drm/vmwgfx: Implement a buffer object synccpu ioctl.

This ioctl enables inter-process synchronization of buffer objects,
which is needed for mesa Guest-Backed objects.

Signed-off-by: Thomas Hellstrom <thellstrom@vmware.com>
Reviewed-by: Brian Paul <brianp@vmware.com>

4 files changed, 211 insertions, 2 deletions

diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c
index cfeaa478dcd..fb56676ed3e 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.c
@@ -124,6 +124,9 @@
 #define DRM_IOCTL_VMW_GB_SURFACE_REF				\
 	DRM_IOWR(DRM_COMMAND_BASE + DRM_VMW_GB_SURFACE_REF,	\
 		 union drm_vmw_gb_surface_reference_arg)
+#define DRM_IOCTL_VMW_SYNCCPU					\
+	DRM_IOW(DRM_COMMAND_BASE + DRM_VMW_SYNCCPU,		\
+		 struct drm_vmw_synccpu_arg)
 
 /**
  * The core DRM version of this macro doesn't account for
@@ -201,6 +204,9 @@ static const struct drm_ioctl_desc vmw_ioctls[] = {
 	VMW_IOCTL_DEF(VMW_GB_SURFACE_REF,
 		      vmw_gb_surface_reference_ioctl,
 		      DRM_AUTH | DRM_UNLOCKED),
+	VMW_IOCTL_DEF(VMW_SYNCCPU,
+		      vmw_user_dmabuf_synccpu_ioctl,
+		      DRM_AUTH | DRM_UNLOCKED),
 };
 
 static struct pci_device_id vmw_pci_id_list[] = {
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
index e0859eebe7b..fe3c2e3e1cb 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_drv.h
@@ -533,6 +533,8 @@ extern int vmw_dmabuf_alloc_ioctl(struct drm_device *dev, void *data,
 				  struct drm_file *file_priv);
 extern int vmw_dmabuf_unref_ioctl(struct drm_device *dev, void *data,
 				  struct drm_file *file_priv);
+extern int vmw_user_dmabuf_synccpu_ioctl(struct drm_device *dev, void *data,
+					 struct drm_file *file_priv);
 extern uint32_t vmw_dmabuf_validate_node(struct ttm_buffer_object *bo,
 					 uint32_t cur_validate_node);
 extern void vmw_dmabuf_validate_clear(struct ttm_buffer_object *bo);
diff --git a/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c b/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
index b40978f0ca9..12e68e58d9e 100644
--- a/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
+++ b/drivers/gpu/drm/vmwgfx/vmwgfx_resource.c
@@ -441,6 +441,21 @@ static void vmw_user_dmabuf_release(struct ttm_base_object **p_base)
 	ttm_bo_unref(&bo);
 }
 
+static void vmw_user_dmabuf_ref_obj_release(struct ttm_base_object *base,
+					    enum ttm_ref_type ref_type)
+{
+	struct vmw_user_dma_buffer *user_bo;
+	user_bo = container_of(base, struct vmw_user_dma_buffer, prime.base);
+
+	switch (ref_type) {
+	case TTM_REF_SYNCCPU_WRITE:
+		ttm_bo_synccpu_write_release(&user_bo->dma.base);
+		break;
+	default:
+		BUG();
+	}
+}
+
 /**
  * vmw_user_dmabuf_alloc - Allocate a user dma buffer
  *
@@ -484,7 +499,8 @@ int vmw_user_dmabuf_alloc(struct vmw_private *dev_priv,
 				    &user_bo->prime,
 				    shareable,
 				    ttm_buffer_type,
-				    &vmw_user_dmabuf_release, NULL);
+				    &vmw_user_dmabuf_release,
+				    &vmw_user_dmabuf_ref_obj_release);
 	if (unlikely(ret != 0)) {
 		ttm_bo_unref(&tmp);
 		goto out_no_base_object;
@@ -517,6 +533,130 @@ int vmw_user_dmabuf_verify_access(struct ttm_buffer_object *bo,
 		vmw_user_bo->prime.base.shareable) ? 0 : -EPERM;
 }
 
+/**
+ * vmw_user_dmabuf_synccpu_grab - Grab a struct vmw_user_dma_buffer for cpu
+ * access, idling previous GPU operations on the buffer and optionally
+ * blocking it for further command submissions.
+ *
+ * @user_bo: Pointer to the buffer object being grabbed for CPU access
+ * @tfile: Identifying the caller.
+ * @flags: Flags indicating how the grab should be performed.
+ *
+ * A blocking grab will be automatically released when @tfile is closed.
+ */
+static int vmw_user_dmabuf_synccpu_grab(struct vmw_user_dma_buffer *user_bo,
+					struct ttm_object_file *tfile,
+					uint32_t flags)
+{
+	struct ttm_buffer_object *bo = &user_bo->dma.base;
+	bool existed;
+	int ret;
+
+	if (flags & drm_vmw_synccpu_allow_cs) {
+		struct ttm_bo_device *bdev = bo->bdev;
+
+		spin_lock(&bdev->fence_lock);
+		ret = ttm_bo_wait(bo, false, true,
+				  !!(flags & drm_vmw_synccpu_dontblock));
+		spin_unlock(&bdev->fence_lock);
+		return ret;
+	}
+
+	ret = ttm_bo_synccpu_write_grab
+		(bo, !!(flags & drm_vmw_synccpu_dontblock));
+	if (unlikely(ret != 0))
+		return ret;
+
+	ret = ttm_ref_object_add(tfile, &user_bo->prime.base,
+				 TTM_REF_SYNCCPU_WRITE, &existed);
+	if (ret != 0 || existed)
+		ttm_bo_synccpu_write_release(&user_bo->dma.base);
+
+	return ret;
+}
+
+/**
+ * vmw_user_dmabuf_synccpu_release - Release a previous grab for CPU access,
+ * and unblock command submission on the buffer if blocked.
+ *
+ * @handle: Handle identifying the buffer object.
+ * @tfile: Identifying the caller.
+ * @flags: Flags indicating the type of release.
+ */
+static int vmw_user_dmabuf_synccpu_release(uint32_t handle,
+					   struct ttm_object_file *tfile,
+					   uint32_t flags)
+{
+	if (!(flags & drm_vmw_synccpu_allow_cs))
+		return ttm_ref_object_base_unref(tfile, handle,
+						 TTM_REF_SYNCCPU_WRITE);
+
+	return 0;
+}
+
+/**
+ * vmw_user_dmabuf_synccpu_release - ioctl function implementing the synccpu
+ * functionality.
+ *
+ * @dev: Identifies the drm device.
+ * @data: Pointer to the ioctl argument.
+ * @file_priv: Identifies the caller.
+ *
+ * This function checks the ioctl arguments for validity and calls the
+ * relevant synccpu functions.
+ */
+int vmw_user_dmabuf_synccpu_ioctl(struct drm_device *dev, void *data,
+				  struct drm_file *file_priv)
+{
+	struct drm_vmw_synccpu_arg *arg =
+		(struct drm_vmw_synccpu_arg *) data;
+	struct vmw_dma_buffer *dma_buf;
+	struct vmw_user_dma_buffer *user_bo;
+	struct ttm_object_file *tfile = vmw_fpriv(file_priv)->tfile;
+	int ret;
+
+	if ((arg->flags & (drm_vmw_synccpu_read | drm_vmw_synccpu_write)) == 0
+	    || (arg->flags & ~(drm_vmw_synccpu_read | drm_vmw_synccpu_write |
+			       drm_vmw_synccpu_dontblock |
+			       drm_vmw_synccpu_allow_cs)) != 0) {
+		DRM_ERROR("Illegal synccpu flags.\n");
+		return -EINVAL;
+	}
+
+	switch (arg->op) {
+	case drm_vmw_synccpu_grab:
+		ret = vmw_user_dmabuf_lookup(tfile, arg->handle, &dma_buf);
+		if (unlikely(ret != 0))
+			return ret;
+
+		user_bo = container_of(dma_buf, struct vmw_user_dma_buffer,
+				       dma);
+		ret = vmw_user_dmabuf_synccpu_grab(user_bo, tfile, arg->flags);
+		vmw_dmabuf_unreference(&dma_buf);
+		if (unlikely(ret != 0 && ret != -ERESTARTSYS &&
+			     ret != -EBUSY)) {
+			DRM_ERROR("Failed synccpu grab on handle 0x%08x.\n",
+				  (unsigned int) arg->handle);
+			return ret;
+		}
+		break;
+	case drm_vmw_synccpu_release:
+		ret = vmw_user_dmabuf_synccpu_release(arg->handle, tfile,
+						      arg->flags);
+		if (unlikely(ret != 0)) {
+			DRM_ERROR("Failed synccpu release on handle 0x%08x.\n",
+				  (unsigned int) arg->handle);
+			return ret;
+		}
+		break;
+	default:
+		DRM_ERROR("Invalid synccpu operation.\n");
+		return -EINVAL;
+	}
+
+	return 0;
+}
+
 int vmw_dmabuf_alloc_ioctl(struct drm_device *dev, void *data,
 			   struct drm_file *file_priv)
 {
diff --git a/include/uapi/drm/vmwgfx_drm.h b/include/uapi/drm/vmwgfx_drm.h
index bb3b91d8491..adb7e0d0d3b 100644
--- a/include/uapi/drm/vmwgfx_drm.h
+++ b/include/uapi/drm/vmwgfx_drm.h
@@ -28,6 +28,10 @@
 #ifndef __VMWGFX_DRM_H__
 #define __VMWGFX_DRM_H__
 
+#ifndef __KERNEL__
+#include <drm.h>
+#endif
+
 #define DRM_VMW_MAX_SURFACE_FACES 6
 #define DRM_VMW_MAX_MIP_LEVELS 24
 
@@ -59,7 +63,7 @@
 #define DRM_VMW_UNREF_SHADER         22
 #define DRM_VMW_GB_SURFACE_CREATE    23
 #define DRM_VMW_GB_SURFACE_REF       24
-
+#define DRM_VMW_SYNCCPU              25
 
 /*************************************************************************/
 /**
@@ -985,5 +989,62 @@ union drm_vmw_gb_surface_reference_arg {
 };
 
 
+/*************************************************************************/
+/**
+ * DRM_VMW_SYNCCPU - Sync a DMA buffer / MOB for CPU access.
+ *
+ * Idles any previously submitted GPU operations on the buffer and
+ * by default blocks command submissions that reference the buffer.
+ * If the file descriptor used to grab a blocking CPU sync is closed, the
+ * cpu sync is released.
+ * The flags argument indicates how the grab / release operation should be
+ * performed:
+ */
+
+/**
+ * enum drm_vmw_synccpu_flags - Synccpu flags:
+ *
+ * @drm_vmw_synccpu_read: Sync for read. If sync is done for read only, it's a
+ * hint to the kernel to allow command submissions that references the buffer
+ * for read-only.
+ * @drm_vmw_synccpu_write: Sync for write. Block all command submissions
+ * referencing this buffer.
+ * @drm_vmw_synccpu_dontblock: Dont wait for GPU idle, but rather return
+ * -EBUSY should the buffer be busy.
+ * @drm_vmw_synccpu_allow_cs: Allow command submission that touches the buffer
+ * while the buffer is synced for CPU. This is similar to the GEM bo idle
+ * behavior.
+ */
+enum drm_vmw_synccpu_flags {
+	drm_vmw_synccpu_read = (1 << 0),
+	drm_vmw_synccpu_write = (1 << 1),
+	drm_vmw_synccpu_dontblock = (1 << 2),
+	drm_vmw_synccpu_allow_cs = (1 << 3)
+};
+
+/**
+ * enum drm_vmw_synccpu_op - Synccpu operations:
+ *
+ * @drm_vmw_synccpu_grab:    Grab the buffer for CPU operations
+ * @drm_vmw_synccpu_release: Release a previous grab.
+ */
+enum drm_vmw_synccpu_op {
+	drm_vmw_synccpu_grab,
+	drm_vmw_synccpu_release
+};
+
+/**
+ * struct drm_vmw_synccpu_arg
+ *
+ * @op:			     The synccpu operation as described above.
+ * @handle:		     Handle identifying the buffer object.
+ * @flags:		     Flags as described above.
+ */
+struct drm_vmw_synccpu_arg {
+	enum drm_vmw_synccpu_op op;
+	enum drm_vmw_synccpu_flags flags;
+	uint32_t handle;
+	uint32_t pad64;
+};
 
 #endif