvm_bo_find(vm, obj);
-	if (vm_bo) {
-		struct drm_gpuva *vma;
-
-		drm_gpuvm_bo_for_each_va (vma, vm_bo) {
-			if (vma->vm != vm)
-				continue;
-			msm_gem_vma_unmap(vma, "detach");
-			msm_gem_vma_close(vma);
-			break;
-		}
-
-		drm_gpuvm_bo_put(vm_bo);
-	}
-}
-
 static void msm_gem_close(struct drm_gem_object *obj, struct drm_file *file)
 {
 	struct msm_context *ctx = file->driver_priv;
 	struct drm_exec exec;
 
 	update_ctx_mem(file, -obj->size);
+	msm_gem_vma_put(obj);
 
 	/*
 	 * If VM isn't created yet, nothing to cleanup.  And in fact calling
@@ -99,7 +81,31 @@ static void msm_gem_close(struct drm_gem_object *obj, struct drm_file *file)
 
 	msm_gem_lock_vm_and_obj(&exec, obj, ctx->vm);
 	put_iova_spaces(obj, ctx->vm, true, "close");
-	detach_vm(obj, ctx->vm);
+	drm_exec_fini(&exec);     /* drop locks */
+}
+
+/*
+ * Get/put for kms->vm VMA
+ */
+
+void msm_gem_vma_get(struct drm_gem_object *obj)
+{
+	atomic_inc(&to_msm_bo(obj)->vma_ref);
+}
+
+void msm_gem_vma_put(struct drm_gem_object *obj)
+{
+	struct msm_drm_private *priv = obj->dev->dev_private;
+	struct drm_exec exec;
+
+	if (atomic_dec_return(&to_msm_bo(obj)->vma_ref))
+		return;
+
+	if (!priv->kms)
+		return;
+
+	msm_gem_lock_vm_and_obj(&exec, obj, priv->kms->vm);
+	put_iova_spaces(obj, priv->kms->vm, true, "vma_put");
 	drm_exec_fini(&exec);     /* drop locks */
 }
 
@@ -656,6 +662,13 @@ int msm_gem_set_iova(struct drm_gem_object *obj,
 	return ret;
 }
 
+static bool is_kms_vm(struct drm_gpuvm *vm)
+{
+	struct msm_drm_private *priv = vm->drm->dev_private;
+
+	return priv->kms && (priv->kms->vm == vm);
+}
+
 /*
  * Unpin a iova by updating the reference counts. The memory isn't actually
  * purged until something else (shrinker, mm_notifier, destroy, etc) decides
@@ -671,7 +684,8 @@ void msm_gem_unpin_iova(struct drm_gem_object *obj, struct drm_gpuvm *vm)
 	if (vma) {
 		msm_gem_unpin_locked(obj);
 	}
-	detach_vm(obj, vm);
+	if (!is_kms_vm(vm))
+		put_iova_spaces(obj, vm, true, "close");
 	drm_exec_fini(&exec);     /* drop locks */
 }