drm/nv40: implement support for on-chip PCIEGART

v2. moved nv44 pciegart table back to instmem, where it's not
    accessible by userspace clients.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>

drivers/gpu/drm/nouveau/nouveau_drv.h

@@ -699,6 +699,13 @@ struct drm_nouveau_private {
 		uint64_t aper_size;
 		uint64_t aper_free;
 
+		struct ttm_backend_func *func;
+
+		struct {
+			struct page *page;
+			dma_addr_t   addr;
+		} dummy;
+
 		struct nouveau_gpuobj *sg_ctxdma;
 		struct nouveau_vma vma;
 	} gart_info;

drivers/gpu/drm/nouveau/nouveau_sgdma.c

@@ -164,6 +164,213 @@ static struct ttm_backend_func nv04_sgdma_backend = {
 	.destroy		= nouveau_sgdma_destroy
 };
 
+static void
+nv41_sgdma_flush(struct nouveau_sgdma_be *nvbe)
+{
+	struct drm_device *dev = nvbe->dev;
+
+	nv_wr32(dev, 0x100810, 0x00000022);
+	if (!nv_wait(dev, 0x100810, 0x00000100, 0x00000100))
+		NV_ERROR(dev, "vm flush timeout: 0x%08x\n",
+			 nv_rd32(dev, 0x100810));
+	nv_wr32(dev, 0x100810, 0x00000000);
+}
+
+static int
+nv41_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;
+	struct nouveau_gpuobj *pgt = dev_priv->gart_info.sg_ctxdma;
+	dma_addr_t *list = nvbe->pages;
+	u32 pte = mem->start << 2;
+	u32 cnt = nvbe->nr_pages;
+
+	nvbe->offset = mem->start << PAGE_SHIFT;
+
+	while (cnt--) {
+		nv_wo32(pgt, pte, (*list++ >> 7) | 1);
+		pte += 4;
+	}
+
+	nv41_sgdma_flush(nvbe);
+	nvbe->bound = true;
+	return 0;
+}
+
+static int
+nv41_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;
+	struct nouveau_gpuobj *pgt = dev_priv->gart_info.sg_ctxdma;
+	u32 pte = (nvbe->offset >> 12) << 2;
+	u32 cnt = nvbe->nr_pages;
+
+	while (cnt--) {
+		nv_wo32(pgt, pte, 0x00000000);
+		pte += 4;
+	}
+
+	nv41_sgdma_flush(nvbe);
+	nvbe->bound = false;
+	return 0;
+}
+
+static struct ttm_backend_func nv41_sgdma_backend = {
+	.populate		= nouveau_sgdma_populate,
+	.clear			= nouveau_sgdma_clear,
+	.bind			= nv41_sgdma_bind,
+	.unbind			= nv41_sgdma_unbind,
+	.destroy		= nouveau_sgdma_destroy
+};
+
+static void
+nv44_sgdma_flush(struct nouveau_sgdma_be *nvbe)
+{
+	struct drm_device *dev = nvbe->dev;
+
+	nv_wr32(dev, 0x100814, (nvbe->nr_pages - 1) << 12);
+	nv_wr32(dev, 0x100808, nvbe->offset | 0x20);
+	if (!nv_wait(dev, 0x100808, 0x00000001, 0x00000001))
+		NV_ERROR(dev, "gart flush timeout: 0x%08x\n",
+			 nv_rd32(dev, 0x100808));
+	nv_wr32(dev, 0x100808, 0x00000000);
+}
+
+static void
+nv44_sgdma_fill(struct nouveau_gpuobj *pgt, dma_addr_t *list, u32 base, u32 cnt)
+{
+	struct drm_nouveau_private *dev_priv = pgt->dev->dev_private;
+	dma_addr_t dummy = dev_priv->gart_info.dummy.addr;
+	u32 pte, tmp[4];
+
+	pte   = base >> 2;
+	base &= ~0x0000000f;
+
+	tmp[0] = nv_ro32(pgt, base + 0x0);
+	tmp[1] = nv_ro32(pgt, base + 0x4);
+	tmp[2] = nv_ro32(pgt, base + 0x8);
+	tmp[3] = nv_ro32(pgt, base + 0xc);
+	while (cnt--) {
+		u32 addr = list ? (*list++ >> 12) : (dummy >> 12);
+		switch (pte++ & 0x3) {
+		case 0:
+			tmp[0] &= ~0x07ffffff;
+			tmp[0] |= addr;
+			break;
+		case 1:
+			tmp[0] &= ~0xf8000000;
+			tmp[0] |= addr << 27;
+			tmp[1] &= ~0x003fffff;
+			tmp[1] |= addr >> 5;
+			break;
+		case 2:
+			tmp[1] &= ~0xffc00000;
+			tmp[1] |= addr << 22;
+			tmp[2] &= ~0x0001ffff;
+			tmp[2] |= addr >> 10;
+			break;
+		case 3:
+			tmp[2] &= ~0xfffe0000;
+			tmp[2] |= addr << 17;
+			tmp[3] &= ~0x00000fff;
+			tmp[3] |= addr >> 15;
+			break;
+		}
+	}
+
+	tmp[3] |= 0x40000000;
+
+	nv_wo32(pgt, base + 0x0, tmp[0]);
+	nv_wo32(pgt, base + 0x4, tmp[1]);
+	nv_wo32(pgt, base + 0x8, tmp[2]);
+	nv_wo32(pgt, base + 0xc, tmp[3]);
+}
+
+static int
+nv44_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;
+	struct nouveau_gpuobj *pgt = dev_priv->gart_info.sg_ctxdma;
+	dma_addr_t *list = nvbe->pages;
+	u32 pte = mem->start << 2, tmp[4];
+	u32 cnt = nvbe->nr_pages;
+	int i;
+
+	nvbe->offset = mem->start << PAGE_SHIFT;
+
+	if (pte & 0x0000000c) {
+		u32  max = 4 - ((pte >> 2) & 0x3);
+		u32 part = (cnt > max) ? max : cnt;
+		nv44_sgdma_fill(pgt, list, pte, part);
+		pte  += (part << 2);
+		list += part;
+		cnt  -= part;
+	}
+
+	while (cnt >= 4) {
+		for (i = 0; i < 4; i++)
+			tmp[i] = *list++ >> 12;
+		nv_wo32(pgt, pte + 0x0, tmp[0] >>  0 | tmp[1] << 27);
+		nv_wo32(pgt, pte + 0x4, tmp[1] >>  5 | tmp[2] << 22);
+		nv_wo32(pgt, pte + 0x8, tmp[2] >> 10 | tmp[3] << 17);
+		nv_wo32(pgt, pte + 0xc, tmp[3] >> 15 | 0x40000000);
+		pte  += 0x10;
+		cnt  -= 4;
+	}
+
+	if (cnt)
+		nv44_sgdma_fill(pgt, list, pte, cnt);
+
+	nv44_sgdma_flush(nvbe);
+	nvbe->bound = true;
+	return 0;
+}
+
+static int
+nv44_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;
+	struct nouveau_gpuobj *pgt = dev_priv->gart_info.sg_ctxdma;
+	u32 pte = (nvbe->offset >> 12) << 2;
+	u32 cnt = nvbe->nr_pages;
+
+	if (pte & 0x0000000c) {
+		u32  max = 4 - ((pte >> 2) & 0x3);
+		u32 part = (cnt > max) ? max : cnt;
+		nv44_sgdma_fill(pgt, NULL, pte, part);
+		pte  += (part << 2);
+		cnt  -= part;
+	}
+
+	while (cnt >= 4) {
+		nv_wo32(pgt, pte + 0x0, 0x00000000);
+		nv_wo32(pgt, pte + 0x4, 0x00000000);
+		nv_wo32(pgt, pte + 0x8, 0x00000000);
+		nv_wo32(pgt, pte + 0xc, 0x00000000);
+		pte  += 0x10;
+		cnt  -= 4;
+	}
+
+	if (cnt)
+		nv44_sgdma_fill(pgt, NULL, pte, cnt);
+
+	nv44_sgdma_flush(nvbe);
+	nvbe->bound = false;
+	return 0;
+}
+
+static struct ttm_backend_func nv44_sgdma_backend = {
+	.populate		= nouveau_sgdma_populate,
+	.clear			= nouveau_sgdma_clear,
+	.bind			= nv44_sgdma_bind,
+	.unbind			= nv44_sgdma_unbind,
+	.destroy		= nouveau_sgdma_destroy
+};
+
 static int
 nv50_sgdma_bind(struct ttm_backend *be, struct ttm_mem_reg *mem)
 {
@@ -213,10 +420,7 @@ nouveau_sgdma_init_ttm(struct drm_device *dev)
 
 	nvbe->dev = dev;
 
-	if (dev_priv->card_type >= NV_50)
-		nvbe->backend.func = &nv50_sgdma_backend;
-	else
-		nvbe->backend.func = &nv04_sgdma_backend;
+	nvbe->backend.func = dev_priv->gart_info.func;
 	return &nvbe->backend;
 }
 
@@ -225,29 +429,69 @@ nouveau_sgdma_init(struct drm_device *dev)
 {
 	struct drm_nouveau_private *dev_priv = dev->dev_private;
 	struct nouveau_gpuobj *gpuobj = NULL;
-	uint32_t aper_size, obj_size;
-	int i, ret;
+	u32 aper_size, align;
+	int ret;
+
+	if (dev_priv->card_type >= NV_50 ||
+	    dev_priv->ramin_rsvd_vram >= 2 * 1024 * 1024)
+		aper_size = 512 * 1024 * 1024;
+	else
+		aper_size = 64 * 1024 * 1024;
+
+	/* Dear NVIDIA, NV44+ would like proper present bits in PTEs for
+	 * christmas.  The cards before it have them, the cards after
+	 * it have them, why is NV44 so unloved?
+	 */
+	dev_priv->gart_info.dummy.page = alloc_page(GFP_DMA32 | GFP_KERNEL);
+	if (!dev_priv->gart_info.dummy.page)
+		return -ENOMEM;
+
+	dev_priv->gart_info.dummy.addr =
+		pci_map_page(dev->pdev, dev_priv->gart_info.dummy.page,
+			     0, PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
+	if (pci_dma_mapping_error(dev->pdev, dev_priv->gart_info.dummy.addr)) {
+		NV_ERROR(dev, "error mapping dummy page\n");
+		__free_page(dev_priv->gart_info.dummy.page);
+		dev_priv->gart_info.dummy.page = NULL;
+		return -ENOMEM;
+	}
 
 	if (dev_priv->card_type >= NV_50) {
-		ret = nouveau_vm_get(dev_priv->chan_vm, 512 * 1024 * 1024,
+		ret = nouveau_vm_get(dev_priv->chan_vm, aper_size,
 				     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->gart_info.aper_size = aper_size;
 		dev_priv->gart_info.type = NOUVEAU_GART_HW;
+		dev_priv->gart_info.func = &nv50_sgdma_backend;
+	} else
+	if (drm_pci_device_is_pcie(dev) &&
+	    dev_priv->chipset != 0x40 && dev_priv->chipset != 0x45) {
+		if (nv44_graph_class(dev)) {
+			dev_priv->gart_info.func = &nv44_sgdma_backend;
+			align = 512 * 1024;
 		} else {
-		if(dev_priv->ramin_rsvd_vram < 2 * 1024 * 1024)
-			aper_size = 64 * 1024 * 1024;
-		else
-			aper_size = 512 * 1024 * 1024;
+			dev_priv->gart_info.func = &nv41_sgdma_backend;
+			align = 16;
+		}
 
-		obj_size  = (aper_size >> NV_CTXDMA_PAGE_SHIFT) * 4;
-		obj_size += 8; /* ctxdma header */
+		ret = nouveau_gpuobj_new(dev, NULL, aper_size / 1024, align,
+					 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,
+		dev_priv->gart_info.sg_ctxdma = gpuobj;
+		dev_priv->gart_info.aper_base = 0;
+		dev_priv->gart_info.aper_size = aper_size;
+		dev_priv->gart_info.type = NOUVEAU_GART_HW;
+	} else {
+		ret = nouveau_gpuobj_new(dev, NULL, (aper_size / 1024) + 8, 16,
 					 NVOBJ_FLAG_ZERO_ALLOC |
 					 NVOBJ_FLAG_ZERO_FREE, &gpuobj);
 		if (ret) {
@@ -261,13 +505,12 @@ nouveau_sgdma_init(struct drm_device *dev)
 				   (0 << 14) /* RW */ |
 				   (2 << 16) /* PCI */);
 		nv_wo32(gpuobj, 4, aper_size - 1);
-		for (i = 2; i < 2 + (aper_size >> 12); i++)
-			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;
 		dev_priv->gart_info.type = NOUVEAU_GART_PDMA;
+		dev_priv->gart_info.func = &nv04_sgdma_backend;
 	}
 
 	return 0;
@@ -280,6 +523,13 @@ nouveau_sgdma_takedown(struct drm_device *dev)
 
 	nouveau_gpuobj_ref(NULL, &dev_priv->gart_info.sg_ctxdma);
 	nouveau_vm_put(&dev_priv->gart_info.vma);
+
+	if (dev_priv->gart_info.dummy.page) {
+		pci_unmap_page(dev->pdev, dev_priv->gart_info.dummy.addr,
+			       PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
+		__free_page(dev_priv->gart_info.dummy.page);
+		dev_priv->gart_info.dummy.page = NULL;
+	}
 }
 
 uint32_t

drivers/gpu/drm/nouveau/nv40_fb.c

@@ -24,6 +24,53 @@ nv40_fb_set_tile_region(struct drm_device *dev, int i)
 	}
 }
 
+static void
+nv40_fb_init_gart(struct drm_device *dev)
+{
+	struct drm_nouveau_private *dev_priv = dev->dev_private;
+	struct nouveau_gpuobj *gart = dev_priv->gart_info.sg_ctxdma;
+
+	if (dev_priv->gart_info.type != NOUVEAU_GART_HW) {
+		nv_wr32(dev, 0x100800, 0x00000001);
+		return;
+	}
+
+	nv_wr32(dev, 0x100800, gart->pinst | 0x00000002);
+	nv_mask(dev, 0x10008c, 0x00000100, 0x00000100);
+	nv_wr32(dev, 0x100820, 0x00000000);
+}
+
+static void
+nv44_fb_init_gart(struct drm_device *dev)
+{
+	struct drm_nouveau_private *dev_priv = dev->dev_private;
+	struct nouveau_gpuobj *gart = dev_priv->gart_info.sg_ctxdma;
+	u32 vinst;
+
+	if (dev_priv->gart_info.type != NOUVEAU_GART_HW) {
+		nv_wr32(dev, 0x100850, 0x80000000);
+		nv_wr32(dev, 0x100800, 0x00000001);
+		return;
+	}
+
+	/* calculate vram address of this PRAMIN block, object
+	 * must be allocated on 512KiB alignment, and not exceed
+	 * a total size of 512KiB for this to work correctly
+	 */
+	vinst  = nv_rd32(dev, 0x10020c);
+	vinst -= ((gart->pinst >> 19) + 1) << 19;
+
+	nv_wr32(dev, 0x100850, 0x80000000);
+	nv_wr32(dev, 0x100818, dev_priv->gart_info.dummy.addr);
+
+	nv_wr32(dev, 0x100804, dev_priv->gart_info.aper_size);
+	nv_wr32(dev, 0x100850, 0x00008000);
+	nv_mask(dev, 0x10008c, 0x00000200, 0x00000200);
+	nv_wr32(dev, 0x100820, 0x00000000);
+	nv_wr32(dev, 0x10082c, 0x00000001);
+	nv_wr32(dev, 0x100800, vinst | 0x00000010);
+}
+
 int
 nv40_fb_init(struct drm_device *dev)
 {
@@ -32,12 +79,12 @@ nv40_fb_init(struct drm_device *dev)
 	uint32_t tmp;
 	int i;
 
-	/* This is strictly a NV4x register (don't know about NV5x). */
-	/* The blob sets these to all kinds of values, and they mess up our setup. */
-	/* I got value 0x52802 instead. For some cards the blob even sets it back to 0x1. */
-	/* Note: the blob doesn't read this value, so i'm pretty sure this is safe for all cards. */
-	/* Any idea what this is? */
-	nv_wr32(dev, NV40_PFB_UNK_800, 0x1);
+	if (dev_priv->chipset != 0x40 && dev_priv->chipset != 0x45) {
+		if (nv44_graph_class(dev))
+			nv44_fb_init_gart(dev);
+		else
+			nv40_fb_init_gart(dev);
+	}
 
 	switch (dev_priv->chipset) {
 	case 0x40: