drm/nv50: import new vm code

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

11 files changed, 714 insertions, 22 deletions

diff --git a/drivers/gpu/drm/nouveau/Makefile b/drivers/gpu/drm/nouveau/Makefile
index 26fdd12561b6..b1d8941e04d8 100644
--- a/drivers/gpu/drm/nouveau/Makefile
+++ b/drivers/gpu/drm/nouveau/Makefile
@@ -9,8 +9,9 @@ nouveau-y := nouveau_drv.o nouveau_state.o nouveau_channel.o nouveau_mem.o \
              nouveau_bo.o nouveau_fence.o nouveau_gem.o nouveau_ttm.o \
              nouveau_hw.o nouveau_calc.o nouveau_bios.o nouveau_i2c.o \
              nouveau_display.o nouveau_connector.o nouveau_fbcon.o \
-             nouveau_dp.o nouveau_ramht.o nouveau_mm.o \
+             nouveau_dp.o nouveau_ramht.o \
              nouveau_pm.o nouveau_volt.o nouveau_perf.o nouveau_temp.o \
+             nouveau_mm.o nouveau_vm.o \
              nv04_timer.o \
              nv04_mc.o nv40_mc.o nv50_mc.o \
              nv04_fb.o nv10_fb.o nv30_fb.o nv40_fb.o nv50_fb.o nvc0_fb.o \
@@ -27,7 +28,7 @@ nouveau-y := nouveau_drv.o nouveau_state.o nouveau_channel.o nouveau_mem.o \
              nv10_gpio.o nv50_gpio.o \
              nv50_calc.o \
              nv04_pm.o nv50_pm.o nva3_pm.o \
-             nv50_vram.o
+             nv50_vram.o nv50_vm.o
 
 nouveau-$(CONFIG_DRM_NOUVEAU_DEBUG) += nouveau_debugfs.o
 nouveau-$(CONFIG_COMPAT) += nouveau_ioc32.o
diff --git a/drivers/gpu/drm/nouveau/nouveau_drv.h b/drivers/gpu/drm/nouveau/nouveau_drv.h
index 1305e2c94201..ce1dde4a65d6 100644
--- a/drivers/gpu/drm/nouveau/nouveau_drv.h
+++ b/drivers/gpu/drm/nouveau/nouveau_drv.h
@@ -153,6 +153,7 @@ enum nouveau_flags {
 #define NVOBJ_ENGINE_DISPLAY    0xcafe0001
 #define NVOBJ_ENGINE_INT        0xdeadbeef
 
+#define NVOBJ_FLAG_DONT_MAP             (1 << 0)
 #define NVOBJ_FLAG_ZERO_ALLOC           (1 << 1)
 #define NVOBJ_FLAG_ZERO_FREE            (1 << 2)
 
@@ -1213,7 +1214,6 @@ extern int  nv50_instmem_map(struct nouveau_gpuobj *);
 extern void nv50_instmem_unmap(struct nouveau_gpuobj *);
 extern void nv50_instmem_flush(struct drm_device *);
 extern void nv84_instmem_flush(struct drm_device *);
-extern void nv50_vm_flush(struct drm_device *, int engine);
 
 /* nvc0_instmem.c */
 extern int  nvc0_instmem_init(struct drm_device *);
@@ -1564,10 +1564,11 @@ nv_match_device(struct drm_device *dev, unsigned device,
 }
 
 /* memory type/access flags, do not match hardware values */
-#define NV_MEM_ACCESS_RO 1
-#define NV_MEM_ACCESS_WO 2
+#define NV_MEM_ACCESS_RO  1
+#define NV_MEM_ACCESS_WO  2
 #define NV_MEM_ACCESS_RW (NV_MEM_ACCESS_RO | NV_MEM_ACCESS_WO)
-#define NV_MEM_ACCESS_VM 4
+#define NV_MEM_ACCESS_SYS 4
+#define NV_MEM_ACCESS_VM  8
 
 #define NV_MEM_TARGET_VRAM        0
 #define NV_MEM_TARGET_PCI         1
diff --git a/drivers/gpu/drm/nouveau/nouveau_mem.c b/drivers/gpu/drm/nouveau/nouveau_mem.c
index dbeb9e5f6b22..2d02401e8227 100644
--- a/drivers/gpu/drm/nouveau/nouveau_mem.c
+++ b/drivers/gpu/drm/nouveau/nouveau_mem.c
@@ -37,6 +37,7 @@
 #include "nouveau_drv.h"
 #include "nouveau_pm.h"
 #include "nouveau_mm.h"
+#include "nouveau_vm.h"
 
 /*
  * NV10-NV40 tiling helpers
@@ -201,7 +202,7 @@ nv50_mem_vm_bind_linear(struct drm_device *dev, uint64_t virt, uint32_t size,
         dev_priv->engine.instmem.flush(dev);
         dev_priv->engine.fifo.tlb_flush(dev);
         dev_priv->engine.graph.tlb_flush(dev);
-        nv50_vm_flush(dev, 6);
+        nv50_vm_flush_engine(dev, 6);
         return 0;
 }
 
@@ -234,7 +235,7 @@ nv50_mem_vm_unbind(struct drm_device *dev, uint64_t virt, uint32_t size)
         dev_priv->engine.instmem.flush(dev);
         dev_priv->engine.fifo.tlb_flush(dev);
         dev_priv->engine.graph.tlb_flush(dev);
-        nv50_vm_flush(dev, 6);
+        nv50_vm_flush_engine(dev, 6);
 }
 
 /*
diff --git a/drivers/gpu/drm/nouveau/nouveau_object.c b/drivers/gpu/drm/nouveau/nouveau_object.c
index 5cc3f7e59fa1..dd1859f7d8b0 100644
--- a/drivers/gpu/drm/nouveau/nouveau_object.c
+++ b/drivers/gpu/drm/nouveau/nouveau_object.c
@@ -213,7 +213,7 @@ nouveau_gpuobj_new(struct drm_device *dev, struct nouveau_channel *chan,
                 }
 
                 ret = -ENOSYS;
-                if (dev_priv->ramin_available)
+                if (!(flags & NVOBJ_FLAG_DONT_MAP))
                         ret = instmem->map(gpuobj);
                 if (ret)
                         gpuobj->pinst = ~0;
diff --git a/drivers/gpu/drm/nouveau/nouveau_vm.c b/drivers/gpu/drm/nouveau/nouveau_vm.c
new file mode 100644
index 000000000000..07ab1749cf7d
--- /dev/null
+++ b/drivers/gpu/drm/nouveau/nouveau_vm.c
@@ -0,0 +1,421 @@
+/*
+ * 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;
+        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;
+
+                        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;
+        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;
+
+                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;
+        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;
+
+                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, 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)
+                        continue;
+
+                list_for_each_entry(vpgd, &vm->pgd_list, head) {
+                        vm->unmap_pgt(vpgd->obj, pde);
+                }
+
+                pgt = vpgt->obj;
+                vpgt->obj = NULL;
+
+                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;
+        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++)) {
+                mutex_unlock(&vm->mm->mutex);
+                nouveau_gpuobj_ref(NULL, &pgt);
+                mutex_lock(&vm->mm->mutex);
+                return 0;
+        }
+
+        list_for_each_entry(vpgd, &vm->pgd_list, head) {
+                vm->map_pgt(vpgd->obj, type, pde, pgt);
+        }
+
+        vpgt->page_shift = type;
+        vpgt->obj = pgt;
+        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];
+
+                if (likely(vpgt->refcount)) {
+                        vpgt->refcount++;
+                        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_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_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;
+        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->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;
+        } 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;
+        vm->spg_shift = spg_shift;
+        vm->lpg_shift = lpg_shift;
+
+        block = (1 << pgt_bits);
+        if (length < block)
+                block = length;
+
+        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++) {
+                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);
+        }
+        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;
+}
diff --git a/drivers/gpu/drm/nouveau/nouveau_vm.h b/drivers/gpu/drm/nouveau/nouveau_vm.h
new file mode 100644
index 000000000000..b6755cfa7b71
--- /dev/null
+++ b/drivers/gpu/drm/nouveau/nouveau_vm.h
@@ -0,0 +1,107 @@
+/*
+ * 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
+ */
+
+#ifndef __NOUVEAU_VM_H__
+#define __NOUVEAU_VM_H__
+
+#include "drmP.h"
+
+#include "nouveau_drv.h"
+#include "nouveau_mm.h"
+
+struct nouveau_vm_pgt {
+        struct nouveau_gpuobj *obj;
+        u32 page_shift;
+        u32 refcount;
+};
+
+struct nouveau_vm_pgd {
+        struct list_head head;
+        struct nouveau_gpuobj *obj;
+};
+
+struct nouveau_vma {
+        struct nouveau_vm *vm;
+        struct nouveau_mm_node *node;
+        u64 offset;
+        u32 access;
+};
+
+struct nouveau_vm {
+        struct drm_device *dev;
+        struct nouveau_mm *mm;
+        int refcount;
+
+        struct list_head pgd_list;
+        atomic_t pgraph_refs;
+        atomic_t pcrypt_refs;
+
+        struct nouveau_vm_pgt *pgt;
+        u32 fpde;
+        u32 lpde;
+
+        u32 pgt_bits;
+        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)(struct nouveau_vma *, struct nouveau_gpuobj *,
+                    struct nouveau_vram *, u32 pte, u32 cnt, u64 phys);
+        void (*map_sg)(struct nouveau_vma *, struct nouveau_gpuobj *,
+                       u32 pte, dma_addr_t *, u32 cnt);
+        void (*unmap)(struct nouveau_gpuobj *pgt, u32 pte, u32 cnt);
+        void (*flush)(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);
+int  nouveau_vm_get(struct nouveau_vm *, u64 size, u32 page_shift,
+                    u32 access, struct nouveau_vma *);
+void nouveau_vm_put(struct nouveau_vma *);
+void nouveau_vm_map(struct nouveau_vma *, struct nouveau_vram *);
+void nouveau_vm_map_at(struct nouveau_vma *, u64 offset, struct nouveau_vram *);
+void nouveau_vm_unmap(struct nouveau_vma *);
+void nouveau_vm_unmap_at(struct nouveau_vma *, u64 offset, u64 length);
+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(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 *,
+                    u32 pte, dma_addr_t *, u32 cnt);
+void nv50_vm_unmap(struct nouveau_gpuobj *, u32 pte, u32 cnt);
+void nv50_vm_flush(struct nouveau_vm *);
+void nv50_vm_flush_engine(struct drm_device *, int engine);
+
+#endif
diff --git a/drivers/gpu/drm/nouveau/nv50_fifo.c b/drivers/gpu/drm/nouveau/nv50_fifo.c
index 7add3dfde3df..8dd04c5dac67 100644
--- a/drivers/gpu/drm/nouveau/nv50_fifo.c
+++ b/drivers/gpu/drm/nouveau/nv50_fifo.c
@@ -28,6 +28,7 @@
 #include "drm.h"
 #include "nouveau_drv.h"
 #include "nouveau_ramht.h"
+#include "nouveau_vm.h"
 
 static void
 nv50_fifo_playlist_update(struct drm_device *dev)
@@ -498,5 +499,5 @@ nv50_fifo_unload_context(struct drm_device *dev)
 void
 nv50_fifo_tlb_flush(struct drm_device *dev)
 {
-        nv50_vm_flush(dev, 5);
+        nv50_vm_flush_engine(dev, 5);
 }
diff --git a/drivers/gpu/drm/nouveau/nv50_graph.c b/drivers/gpu/drm/nouveau/nv50_graph.c
index b3900788c66d..f5fd1b296d27 100644
--- a/drivers/gpu/drm/nouveau/nv50_graph.c
+++ b/drivers/gpu/drm/nouveau/nv50_graph.c
@@ -30,6 +30,7 @@
 #include "nouveau_ramht.h"
 #include "nouveau_grctx.h"
 #include "nouveau_dma.h"
+#include "nouveau_vm.h"
 #include "nv50_evo.h"
 
 static int  nv50_graph_register(struct drm_device *);
@@ -468,7 +469,7 @@ nv50_graph_register(struct drm_device *dev)
 void
 nv50_graph_tlb_flush(struct drm_device *dev)
 {
-        nv50_vm_flush(dev, 0);
+        nv50_vm_flush_engine(dev, 0);
 }
 
 void
@@ -511,7 +512,7 @@ nv86_graph_tlb_flush(struct drm_device *dev)
                          nv_rd32(dev, 0x400384), nv_rd32(dev, 0x400388));
         }
 
-        nv50_vm_flush(dev, 0);
+        nv50_vm_flush_engine(dev, 0);
 
         nv_mask(dev, 0x400500, 0x00000001, 0x00000001);
         spin_unlock_irqrestore(&dev_priv->context_switch_lock, flags);
diff --git a/drivers/gpu/drm/nouveau/nv50_instmem.c b/drivers/gpu/drm/nouveau/nv50_instmem.c
index 1e7d50397e4a..4eb2f0835e27 100644
--- a/drivers/gpu/drm/nouveau/nv50_instmem.c
+++ b/drivers/gpu/drm/nouveau/nv50_instmem.c
@@ -27,7 +27,9 @@
 
 #include "drmP.h"
 #include "drm.h"
+
 #include "nouveau_drv.h"
+#include "nouveau_vm.h"
 
 struct nv50_instmem_priv {
         uint32_t save1700[5]; /* 0x1700->0x1710 */
@@ -404,7 +406,7 @@ nv50_instmem_map(struct nouveau_gpuobj *gpuobj)
         }
         dev_priv->engine.instmem.flush(dev);
 
-        nv50_vm_flush(dev, 6);
+        nv50_vm_flush_engine(dev, 6);
 
         node->ramin   = ramin;
         gpuobj->pinst = ramin->start;
@@ -454,11 +456,3 @@ nv84_instmem_flush(struct drm_device *dev)
                 NV_ERROR(dev, "PRAMIN flush timeout\n");
 }
 
-void
-nv50_vm_flush(struct drm_device *dev, int engine)
-{
-        nv_wr32(dev, 0x100c80, (engine << 16) | 1);
-        if (!nv_wait(dev, 0x100c80, 0x00000001, 0x00000000))
-                NV_ERROR(dev, "vm flush timeout: engine %d\n", engine);
-}
-
diff --git a/drivers/gpu/drm/nouveau/nv50_vm.c b/drivers/gpu/drm/nouveau/nv50_vm.c
new file mode 100644
index 000000000000..ab6c3d0ce32e
--- /dev/null
+++ b/drivers/gpu/drm/nouveau/nv50_vm.c
@@ -0,0 +1,164 @@
+/*
+ * 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_vm.h"
+
+void
+nv50_vm_map_pgt(struct nouveau_gpuobj *pgd, u32 type, u32 pde,
+                struct nouveau_gpuobj *pgt)
+{
+        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;
+        }
+
+        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)
+{
+        struct drm_nouveau_private *dev_priv = pgt->dev->dev_private;
+
+        phys |= 1; /* present */
+        phys |= (u64)memtype << 40;
+
+        /* IGPs don't have real VRAM, re-target to stolen system memory */
+        if (target == 0 && dev_priv->vram_sys_base) {
+                phys  += dev_priv->vram_sys_base;
+                target = 3;
+        }
+
+        phys |= target << 4;
+
+        if (vma->access & NV_MEM_ACCESS_SYS)
+                phys |= (1 << 6);
+
+        if (!(vma->access & NV_MEM_ACCESS_WO))
+                phys |= (1 << 3);
+
+        return phys;
+}
+
+void
+nv50_vm_map(struct nouveau_vma *vma, struct nouveau_gpuobj *pgt,
+            struct nouveau_vram *mem, u32 pte, u32 cnt, u64 phys)
+{
+        u32 block, i;
+
+        phys  = nv50_vm_addr(vma, pgt, phys, mem->memtype, 0);
+        pte <<= 3;
+        cnt <<= 3;
+
+        while (cnt) {
+                u32 offset_h = upper_32_bits(phys);
+                u32 offset_l = lower_32_bits(phys);
+
+                for (i = 7; i >= 0; i--) {
+                        block = 1 << (i + 3);
+                        if (cnt >= block && !(pte & (block - 1)))
+                                break;
+                }
+                offset_l |= (i << 7);
+
+                phys += block << (vma->node->type - 3);
+                cnt  -= block;
+
+                while (block) {
+                        nv_wo32(pgt, pte + 0, offset_l);
+                        nv_wo32(pgt, pte + 4, offset_h);
+                        pte += 8;
+                        block -= 8;
+                }
+        }
+}
+
+void
+nv50_vm_map_sg(struct nouveau_vma *vma, struct nouveau_gpuobj *pgt,
+               u32 pte, dma_addr_t *list, u32 cnt)
+{
+        pte <<= 3;
+        while (cnt--) {
+                u64 phys = nv50_vm_addr(vma, pgt, (u64)*list++, 0, 2);
+                nv_wo32(pgt, pte + 0, lower_32_bits(phys));
+                nv_wo32(pgt, pte + 4, upper_32_bits(phys));
+                pte += 8;
+        }
+}
+
+void
+nv50_vm_unmap(struct nouveau_gpuobj *pgt, u32 pte, u32 cnt)
+{
+        pte <<= 3;
+        while (cnt--) {
+                nv_wo32(pgt, pte + 0, 0x00000000);
+                nv_wo32(pgt, pte + 4, 0x00000000);
+                pte += 8;
+        }
+}
+
+void
+nv50_vm_flush(struct nouveau_vm *vm)
+{
+        struct drm_nouveau_private *dev_priv = vm->dev->dev_private;
+        struct nouveau_instmem_engine *pinstmem = &dev_priv->engine.instmem;
+
+        pinstmem->flush(vm->dev);
+
+        nv50_vm_flush_engine(vm->dev, 6);
+}
+
+void
+nv50_vm_flush_engine(struct drm_device *dev, int engine)
+{
+        nv_wr32(dev, 0x100c80, (engine << 16) | 1);
+        if (!nv_wait(dev, 0x100c80, 0x00000001, 0x00000000))
+                NV_ERROR(dev, "vm flush timeout: engine %d\n", engine);
+}
diff --git a/drivers/gpu/drm/nouveau/nv84_crypt.c b/drivers/gpu/drm/nouveau/nv84_crypt.c
index 780bb1d66f95..a333e5905346 100644
--- a/drivers/gpu/drm/nouveau/nv84_crypt.c
+++ b/drivers/gpu/drm/nouveau/nv84_crypt.c
@@ -25,6 +25,7 @@
 #include "drmP.h"
 #include "nouveau_drv.h"
 #include "nouveau_util.h"
+#include "nouveau_vm.h"
 
 static void nv84_crypt_isr(struct drm_device *);
 
@@ -84,7 +85,7 @@ nv84_crypt_destroy_context(struct nouveau_channel *chan)
 void
 nv84_crypt_tlb_flush(struct drm_device *dev)
 {
-        nv50_vm_flush(dev, 0x0a);
+        nv50_vm_flush_engine(dev, 0x0a);
 }
 
 int