drm/nouveau: rework init ordering so nv50_instmem.c can be less bad

Reviewed-by: Francisco Jerez <currojerez@riseup.net>
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>

drivers/gpu/drm/nouveau/nouveau_bo.c

@@ -403,7 +403,10 @@ 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 = dev_priv->vm_vram_base;
+		if (dev_priv->card_type == NV_50)
+			man->gpu_offset = 0x40000000;
+		else
+			man->gpu_offset = 0;
 		break;
 	case TTM_PL_TT:
 		switch (dev_priv->gart_info.type) {

drivers/gpu/drm/nouveau/nouveau_drv.h

@@ -700,8 +700,10 @@ extern bool nouveau_wait_for_idle(struct drm_device *);
 extern int  nouveau_card_init(struct drm_device *);
 
 /* nouveau_mem.c */
-extern int  nouveau_mem_detect(struct drm_device *dev);
-extern int  nouveau_mem_init(struct drm_device *);
+extern int  nouveau_mem_vram_init(struct drm_device *);
+extern void nouveau_mem_vram_fini(struct drm_device *);
+extern int  nouveau_mem_gart_init(struct drm_device *);
+extern void nouveau_mem_gart_fini(struct drm_device *);
 extern int  nouveau_mem_init_agp(struct drm_device *);
 extern int  nouveau_mem_reset_agp(struct drm_device *);
 extern void nouveau_mem_close(struct drm_device *);

drivers/gpu/drm/nouveau/nouveau_mem.c

@@ -221,7 +221,7 @@ nv50_mem_vm_unbind(struct drm_device *dev, uint64_t virt, uint32_t size)
  * Cleanup everything
  */
 void
-nouveau_mem_close(struct drm_device *dev)
+nouveau_mem_vram_fini(struct drm_device *dev)
 {
 	struct drm_nouveau_private *dev_priv = dev->dev_private;
 
@@ -232,6 +232,19 @@ nouveau_mem_close(struct drm_device *dev)
 
 	nouveau_ttm_global_release(dev_priv);
 
+	if (dev_priv->fb_mtrr >= 0) {
+		drm_mtrr_del(dev_priv->fb_mtrr,
+			     pci_resource_start(dev->pdev, 1),
+			     pci_resource_len(dev->pdev, 1), DRM_MTRR_WC);
+		dev_priv->fb_mtrr = -1;
+	}
+}
+
+void
+nouveau_mem_gart_fini(struct drm_device *dev)
+{
+	nouveau_sgdma_takedown(dev);
+
 	if (drm_core_has_AGP(dev) && dev->agp) {
 		struct drm_agp_mem *entry, *tempe;
 
@@ -251,13 +264,6 @@ nouveau_mem_close(struct drm_device *dev)
 		dev->agp->acquired = 0;
 		dev->agp->enabled = 0;
 	}
-
-	if (dev_priv->fb_mtrr) {
-		drm_mtrr_del(dev_priv->fb_mtrr,
-			     pci_resource_start(dev->pdev, 1),
-			     pci_resource_len(dev->pdev, 1), DRM_MTRR_WC);
-		dev_priv->fb_mtrr = -1;
-	}
 }
 
 static uint32_t
@@ -363,7 +369,7 @@ nvaa_vram_preinit(struct drm_device *dev)
 	dev_priv->vram_rblock_size = 1;
 }
 
-int
+static int
 nouveau_mem_detect(struct drm_device *dev)
 {
 	struct drm_nouveau_private *dev_priv = dev->dev_private;
@@ -500,24 +506,27 @@ nouveau_mem_init_agp(struct drm_device *dev)
 }
 
 int
-nouveau_mem_init(struct drm_device *dev)
+nouveau_mem_vram_init(struct drm_device *dev)
 {
 	struct drm_nouveau_private *dev_priv = dev->dev_private;
 	struct ttm_bo_device *bdev = &dev_priv->ttm.bdev;
-	int ret, dma_bits = 32;
-
-	dev_priv->fb_phys = pci_resource_start(dev->pdev, 1);
-	dev_priv->gart_info.type = NOUVEAU_GART_NONE;
+	int ret, dma_bits;
 
 	if (dev_priv->card_type >= NV_50 &&
 	    pci_dma_supported(dev->pdev, DMA_BIT_MASK(40)))
 		dma_bits = 40;
+	else
+		dma_bits = 32;
 
 	ret = pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(dma_bits));
-	if (ret) {
-		NV_ERROR(dev, "Error setting DMA mask: %d\n", ret);
+	if (ret)
 		return ret;
-	}
+
+	ret = nouveau_mem_detect(dev);
+	if (ret)
+		return ret;
+
+	dev_priv->fb_phys = pci_resource_start(dev->pdev, 1);
 
 	ret = nouveau_ttm_global_init(dev_priv);
 	if (ret)
@@ -541,7 +550,16 @@ nouveau_mem_init(struct drm_device *dev)
 			pci_resource_len(dev->pdev, 1);
 	dev_priv->fb_mappable_pages >>= PAGE_SHIFT;
 
-	/* remove reserved space at end of vram from available amount */
+	/* reserve space at end of VRAM for PRAMIN */
+	if (dev_priv->chipset == 0x40 || dev_priv->chipset == 0x47 ||
+	    dev_priv->chipset == 0x49 || dev_priv->chipset == 0x4b)
+		dev_priv->ramin_rsvd_vram = (2 * 1024 * 1024);
+	else
+	if (dev_priv->card_type >= NV_40)
+		dev_priv->ramin_rsvd_vram = (1 * 1024 * 1024);
+	else
+		dev_priv->ramin_rsvd_vram = (512 * 1024);
+
 	dev_priv->fb_available_size -= dev_priv->ramin_rsvd_vram;
 	dev_priv->fb_aper_free = dev_priv->fb_available_size;
 
@@ -562,7 +580,21 @@ nouveau_mem_init(struct drm_device *dev)
 		nouveau_bo_ref(NULL, &dev_priv->vga_ram);
 	}
 
-	/* GART */
+	dev_priv->fb_mtrr = drm_mtrr_add(pci_resource_start(dev->pdev, 1),
+					 pci_resource_len(dev->pdev, 1),
+					 DRM_MTRR_WC);
+	return 0;
+}
+
+int
+nouveau_mem_gart_init(struct drm_device *dev)
+{
+	struct drm_nouveau_private *dev_priv = dev->dev_private;
+	struct ttm_bo_device *bdev = &dev_priv->ttm.bdev;
+	int ret;
+
+	dev_priv->gart_info.type = NOUVEAU_GART_NONE;
+
 #if !defined(__powerpc__) && !defined(__ia64__)
 	if (drm_device_is_agp(dev) && dev->agp && !nouveau_noagp) {
 		ret = nouveau_mem_init_agp(dev);
@@ -590,11 +622,6 @@ nouveau_mem_init(struct drm_device *dev)
 		return ret;
 	}
 
-	dev_priv->fb_mtrr = drm_mtrr_add(pci_resource_start(dev->pdev, 1),
-					 pci_resource_len(dev->pdev, 1),
-					 DRM_MTRR_WC);
-
 	return 0;
 }
 
-

drivers/gpu/drm/nouveau/nouveau_object.c

@@ -130,7 +130,7 @@ nouveau_gpuobj_new(struct drm_device *dev, struct nouveau_channel *chan,
 
 	/* if we got a chunk of the aperture, map pages into it */
 	gpuobj->im_pramin = ramin;
-	if (!chan && gpuobj->im_pramin) {
+	if (!chan && gpuobj->im_pramin && dev_priv->ramin_available) {
 		ret = engine->instmem.bind(dev, gpuobj);
 		if (ret) {
 			nouveau_gpuobj_ref(NULL, &gpuobj);
@@ -173,7 +173,7 @@ nouveau_gpuobj_new(struct drm_device *dev, struct nouveau_channel *chan,
 }
 
 int
-nouveau_gpuobj_early_init(struct drm_device *dev)
+nouveau_gpuobj_init(struct drm_device *dev)
 {
 	struct drm_nouveau_private *dev_priv = dev->dev_private;
 
@@ -186,29 +186,6 @@ nouveau_gpuobj_early_init(struct drm_device *dev)
 	return 0;
 }
 
-int
-nouveau_gpuobj_init(struct drm_device *dev)
-{
-	struct drm_nouveau_private *dev_priv = dev->dev_private;
-	struct nouveau_gpuobj *ramht = NULL;
-	int ret;
-
-	NV_DEBUG(dev, "\n");
-
-	if (dev_priv->card_type >= NV_50)
-		return 0;
-
-	ret = nouveau_gpuobj_new_fake(dev, dev_priv->ramht_offset, ~0,
-				      dev_priv->ramht_size,
-				      NVOBJ_FLAG_ZERO_ALLOC, &ramht);
-	if (ret)
-		return ret;
-
-	ret = nouveau_ramht_new(dev, ramht, &dev_priv->ramht);
-	nouveau_gpuobj_ref(NULL, &ramht);
-	return ret;
-}
-
 void
 nouveau_gpuobj_takedown(struct drm_device *dev)
 {

drivers/gpu/drm/nouveau/nouveau_state.c

@@ -532,35 +532,26 @@ nouveau_card_init(struct drm_device *dev)
 	if (ret)
 		goto out_display_early;
 
-	ret = nouveau_mem_detect(dev);
+	ret = nouveau_mem_vram_init(dev);
 	if (ret)
 		goto out_bios;
 
-	ret = nouveau_gpuobj_early_init(dev);
+	ret = nouveau_gpuobj_init(dev);
 	if (ret)
-		goto out_bios;
+		goto out_vram;
 
-	/* Initialise instance memory, must happen before mem_init so we
-	 * know exactly how much VRAM we're able to use for "normal"
-	 * purposes.
-	 */
 	ret = engine->instmem.init(dev);
 	if (ret)
-		goto out_gpuobj_early;
+		goto out_gpuobj;
 
-	/* Setup the memory manager */
-	ret = nouveau_mem_init(dev);
+	ret = nouveau_mem_gart_init(dev);
 	if (ret)
 		goto out_instmem;
 
-	ret = nouveau_gpuobj_init(dev);
-	if (ret)
-		goto out_mem;
-
 	/* PMC */
 	ret = engine->mc.init(dev);
 	if (ret)
-		goto out_gpuobj;
+		goto out_gart;
 
 	/* PGPIO */
 	ret = engine->gpio.init(dev);
@@ -640,15 +631,14 @@ nouveau_card_init(struct drm_device *dev)
 	engine->gpio.takedown(dev);
 out_mc:
 	engine->mc.takedown(dev);
-out_gpuobj:
-	nouveau_gpuobj_takedown(dev);
-out_mem:
-	nouveau_sgdma_takedown(dev);
-	nouveau_mem_close(dev);
+out_gart:
+	nouveau_mem_gart_fini(dev);
 out_instmem:
 	engine->instmem.takedown(dev);
-out_gpuobj_early:
-	nouveau_gpuobj_late_takedown(dev);
+out_gpuobj:
+	nouveau_gpuobj_takedown(dev);
+out_vram:
+	nouveau_mem_vram_fini(dev);
 out_bios:
 	nouveau_bios_takedown(dev);
 out_display_early:
@@ -684,15 +674,14 @@ static void nouveau_card_takedown(struct drm_device *dev)
 	ttm_bo_clean_mm(&dev_priv->ttm.bdev, TTM_PL_VRAM);
 	ttm_bo_clean_mm(&dev_priv->ttm.bdev, TTM_PL_TT);
 	mutex_unlock(&dev->struct_mutex);
-	nouveau_sgdma_takedown(dev);
+	nouveau_mem_gart_fini(dev);
 
-	nouveau_gpuobj_takedown(dev);
-	nouveau_mem_close(dev);
 	engine->instmem.takedown(dev);
+	nouveau_gpuobj_takedown(dev);
+	nouveau_mem_vram_fini(dev);
 
 	drm_irq_uninstall(dev);
 
-	nouveau_gpuobj_late_takedown(dev);
 	nouveau_bios_takedown(dev);
 
 	vga_client_register(dev->pdev, NULL, NULL, NULL);

drivers/gpu/drm/nouveau/nv04_instmem.c

 #include "drmP.h"
 #include "drm.h"
 #include "nouveau_drv.h"
+#include "nouveau_ramht.h"
 
 /* returns the size of fifo context */
 static int
@@ -17,42 +18,6 @@ nouveau_fifo_ctx_size(struct drm_device *dev)
 	return 32;
 }
 
-static void
-nv04_instmem_determine_amount(struct drm_device *dev)
-{
-	struct drm_nouveau_private *dev_priv = dev->dev_private;
-	int i;
-
-	/* Figure out how much instance memory we need */
-	if (dev_priv->card_type >= NV_40) {
-		/* We'll want more instance memory than this on some NV4x cards.
-		 * There's a 16MB aperture to play with that maps onto the end
-		 * of vram.  For now, only reserve a small piece until we know
-		 * more about what each chipset requires.
-		 */
-		switch (dev_priv->chipset) {
-		case 0x40:
-		case 0x47:
-		case 0x49:
-		case 0x4b:
-			dev_priv->ramin_rsvd_vram = (2 * 1024 * 1024);
-			break;
-		default:
-			dev_priv->ramin_rsvd_vram = (1 * 1024 * 1024);
-			break;
-		}
-	} else {
-		/*XXX: what *are* the limits on <NV40 cards?
-		 */
-		dev_priv->ramin_rsvd_vram = (512 * 1024);
-	}
-	NV_DEBUG(dev, "RAMIN size: %dKiB\n", dev_priv->ramin_rsvd_vram >> 10);
-
-	/* Clear all of it, except the BIOS image that's in the first 64KiB */
-	for (i = 64 * 1024; i < dev_priv->ramin_rsvd_vram; i += 4)
-		nv_wi32(dev, i, 0x00000000);
-}
-
 static void
 nv04_instmem_configure_fixed_tables(struct drm_device *dev)
 {
@@ -103,12 +68,24 @@ nv04_instmem_configure_fixed_tables(struct drm_device *dev)
 int nv04_instmem_init(struct drm_device *dev)
 {
 	struct drm_nouveau_private *dev_priv = dev->dev_private;
+	struct nouveau_gpuobj *ramht = NULL;
 	uint32_t offset;
 	int ret;
 
-	nv04_instmem_determine_amount(dev);
 	nv04_instmem_configure_fixed_tables(dev);
 
+	/* Setup shared RAMHT */
+	ret = nouveau_gpuobj_new_fake(dev, dev_priv->ramht_offset, ~0,
+				      dev_priv->ramht_size,
+				      NVOBJ_FLAG_ZERO_ALLOC, &ramht);
+	if (ret)
+		return ret;
+
+	ret = nouveau_ramht_new(dev, ramht, &dev_priv->ramht);
+	nouveau_gpuobj_ref(NULL, &ramht);
+	if (ret)
+		return ret;
+
 	/* Create a heap to manage RAMIN allocations, we don't allocate
 	 * the space that was reserved for RAMHT/FC/RO.
 	 */