drivers/gpu/drm/nouveau/nouveau_vm.c - nest-cam/v179/linux - Git at Google

 /*
  * Copyright 2010 Red Hat Inc.
  *
  * Permission is hereby granted, free of charge, to any person obtaining a
  * copy of this software and associated documentation files (the "Software"),
  * to deal in the Software without restriction, including without limitation
  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
  * and/or sell copies of the Software, and to permit persons to whom the
  * Software is furnished to do so, subject to the following conditions:
  *
  * The above copyright notice and this permission notice shall be included in
  * all copies or substantial portions of the Software.
  *
  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  * OTHER DEALINGS IN THE SOFTWARE.
  *
  * Authors: Ben Skeggs
  */

 #include "drmP.h"
 #include "nouveau_drv.h"
 #include "nouveau_mm.h"
 #include "nouveau_vm.h"

 void
 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;
 	u32 pte  = (offset & ((1 << vm->pgt_bits) - 1)) >> bits;
 	u32 max  = 1 << (vm->pgt_bits - bits);
 	u32 end, len;

 	list_for_each_entry(r, &vram->regions, rl_entry) {
 		u64 phys = (u64)r->offset << 12;
 		u32 num  = r->length >> bits;

 		while (num) {
 			struct nouveau_gpuobj *pgt = vm->pgt[pde].obj[big];

 			end = (pte + num);
 			if (unlikely(end >= max))
 				end = max;
 			len = end - pte;

 			vm->map(vma, pgt, vram, pte, len, phys);

 			num -= len;
 			pte += len;
 			if (unlikely(end >= max)) {
 				pde++;
 				pte = 0;
 			}
 		}
 	}

 	vm->flush(vm);
 }

 void
 nouveau_vm_map(struct nouveau_vma *vma, struct nouveau_vram *vram)
 {
 	nouveau_vm_map_at(vma, 0, vram);
 }

 void
 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;
 	u32 pde  = (offset >> vm->pgt_bits) - vm->fpde;
 	u32 pte  = (offset & ((1 << vm->pgt_bits) - 1)) >> bits;
 	u32 max  = 1 << (vm->pgt_bits - bits);
 	u32 end, len;

 	while (num) {
 		struct nouveau_gpuobj *pgt = vm->pgt[pde].obj[big];

 		end = (pte + num);
 		if (unlikely(end >= max))
 			end = max;
 		len = end - pte;

 		vm->map_sg(vma, pgt, pte, list, len);

 		num  -= len;
 		pte  += len;
 		list += len;
 		if (unlikely(end >= max)) {
 			pde++;
 			pte = 0;
 		}
 	}

 	vm->flush(vm);
 }

 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;
 	u32 pde  = (offset >> vm->pgt_bits) - vm->fpde;
 	u32 pte  = (offset & ((1 << vm->pgt_bits) - 1)) >> bits;
 	u32 max  = 1 << (vm->pgt_bits - bits);
 	u32 end, len;

 	while (num) {
 		struct nouveau_gpuobj *pgt = vm->pgt[pde].obj[big];

 		end = (pte + num);
 		if (unlikely(end >= max))
 			end = max;
 		len = end - pte;

 		vm->unmap(pgt, pte, len);

 		num -= len;
 		pte += len;
 		if (unlikely(end >= max)) {
 			pde++;
 			pte = 0;
 		}
 	}

 	vm->flush(vm);
 }

 void
 nouveau_vm_unmap(struct nouveau_vma *vma)
 {
 	nouveau_vm_unmap_at(vma, 0, (u64)vma->node->length << 12);
 }

 static void
 nouveau_vm_unmap_pgt(struct nouveau_vm *vm, int big, u32 fpde, u32 lpde)
 {
 	struct nouveau_vm_pgd *vpgd;
 	struct nouveau_vm_pgt *vpgt;
 	struct nouveau_gpuobj *pgt;
 	u32 pde;

 	for (pde = fpde; pde <= lpde; pde++) {
 		vpgt = &vm->pgt[pde - vm->fpde];
 		if (--vpgt->refcount[big])
 			continue;

 		pgt = vpgt->obj[big];
 		vpgt->obj[big] = NULL;

 		list_for_each_entry(vpgd, &vm->pgd_list, head) {
 			vm->map_pgt(vpgd->obj, pde, vpgt->obj);
 		}

 		mutex_unlock(&vm->mm->mutex);
 		nouveau_gpuobj_ref(NULL, &pgt);
 		mutex_lock(&vm->mm->mutex);
 	}
 }

 static int
 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;

 	pgt_size  = (1 << (vm->pgt_bits + 12)) >> type;
 	pgt_size *= 8;

 	mutex_unlock(&vm->mm->mutex);
 	ret = nouveau_gpuobj_new(vm->dev, NULL, pgt_size, 0x1000,
 				 NVOBJ_FLAG_ZERO_ALLOC, &pgt);
 	mutex_lock(&vm->mm->mutex);
 	if (unlikely(ret))
 		return ret;

 	/* someone beat us to filling the PDE while we didn't have the lock */
 	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, pde, vpgt->obj);
 	}

 	return 0;
 }

 int
 nouveau_vm_get(struct nouveau_vm *vm, u64 size, u32 page_shift,
 	       u32 access, struct nouveau_vma *vma)
 {
 	u32 align = (1 << page_shift) >> 12;
 	u32 msize = size >> 12;
 	u32 fpde, lpde, pde;
 	int ret;

 	mutex_lock(&vm->mm->mutex);
 	ret = nouveau_mm_get(vm->mm, page_shift, msize, 0, align, &vma->node);
 	if (unlikely(ret != 0)) {
 		mutex_unlock(&vm->mm->mutex);
 		return ret;
 	}

 	fpde = (vma->node->offset >> vm->pgt_bits);
 	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[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, big, fpde, pde - 1);
 			nouveau_mm_put(vm->mm, vma->node);
 			mutex_unlock(&vm->mm->mutex);
 			vma->node = NULL;
 			return ret;
 		}
 	}
 	mutex_unlock(&vm->mm->mutex);

 	vma->vm     = vm;
 	vma->offset = (u64)vma->node->offset << 12;
 	vma->access = access;
 	return 0;
 }

 void
 nouveau_vm_put(struct nouveau_vma *vma)
 {
 	struct nouveau_vm *vm = vma->vm;
 	u32 fpde, lpde;

 	if (unlikely(vma->node == NULL))
 		return;
 	fpde = (vma->node->offset >> vm->pgt_bits);
 	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;
 	mutex_unlock(&vm->mm->mutex);
 }

 int
 nouveau_vm_new(struct drm_device *dev, u64 offset, u64 length, u64 mm_offset,
 	       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, pgt_bits;
 	int ret;

 	vm = kzalloc(sizeof(*vm), GFP_KERNEL);
 	if (!vm)
 		return -ENOMEM;

 	if (dev_priv->card_type == NV_50) {
 		vm->map_pgt = nv50_vm_map_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;
 		block = (1 << pgt_bits);
 		if (length < block)
 			block = length;

 	} else
 	if (dev_priv->card_type == NV_C0) {
 		vm->map_pgt = nvc0_vm_map_pgt;
 		vm->map = nvc0_vm_map;
 		vm->map_sg = nvc0_vm_map_sg;
 		vm->unmap = nvc0_vm_unmap;
 		vm->flush = nvc0_vm_flush;
 		vm->spg_shift = 12;
 		vm->lpg_shift = 17;
 		pgt_bits = 27;

 		/* Should be 4096 everywhere, this is a hack that's
 		 * currently necessary to avoid an elusive bug that
 		 * causes corruption when mixing small/large pages
 		 */
 		if (length < (1ULL << 40))
 			block = 4096;
 		else {
 			block = (1 << pgt_bits);
 			if (length < block)
 				block = length;
 		}
 	} else {
 		kfree(vm);
 		return -ENOSYS;
 	}

 	vm->fpde   = offset >> pgt_bits;
 	vm->lpde   = (offset + length - 1) >> pgt_bits;
 	vm->pgt = kcalloc(vm->lpde - vm->fpde + 1, sizeof(*vm->pgt), GFP_KERNEL);
 	if (!vm->pgt) {
 		kfree(vm);
 		return -ENOMEM;
 	}

 	INIT_LIST_HEAD(&vm->pgd_list);
 	vm->dev = dev;
 	vm->refcount = 1;
 	vm->pgt_bits = pgt_bits - 12;

 	ret = nouveau_mm_init(&vm->mm, mm_offset >> 12, mm_length >> 12,
 			      block >> 12);
 	if (ret) {
 		kfree(vm);
 		return ret;
 	}

 	*pvm = vm;
 	return 0;
 }

 static int
 nouveau_vm_link(struct nouveau_vm *vm, struct nouveau_gpuobj *pgd)
 {
 	struct nouveau_vm_pgd *vpgd;
 	int i;

 	if (!pgd)
 		return 0;

 	vpgd = kzalloc(sizeof(*vpgd), GFP_KERNEL);
 	if (!vpgd)
 		return -ENOMEM;

 	nouveau_gpuobj_ref(pgd, &vpgd->obj);

 	mutex_lock(&vm->mm->mutex);
 	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;
 }

 static void
 nouveau_vm_unlink(struct nouveau_vm *vm, struct nouveau_gpuobj *pgd)
 {
 	struct nouveau_vm_pgd *vpgd, *tmp;

 	if (!pgd)
 		return;

 	mutex_lock(&vm->mm->mutex);
 	list_for_each_entry_safe(vpgd, tmp, &vm->pgd_list, head) {
 		if (vpgd->obj != pgd)
 			continue;

 		list_del(&vpgd->head);
 		nouveau_gpuobj_ref(NULL, &vpgd->obj);
 		kfree(vpgd);
 	}
 	mutex_unlock(&vm->mm->mutex);
 }

 static void
 nouveau_vm_del(struct nouveau_vm *vm)
 {
 	struct nouveau_vm_pgd *vpgd, *tmp;

 	list_for_each_entry_safe(vpgd, tmp, &vm->pgd_list, head) {
 		nouveau_vm_unlink(vm, vpgd->obj);
 	}
 	WARN_ON(nouveau_mm_fini(&vm->mm) != 0);

 	kfree(vm->pgt);
 	kfree(vm);
 }

 int
 nouveau_vm_ref(struct nouveau_vm *ref, struct nouveau_vm **ptr,
 	       struct nouveau_gpuobj *pgd)
 {
 	struct nouveau_vm *vm;
 	int ret;

 	vm = ref;
 	if (vm) {
 		ret = nouveau_vm_link(vm, pgd);
 		if (ret)
 			return ret;

 		vm->refcount++;
 	}

 	vm = *ptr;
 	*ptr = ref;

 	if (vm) {
 		nouveau_vm_unlink(vm, pgd);

 		if (--vm->refcount == 0)
 			nouveau_vm_del(vm);
 	}

 	return 0;
 }
	/*
	* Copyright 2010 Red Hat Inc.
	*
	* Permission is hereby granted, free of charge, to any person obtaining a
	* copy of this software and associated documentation files (the "Software"),
	* to deal in the Software without restriction, including without limitation
	* the rights to use, copy, modify, merge, publish, distribute, sublicense,
	* and/or sell copies of the Software, and to permit persons to whom the
	* Software is furnished to do so, subject to the following conditions:
	*
	* The above copyright notice and this permission notice shall be included in
	* all copies or substantial portions of the Software.
	*
	* THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
	* IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
	* FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
	* THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
	* OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
	* ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
	* OTHER DEALINGS IN THE SOFTWARE.
	*
	* Authors: Ben Skeggs
	*/

	#include "drmP.h"
	#include "nouveau_drv.h"
	#include "nouveau_mm.h"
	#include "nouveau_vm.h"

	void
	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;
	u32 pte = (offset & ((1 << vm->pgt_bits) - 1)) >> bits;
	u32 max = 1 << (vm->pgt_bits - bits);
	u32 end, len;

	list_for_each_entry(r, &vram->regions, rl_entry) {
	u64 phys = (u64)r->offset << 12;
	u32 num = r->length >> bits;

	while (num) {
	struct nouveau_gpuobj *pgt = vm->pgt[pde].obj[big];

	end = (pte + num);
	if (unlikely(end >= max))
	end = max;
	len = end - pte;

	vm->map(vma, pgt, vram, pte, len, phys);

	num -= len;
	pte += len;
	if (unlikely(end >= max)) {
	pde++;
	pte = 0;
	}
	}
	}

	vm->flush(vm);
	}

	void
	nouveau_vm_map(struct nouveau_vma vma, struct nouveau_vram vram)
	{
	nouveau_vm_map_at(vma, 0, vram);
	}

	void
	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;
	u32 pde = (offset >> vm->pgt_bits) - vm->fpde;
	u32 pte = (offset & ((1 << vm->pgt_bits) - 1)) >> bits;
	u32 max = 1 << (vm->pgt_bits - bits);
	u32 end, len;

	while (num) {
	struct nouveau_gpuobj *pgt = vm->pgt[pde].obj[big];

	end = (pte + num);
	if (unlikely(end >= max))
	end = max;
	len = end - pte;

	vm->map_sg(vma, pgt, pte, list, len);

	num -= len;
	pte += len;
	list += len;
	if (unlikely(end >= max)) {
	pde++;
	pte = 0;
	}
	}

	vm->flush(vm);
	}

	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;
	u32 pde = (offset >> vm->pgt_bits) - vm->fpde;
	u32 pte = (offset & ((1 << vm->pgt_bits) - 1)) >> bits;
	u32 max = 1 << (vm->pgt_bits - bits);
	u32 end, len;

	while (num) {
	struct nouveau_gpuobj *pgt = vm->pgt[pde].obj[big];

	end = (pte + num);
	if (unlikely(end >= max))
	end = max;
	len = end - pte;

	vm->unmap(pgt, pte, len);

	num -= len;
	pte += len;
	if (unlikely(end >= max)) {
	pde++;
	pte = 0;
	}
	}

	vm->flush(vm);
	}

	void
	nouveau_vm_unmap(struct nouveau_vma *vma)
	{
	nouveau_vm_unmap_at(vma, 0, (u64)vma->node->length << 12);
	}

	static void
	nouveau_vm_unmap_pgt(struct nouveau_vm *vm, int big, u32 fpde, u32 lpde)
	{
	struct nouveau_vm_pgd *vpgd;
	struct nouveau_vm_pgt *vpgt;
	struct nouveau_gpuobj *pgt;
	u32 pde;

	for (pde = fpde; pde <= lpde; pde++) {
	vpgt = &vm->pgt[pde - vm->fpde];
	if (--vpgt->refcount[big])
	continue;

	pgt = vpgt->obj[big];
	vpgt->obj[big] = NULL;

	list_for_each_entry(vpgd, &vm->pgd_list, head) {
	vm->map_pgt(vpgd->obj, pde, vpgt->obj);
	}

	mutex_unlock(&vm->mm->mutex);
	nouveau_gpuobj_ref(NULL, &pgt);
	mutex_lock(&vm->mm->mutex);
	}
	}

	static int
	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;

	pgt_size = (1 << (vm->pgt_bits + 12)) >> type;
	pgt_size *= 8;

	mutex_unlock(&vm->mm->mutex);
	ret = nouveau_gpuobj_new(vm->dev, NULL, pgt_size, 0x1000,
	NVOBJ_FLAG_ZERO_ALLOC, &pgt);
	mutex_lock(&vm->mm->mutex);
	if (unlikely(ret))
	return ret;

	/* someone beat us to filling the PDE while we didn't have the lock */
	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, pde, vpgt->obj);
	}

	return 0;
	}

	int
	nouveau_vm_get(struct nouveau_vm *vm, u64 size, u32 page_shift,
	u32 access, struct nouveau_vma *vma)
	{
	u32 align = (1 << page_shift) >> 12;
	u32 msize = size >> 12;
	u32 fpde, lpde, pde;
	int ret;

	mutex_lock(&vm->mm->mutex);
	ret = nouveau_mm_get(vm->mm, page_shift, msize, 0, align, &vma->node);
	if (unlikely(ret != 0)) {
	mutex_unlock(&vm->mm->mutex);
	return ret;
	}

	fpde = (vma->node->offset >> vm->pgt_bits);
	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[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, big, fpde, pde - 1);
	nouveau_mm_put(vm->mm, vma->node);
	mutex_unlock(&vm->mm->mutex);
	vma->node = NULL;
	return ret;
	}
	}
	mutex_unlock(&vm->mm->mutex);

	vma->vm = vm;
	vma->offset = (u64)vma->node->offset << 12;
	vma->access = access;
	return 0;
	}

	void
	nouveau_vm_put(struct nouveau_vma *vma)
	{
	struct nouveau_vm *vm = vma->vm;
	u32 fpde, lpde;

	if (unlikely(vma->node == NULL))
	return;
	fpde = (vma->node->offset >> vm->pgt_bits);
	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;
	mutex_unlock(&vm->mm->mutex);
	}

	int
	nouveau_vm_new(struct drm_device *dev, u64 offset, u64 length, u64 mm_offset,
	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, pgt_bits;
	int ret;

	vm = kzalloc(sizeof(*vm), GFP_KERNEL);
	if (!vm)
	return -ENOMEM;

	if (dev_priv->card_type == NV_50) {
	vm->map_pgt = nv50_vm_map_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;
	block = (1 << pgt_bits);
	if (length < block)
	block = length;

	} else
	if (dev_priv->card_type == NV_C0) {
	vm->map_pgt = nvc0_vm_map_pgt;
	vm->map = nvc0_vm_map;
	vm->map_sg = nvc0_vm_map_sg;
	vm->unmap = nvc0_vm_unmap;
	vm->flush = nvc0_vm_flush;
	vm->spg_shift = 12;
	vm->lpg_shift = 17;
	pgt_bits = 27;

	/* Should be 4096 everywhere, this is a hack that's
	* currently necessary to avoid an elusive bug that
	* causes corruption when mixing small/large pages
	*/
	if (length < (1ULL << 40))
	block = 4096;
	else {
	block = (1 << pgt_bits);
	if (length < block)
	block = length;
	}
	} else {
	kfree(vm);
	return -ENOSYS;
	}

	vm->fpde = offset >> pgt_bits;
	vm->lpde = (offset + length - 1) >> pgt_bits;
	vm->pgt = kcalloc(vm->lpde - vm->fpde + 1, sizeof(*vm->pgt), GFP_KERNEL);
	if (!vm->pgt) {
	kfree(vm);
	return -ENOMEM;
	}

	INIT_LIST_HEAD(&vm->pgd_list);
	vm->dev = dev;
	vm->refcount = 1;
	vm->pgt_bits = pgt_bits - 12;

	ret = nouveau_mm_init(&vm->mm, mm_offset >> 12, mm_length >> 12,
	block >> 12);
	if (ret) {
	kfree(vm);
	return ret;
	}

	*pvm = vm;
	return 0;
	}

	static int
	nouveau_vm_link(struct nouveau_vm vm, struct nouveau_gpuobj pgd)
	{
	struct nouveau_vm_pgd *vpgd;
	int i;

	if (!pgd)
	return 0;

	vpgd = kzalloc(sizeof(*vpgd), GFP_KERNEL);
	if (!vpgd)
	return -ENOMEM;

	nouveau_gpuobj_ref(pgd, &vpgd->obj);

	mutex_lock(&vm->mm->mutex);
	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;
	}

	static void
	nouveau_vm_unlink(struct nouveau_vm vm, struct nouveau_gpuobj pgd)
	{
	struct nouveau_vm_pgd vpgd, tmp;

	if (!pgd)
	return;

	mutex_lock(&vm->mm->mutex);
	list_for_each_entry_safe(vpgd, tmp, &vm->pgd_list, head) {
	if (vpgd->obj != pgd)
	continue;

	list_del(&vpgd->head);
	nouveau_gpuobj_ref(NULL, &vpgd->obj);
	kfree(vpgd);
	}
	mutex_unlock(&vm->mm->mutex);
	}

	static void
	nouveau_vm_del(struct nouveau_vm *vm)
	{
	struct nouveau_vm_pgd vpgd, tmp;

	list_for_each_entry_safe(vpgd, tmp, &vm->pgd_list, head) {
	nouveau_vm_unlink(vm, vpgd->obj);
	}
	WARN_ON(nouveau_mm_fini(&vm->mm) != 0);

	kfree(vm->pgt);
	kfree(vm);
	}

	int
	nouveau_vm_ref(struct nouveau_vm ref, struct nouveau_vm *ptr,
	struct nouveau_gpuobj *pgd)
	{
	struct nouveau_vm *vm;
	int ret;

	vm = ref;
	if (vm) {
	ret = nouveau_vm_link(vm, pgd);
	if (ret)
	return ret;

	vm->refcount++;
	}

	vm = *ptr;
	*ptr = ref;

	if (vm) {
	nouveau_vm_unlink(vm, pgd);

	if (--vm->refcount == 0)
	nouveau_vm_del(vm);
	}

	return 0;
	}