vmwgfx: Implement fence objects

Will be needed for queries and drm event-driven throttling.

As a benefit, they help avoid stale user-space fence handles.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Jakob Bornecrantz <jakob@vmware.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drivers/gpu/drm/vmwgfx/Makefile

@@ -4,6 +4,7 @@ ccflags-y := -Iinclude/drm
 vmwgfx-y := vmwgfx_execbuf.o vmwgfx_gmr.o vmwgfx_kms.o vmwgfx_drv.o \
 	    vmwgfx_fb.o vmwgfx_ioctl.o vmwgfx_resource.o vmwgfx_buffer.o \
 	    vmwgfx_fifo.o vmwgfx_irq.o vmwgfx_ldu.o vmwgfx_ttm_glue.o \
-	    vmwgfx_overlay.o vmwgfx_marker.o vmwgfx_gmrid_manager.o
+	    vmwgfx_overlay.o vmwgfx_marker.o vmwgfx_gmrid_manager.o \
+	    vmwgfx_fence.o
 
 obj-$(CONFIG_DRM_VMWGFX) := vmwgfx.o

drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c

@@ -274,39 +274,39 @@ static int vmw_ttm_fault_reserve_notify(struct ttm_buffer_object *bo)
 
 static void *vmw_sync_obj_ref(void *sync_obj)
 {
-	return sync_obj;
+
+	return (void *)
+		vmw_fence_obj_reference((struct vmw_fence_obj *) sync_obj);
 }
 
 static void vmw_sync_obj_unref(void **sync_obj)
 {
-	*sync_obj = NULL;
+	vmw_fence_obj_unreference((struct vmw_fence_obj **) sync_obj);
 }
 
 static int vmw_sync_obj_flush(void *sync_obj, void *sync_arg)
 {
-	struct vmw_private *dev_priv = (struct vmw_private *)sync_arg;
-
-	mutex_lock(&dev_priv->hw_mutex);
-	vmw_write(dev_priv, SVGA_REG_SYNC, SVGA_SYNC_GENERIC);
-	mutex_unlock(&dev_priv->hw_mutex);
+	vmw_fence_obj_flush((struct vmw_fence_obj *) sync_obj);
 	return 0;
 }
 
 static bool vmw_sync_obj_signaled(void *sync_obj, void *sync_arg)
 {
-	struct vmw_private *dev_priv = (struct vmw_private *)sync_arg;
-	uint32_t seqno = (unsigned long) sync_obj;
+	unsigned long flags = (unsigned long) sync_arg;
+	return	vmw_fence_obj_signaled((struct vmw_fence_obj *) sync_obj,
+				       (uint32_t) flags);
 
-	return vmw_seqno_passed(dev_priv, seqno);
 }
 
 static int vmw_sync_obj_wait(void *sync_obj, void *sync_arg,
 			     bool lazy, bool interruptible)
 {
-	struct vmw_private *dev_priv = (struct vmw_private *)sync_arg;
-	uint32_t seqno = (unsigned long) sync_obj;
+	unsigned long flags = (unsigned long) sync_arg;
 
-	return vmw_wait_seqno(dev_priv, false, seqno, false, 3*HZ);
+	return vmw_fence_obj_wait((struct vmw_fence_obj *) sync_obj,
+				  (uint32_t) flags,
+				  lazy, interruptible,
+				  VMW_FENCE_WAIT_TIMEOUT);
 }
 
 struct ttm_bo_driver vmw_bo_driver = {

drivers/gpu/drm/vmwgfx/vmwgfx_drv.c

@@ -82,12 +82,18 @@
 #define DRM_IOCTL_VMW_EXECBUF					\
 	DRM_IOW(DRM_COMMAND_BASE + DRM_VMW_EXECBUF,		\
 		struct drm_vmw_execbuf_arg)
+#define DRM_IOCTL_VMW_GET_3D_CAP				\
+	DRM_IOW(DRM_COMMAND_BASE + DRM_VMW_GET_3D_CAP,		\
+		 struct drm_vmw_get_3d_cap_arg)
 #define DRM_IOCTL_VMW_FENCE_WAIT				\
 	DRM_IOWR(DRM_COMMAND_BASE + DRM_VMW_FENCE_WAIT,		\
 		 struct drm_vmw_fence_wait_arg)
-#define DRM_IOCTL_VMW_GET_3D_CAP				\
-	DRM_IOW(DRM_COMMAND_BASE + DRM_VMW_GET_3D_CAP,		\
-		struct drm_vmw_get_3d_cap_arg)
+#define DRM_IOCTL_VMW_FENCE_SIGNALED				\
+	DRM_IOWR(DRM_COMMAND_BASE + DRM_VMW_FENCE_SIGNALED,	\
+		 struct drm_vmw_fence_signaled_arg)
+#define DRM_IOCTL_VMW_FENCE_UNREF				\
+	DRM_IOW(DRM_COMMAND_BASE + DRM_VMW_FENCE_UNREF,		\
+		 struct drm_vmw_fence_arg)
 
 /**
  * The core DRM version of this macro doesn't account for
@@ -131,7 +137,12 @@ static struct drm_ioctl_desc vmw_ioctls[] = {
 		      DRM_AUTH | DRM_UNLOCKED),
 	VMW_IOCTL_DEF(VMW_EXECBUF, vmw_execbuf_ioctl,
 		      DRM_AUTH | DRM_UNLOCKED),
-	VMW_IOCTL_DEF(VMW_FENCE_WAIT, vmw_fence_wait_ioctl,
+	VMW_IOCTL_DEF(VMW_FENCE_WAIT, vmw_fence_obj_wait_ioctl,
+		      DRM_AUTH | DRM_UNLOCKED),
+	VMW_IOCTL_DEF(VMW_FENCE_SIGNALED,
+		      vmw_fence_obj_signaled_ioctl,
+		      DRM_AUTH | DRM_UNLOCKED),
+	VMW_IOCTL_DEF(VMW_FENCE_UNREF, vmw_fence_obj_unref_ioctl,
 		      DRM_AUTH | DRM_UNLOCKED),
 	VMW_IOCTL_DEF(VMW_GET_3D_CAP, vmw_get_cap_3d_ioctl,
 		      DRM_AUTH | DRM_UNLOCKED),
@@ -198,12 +209,14 @@ static int vmw_request_device(struct vmw_private *dev_priv)
 		DRM_ERROR("Unable to initialize FIFO.\n");
 		return ret;
 	}
+	vmw_fence_fifo_up(dev_priv->fman);
 
 	return 0;
 }
 
 static void vmw_release_device(struct vmw_private *dev_priv)
 {
+	vmw_fence_fifo_down(dev_priv->fman);
 	vmw_fifo_release(dev_priv, &dev_priv->fifo);
 }
 
@@ -434,6 +447,10 @@ static int vmw_driver_load(struct drm_device *dev, unsigned long chipset)
 			goto out_no_device;
 		}
 	}
+
+	dev_priv->fman = vmw_fence_manager_init(dev_priv);
+	if (unlikely(dev_priv->fman == NULL))
+		goto out_no_fman;
 	ret = vmw_kms_init(dev_priv);
 	if (unlikely(ret != 0))
 		goto out_no_kms;
@@ -475,6 +492,8 @@ static int vmw_driver_load(struct drm_device *dev, unsigned long chipset)
 	vmw_overlay_close(dev_priv);
 	vmw_kms_close(dev_priv);
 out_no_kms:
+	vmw_fence_manager_takedown(dev_priv->fman);
+out_no_fman:
 	if (dev_priv->stealth)
 		pci_release_region(dev->pdev, 2);
 	else
@@ -518,6 +537,7 @@ static int vmw_driver_unload(struct drm_device *dev)
 	}
 	vmw_kms_close(dev_priv);
 	vmw_overlay_close(dev_priv);
+	vmw_fence_manager_takedown(dev_priv->fman);
 	if (dev_priv->stealth)
 		pci_release_region(dev->pdev, 2);
 	else

drivers/gpu/drm/vmwgfx/vmwgfx_drv.h

@@ -38,6 +38,7 @@
 #include "ttm/ttm_lock.h"
 #include "ttm/ttm_execbuf_util.h"
 #include "ttm/ttm_module.h"
+#include "vmwgfx_fence.h"
 
 #define VMWGFX_DRIVER_DATE "20100927"
 #define VMWGFX_DRIVER_MAJOR 1
@@ -53,6 +54,11 @@
 #define VMW_PL_GMR TTM_PL_PRIV0
 #define VMW_PL_FLAG_GMR TTM_PL_FLAG_PRIV0
 
+#define VMW_RES_CONTEXT ttm_driver_type0
+#define VMW_RES_SURFACE ttm_driver_type1
+#define VMW_RES_STREAM ttm_driver_type2
+#define VMW_RES_FENCE ttm_driver_type3
+
 struct vmw_fpriv {
 	struct drm_master *locked_master;
 	struct ttm_object_file *tfile;
@@ -245,6 +251,7 @@ struct vmw_private {
 	atomic_t fifo_queue_waiters;
 	uint32_t last_read_seqno;
 	spinlock_t irq_lock;
+	struct vmw_fence_manager *fman;
 
 	/*
 	 * Device state
@@ -456,8 +463,6 @@ extern int vmw_irq_postinstall(struct drm_device *dev);
 extern void vmw_irq_uninstall(struct drm_device *dev);
 extern bool vmw_seqno_passed(struct vmw_private *dev_priv,
 				uint32_t seqno);
-extern int vmw_fence_wait_ioctl(struct drm_device *dev, void *data,
-				struct drm_file *file_priv);
 extern int vmw_fallback_wait(struct vmw_private *dev_priv,
 			     bool lazy,
 			     bool fifo_idle,
@@ -466,7 +471,8 @@ extern int vmw_fallback_wait(struct vmw_private *dev_priv,
 			     unsigned long timeout);
 extern void vmw_update_seqno(struct vmw_private *dev_priv,
 				struct vmw_fifo_state *fifo_state);
-
+extern void vmw_seqno_waiter_add(struct vmw_private *dev_priv);
+extern void vmw_seqno_waiter_remove(struct vmw_private *dev_priv);
 
 /**
  * Rudimentary fence-like objects currently used only for throttling -
@@ -572,4 +578,8 @@ static inline struct vmw_dma_buffer *vmw_dmabuf_reference(struct vmw_dma_buffer
 	return NULL;
 }
 
+static inline struct ttm_mem_global *vmw_mem_glob(struct vmw_private *dev_priv)
+{
+	return (struct ttm_mem_global *) dev_priv->mem_global_ref.object;
+}
 #endif

drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c

@@ -256,7 +256,7 @@ static int vmw_translate_guest_ptr(struct vmw_private *dev_priv,
 		val_buf = &sw_context->val_bufs[cur_validate_node];
 		val_buf->bo = ttm_bo_reference(bo);
 		val_buf->usage = TTM_USAGE_READWRITE;
-		val_buf->new_sync_obj_arg = (void *) dev_priv;
+		val_buf->new_sync_obj_arg = (void *) DRM_VMW_FENCE_FLAG_EXEC;
 		list_add_tail(&val_buf->head, &sw_context->validate_nodes);
 		++sw_context->cur_val_buf;
 	}
@@ -321,7 +321,6 @@ static int vmw_cmd_wait_query(struct vmw_private *dev_priv,
 	return 0;
 }
 
-
 static int vmw_cmd_dma(struct vmw_private *dev_priv,
 		       struct vmw_sw_context *sw_context,
 		       SVGA3dCmdHeader *header)
@@ -676,6 +675,50 @@ static int vmw_resize_cmd_bounce(struct vmw_sw_context *sw_context,
 	return 0;
 }
 
+/**
+ * vmw_execbuf_fence_commands - create and submit a command stream fence
+ *
+ * Creates a fence object and submits a command stream marker.
+ * If this fails for some reason, We sync the fifo and return NULL.
+ * It is then safe to fence buffers with a NULL pointer.
+ */
+
+int vmw_execbuf_fence_commands(struct drm_file *file_priv,
+			       struct vmw_private *dev_priv,
+			       struct vmw_fence_obj **p_fence,
+			       uint32_t *p_handle)
+{
+	uint32_t sequence;
+	int ret;
+	bool synced = false;
+
+
+	ret = vmw_fifo_send_fence(dev_priv, &sequence);
+	if (unlikely(ret != 0)) {
+		DRM_ERROR("Fence submission error. Syncing.\n");
+		synced = true;
+	}
+
+	if (p_handle != NULL)
+		ret = vmw_user_fence_create(file_priv, dev_priv->fman,
+					    sequence,
+					    DRM_VMW_FENCE_FLAG_EXEC,
+					    p_fence, p_handle);
+	else
+		ret = vmw_fence_create(dev_priv->fman, sequence,
+				       DRM_VMW_FENCE_FLAG_EXEC,
+				       p_fence);
+
+	if (unlikely(ret != 0 && !synced)) {
+		(void) vmw_fallback_wait(dev_priv, false, false,
+					 sequence, false,
+					 VMW_FENCE_WAIT_TIMEOUT);
+		*p_fence = NULL;
+	}
+
+	return 0;
+}
+
 int vmw_execbuf_ioctl(struct drm_device *dev, void *data,
 		      struct drm_file *file_priv)
 {
@@ -686,9 +729,10 @@ int vmw_execbuf_ioctl(struct drm_device *dev, void *data,
 	int ret;
 	void *user_cmd;
 	void *cmd;
-	uint32_t seqno;
 	struct vmw_sw_context *sw_context = &dev_priv->ctx;
 	struct vmw_master *vmaster = vmw_master(file_priv->master);
+	struct vmw_fence_obj *fence;
+	uint32_t handle;
 
 	ret = ttm_read_lock(&vmaster->lock, true);
 	if (unlikely(ret != 0))
@@ -755,34 +799,60 @@ int vmw_execbuf_ioctl(struct drm_device *dev, void *data,
 	memcpy(cmd, sw_context->cmd_bounce, arg->command_size);
 	vmw_fifo_commit(dev_priv, arg->command_size);
 
-	ret = vmw_fifo_send_fence(dev_priv, &seqno);
-
-	ttm_eu_fence_buffer_objects(&sw_context->validate_nodes,
-				    (void *)(unsigned long) seqno);
-	vmw_clear_validations(sw_context);
-	mutex_unlock(&dev_priv->cmdbuf_mutex);
-
+	user_fence_rep = (struct drm_vmw_fence_rep __user *)
+		(unsigned long)arg->fence_rep;
+	ret = vmw_execbuf_fence_commands(file_priv, dev_priv,
+					 &fence,
+					 (user_fence_rep) ? &handle : NULL);
 	/*
 	 * This error is harmless, because if fence submission fails,
-	 * vmw_fifo_send_fence will sync.
+	 * vmw_fifo_send_fence will sync. The error will be propagated to
+	 * user-space in @fence_rep
 	 */
 
 	if (ret != 0)
 		DRM_ERROR("Fence submission error. Syncing.\n");
 
-	fence_rep.error = ret;
-	fence_rep.fence_seq = (uint64_t) seqno;
-	fence_rep.pad64 = 0;
+	ttm_eu_fence_buffer_objects(&sw_context->validate_nodes,
+				    (void *) fence);
 
-	user_fence_rep = (struct drm_vmw_fence_rep __user *)
-	    (unsigned long)arg->fence_rep;
+	vmw_clear_validations(sw_context);
+	mutex_unlock(&dev_priv->cmdbuf_mutex);
 
-	/*
-	 * copy_to_user errors will be detected by user space not
-	 * seeing fence_rep::error filled in.
-	 */
+	if (user_fence_rep) {
+		fence_rep.error = ret;
+		fence_rep.handle = handle;
+		fence_rep.seqno = fence->seqno;
+		vmw_update_seqno(dev_priv, &dev_priv->fifo);
+		fence_rep.passed_seqno = dev_priv->last_read_seqno;
+
+		/*
+		 * copy_to_user errors will be detected by user space not
+		 * seeing fence_rep::error filled in. Typically
+		 * user-space would have pre-set that member to -EFAULT.
+		 */
+		ret = copy_to_user(user_fence_rep, &fence_rep,
+				   sizeof(fence_rep));
+
+		/*
+		 * User-space lost the fence object. We need to sync
+		 * and unreference the handle.
+		 */
+		if (unlikely(ret != 0) && (fence_rep.error == 0)) {
+			BUG_ON(fence == NULL);
+
+			ttm_ref_object_base_unref(vmw_fpriv(file_priv)->tfile,
+						  handle, TTM_REF_USAGE);
+			DRM_ERROR("Fence copy error. Syncing.\n");
+			(void) vmw_fence_obj_wait(fence,
+						  fence->signal_mask,
+						  false, false,
+						  VMW_FENCE_WAIT_TIMEOUT);
+		}
+	}
 
-	ret = copy_to_user(user_fence_rep, &fence_rep, sizeof(fence_rep));
+	if (likely(fence != NULL))
+		vmw_fence_obj_unreference(&fence);
 
 	vmw_kms_cursor_post_execbuf(dev_priv);
 	ttm_read_unlock(&vmaster->lock);