drm/nv50: tidy up PCIEGART implementation

Signed-off-by: Ben Skeggs <bskeggs@redhat.com>

5 files changed, 79 insertions, 95 deletions

diff --git a/drivers/gpu/drm/nouveau/nouveau_bo.c b/drivers/gpu/drm/nouveau/nouveau_bo.c
index bc28aea..42d1ad6 100644
--- a/drivers/gpu/drm/nouveau/nouveau_bo.c
+++ b/drivers/gpu/drm/nouveau/nouveau_bo.c
@@ -425,7 +425,6 @@ nouveau_bo_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
 		man->available_caching = TTM_PL_FLAG_UNCACHED |
 					 TTM_PL_FLAG_WC;
 		man->default_caching = TTM_PL_FLAG_WC;
-		man->gpu_offset = 0;
 		break;
 	case TTM_PL_TT:
 		man->func = &ttm_bo_manager_func;
@@ -441,13 +440,13 @@ nouveau_bo_init_mem_type(struct ttm_bo_device *bdev, uint32_t type,
 				     TTM_MEMTYPE_FLAG_CMA;
 			man->available_caching = TTM_PL_MASK_CACHING;
 			man->default_caching = TTM_PL_FLAG_CACHED;
+			man->gpu_offset = dev_priv->gart_info.aper_base;
 			break;
 		default:
 			NV_ERROR(dev, "Unknown GART type: %d\n",
 				 dev_priv->gart_info.type);
 			return -EINVAL;
 		}
-		man->gpu_offset = dev_priv->vm_gart_base;
 		break;
 	default:
 		NV_ERROR(dev, "Unsupported memory type %u\n", (unsigned)type);
@@ -531,12 +530,12 @@ nv50_bo_move_m2mf(struct nouveau_channel *chan, struct ttm_buffer_object *bo,
 		if (old_mem->mem_type == TTM_PL_VRAM)
 			src_offset  = nvbo->vma.offset;
 		else
-			src_offset += dev_priv->vm_gart_base;
+			src_offset += dev_priv->gart_info.aper_base;
 
 		if (new_mem->mem_type == TTM_PL_VRAM)
 			dst_offset  = nvbo->vma.offset;
 		else
-			dst_offset += dev_priv->vm_gart_base;
+			dst_offset += dev_priv->gart_info.aper_base;
 	}
 
 	ret = RING_SPACE(chan, 3);
diff --git a/drivers/gpu/drm/nouveau/nouveau_drv.h b/drivers/gpu/drm/nouveau/nouveau_drv.h
index dce9a5f..4c5fc9c 100644
--- a/drivers/gpu/drm/nouveau/nouveau_drv.h
+++ b/drivers/gpu/drm/nouveau/nouveau_drv.h
@@ -248,7 +248,6 @@ struct nouveau_channel {
 	/* NV50 VM */
 	struct nouveau_vm     *vm;
 	struct nouveau_gpuobj *vm_pd;
-	struct nouveau_gpuobj *vm_gart_pt;
 
 	/* Objects */
 	struct nouveau_gpuobj *ramin; /* Private instmem */
@@ -684,6 +683,7 @@ struct drm_nouveau_private {
 		uint64_t aper_free;
 
 		struct nouveau_gpuobj *sg_ctxdma;
+		struct nouveau_vma vma;
 	} gart_info;
 
 	/* nv10-nv40 tiling regions */
@@ -709,8 +709,6 @@ struct drm_nouveau_private {
 
 	/* G8x/G9x virtual address space */
 	struct nouveau_vm *chan_vm;
-	uint64_t vm_gart_base;
-	uint64_t vm_gart_size;
 
 	struct nvbios vbios;
 
diff --git a/drivers/gpu/drm/nouveau/nouveau_object.c b/drivers/gpu/drm/nouveau/nouveau_object.c
index 573fd73..d1bed40 100644
--- a/drivers/gpu/drm/nouveau/nouveau_object.c
+++ b/drivers/gpu/drm/nouveau/nouveau_object.c
@@ -433,7 +433,7 @@ nv50_gpuobj_dma_init(struct nouveau_gpuobj *obj, u32 offset, int class,
 		flags0 |= 0x00030000;
 		break;
 	case NV_MEM_TARGET_GART:
-		base += dev_priv->vm_gart_base;
+		base += dev_priv->gart_info.aper_base;
 	default:
 		flags0 &= ~0x00100000;
 		break;
@@ -801,7 +801,6 @@ nouveau_gpuobj_channel_init(struct nouveau_channel *chan,
 			return ret;
 
 		nouveau_vm_ref(dev_priv->chan_vm, &chan->vm, chan->vm_pd);
-		chan->vm->map_pgt(chan->vm_pd, 12, 1, dev_priv->gart_info.sg_ctxdma);
 	}
 
 	/* RAMHT */
@@ -889,7 +888,6 @@ nouveau_gpuobj_channel_takedown(struct nouveau_channel *chan)
 
 	nouveau_vm_ref(NULL, &chan->vm, chan->vm_pd);
 	nouveau_gpuobj_ref(NULL, &chan->vm_pd);
-	nouveau_gpuobj_ref(NULL, &chan->vm_gart_pt);
 
 	if (chan->ramin_heap.free_stack.next)
 		drm_mm_takedown(&chan->ramin_heap);
diff --git a/drivers/gpu/drm/nouveau/nouveau_sgdma.c b/drivers/gpu/drm/nouveau/nouveau_sgdma.c
index a0bf130..b57201a 100644
--- a/drivers/gpu/drm/nouveau/nouveau_sgdma.c
+++ b/drivers/gpu/drm/nouveau/nouveau_sgdma.c
@@ -14,7 +14,7 @@ struct nouveau_sgdma_be {
 	dma_addr_t *pages;
 	unsigned nr_pages;
 
-	unsigned pte_start;
+	u64 offset;
 	bool bound;
 };
 
@@ -74,18 +74,6 @@ nouveau_sgdma_clear(struct ttm_backend *be)
 	}
 }
 
-static inline unsigned
-nouveau_sgdma_pte(struct drm_device *dev, uint64_t offset)
-{
-	struct drm_nouveau_private *dev_priv = dev->dev_private;
-	unsigned pte = (offset >> NV_CTXDMA_PAGE_SHIFT);
-
-	if (dev_priv->card_type < NV_50)
-		return pte + 2;
-
-	return pte << 1;
-}
-
 static int
 nouveau_sgdma_bind(struct ttm_backend *be, struct ttm_mem_reg *mem)
 {
@@ -97,32 +85,17 @@ nouveau_sgdma_bind(struct ttm_backend *be, struct ttm_mem_reg *mem)
 
 	NV_DEBUG(dev, "pg=0x%lx\n", mem->start);
 
-	pte = nouveau_sgdma_pte(nvbe->dev, mem->start << PAGE_SHIFT);
-	nvbe->pte_start = pte;
+	nvbe->offset = mem->start << PAGE_SHIFT;
+	pte = (nvbe->offset >> NV_CTXDMA_PAGE_SHIFT) + 2;
 	for (i = 0; i < nvbe->nr_pages; i++) {
 		dma_addr_t dma_offset = nvbe->pages[i];
 		uint32_t offset_l = lower_32_bits(dma_offset);
-		uint32_t offset_h = upper_32_bits(dma_offset);
-
-		for (j = 0; j < PAGE_SIZE / NV_CTXDMA_PAGE_SIZE; j++) {
-			if (dev_priv->card_type < NV_50) {
-				nv_wo32(gpuobj, (pte * 4) + 0, offset_l | 3);
-				pte += 1;
-			} else {
-				nv_wo32(gpuobj, (pte * 4) + 0, offset_l | 0x21);
-				nv_wo32(gpuobj, (pte * 4) + 4, offset_h & 0xff);
-				pte += 2;
-			}
 
+		for (j = 0; j < PAGE_SIZE / NV_CTXDMA_PAGE_SIZE; j++, pte++) {
+			nv_wo32(gpuobj, (pte * 4) + 0, offset_l | 3);
 			dma_offset += NV_CTXDMA_PAGE_SIZE;
 		}
 	}
-	dev_priv->engine.instmem.flush(nvbe->dev);
-
-	if (dev_priv->card_type == NV_50) {
-		dev_priv->engine.fifo.tlb_flush(dev);
-		dev_priv->engine.graph.tlb_flush(dev);
-	}
 
 	nvbe->bound = true;
 	return 0;
@@ -142,24 +115,10 @@ nouveau_sgdma_unbind(struct ttm_backend *be)
 	if (!nvbe->bound)
 		return 0;
 
-	pte = nvbe->pte_start;
+	pte = (nvbe->offset >> NV_CTXDMA_PAGE_SHIFT) + 2;
 	for (i = 0; i < nvbe->nr_pages; i++) {
-		for (j = 0; j < PAGE_SIZE / NV_CTXDMA_PAGE_SIZE; j++) {
-			if (dev_priv->card_type < NV_50) {
-				nv_wo32(gpuobj, (pte * 4) + 0, 0x00000000);
-				pte += 1;
-			} else {
-				nv_wo32(gpuobj, (pte * 4) + 0, 0x00000000);
-				nv_wo32(gpuobj, (pte * 4) + 4, 0x00000000);
-				pte += 2;
-			}
-		}
-	}
-	dev_priv->engine.instmem.flush(nvbe->dev);
-
-	if (dev_priv->card_type == NV_50) {
-		dev_priv->engine.fifo.tlb_flush(dev);
-		dev_priv->engine.graph.tlb_flush(dev);
+		for (j = 0; j < PAGE_SIZE / NV_CTXDMA_PAGE_SIZE; j++, pte++)
+			nv_wo32(gpuobj, (pte * 4) + 0, 0x00000000);
 	}
 
 	nvbe->bound = false;
@@ -182,6 +141,35 @@ nouveau_sgdma_destroy(struct ttm_backend *be)
 	}
 }
 
+static int
+nv50_sgdma_bind(struct ttm_backend *be, struct ttm_mem_reg *mem)
+{
+	struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
+	struct drm_nouveau_private *dev_priv = nvbe->dev->dev_private;
+
+	nvbe->offset = mem->start << PAGE_SHIFT;
+
+	nouveau_vm_map_sg(&dev_priv->gart_info.vma, nvbe->offset,
+			  nvbe->nr_pages << PAGE_SHIFT, nvbe->pages);
+	nvbe->bound = true;
+	return 0;
+}
+
+static int
+nv50_sgdma_unbind(struct ttm_backend *be)
+{
+	struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
+	struct drm_nouveau_private *dev_priv = nvbe->dev->dev_private;
+
+	if (!nvbe->bound)
+		return 0;
+
+	nouveau_vm_unmap_at(&dev_priv->gart_info.vma, nvbe->offset,
+			    nvbe->nr_pages << PAGE_SHIFT);
+	nvbe->bound = false;
+	return 0;
+}
+
 static struct ttm_backend_func nouveau_sgdma_backend = {
 	.populate		= nouveau_sgdma_populate,
 	.clear			= nouveau_sgdma_clear,
@@ -190,23 +178,30 @@ static struct ttm_backend_func nouveau_sgdma_backend = {
 	.destroy		= nouveau_sgdma_destroy
 };
 
+static struct ttm_backend_func nv50_sgdma_backend = {
+	.populate		= nouveau_sgdma_populate,
+	.clear			= nouveau_sgdma_clear,
+	.bind			= nv50_sgdma_bind,
+	.unbind			= nv50_sgdma_unbind,
+	.destroy		= nouveau_sgdma_destroy
+};
+
 struct ttm_backend *
 nouveau_sgdma_init_ttm(struct drm_device *dev)
 {
 	struct drm_nouveau_private *dev_priv = dev->dev_private;
 	struct nouveau_sgdma_be *nvbe;
 
-	if (!dev_priv->gart_info.sg_ctxdma)
-		return NULL;
-
 	nvbe = kzalloc(sizeof(*nvbe), GFP_KERNEL);
 	if (!nvbe)
 		return NULL;
 
 	nvbe->dev = dev;
 
-	nvbe->backend.func	= &nouveau_sgdma_backend;
-
+	if (dev_priv->card_type < NV_50)
+		nvbe->backend.func = &nouveau_sgdma_backend;
+	else
+		nvbe->backend.func = &nv50_sgdma_backend;
 	return &nvbe->backend;
 }
 
@@ -226,21 +221,15 @@ nouveau_sgdma_init(struct drm_device *dev)
 
 		obj_size  = (aper_size >> NV_CTXDMA_PAGE_SHIFT) * 4;
 		obj_size += 8; /* ctxdma header */
-	} else {
-		/* 1 entire VM page table */
-		aper_size = (512 * 1024 * 1024);
-		obj_size  = (aper_size >> NV_CTXDMA_PAGE_SHIFT) * 8;
-	}
 
-	ret = nouveau_gpuobj_new(dev, NULL, obj_size, 16,
-				      NVOBJ_FLAG_ZERO_ALLOC |
-				      NVOBJ_FLAG_ZERO_FREE, &gpuobj);
-	if (ret) {
-		NV_ERROR(dev, "Error creating sgdma object: %d\n", ret);
-		return ret;
-	}
+		ret = nouveau_gpuobj_new(dev, NULL, obj_size, 16,
+					      NVOBJ_FLAG_ZERO_ALLOC |
+					      NVOBJ_FLAG_ZERO_FREE, &gpuobj);
+		if (ret) {
+			NV_ERROR(dev, "Error creating sgdma object: %d\n", ret);
+			return ret;
+		}
 
-	if (dev_priv->card_type < NV_50) {
 		nv_wo32(gpuobj, 0, NV_CLASS_DMA_IN_MEMORY |
 				   (1 << 12) /* PT present */ |
 				   (0 << 13) /* PT *not* linear */ |
@@ -249,18 +238,23 @@ nouveau_sgdma_init(struct drm_device *dev)
 		nv_wo32(gpuobj, 4, aper_size - 1);
 		for (i = 2; i < 2 + (aper_size >> 12); i++)
 			nv_wo32(gpuobj, i * 4, 0x00000000);
-	} else {
-		for (i = 0; i < obj_size; i += 8) {
-			nv_wo32(gpuobj, i + 0, 0x00000000);
-			nv_wo32(gpuobj, i + 4, 0x00000000);
-		}
+
+		dev_priv->gart_info.sg_ctxdma = gpuobj;
+		dev_priv->gart_info.aper_base = 0;
+		dev_priv->gart_info.aper_size = aper_size;
+	} else
+	if (dev_priv->chan_vm) {
+		ret = nouveau_vm_get(dev_priv->chan_vm, 512 * 1024 * 1024,
+				     12, NV_MEM_ACCESS_RW,
+				     &dev_priv->gart_info.vma);
+		if (ret)
+			return ret;
+
+		dev_priv->gart_info.aper_base = dev_priv->gart_info.vma.offset;
+		dev_priv->gart_info.aper_size = 512 * 1024 * 1024;
 	}
-	dev_priv->engine.instmem.flush(dev);
 
 	dev_priv->gart_info.type      = NOUVEAU_GART_SGDMA;
-	dev_priv->gart_info.aper_base = 0;
-	dev_priv->gart_info.aper_size = aper_size;
-	dev_priv->gart_info.sg_ctxdma = gpuobj;
 	return 0;
 }
 
@@ -270,6 +264,7 @@ nouveau_sgdma_takedown(struct drm_device *dev)
 	struct drm_nouveau_private *dev_priv = dev->dev_private;
 
 	nouveau_gpuobj_ref(NULL, &dev_priv->gart_info.sg_ctxdma);
+	nouveau_vm_put(&dev_priv->gart_info.vma);
 }
 
 int
diff --git a/drivers/gpu/drm/nouveau/nv50_instmem.c b/drivers/gpu/drm/nouveau/nv50_instmem.c
index 08202fd..ec102bd 100644
--- a/drivers/gpu/drm/nouveau/nv50_instmem.c
+++ b/drivers/gpu/drm/nouveau/nv50_instmem.c
@@ -131,7 +131,6 @@ nv50_instmem_init(struct drm_device *dev)
 	struct nouveau_channel *chan;
 	struct nouveau_vm *vm;
 	int ret, i;
-	u64 nongart_o;
 	u32 tmp;
 
 	priv = kzalloc(sizeof(*priv), GFP_KERNEL);
@@ -216,15 +215,10 @@ nv50_instmem_init(struct drm_device *dev)
 	for (i = 0; i < 8; i++)
 		nv_wr32(dev, 0x1900 + (i*4), 0);
 
-	/* Create shared channel VM, space is reserved for GART mappings at
-	 * the beginning of this address space, it's managed separately
-	 * because TTM makes life painful
+	/* Create shared channel VM, space is reserved at the beginning
+	 * to catch "NULL pointer" references
 	 */
-	dev_priv->vm_gart_base = 0x0020000000ULL;
-	dev_priv->vm_gart_size = 512 * 1024 * 1024;
-	nongart_o = dev_priv->vm_gart_base + dev_priv->vm_gart_size;
-
-	ret = nouveau_vm_new(dev, 0, (1ULL << 40), nongart_o,
+	ret = nouveau_vm_new(dev, 0, (1ULL << 40), 0x0020000000ULL,
 			     29, 12, 16, &dev_priv->chan_vm);
 	if (ret)
 		return ret;