drm/nouveau: modify vm to accomodate dual page tables for nvc0

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

drivers/gpu/drm/nouveau/nouveau_vm.c

@@ -32,6 +32,7 @@ nouveau_vm_map_at(struct nouveau_vma *vma, u64 delta, struct nouveau_vram *vram)
 {
 	struct nouveau_vm *vm = vma->vm;
 	struct nouveau_mm_node *r;
+	int big = vma->node->type != vm->spg_shift;
 	u32 offset = vma->node->offset + (delta >> 12);
 	u32 bits = vma->node->type - 12;
 	u32 pde  = (offset >> vm->pgt_bits) - vm->fpde;
@@ -44,7 +45,7 @@ nouveau_vm_map_at(struct nouveau_vma *vma, u64 delta, struct nouveau_vram *vram)
 		u32 num  = r->length >> bits;
 
 		while (num) {
-			struct nouveau_gpuobj *pgt = vm->pgt[pde].obj;
+			struct nouveau_gpuobj *pgt = vm->pgt[pde].obj[big];
 
 			end = (pte + num);
 			if (unlikely(end >= max))
@@ -76,6 +77,7 @@ nouveau_vm_map_sg(struct nouveau_vma *vma, u64 delta, u64 length,
 		  dma_addr_t *list)
 {
 	struct nouveau_vm *vm = vma->vm;
+	int big = vma->node->type != vm->spg_shift;
 	u32 offset = vma->node->offset + (delta >> 12);
 	u32 bits = vma->node->type - 12;
 	u32 num  = length >> vma->node->type;
@@ -85,7 +87,7 @@ nouveau_vm_map_sg(struct nouveau_vma *vma, u64 delta, u64 length,
 	u32 end, len;
 
 	while (num) {
-		struct nouveau_gpuobj *pgt = vm->pgt[pde].obj;
+		struct nouveau_gpuobj *pgt = vm->pgt[pde].obj[big];
 
 		end = (pte + num);
 		if (unlikely(end >= max))
@@ -110,6 +112,7 @@ void
 nouveau_vm_unmap_at(struct nouveau_vma *vma, u64 delta, u64 length)
 {
 	struct nouveau_vm *vm = vma->vm;
+	int big = vma->node->type != vm->spg_shift;
 	u32 offset = vma->node->offset + (delta >> 12);
 	u32 bits = vma->node->type - 12;
 	u32 num  = length >> vma->node->type;
@@ -119,7 +122,7 @@ nouveau_vm_unmap_at(struct nouveau_vma *vma, u64 delta, u64 length)
 	u32 end, len;
 
 	while (num) {
-		struct nouveau_gpuobj *pgt = vm->pgt[pde].obj;
+		struct nouveau_gpuobj *pgt = vm->pgt[pde].obj[big];
 
 		end = (pte + num);
 		if (unlikely(end >= max))
@@ -146,7 +149,7 @@ nouveau_vm_unmap(struct nouveau_vma *vma)
 }
 
 static void
-nouveau_vm_unmap_pgt(struct nouveau_vm *vm, u32 fpde, u32 lpde)
+nouveau_vm_unmap_pgt(struct nouveau_vm *vm, int big, u32 fpde, u32 lpde)
 {
 	struct nouveau_vm_pgd *vpgd;
 	struct nouveau_vm_pgt *vpgt;
@@ -155,16 +158,16 @@ nouveau_vm_unmap_pgt(struct nouveau_vm *vm, u32 fpde, u32 lpde)
 
 	for (pde = fpde; pde <= lpde; pde++) {
 		vpgt = &vm->pgt[pde - vm->fpde];
-		if (--vpgt->refcount)
+		if (--vpgt->refcount[big])
 			continue;
 
+		pgt = vpgt->obj[big];
+		vpgt->obj[big] = NULL;
+
 		list_for_each_entry(vpgd, &vm->pgd_list, head) {
-			vm->unmap_pgt(vpgd->obj, pde);
+			vm->map_pgt(vpgd->obj, pde, vpgt->obj);
 		}
 
-		pgt = vpgt->obj;
-		vpgt->obj = NULL;
-
 		mutex_unlock(&vm->mm->mutex);
 		nouveau_gpuobj_ref(NULL, &pgt);
 		mutex_lock(&vm->mm->mutex);
@@ -177,6 +180,7 @@ nouveau_vm_map_pgt(struct nouveau_vm *vm, u32 pde, u32 type)
 	struct nouveau_vm_pgt *vpgt = &vm->pgt[pde - vm->fpde];
 	struct nouveau_vm_pgd *vpgd;
 	struct nouveau_gpuobj *pgt;
+	int big = (type != vm->spg_shift);
 	u32 pgt_size;
 	int ret;
 
@@ -191,19 +195,18 @@ nouveau_vm_map_pgt(struct nouveau_vm *vm, u32 pde, u32 type)
 		return ret;
 
 	/* someone beat us to filling the PDE while we didn't have the lock */
-	if (unlikely(vpgt->refcount++)) {
+	if (unlikely(vpgt->refcount[big]++)) {
 		mutex_unlock(&vm->mm->mutex);
 		nouveau_gpuobj_ref(NULL, &pgt);
 		mutex_lock(&vm->mm->mutex);
 		return 0;
 	}
 
+	vpgt->obj[big] = pgt;
 	list_for_each_entry(vpgd, &vm->pgd_list, head) {
-		vm->map_pgt(vpgd->obj, type, pde, pgt);
+		vm->map_pgt(vpgd->obj, pde, vpgt->obj);
 	}
 
-	vpgt->page_shift = type;
-	vpgt->obj = pgt;
 	return 0;
 }
 
@@ -227,16 +230,17 @@ nouveau_vm_get(struct nouveau_vm *vm, u64 size, u32 page_shift,
 	lpde = (vma->node->offset + vma->node->length - 1) >> vm->pgt_bits;
 	for (pde = fpde; pde <= lpde; pde++) {
 		struct nouveau_vm_pgt *vpgt = &vm->pgt[pde - vm->fpde];
+		int big = (vma->node->type != vm->spg_shift);
 
-		if (likely(vpgt->refcount)) {
-			vpgt->refcount++;
+		if (likely(vpgt->refcount[big])) {
+			vpgt->refcount[big]++;
 			continue;
 		}
 
 		ret = nouveau_vm_map_pgt(vm, pde, vma->node->type);
 		if (ret) {
 			if (pde != fpde)
-				nouveau_vm_unmap_pgt(vm, fpde, pde - 1);
+				nouveau_vm_unmap_pgt(vm, big, fpde, pde - 1);
 			nouveau_mm_put(vm->mm, vma->node);
 			mutex_unlock(&vm->mm->mutex);
 			vma->node = NULL;
@@ -263,21 +267,20 @@ nouveau_vm_put(struct nouveau_vma *vma)
 	lpde = (vma->node->offset + vma->node->length - 1) >> vm->pgt_bits;
 
 	mutex_lock(&vm->mm->mutex);
+	nouveau_vm_unmap_pgt(vm, vma->node->type != vm->spg_shift, fpde, lpde);
 	nouveau_mm_put(vm->mm, vma->node);
 	vma->node = NULL;
-	nouveau_vm_unmap_pgt(vm, fpde, lpde);
 	mutex_unlock(&vm->mm->mutex);
 }
 
 int
 nouveau_vm_new(struct drm_device *dev, u64 offset, u64 length, u64 mm_offset,
-	       u8 pgt_bits, u8 spg_shift, u8 lpg_shift,
 	       struct nouveau_vm **pvm)
 {
 	struct drm_nouveau_private *dev_priv = dev->dev_private;
 	struct nouveau_vm *vm;
 	u64 mm_length = (offset + length) - mm_offset;
-	u32 block;
+	u32 block, pgt_bits;
 	int ret;
 
 	vm = kzalloc(sizeof(*vm), GFP_KERNEL);
@@ -286,11 +289,13 @@ nouveau_vm_new(struct drm_device *dev, u64 offset, u64 length, u64 mm_offset,
 
 	if (dev_priv->card_type == NV_50) {
 		vm->map_pgt = nv50_vm_map_pgt;
-		vm->unmap_pgt = nv50_vm_unmap_pgt;
 		vm->map = nv50_vm_map;
 		vm->map_sg = nv50_vm_map_sg;
 		vm->unmap = nv50_vm_unmap;
 		vm->flush = nv50_vm_flush;
+		vm->spg_shift = 12;
+		vm->lpg_shift = 16;
+		pgt_bits = 29;
 	} else {
 		kfree(vm);
 		return -ENOSYS;
@@ -308,8 +313,6 @@ nouveau_vm_new(struct drm_device *dev, u64 offset, u64 length, u64 mm_offset,
 	vm->dev = dev;
 	vm->refcount = 1;
 	vm->pgt_bits = pgt_bits - 12;
-	vm->spg_shift = spg_shift;
-	vm->lpg_shift = lpg_shift;
 
 	block = (1 << pgt_bits);
 	if (length < block)
@@ -342,16 +345,8 @@ nouveau_vm_link(struct nouveau_vm *vm, struct nouveau_gpuobj *pgd)
 	nouveau_gpuobj_ref(pgd, &vpgd->obj);
 
 	mutex_lock(&vm->mm->mutex);
-	for (i = vm->fpde; i <= vm->lpde; i++) {
-		struct nouveau_vm_pgt *vpgt = &vm->pgt[i - vm->fpde];
-
-		if (!vpgt->obj) {
-			vm->unmap_pgt(pgd, i);
-			continue;
-		}
-
-		vm->map_pgt(pgd, vpgt->page_shift, i, vpgt->obj);
-	}
+	for (i = vm->fpde; i <= vm->lpde; i++)
+		vm->map_pgt(pgd, i, vm->pgt[i - vm->fpde].obj);
 	list_add(&vpgd->head, &vm->pgd_list);
 	mutex_unlock(&vm->mm->mutex);
 	return 0;

drivers/gpu/drm/nouveau/nouveau_vm.h

@@ -31,9 +31,8 @@
 #include "nouveau_mm.h"
 
 struct nouveau_vm_pgt {
-	struct nouveau_gpuobj *obj;
-	u32 page_shift;
-	u32 refcount;
+	struct nouveau_gpuobj *obj[2];
+	u32 refcount[2];
 };
 
 struct nouveau_vm_pgd {
@@ -65,9 +64,8 @@ struct nouveau_vm {
 	u8  spg_shift;
 	u8  lpg_shift;
 
-	void (*map_pgt)(struct nouveau_gpuobj *pgd, u32 type, u32 pde,
-			struct nouveau_gpuobj *pgt);
-	void (*unmap_pgt)(struct nouveau_gpuobj *pgd, u32 pde);
+	void (*map_pgt)(struct nouveau_gpuobj *pgd, u32 pde,
+			struct nouveau_gpuobj *pgt[2]);
 	void (*map)(struct nouveau_vma *, struct nouveau_gpuobj *,
 		    struct nouveau_vram *, u32 pte, u32 cnt, u64 phys);
 	void (*map_sg)(struct nouveau_vma *, struct nouveau_gpuobj *,
@@ -78,7 +76,6 @@ struct nouveau_vm {
 
 /* nouveau_vm.c */
 int  nouveau_vm_new(struct drm_device *, u64 offset, u64 length, u64 mm_offset,
-		    u8 pgt_bits, u8 spg_shift, u8 lpg_shift,
 		    struct nouveau_vm **);
 int  nouveau_vm_ref(struct nouveau_vm *, struct nouveau_vm **,
 		    struct nouveau_gpuobj *pgd);
@@ -93,9 +90,8 @@ void nouveau_vm_map_sg(struct nouveau_vma *, u64 offset, u64 length,
 		       dma_addr_t *);
 
 /* nv50_vm.c */
-void nv50_vm_map_pgt(struct nouveau_gpuobj *pgd, u32 type, u32 pde,
-		     struct nouveau_gpuobj *pgt);
-void nv50_vm_unmap_pgt(struct nouveau_gpuobj *pgd, u32 pde);
+void nv50_vm_map_pgt(struct nouveau_gpuobj *pgd, u32 pde,
+		     struct nouveau_gpuobj *pgt[2]);
 void nv50_vm_map(struct nouveau_vma *, struct nouveau_gpuobj *,
 		 struct nouveau_vram *, u32 pte, u32 cnt, u64 phys);
 void nv50_vm_map_sg(struct nouveau_vma *, struct nouveau_gpuobj *,

drivers/gpu/drm/nouveau/nv50_instmem.c

@@ -151,20 +151,19 @@ nv50_instmem_init(struct drm_device *dev)
 
 	/* BAR3 */
 	ret = nouveau_vm_new(dev, BAR3_VM_BASE, BAR3_VM_SIZE, BAR3_VM_BASE,
-			     29, 12, 16, &dev_priv->bar3_vm);
+			     &dev_priv->bar3_vm);
 	if (ret)
 		goto error;
 
 	ret = nouveau_gpuobj_new(dev, NULL, (BAR3_VM_SIZE >> 12) * 8,
 				 0x1000, NVOBJ_FLAG_DONT_MAP |
 				 NVOBJ_FLAG_ZERO_ALLOC,
-				 &dev_priv->bar3_vm->pgt[0].obj);
+				 &dev_priv->bar3_vm->pgt[0].obj[0]);
 	if (ret)
 		goto error;
-	dev_priv->bar3_vm->pgt[0].page_shift = 12;
-	dev_priv->bar3_vm->pgt[0].refcount = 1;
+	dev_priv->bar3_vm->pgt[0].refcount[0] = 1;
 
-	nv50_instmem_map(dev_priv->bar3_vm->pgt[0].obj);
+	nv50_instmem_map(dev_priv->bar3_vm->pgt[0].obj[0]);
 
 	ret = nv50_channel_new(dev, 128 * 1024, dev_priv->bar3_vm, &chan);
 	if (ret)
@@ -195,8 +194,7 @@ nv50_instmem_init(struct drm_device *dev)
 	nv_wo32(chan->ramin, 0, tmp);
 
 	/* BAR1 */
-	ret = nouveau_vm_new(dev, BAR1_VM_BASE, BAR1_VM_SIZE, BAR1_VM_BASE,
-			     29, 12, 16, &vm);
+	ret = nouveau_vm_new(dev, BAR1_VM_BASE, BAR1_VM_SIZE, BAR1_VM_BASE, &vm);
 	if (ret)
 		goto error;
 
@@ -220,7 +218,7 @@ nv50_instmem_init(struct drm_device *dev)
 	 * to catch "NULL pointer" references
 	 */
 	ret = nouveau_vm_new(dev, 0, (1ULL << 40), 0x0020000000ULL,
-			     29, 12, 16, &dev_priv->chan_vm);
+			     &dev_priv->chan_vm);
 	if (ret)
 		return ret;
 
@@ -258,7 +256,7 @@ nv50_instmem_takedown(struct drm_device *dev)
 	dev_priv->channels.ptr[127] = 0;
 	nv50_channel_del(&dev_priv->channels.ptr[0]);
 
-	nouveau_gpuobj_ref(NULL, &dev_priv->bar3_vm->pgt[0].obj);
+	nouveau_gpuobj_ref(NULL, &dev_priv->bar3_vm->pgt[0].obj[0]);
 	nouveau_vm_ref(NULL, &dev_priv->bar3_vm, NULL);
 
 	if (dev_priv->ramin_heap.free_stack.next)

drivers/gpu/drm/nouveau/nv50_vm.c

@@ -28,39 +28,40 @@
 #include "nouveau_vm.h"
 
 void
-nv50_vm_map_pgt(struct nouveau_gpuobj *pgd, u32 type, u32 pde,
-		struct nouveau_gpuobj *pgt)
+nv50_vm_map_pgt(struct nouveau_gpuobj *pgd, u32 pde,
+		struct nouveau_gpuobj *pgt[2])
 {
 	struct drm_nouveau_private *dev_priv = pgd->dev->dev_private;
-	u32 coverage = (pgt->size >> 3) << type;
-	u64 phys;
-
-	phys  = pgt->vinst;
-	phys |= 0x01; /* present */
-	phys |= (type == 12) ? 0x02 : 0x00; /* 4KiB pages */
-	if (dev_priv->vram_sys_base) {
-		phys += dev_priv->vram_sys_base;
-		phys |= 0x30;
+	u64 phys = 0xdeadcafe00000000ULL;
+	u32 coverage = 0;
+
+	if (pgt[0]) {
+		phys = 0x00000003 | pgt[0]->vinst; /* present, 4KiB pages */
+		coverage = (pgt[0]->size >> 3) << 12;
+	} else
+	if (pgt[1]) {
+		phys = 0x00000001 | pgt[1]->vinst; /* present */
+		coverage = (pgt[1]->size >> 3) << 16;
 	}
 
-	if (coverage <= 32 * 1024 * 1024)
-		phys |= 0x60;
-	else if (coverage <= 64 * 1024 * 1024)
-		phys |= 0x40;
-	else if (coverage < 128 * 1024 * 1024)
-		phys |= 0x20;
+	if (phys & 1) {
+		if (dev_priv->vram_sys_base) {
+			phys += dev_priv->vram_sys_base;
+			phys |= 0x30;
+		}
+
+		if (coverage <= 32 * 1024 * 1024)
+			phys |= 0x60;
+		else if (coverage <= 64 * 1024 * 1024)
+			phys |= 0x40;
+		else if (coverage < 128 * 1024 * 1024)
+			phys |= 0x20;
+	}
 
 	nv_wo32(pgd, (pde * 8) + 0, lower_32_bits(phys));
 	nv_wo32(pgd, (pde * 8) + 4, upper_32_bits(phys));
 }
 
-void
-nv50_vm_unmap_pgt(struct nouveau_gpuobj *pgd, u32 pde)
-{
-	nv_wo32(pgd, (pde * 8) + 0, 0x00000000);
-	nv_wo32(pgd, (pde * 8) + 4, 0xdeadcafe);
-}
-
 static inline u64
 nv50_vm_addr(struct nouveau_vma *vma, struct nouveau_gpuobj *pgt,
 	     u64 phys, u32 memtype, u32 target)