vmwgfx: Implement a proper GMR eviction mechanism

Use Ben's new range manager hooks to implement a manager for
GMRs that manages ids rather than ranges.
This means we can use the standard TTM code for binding, unbinding and
eviction.

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

drivers/gpu/drm/vmwgfx/Makefile

@@ -4,6 +4,6 @@ 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_fence.o
+	    vmwgfx_overlay.o vmwgfx_fence.o vmwgfx_gmrid_manager.o
 
 obj-$(CONFIG_DRM_VMWGFX) := vmwgfx.o

drivers/gpu/drm/vmwgfx/vmwgfx_buffer.c

@@ -39,6 +39,9 @@ static uint32_t vram_ne_placement_flags = TTM_PL_FLAG_VRAM |
 static uint32_t sys_placement_flags = TTM_PL_FLAG_SYSTEM |
 	TTM_PL_FLAG_CACHED;
 
+static uint32_t gmr_placement_flags = VMW_PL_FLAG_GMR |
+	TTM_PL_FLAG_CACHED;
+
 struct ttm_placement vmw_vram_placement = {
 	.fpfn = 0,
 	.lpfn = 0,
@@ -48,6 +51,20 @@ struct ttm_placement vmw_vram_placement = {
 	.busy_placement = &vram_placement_flags
 };
 
+static uint32_t vram_gmr_placement_flags[] = {
+	TTM_PL_FLAG_VRAM | TTM_PL_FLAG_CACHED,
+	VMW_PL_FLAG_GMR | TTM_PL_FLAG_CACHED
+};
+
+struct ttm_placement vmw_vram_gmr_placement = {
+	.fpfn = 0,
+	.lpfn = 0,
+	.num_placement = 2,
+	.placement = vram_gmr_placement_flags,
+	.num_busy_placement = 1,
+	.busy_placement = &gmr_placement_flags
+};
+
 struct ttm_placement vmw_vram_sys_placement = {
 	.fpfn = 0,
 	.lpfn = 0,
@@ -77,27 +94,52 @@ struct ttm_placement vmw_sys_placement = {
 
 struct vmw_ttm_backend {
 	struct ttm_backend backend;
+	struct page **pages;
+	unsigned long num_pages;
+	struct vmw_private *dev_priv;
+	int gmr_id;
 };
 
 static int vmw_ttm_populate(struct ttm_backend *backend,
 			    unsigned long num_pages, struct page **pages,
 			    struct page *dummy_read_page)
 {
+	struct vmw_ttm_backend *vmw_be =
+	    container_of(backend, struct vmw_ttm_backend, backend);
+
+	vmw_be->pages = pages;
+	vmw_be->num_pages = num_pages;
+
 	return 0;
 }
 
 static int vmw_ttm_bind(struct ttm_backend *backend, struct ttm_mem_reg *bo_mem)
 {
-	return 0;
+	struct vmw_ttm_backend *vmw_be =
+	    container_of(backend, struct vmw_ttm_backend, backend);
+
+	vmw_be->gmr_id = bo_mem->start;
+
+	return vmw_gmr_bind(vmw_be->dev_priv, vmw_be->pages,
+			    vmw_be->num_pages, vmw_be->gmr_id);
 }
 
 static int vmw_ttm_unbind(struct ttm_backend *backend)
 {
+	struct vmw_ttm_backend *vmw_be =
+	    container_of(backend, struct vmw_ttm_backend, backend);
+
+	vmw_gmr_unbind(vmw_be->dev_priv, vmw_be->gmr_id);
 	return 0;
 }
 
 static void vmw_ttm_clear(struct ttm_backend *backend)
 {
+	struct vmw_ttm_backend *vmw_be =
+		container_of(backend, struct vmw_ttm_backend, backend);
+
+	vmw_be->pages = NULL;
+	vmw_be->num_pages = 0;
 }
 
 static void vmw_ttm_destroy(struct ttm_backend *backend)
@@ -125,6 +167,7 @@ struct ttm_backend *vmw_ttm_backend_init(struct ttm_bo_device *bdev)
 		return NULL;
 
 	vmw_be->backend.func = &vmw_ttm_func;
+	vmw_be->dev_priv = container_of(bdev, struct vmw_private, bdev);
 
 	return &vmw_be->backend;
 }
@@ -142,16 +185,28 @@ int vmw_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
 		/* System memory */
 
 		man->flags = TTM_MEMTYPE_FLAG_MAPPABLE;
-		man->available_caching = TTM_PL_MASK_CACHING;
+		man->available_caching = TTM_PL_FLAG_CACHED;
 		man->default_caching = TTM_PL_FLAG_CACHED;
 		break;
 	case TTM_PL_VRAM:
 		/* "On-card" video ram */
 		man->func = &ttm_bo_manager_func;
 		man->gpu_offset = 0;
 		man->flags = TTM_MEMTYPE_FLAG_FIXED | TTM_MEMTYPE_FLAG_MAPPABLE;
-		man->available_caching = TTM_PL_MASK_CACHING;
-		man->default_caching = TTM_PL_FLAG_WC;
+		man->available_caching = TTM_PL_FLAG_CACHED;
+		man->default_caching = TTM_PL_FLAG_CACHED;
+		break;
+	case VMW_PL_GMR:
+		/*
+		 * "Guest Memory Regions" is an aperture like feature with
+		 *  one slot per bo. There is an upper limit of the number of
+		 *  slots as well as the bo size.
+		 */
+		man->func = &vmw_gmrid_manager_func;
+		man->gpu_offset = 0;
+		man->flags = TTM_MEMTYPE_FLAG_CMA | TTM_MEMTYPE_FLAG_MAPPABLE;
+		man->available_caching = TTM_PL_FLAG_CACHED;
+		man->default_caching = TTM_PL_FLAG_CACHED;
 		break;
 	default:
 		DRM_ERROR("Unsupported memory type %u\n", (unsigned)type);
@@ -175,18 +230,6 @@ static int vmw_verify_access(struct ttm_buffer_object *bo, struct file *filp)
 	return 0;
 }
 
-static void vmw_move_notify(struct ttm_buffer_object *bo,
-		     struct ttm_mem_reg *new_mem)
-{
-	if (new_mem->mem_type != TTM_PL_SYSTEM)
-		vmw_dmabuf_gmr_unbind(bo);
-}
-
-static void vmw_swap_notify(struct ttm_buffer_object *bo)
-{
-	vmw_dmabuf_gmr_unbind(bo);
-}
-
 static int vmw_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg *mem)
 {
 	struct ttm_mem_type_manager *man = &bdev->man[mem->mem_type];
@@ -201,7 +244,7 @@ static int vmw_ttm_io_mem_reserve(struct ttm_bo_device *bdev, struct ttm_mem_reg
 		return -EINVAL;
 	switch (mem->mem_type) {
 	case TTM_PL_SYSTEM:
-		/* System memory */
+	case VMW_PL_GMR:
 		return 0;
 	case TTM_PL_VRAM:
 		mem->bus.offset = mem->start << PAGE_SHIFT;
@@ -277,8 +320,8 @@ struct ttm_bo_driver vmw_bo_driver = {
 	.sync_obj_flush = vmw_sync_obj_flush,
 	.sync_obj_unref = vmw_sync_obj_unref,
 	.sync_obj_ref = vmw_sync_obj_ref,
-	.move_notify = vmw_move_notify,
-	.swap_notify = vmw_swap_notify,
+	.move_notify = NULL,
+	.swap_notify = NULL,
 	.fault_reserve_notify = &vmw_ttm_fault_reserve_notify,
 	.io_mem_reserve = &vmw_ttm_io_mem_reserve,
 	.io_mem_free = &vmw_ttm_io_mem_free,

drivers/gpu/drm/vmwgfx/vmwgfx_drv.c

@@ -260,13 +260,11 @@ static int vmw_driver_load(struct drm_device *dev, unsigned long chipset)
 	idr_init(&dev_priv->context_idr);
 	idr_init(&dev_priv->surface_idr);
 	idr_init(&dev_priv->stream_idr);
-	ida_init(&dev_priv->gmr_ida);
 	mutex_init(&dev_priv->init_mutex);
 	init_waitqueue_head(&dev_priv->fence_queue);
 	init_waitqueue_head(&dev_priv->fifo_queue);
 	atomic_set(&dev_priv->fence_queue_waiters, 0);
 	atomic_set(&dev_priv->fifo_queue_waiters, 0);
-	INIT_LIST_HEAD(&dev_priv->gmr_lru);
 
 	dev_priv->io_start = pci_resource_start(dev->pdev, 0);
 	dev_priv->vram_start = pci_resource_start(dev->pdev, 1);
@@ -341,6 +339,14 @@ static int vmw_driver_load(struct drm_device *dev, unsigned long chipset)
 		goto out_err2;
 	}
 
+	dev_priv->has_gmr = true;
+	if (ttm_bo_init_mm(&dev_priv->bdev, VMW_PL_GMR,
+			   dev_priv->max_gmr_ids) != 0) {
+		DRM_INFO("No GMR memory available. "
+			 "Graphics memory resources are very limited.\n");
+		dev_priv->has_gmr = false;
+	}
+
 	dev_priv->mmio_mtrr = drm_mtrr_add(dev_priv->mmio_start,
 					   dev_priv->mmio_size, DRM_MTRR_WC);
 
@@ -440,13 +446,14 @@ static int vmw_driver_load(struct drm_device *dev, unsigned long chipset)
 out_err3:
 	drm_mtrr_del(dev_priv->mmio_mtrr, dev_priv->mmio_start,
 		     dev_priv->mmio_size, DRM_MTRR_WC);
+	if (dev_priv->has_gmr)
+		(void) ttm_bo_clean_mm(&dev_priv->bdev, VMW_PL_GMR);
 	(void)ttm_bo_clean_mm(&dev_priv->bdev, TTM_PL_VRAM);
 out_err2:
 	(void)ttm_bo_device_release(&dev_priv->bdev);
 out_err1:
 	vmw_ttm_global_release(dev_priv);
 out_err0:
-	ida_destroy(&dev_priv->gmr_ida);
 	idr_destroy(&dev_priv->surface_idr);
 	idr_destroy(&dev_priv->context_idr);
 	idr_destroy(&dev_priv->stream_idr);
@@ -478,10 +485,11 @@ static int vmw_driver_unload(struct drm_device *dev)
 	iounmap(dev_priv->mmio_virt);
 	drm_mtrr_del(dev_priv->mmio_mtrr, dev_priv->mmio_start,
 		     dev_priv->mmio_size, DRM_MTRR_WC);
+	if (dev_priv->has_gmr)
+		(void)ttm_bo_clean_mm(&dev_priv->bdev, VMW_PL_GMR);
 	(void)ttm_bo_clean_mm(&dev_priv->bdev, TTM_PL_VRAM);
 	(void)ttm_bo_device_release(&dev_priv->bdev);
 	vmw_ttm_global_release(dev_priv);
-	ida_destroy(&dev_priv->gmr_ida);
 	idr_destroy(&dev_priv->surface_idr);
 	idr_destroy(&dev_priv->context_idr);
 	idr_destroy(&dev_priv->stream_idr);

drivers/gpu/drm/vmwgfx/vmwgfx_drv.h

@@ -49,6 +49,9 @@
 #define VMWGFX_MAX_GMRS 2048
 #define VMWGFX_MAX_DISPLAYS 16
 
+#define VMW_PL_GMR TTM_PL_PRIV0
+#define VMW_PL_FLAG_GMR TTM_PL_FLAG_PRIV0
+
 struct vmw_fpriv {
 	struct drm_master *locked_master;
 	struct ttm_object_file *tfile;
@@ -57,8 +60,6 @@ struct vmw_fpriv {
 struct vmw_dma_buffer {
 	struct ttm_buffer_object base;
 	struct list_head validate_list;
-	struct list_head gmr_lru;
-	uint32_t gmr_id;
 	bool gmr_bound;
 	uint32_t cur_validate_node;
 	bool on_validate_list;
@@ -184,6 +185,7 @@ struct vmw_private {
 	uint32_t capabilities;
 	uint32_t max_gmr_descriptors;
 	uint32_t max_gmr_ids;
+	bool has_gmr;
 	struct mutex hw_mutex;
 
 	/*
@@ -265,14 +267,6 @@ struct vmw_private {
 	uint32_t val_seq;
 	struct mutex cmdbuf_mutex;
 
-	/**
-	 * GMR management. Protected by the lru spinlock.
-	 */
-
-	struct ida gmr_ida;
-	struct list_head gmr_lru;
-
-
 	/**
 	 * Operating mode.
 	 */
@@ -334,7 +328,9 @@ void vmw_3d_resource_dec(struct vmw_private *dev_priv);
  */
 
 extern int vmw_gmr_bind(struct vmw_private *dev_priv,
-			struct ttm_buffer_object *bo);
+			struct page *pages[],
+			unsigned long num_pages,
+			int gmr_id);
 extern void vmw_gmr_unbind(struct vmw_private *dev_priv, int gmr_id);
 
 /**
@@ -383,14 +379,10 @@ extern uint32_t vmw_dmabuf_validate_node(struct ttm_buffer_object *bo,
 extern void vmw_dmabuf_validate_clear(struct ttm_buffer_object *bo);
 extern int vmw_user_dmabuf_lookup(struct ttm_object_file *tfile,
 				  uint32_t id, struct vmw_dma_buffer **out);
-extern uint32_t vmw_dmabuf_gmr(struct ttm_buffer_object *bo);
-extern void vmw_dmabuf_set_gmr(struct ttm_buffer_object *bo, uint32_t id);
-extern int vmw_gmr_id_alloc(struct vmw_private *dev_priv, uint32_t *p_id);
 extern int vmw_dmabuf_to_start_of_vram(struct vmw_private *vmw_priv,
 				       struct vmw_dma_buffer *bo);
 extern int vmw_dmabuf_from_vram(struct vmw_private *vmw_priv,
 				struct vmw_dma_buffer *bo);
-extern void vmw_dmabuf_gmr_unbind(struct ttm_buffer_object *bo);
 extern int vmw_stream_claim_ioctl(struct drm_device *dev, void *data,
 				  struct drm_file *file_priv);
 extern int vmw_stream_unref_ioctl(struct drm_device *dev, void *data,
@@ -442,6 +434,7 @@ extern int vmw_mmap(struct file *filp, struct vm_area_struct *vma);
 extern struct ttm_placement vmw_vram_placement;
 extern struct ttm_placement vmw_vram_ne_placement;
 extern struct ttm_placement vmw_vram_sys_placement;
+extern struct ttm_placement vmw_vram_gmr_placement;
 extern struct ttm_placement vmw_sys_placement;
 extern struct ttm_bo_driver vmw_bo_driver;
 extern int vmw_dma_quiescent(struct drm_device *dev);
@@ -543,6 +536,12 @@ int vmw_overlay_unref(struct vmw_private *dev_priv, uint32_t stream_id);
 int vmw_overlay_num_overlays(struct vmw_private *dev_priv);
 int vmw_overlay_num_free_overlays(struct vmw_private *dev_priv);
 
+/**
+ * GMR Id manager
+ */
+
+extern const struct ttm_mem_type_manager_func vmw_gmrid_manager_func;
+
 /**
  * Inline helper functions
  */

drivers/gpu/drm/vmwgfx/vmwgfx_execbuf.c

@@ -538,8 +538,11 @@ static void vmw_apply_relocations(struct vmw_sw_context *sw_context)
 		reloc = &sw_context->relocs[i];
 		validate = &sw_context->val_bufs[reloc->index];
 		bo = validate->bo;
-		reloc->location->offset += bo->offset;
-		reloc->location->gmrId = vmw_dmabuf_gmr(bo);
+		if (bo->mem.mem_type == TTM_PL_VRAM) {
+			reloc->location->offset += bo->offset;
+			reloc->location->gmrId = SVGA_GMR_FRAMEBUFFER;
+		} else
+			reloc->location->gmrId = bo->mem.start;
 	}
 	vmw_free_relocations(sw_context);
 }
@@ -563,25 +566,14 @@ static int vmw_validate_single_buffer(struct vmw_private *dev_priv,
 {
 	int ret;
 
-	if (vmw_dmabuf_gmr(bo) != SVGA_GMR_NULL)
-		return 0;
-
 	/**
-	 * Put BO in VRAM, only if there is space.
+	 * Put BO in VRAM if there is space, otherwise as a GMR.
+	 * If there is no space in VRAM and GMR ids are all used up,
+	 * start evicting GMRs to make room. If the DMA buffer can't be
+	 * used as a GMR, this will return -ENOMEM.
 	 */
 
-	ret = ttm_bo_validate(bo, &vmw_vram_sys_placement, true, false, false);
-	if (unlikely(ret == -ERESTARTSYS))
-		return ret;
-
-	/**
-	 * Otherwise, set it up as GMR.
-	 */
-
-	if (vmw_dmabuf_gmr(bo) != SVGA_GMR_NULL)
-		return 0;
-
-	ret = vmw_gmr_bind(dev_priv, bo);
+	ret = ttm_bo_validate(bo, &vmw_vram_gmr_placement, true, false, false);
 	if (likely(ret == 0 || ret == -ERESTARTSYS))
 		return ret;
 
@@ -590,6 +582,7 @@ static int vmw_validate_single_buffer(struct vmw_private *dev_priv,
 	 * previous contents.
 	 */
 
+	DRM_INFO("Falling through to VRAM.\n");
 	ret = ttm_bo_validate(bo, &vmw_vram_placement, true, false, false);
 	return ret;
 }

drivers/gpu/drm/vmwgfx/vmwgfx_fb.c

@@ -612,7 +612,6 @@ int vmw_dmabuf_to_start_of_vram(struct vmw_private *vmw_priv,
 {
 	struct ttm_buffer_object *bo = &vmw_bo->base;
 	struct ttm_placement ne_placement = vmw_vram_ne_placement;
-	struct drm_mm_node *mm_node;
 	int ret = 0;
 
 	ne_placement.lpfn = bo->num_pages;
@@ -626,9 +625,9 @@ int vmw_dmabuf_to_start_of_vram(struct vmw_private *vmw_priv,
 	if (unlikely(ret != 0))
 		goto err_unlock;
 
-	mm_node = bo->mem.mm_node;
 	if (bo->mem.mem_type == TTM_PL_VRAM &&
-	    mm_node->start < bo->num_pages)
+	    bo->mem.start < bo->num_pages &&
+	    bo->mem.start > 0)
 		(void) ttm_bo_validate(bo, &vmw_sys_placement, false,
 				       false, false);