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/

drivers/gpu/drm/msm/msm_drv.c

@@ -837,6 +837,7 @@ static const struct drm_driver msm_driver = {
 	.postclose          = msm_postclose,
 	.dumb_create        = msm_gem_dumb_create,
 	.dumb_map_offset    = msm_gem_dumb_map_offset,
+	.gem_prime_import   = msm_gem_prime_import,
 	.gem_prime_import_sg_table = msm_gem_prime_import_sg_table,
 #ifdef CONFIG_DEBUG_FS
 	.debugfs_init       = msm_debugfs_init,

drivers/gpu/drm/msm/msm_drv.h

@@ -269,6 +269,7 @@ void msm_gem_shrinker_cleanup(struct drm_device *dev);
 struct sg_table *msm_gem_prime_get_sg_table(struct drm_gem_object *obj);
 int msm_gem_prime_vmap(struct drm_gem_object *obj, struct iosys_map *map);
 void msm_gem_prime_vunmap(struct drm_gem_object *obj, struct iosys_map *map);
+struct drm_gem_object *msm_gem_prime_import(struct drm_device *dev, struct dma_buf *buf);
 struct drm_gem_object *msm_gem_prime_import_sg_table(struct drm_device *dev,
 		struct dma_buf_attachment *attach, struct sg_table *sg);
 struct dma_buf *msm_gem_prime_export(struct drm_gem_object *obj, int flags);

drivers/gpu/drm/msm/msm_fb.c

@@ -89,6 +89,7 @@ int msm_framebuffer_prepare(struct drm_framebuffer *fb, bool needs_dirtyfb)
 		return 0;
 
 	for (i = 0; i < n; i++) {
+		msm_gem_vma_get(fb->obj[i]);
 		ret = msm_gem_get_and_pin_iova(fb->obj[i], vm, &msm_fb->iova[i]);
 		drm_dbg_state(fb->dev, "FB[%u]: iova[%d]: %08llx (%d)\n",
 			      fb->base.id, i, msm_fb->iova[i], ret);
@@ -114,8 +115,10 @@ void msm_framebuffer_cleanup(struct drm_framebuffer *fb, bool needed_dirtyfb)
 
 	memset(msm_fb->iova, 0, sizeof(msm_fb->iova));
 
-	for (i = 0; i < n; i++)
+	for (i = 0; i < n; i++) {
 		msm_gem_unpin_iova(fb->obj[i], vm);
+		msm_gem_vma_put(fb->obj[i]);
+	}
 }
 
 uint32_t msm_framebuffer_iova(struct drm_framebuffer *fb, int plane)

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 */
 }
 

drivers/gpu/drm/msm/msm_gem.h

@@ -211,9 +211,37 @@ struct msm_gem_object {
 	 * Protected by LRU lock.
 	 */
 	int pin_count;
+
+	/**
+	 * @vma_ref: Reference count of VMA users.
+	 *
+	 * With the vm_bo/vma holding a reference to the GEM object, we'd
+	 * otherwise have to actively tear down a VMA when, for example,
+	 * a buffer is unpinned for scanout, vs. the pre-drm_gpuvm approach
+	 * where a VMA did not hold a reference to the BO, but instead was
+	 * implicitly torn down when the BO was freed.
+	 *
+	 * To regain the lazy VMA teardown, we use the @vma_ref.  It is
+	 * incremented for any of the following:
+	 *
+	 * 1) the BO is exported as a dma_buf
+	 * 2) the BO has open userspace handle
+	 *
+	 * All of those conditions will hold an reference to the BO,
+	 * preventing it from being freed.  So lazily keeping around the
+	 * VMA will not prevent the BO from being freed.  (Or rather, the
+	 * reference loop is harmless in this case.)
+	 *
+	 * When the @vma_ref drops to zero, then kms->vm VMA will be
+	 * torn down.
+	 */
+	atomic_t vma_ref;
 };
 #define to_msm_bo(x) container_of(x, struct msm_gem_object, base)
 
+void msm_gem_vma_get(struct drm_gem_object *obj);
+void msm_gem_vma_put(struct drm_gem_object *obj);
+
 uint64_t msm_gem_mmap_offset(struct drm_gem_object *obj);
 int msm_gem_prot(struct drm_gem_object *obj);
 int msm_gem_pin_vma_locked(struct drm_gem_object *obj, struct drm_gpuva *vma);

drivers/gpu/drm/msm/msm_gem_prime.c

@@ -6,6 +6,7 @@
 
 #include <linux/dma-buf.h>
 
+#include <drm/drm_drv.h>
 #include <drm/drm_prime.h>
 
 #include "msm_drv.h"
@@ -42,19 +43,68 @@ void msm_gem_prime_vunmap(struct drm_gem_object *obj, struct iosys_map *map)
 	msm_gem_put_vaddr_locked(obj);
 }
 
+static void msm_gem_dmabuf_release(struct dma_buf *dma_buf)
+{
+	struct drm_gem_object *obj = dma_buf->priv;
+
+	msm_gem_vma_put(obj);
+	drm_gem_dmabuf_release(dma_buf);
+}
+
+static const struct dma_buf_ops msm_gem_prime_dmabuf_ops =  {
+	.attach = drm_gem_map_attach,
+	.detach = drm_gem_map_detach,
+	.map_dma_buf = drm_gem_map_dma_buf,
+	.unmap_dma_buf = drm_gem_unmap_dma_buf,
+	.release = msm_gem_dmabuf_release,
+	.mmap = drm_gem_dmabuf_mmap,
+	.vmap = drm_gem_dmabuf_vmap,
+	.vunmap = drm_gem_dmabuf_vunmap,
+};
+
+struct drm_gem_object *msm_gem_prime_import(struct drm_device *dev,
+					    struct dma_buf *buf)
+{
+	if (buf->ops == &msm_gem_prime_dmabuf_ops) {
+		struct drm_gem_object *obj = buf->priv;
+		if (obj->dev == dev) {
+			/*
+			 * Importing dmabuf exported from our own gem increases
+			 * refcount on gem itself instead of f_count of dmabuf.
+			 */
+			drm_gem_object_get(obj);
+			return obj;
+		}
+	}
+
+	return drm_gem_prime_import(dev, buf);
+}
+
 struct drm_gem_object *msm_gem_prime_import_sg_table(struct drm_device *dev,
 		struct dma_buf_attachment *attach, struct sg_table *sg)
 {
 	return msm_gem_import(dev, attach->dmabuf, sg);
 }
 
-
 struct dma_buf *msm_gem_prime_export(struct drm_gem_object *obj, int flags)
 {
 	if (to_msm_bo(obj)->flags & MSM_BO_NO_SHARE)
 		return ERR_PTR(-EPERM);
 
-	return drm_gem_prime_export(obj, flags);
+	msm_gem_vma_get(obj);
+
+	struct drm_device *dev = obj->dev;
+	struct dma_buf_export_info exp_info = {
+		.exp_name = KBUILD_MODNAME, /* white lie for debug */
+		.owner = dev->driver->fops->owner,
+		.ops = &msm_gem_prime_dmabuf_ops,
+		.size = obj->size,
+		.flags = flags,
+		.priv = obj,
+		.resv = obj->resv,
+	};
+
+	return drm_gem_dmabuf_export(dev, &exp_info);
 }
 
 int msm_gem_prime_pin(struct drm_gem_object *obj)