drm/nouveau/mmu: handle instance block setup

We previously required each VMM user to allocate their own page directory
and fill in the instance block themselves.

It makes more sense to handle this in a common location.

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

20 files changed, 93 insertions, 192 deletions

diff --git a/drivers/gpu/drm/nouveau/include/nvkm/engine/falcon.h b/drivers/gpu/drm/nouveau/include/nvkm/engine/falcon.h
index 8f9d132fbedf..f0024fb5a5af 100644
--- a/drivers/gpu/drm/nouveau/include/nvkm/engine/falcon.h
+++ b/drivers/gpu/drm/nouveau/include/nvkm/engine/falcon.h
@@ -78,7 +78,7 @@ struct nvkm_falcon_func {
 	void (*load_imem)(struct nvkm_falcon *, void *, u32, u32, u16, u8, bool);
 	void (*load_dmem)(struct nvkm_falcon *, void *, u32, u32, u8);
 	void (*read_dmem)(struct nvkm_falcon *, u32, u32, u8, void *);
-	void (*bind_context)(struct nvkm_falcon *, struct nvkm_gpuobj *);
+	void (*bind_context)(struct nvkm_falcon *, struct nvkm_memory *);
 	int (*wait_for_halt)(struct nvkm_falcon *, u32);
 	int (*clear_interrupt)(struct nvkm_falcon *, u32);
 	void (*set_start_addr)(struct nvkm_falcon *, u32 start_addr);
@@ -113,7 +113,7 @@ void nvkm_falcon_load_imem(struct nvkm_falcon *, void *, u32, u32, u16, u8,
 			   bool);
 void nvkm_falcon_load_dmem(struct nvkm_falcon *, void *, u32, u32, u8);
 void nvkm_falcon_read_dmem(struct nvkm_falcon *, u32, u32, u8, void *);
-void nvkm_falcon_bind_context(struct nvkm_falcon *, struct nvkm_gpuobj *);
+void nvkm_falcon_bind_context(struct nvkm_falcon *, struct nvkm_memory *);
 void nvkm_falcon_set_start_addr(struct nvkm_falcon *, u32);
 void nvkm_falcon_start(struct nvkm_falcon *);
 int nvkm_falcon_wait_for_halt(struct nvkm_falcon *, u32);
diff --git a/drivers/gpu/drm/nouveau/include/nvkm/subdev/mmu.h b/drivers/gpu/drm/nouveau/include/nvkm/subdev/mmu.h
index 71d4863378f4..eb6704a2fc8f 100644
--- a/drivers/gpu/drm/nouveau/include/nvkm/subdev/mmu.h
+++ b/drivers/gpu/drm/nouveau/include/nvkm/subdev/mmu.h
@@ -10,11 +10,6 @@ struct nvkm_vm_pgt {
 	u32 refcount[2];
 };
 
-struct nvkm_vm_pgd {
-	struct list_head head;
-	struct nvkm_gpuobj *obj;
-};
-
 struct nvkm_vma {
 	struct nvkm_vm *vm;
 	struct nvkm_mm_node *node;
@@ -40,7 +35,6 @@ struct nvkm_vm {
 
 	struct nvkm_mm mm;
 	struct kref refcount;
-	struct list_head pgd_list;
 	struct nvkm_vm_pgt *pgt;
 	u32 fpde;
 	u32 lpde;
@@ -54,7 +48,7 @@ struct nvkm_vm {
 
 int  nvkm_vm_new(struct nvkm_device *, u64 offset, u64 length, u64 mm_offset,
 		 struct lock_class_key *, struct nvkm_vm **);
-int  nvkm_vm_ref(struct nvkm_vm *, struct nvkm_vm **, struct nvkm_gpuobj *pgd);
+int  nvkm_vm_ref(struct nvkm_vm *, struct nvkm_vm **, struct nvkm_memory *inst);
 int  nvkm_vm_boot(struct nvkm_vm *, u64 size);
 int  nvkm_vm_get(struct nvkm_vm *, u64 size, u32 page_shift, u32 access,
 		 struct nvkm_vma *);
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/chang84.c b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/chang84.c
index 61797c4dd07a..705a6d951d32 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/chang84.c
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/chang84.c
@@ -281,5 +281,5 @@ g84_fifo_chan_ctor(struct nv50_fifo *fifo, u64 vm, u64 push,
 	if (ret)
 		return ret;
 
-	return nvkm_vm_ref(chan->base.vm, &chan->vm, chan->pgd);
+	return nvkm_vm_ref(chan->base.vm, &chan->vm, chan->base.inst->memory);
 }
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/changf100.h b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/changf100.h
index 7d697e2dce1a..a902f848a878 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/changf100.h
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/changf100.h
@@ -11,7 +11,6 @@ struct gf100_fifo_chan {
 	struct list_head head;
 	bool killed;
 
-	struct nvkm_gpuobj *pgd;
 	struct nvkm_vm *vm;
 
 	struct {
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/changk104.h b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/changk104.h
index 230f64e5f731..1cd5b8d3fb2e 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/changk104.h
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/changk104.h
@@ -12,7 +12,6 @@ struct gk104_fifo_chan {
 	struct list_head head;
 	bool killed;
 
-	struct nvkm_gpuobj *pgd;
 	struct nvkm_vm *vm;
 
 	struct {
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/channv50.c b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/channv50.c
index 25b60aff40e4..161aa9ec73bc 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/channv50.c
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/channv50.c
@@ -206,7 +206,8 @@ void *
 nv50_fifo_chan_dtor(struct nvkm_fifo_chan *base)
 {
 	struct nv50_fifo_chan *chan = nv50_fifo_chan(base);
-	nvkm_vm_ref(NULL, &chan->vm, chan->pgd);
+	if (chan->base.inst)
+		nvkm_vm_ref(NULL, &chan->vm, chan->base.inst->memory);
 	nvkm_ramht_del(&chan->ramht);
 	nvkm_gpuobj_del(&chan->pgd);
 	nvkm_gpuobj_del(&chan->eng);
@@ -266,5 +267,5 @@ nv50_fifo_chan_ctor(struct nv50_fifo *fifo, u64 vm, u64 push,
 	if (ret)
 		return ret;
 
-	return nvkm_vm_ref(chan->base.vm, &chan->vm, chan->pgd);
+	return nvkm_vm_ref(chan->base.vm, &chan->vm, chan->base.inst->memory);
 }
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/gpfifogf100.c b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/gpfifogf100.c
index 6343e0c84d96..695d0721b1ad 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/gpfifogf100.c
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/gpfifogf100.c
@@ -200,8 +200,8 @@ static void *
 gf100_fifo_gpfifo_dtor(struct nvkm_fifo_chan *base)
 {
 	struct gf100_fifo_chan *chan = gf100_fifo_chan(base);
-	nvkm_vm_ref(NULL, &chan->vm, chan->pgd);
-	nvkm_gpuobj_del(&chan->pgd);
+	if (chan->base.inst)
+		nvkm_vm_ref(NULL, &chan->vm, chan->base.inst->memory);
 	return chan;
 }
 
@@ -225,7 +225,6 @@ gf100_fifo_gpfifo_new(struct nvkm_fifo *base, const struct nvkm_oclass *oclass,
 		struct fermi_channel_gpfifo_v0 v0;
 	} *args = data;
 	struct gf100_fifo *fifo = gf100_fifo(base);
-	struct nvkm_device *device = fifo->base.engine.subdev.device;
 	struct nvkm_object *parent = oclass->parent;
 	struct gf100_fifo_chan *chan;
 	u64 usermem, ioffset, ilength;
@@ -263,19 +262,7 @@ gf100_fifo_gpfifo_new(struct nvkm_fifo *base, const struct nvkm_oclass *oclass,
 
 	args->v0.chid = chan->base.chid;
 
-	/* page directory */
-	ret = nvkm_gpuobj_new(device, 0x10000, 0x1000, false, NULL, &chan->pgd);
-	if (ret)
-		return ret;
-
-	nvkm_kmap(chan->base.inst);
-	nvkm_wo32(chan->base.inst, 0x0200, lower_32_bits(chan->pgd->addr));
-	nvkm_wo32(chan->base.inst, 0x0204, upper_32_bits(chan->pgd->addr));
-	nvkm_wo32(chan->base.inst, 0x0208, 0xffffffff);
-	nvkm_wo32(chan->base.inst, 0x020c, 0x000000ff);
-	nvkm_done(chan->base.inst);
-
-	ret = nvkm_vm_ref(chan->base.vm, &chan->vm, chan->pgd);
+	ret = nvkm_vm_ref(chan->base.vm, &chan->vm, chan->base.inst->memory);
 	if (ret)
 		return ret;
 
diff --git a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/gpfifogk104.c b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/gpfifogk104.c
index 6e93b8652c66..978c9ad952b3 100644
--- a/drivers/gpu/drm/nouveau/nvkm/engine/fifo/gpfifogk104.c
+++ b/drivers/gpu/drm/nouveau/nvkm/engine/fifo/gpfifogk104.c
@@ -213,8 +213,8 @@ static void *
 gk104_fifo_gpfifo_dtor(struct nvkm_fifo_chan *base)
 {
 	struct gk104_fifo_chan *chan = gk104_fifo_chan(base);
-	nvkm_vm_ref(NULL, &chan->vm, chan->pgd);
-	nvkm_gpuobj_del(&chan->pgd);
+	if (chan->base.inst)
+		nvkm_vm_ref(NULL, &chan->vm, chan->base.inst->memory);
 	return chan;
 }
 
@@ -242,7 +242,6 @@ gk104_fifo_gpfifo_new_(const struct gk104_fifo_chan_func *func,
 		       const struct nvkm_oclass *oclass,
 		       struct nvkm_object **pobject)
 {
-	struct nvkm_device *device = fifo->base.engine.subdev.device;
 	struct gk104_fifo_chan *chan;
 	int runlist = -1, ret = -ENOSYS, i, j;
 	u32 engines = 0, present = 0;
@@ -302,19 +301,7 @@ gk104_fifo_gpfifo_new_(const struct gk104_fifo_chan_func *func,
 
 	*chid = chan->base.chid;
 
-	/* Page directory. */
-	ret = nvkm_gpuobj_new(device, 0x10000, 0x1000, false, NULL, &chan->pgd);
-	if (ret)
-		return ret;
-
-	nvkm_kmap(chan->base.inst);
-	nvkm_wo32(chan->base.inst, 0x0200, lower_32_bits(chan->pgd->addr));
-	nvkm_wo32(chan->base.inst, 0x0204, upper_32_bits(chan->pgd->addr));
-	nvkm_wo32(chan->base.inst, 0x0208, 0xffffffff);
-	nvkm_wo32(chan->base.inst, 0x020c, 0x000000ff);
-	nvkm_done(chan->base.inst);
-
-	ret = nvkm_vm_ref(chan->base.vm, &chan->vm, chan->pgd);
+	ret = nvkm_vm_ref(chan->base.vm, &chan->vm, chan->base.inst->memory);
 	if (ret)
 		return ret;
 
diff --git a/drivers/gpu/drm/nouveau/nvkm/falcon/base.c b/drivers/gpu/drm/nouveau/nvkm/falcon/base.c
index 1b7f48efd8b1..14be41f24155 100644
--- a/drivers/gpu/drm/nouveau/nvkm/falcon/base.c
+++ b/drivers/gpu/drm/nouveau/nvkm/falcon/base.c
@@ -60,7 +60,7 @@ nvkm_falcon_read_dmem(struct nvkm_falcon *falcon, u32 start, u32 size, u8 port,
 }
 
 void
-nvkm_falcon_bind_context(struct nvkm_falcon *falcon, struct nvkm_gpuobj *inst)
+nvkm_falcon_bind_context(struct nvkm_falcon *falcon, struct nvkm_memory *inst)
 {
 	if (!falcon->func->bind_context) {
 		nvkm_error(falcon->user,
diff --git a/drivers/gpu/drm/nouveau/nvkm/falcon/v1.c b/drivers/gpu/drm/nouveau/nvkm/falcon/v1.c
index 669c24028470..9def926f24d4 100644
--- a/drivers/gpu/drm/nouveau/nvkm/falcon/v1.c
+++ b/drivers/gpu/drm/nouveau/nvkm/falcon/v1.c
@@ -180,7 +180,7 @@ nvkm_falcon_v1_read_dmem(struct nvkm_falcon *falcon, u32 start, u32 size,
 }
 
 static void
-nvkm_falcon_v1_bind_context(struct nvkm_falcon *falcon, struct nvkm_gpuobj *ctx)
+nvkm_falcon_v1_bind_context(struct nvkm_falcon *falcon, struct nvkm_memory *ctx)
 {
 	u32 inst_loc;
 	u32 fbif;
@@ -216,7 +216,7 @@ nvkm_falcon_v1_bind_context(struct nvkm_falcon *falcon, struct nvkm_gpuobj *ctx)
 	nvkm_falcon_wr32(falcon, fbif + 4 * FALCON_DMAIDX_PHYS_SYS_NCOH, 0x6);
 
 	/* Set context */
-	switch (nvkm_memory_target(ctx->memory)) {
+	switch (nvkm_memory_target(ctx)) {
 	case NVKM_MEM_TARGET_VRAM: inst_loc = 0; break;
 	case NVKM_MEM_TARGET_HOST: inst_loc = 2; break;
 	case NVKM_MEM_TARGET_NCOH: inst_loc = 3; break;
@@ -228,7 +228,7 @@ nvkm_falcon_v1_bind_context(struct nvkm_falcon *falcon, struct nvkm_gpuobj *ctx)
 	/* Enable context */
 	nvkm_falcon_mask(falcon, 0x048, 0x1, 0x1);
 	nvkm_falcon_wr32(falcon, 0x054,
-			 ((ctx->addr >> 12) & 0xfffffff) |
+			 ((nvkm_memory_addr(ctx) >> 12) & 0xfffffff) |
 			 (inst_loc << 28) | (1 << 30));
 
 	nvkm_falcon_mask(falcon, 0x090, 0x10000, 0x10000);
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/bar/gf100.c b/drivers/gpu/drm/nouveau/nvkm/subdev/bar/gf100.c
index 8b588d1c776b..8077e1a5017a 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/bar/gf100.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/bar/gf100.c
@@ -23,7 +23,7 @@
  */
 #include "gf100.h"
 
-#include <core/gpuobj.h>
+#include <core/memory.h>
 #include <core/option.h>
 #include <subdev/fb.h>
 #include <subdev/mmu.h>
@@ -53,7 +53,7 @@ gf100_bar_bar1_init(struct nvkm_bar *base)
 {
 	struct nvkm_device *device = base->subdev.device;
 	struct gf100_bar *bar = gf100_bar(base);
-	const u32 addr = nvkm_memory_addr(bar->bar[1].mem) >> 12;
+	const u32 addr = nvkm_memory_addr(bar->bar[1].inst) >> 12;
 	nvkm_wr32(device, 0x001704, 0x80000000 | addr);
 }
 
@@ -74,7 +74,7 @@ gf100_bar_bar2_init(struct nvkm_bar *base)
 {
 	struct nvkm_device *device = base->subdev.device;
 	struct gf100_bar *bar = gf100_bar(base);
-	u32 addr = nvkm_memory_addr(bar->bar[0].mem) >> 12;
+	u32 addr = nvkm_memory_addr(bar->bar[0].inst) >> 12;
 	if (bar->bar2_halve)
 		addr |= 0x40000000;
 	nvkm_wr32(device, 0x001714, 0x80000000 | addr);
@@ -90,11 +90,7 @@ gf100_bar_oneinit_bar(struct gf100_bar *bar, struct gf100_barN *bar_vm,
 	int ret;
 
 	ret = nvkm_memory_new(device, NVKM_MEM_TARGET_INST, 0x1000, 0, false,
-			      &bar_vm->mem);
-	if (ret)
-		return ret;
-
-	ret = nvkm_gpuobj_new(device, 0x8000, 0, false, NULL, &bar_vm->pgd);
+			      &bar_vm->inst);
 	if (ret)
 		return ret;
 
@@ -119,17 +115,11 @@ gf100_bar_oneinit_bar(struct gf100_bar *bar, struct gf100_barN *bar_vm,
 		}
 	}
 
-	ret = nvkm_vm_ref(vm, &bar_vm->vm, bar_vm->pgd);
+	ret = nvkm_vm_ref(vm, &bar_vm->vm, bar_vm->inst);
 	nvkm_vm_ref(NULL, &vm, NULL);
 	if (ret)
 		return ret;
 
-	nvkm_kmap(bar_vm->mem);
-	nvkm_wo32(bar_vm->mem, 0x0200, lower_32_bits(bar_vm->pgd->addr));
-	nvkm_wo32(bar_vm->mem, 0x0204, upper_32_bits(bar_vm->pgd->addr));
-	nvkm_wo32(bar_vm->mem, 0x0208, lower_32_bits(bar_len - 1));
-	nvkm_wo32(bar_vm->mem, 0x020c, upper_32_bits(bar_len - 1));
-	nvkm_done(bar_vm->mem);
 	return 0;
 }
 
@@ -164,13 +154,11 @@ gf100_bar_dtor(struct nvkm_bar *base)
 {
 	struct gf100_bar *bar = gf100_bar(base);
 
-	nvkm_vm_ref(NULL, &bar->bar[1].vm, bar->bar[1].pgd);
-	nvkm_gpuobj_del(&bar->bar[1].pgd);
-	nvkm_memory_unref(&bar->bar[1].mem);
+	nvkm_vm_ref(NULL, &bar->bar[1].vm, bar->bar[1].inst);
+	nvkm_memory_unref(&bar->bar[1].inst);
 
-	nvkm_vm_ref(NULL, &bar->bar[0].vm, bar->bar[0].pgd);
-	nvkm_gpuobj_del(&bar->bar[0].pgd);
-	nvkm_memory_unref(&bar->bar[0].mem);
+	nvkm_vm_ref(NULL, &bar->bar[0].vm, bar->bar[0].inst);
+	nvkm_memory_unref(&bar->bar[0].inst);
 	return bar;
 }
 
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/bar/gf100.h b/drivers/gpu/drm/nouveau/nvkm/subdev/bar/gf100.h
index fc1a7bd975a7..9ce80c6b6fec 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/bar/gf100.h
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/bar/gf100.h
@@ -4,8 +4,7 @@
 #include "priv.h"
 
 struct gf100_barN {
-	struct nvkm_memory *mem;
-	struct nvkm_gpuobj *pgd;
+	struct nvkm_memory *inst;
 	struct nvkm_vm *vm;
 };
 
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/bar/nv50.c b/drivers/gpu/drm/nouveau/nvkm/subdev/bar/nv50.c
index b40c131af20e..9300529917fc 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/bar/nv50.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/bar/nv50.c
@@ -140,7 +140,7 @@ nv50_bar_oneinit(struct nvkm_bar *base)
 	if (ret)
 		return ret;
 
-	ret = nvkm_vm_ref(vm, &bar->bar2_vm, bar->pgd);
+	ret = nvkm_vm_ref(vm, &bar->bar2_vm, bar->mem->memory);
 	nvkm_vm_ref(NULL, &vm, NULL);
 	if (ret)
 		return ret;
@@ -172,7 +172,7 @@ nv50_bar_oneinit(struct nvkm_bar *base)
 
 	atomic_inc(&vm->engref[NVKM_SUBDEV_BAR]);
 
-	ret = nvkm_vm_ref(vm, &bar->bar1_vm, bar->pgd);
+	ret = nvkm_vm_ref(vm, &bar->bar1_vm, bar->mem->memory);
 	nvkm_vm_ref(NULL, &vm, NULL);
 	if (ret)
 		return ret;
@@ -197,13 +197,15 @@ void *
 nv50_bar_dtor(struct nvkm_bar *base)
 {
 	struct nv50_bar *bar = nv50_bar(base);
-	nvkm_gpuobj_del(&bar->bar1);
-	nvkm_vm_ref(NULL, &bar->bar1_vm, bar->pgd);
-	nvkm_gpuobj_del(&bar->bar2);
-	nvkm_vm_ref(NULL, &bar->bar2_vm, bar->pgd);
-	nvkm_gpuobj_del(&bar->pgd);
-	nvkm_gpuobj_del(&bar->pad);
-	nvkm_gpuobj_del(&bar->mem);
+	if (bar->mem) {
+		nvkm_gpuobj_del(&bar->bar1);
+		nvkm_vm_ref(NULL, &bar->bar1_vm, bar->mem->memory);
+		nvkm_gpuobj_del(&bar->bar2);
+		nvkm_vm_ref(NULL, &bar->bar2_vm, bar->mem->memory);
+		nvkm_gpuobj_del(&bar->pgd);
+		nvkm_gpuobj_del(&bar->pad);
+		nvkm_gpuobj_del(&bar->mem);
+	}
 	return bar;
 }
 
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/base.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/base.c
index 9fdd1446da5f..22264d3db22f 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/base.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/base.c
@@ -446,7 +446,6 @@ static void
 nvkm_vm_unmap_pgt(struct nvkm_vm *vm, int big, u32 fpde, u32 lpde)
 {
 	struct nvkm_mmu *mmu = vm->mmu;
-	struct nvkm_vm_pgd *vpgd;
 	struct nvkm_vm_pgt *vpgt;
 	struct nvkm_memory *pgt;
 	u32 pde;
@@ -459,9 +458,8 @@ nvkm_vm_unmap_pgt(struct nvkm_vm *vm, int big, u32 fpde, u32 lpde)
 		pgt = vpgt->mem[big];
 		vpgt->mem[big] = NULL;
 
-		list_for_each_entry(vpgd, &vm->pgd_list, head) {
-			mmu->func->map_pgt(vpgd->obj, pde, vpgt->mem);
-		}
+		if (mmu->func->map_pgt)
+			mmu->func->map_pgt(vm, pde, vpgt->mem);
 
 		mmu->func->flush(vm);
 
@@ -474,7 +472,6 @@ nvkm_vm_map_pgt(struct nvkm_vm *vm, u32 pde, u32 type)
 {
 	struct nvkm_mmu *mmu = vm->mmu;
 	struct nvkm_vm_pgt *vpgt = &vm->pgt[pde - vm->fpde];
-	struct nvkm_vm_pgd *vpgd;
 	int big = (type != mmu->func->spg_shift);
 	u32 pgt_size;
 	int ret;
@@ -487,9 +484,8 @@ nvkm_vm_map_pgt(struct nvkm_vm *vm, u32 pde, u32 type)
 	if (unlikely(ret))
 		return ret;
 
-	list_for_each_entry(vpgd, &vm->pgd_list, head) {
-		mmu->func->map_pgt(vpgd->obj, pde, vpgt->mem);
-	}
+	if (mmu->func->map_pgt)
+		mmu->func->map_pgt(vm, pde, vpgt->mem);
 
 	vpgt->refcount[big]++;
 	return 0;
@@ -592,7 +588,6 @@ nvkm_vm_legacy(struct nvkm_mmu *mmu, u64 offset, u64 length, u64 mm_offset,
 	u64 mm_length = (offset + length) - mm_offset;
 	int ret;
 
-	INIT_LIST_HEAD(&vm->pgd_list);
 	kref_init(&vm->refcount);
 	vm->fpde = offset >> (mmu->func->pgt_bits + 12);
 	vm->lpde = (offset + length - 1) >> (mmu->func->pgt_bits + 12);
@@ -644,58 +639,10 @@ nvkm_vm_new(struct nvkm_device *device, u64 offset, u64 length, u64 mm_offset,
 	return -EINVAL;
 }
 
-static int
-nvkm_vm_link(struct nvkm_vm *vm, struct nvkm_gpuobj *pgd)
-{
-	struct nvkm_mmu *mmu = vm->mmu;
-	struct nvkm_vm_pgd *vpgd;
-	int i;
-
-	if (!pgd)
-		return 0;
-
-	vpgd = kzalloc(sizeof(*vpgd), GFP_KERNEL);
-	if (!vpgd)
-		return -ENOMEM;
-
-	vpgd->obj = pgd;
-
-	mutex_lock(&vm->mutex);
-	for (i = vm->fpde; i <= vm->lpde; i++)
-		mmu->func->map_pgt(pgd, i, vm->pgt[i - vm->fpde].mem);
-	list_add(&vpgd->head, &vm->pgd_list);
-	mutex_unlock(&vm->mutex);
-	return 0;
-}
-
-static void
-nvkm_vm_unlink(struct nvkm_vm *vm, struct nvkm_gpuobj *mpgd)
-{
-	struct nvkm_vm_pgd *vpgd, *tmp;
-
-	if (!mpgd)
-		return;
-
-	mutex_lock(&vm->mutex);
-	list_for_each_entry_safe(vpgd, tmp, &vm->pgd_list, head) {
-		if (vpgd->obj == mpgd) {
-			list_del(&vpgd->head);
-			kfree(vpgd);
-			break;
-		}
-	}
-	mutex_unlock(&vm->mutex);
-}
-
 static void
 nvkm_vm_del(struct kref *kref)
 {
 	struct nvkm_vm *vm = container_of(kref, typeof(*vm), refcount);
-	struct nvkm_vm_pgd *vpgd, *tmp;
-
-	list_for_each_entry_safe(vpgd, tmp, &vm->pgd_list, head) {
-		nvkm_vm_unlink(vm, vpgd->obj);
-	}
 
 	nvkm_mm_fini(&vm->mm);
 	vfree(vm->pgt);
@@ -705,20 +652,28 @@ nvkm_vm_del(struct kref *kref)
 }
 
 int
-nvkm_vm_ref(struct nvkm_vm *ref, struct nvkm_vm **ptr, struct nvkm_gpuobj *pgd)
+nvkm_vm_ref(struct nvkm_vm *ref, struct nvkm_vm **ptr, struct nvkm_memory *inst)
 {
 	if (ref) {
-		int ret = nvkm_vm_link(ref, pgd);
-		if (ret)
-			return ret;
+		if (ref->func->join && inst) {
+			int ret = ref->func->join(ref, inst), i;
+			if (ret)
+				return ret;
+
+			if (ref->mmu->func->map_pgt) {
+				for (i = ref->fpde; i <= ref->lpde; i++)
+					ref->mmu->func->map_pgt(ref, i, ref->pgt[i - ref->fpde].mem);
+			}
+		}
 
 		kref_get(&ref->refcount);
 	}
 
 	if (*ptr) {
-		if ((*ptr)->bootstrapped && pgd)
+		if ((*ptr)->func->part && inst)
+			(*ptr)->func->part(*ptr, inst);
+		if ((*ptr)->bootstrapped && inst)
 			nvkm_memory_unref(&(*ptr)->pgt[0].mem[0]);
-		nvkm_vm_unlink(*ptr, pgd);
 		kref_put(&(*ptr)->refcount, nvkm_vm_del);
 	}
 
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/gf100.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/gf100.c
index 536260782cff..167c201d29e3 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/gf100.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/gf100.c
@@ -70,8 +70,9 @@ const u8 gf100_pte_storage_type_map[256] =
 
 
 void
-gf100_vm_map_pgt(struct nvkm_gpuobj *pgd, u32 index, struct nvkm_memory *pgt[2])
+gf100_vm_map_pgt(struct nvkm_vmm *vmm, u32 index, struct nvkm_memory *pgt[2])
 {
+	struct nvkm_memory *pgd = vmm->pd->pt[0]->memory;
 	u32 pde[2] = { 0, 0 };
 
 	if (pgt[0])
@@ -161,7 +162,6 @@ gf100_vm_flush(struct nvkm_vm *vm)
 {
 	struct nvkm_mmu *mmu = vm->mmu;
 	struct nvkm_device *device = mmu->subdev.device;
-	struct nvkm_vm_pgd *vpgd;
 	u32 type;
 
 	type = 0x00000001; /* PAGE_ALL */
@@ -169,24 +169,22 @@ gf100_vm_flush(struct nvkm_vm *vm)
 		type |= 0x00000004; /* HUB_ONLY */
 
 	mutex_lock(&mmu->subdev.mutex);
-	list_for_each_entry(vpgd, &vm->pgd_list, head) {
-		/* looks like maybe a "free flush slots" counter, the
-		 * faster you write to 0x100cbc to more it decreases
-		 */
-		nvkm_msec(device, 2000,
-			if (nvkm_rd32(device, 0x100c80) & 0x00ff0000)
-				break;
-		);
-
-		nvkm_wr32(device, 0x100cb8, vpgd->obj->addr >> 8);
-		nvkm_wr32(device, 0x100cbc, 0x80000000 | type);
-
-		/* wait for flush to be queued? */
-		nvkm_msec(device, 2000,
-			if (nvkm_rd32(device, 0x100c80) & 0x00008000)
-				break;
-		);
-	}
+	/* looks like maybe a "free flush slots" counter, the
+	 * faster you write to 0x100cbc to more it decreases
+	 */
+	nvkm_msec(device, 2000,
+		if (nvkm_rd32(device, 0x100c80) & 0x00ff0000)
+			break;
+	);
+
+	nvkm_wr32(device, 0x100cb8, vm->pd->pt[0]->addr >> 8);
+	nvkm_wr32(device, 0x100cbc, 0x80000000 | type);
+
+	/* wait for flush to be queued? */
+	nvkm_msec(device, 2000,
+		if (nvkm_rd32(device, 0x100c80) & 0x00008000)
+			break;
+	);
 	mutex_unlock(&mmu->subdev.mutex);
 }
 
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/nv50.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/nv50.c
index ba1c04e00701..da75d1da0658 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/nv50.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/nv50.c
@@ -31,8 +31,10 @@
 #include <nvif/class.h>
 
 void
-nv50_vm_map_pgt(struct nvkm_gpuobj *pgd, u32 pde, struct nvkm_memory *pgt[2])
+nv50_vm_map_pgt(struct nvkm_vmm *vmm, u32 pde, struct nvkm_memory *pgt[2])
 {
+	struct nvkm_vmm_join *join;
+	u32 pdeo = vmm->mmu->func->vmm.pd_offset + (pde * 8);
 	u64 phys = 0xdeadcafe00000000ULL;
 	u32 coverage = 0;
 
@@ -56,10 +58,12 @@ nv50_vm_map_pgt(struct nvkm_gpuobj *pgd, u32 pde, struct nvkm_memory *pgt[2])
 			phys |= 0x20;
 	}
 
-	nvkm_kmap(pgd);
-	nvkm_wo32(pgd, (pde * 8) + 0, lower_32_bits(phys));
-	nvkm_wo32(pgd, (pde * 8) + 4, upper_32_bits(phys));
-	nvkm_done(pgd);
+	list_for_each_entry(join, &vmm->join, head) {
+		nvkm_kmap(join->inst);
+		nvkm_wo32(join->inst, pdeo + 0, lower_32_bits(phys));
+		nvkm_wo32(join->inst, pdeo + 4, upper_32_bits(phys));
+		nvkm_done(join->inst);
+	}
 }
 
 static inline u64
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/priv.h b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/priv.h
index f490a636e9f0..3271ec1c3c83 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/priv.h
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/priv.h
@@ -18,7 +18,7 @@ struct nvkm_mmu_func {
 	u8  spg_shift;
 	u8  lpg_shift;
 
-	void (*map_pgt)(struct nvkm_gpuobj *pgd, u32 pde,
+	void (*map_pgt)(struct nvkm_vmm *, u32 pde,
 			struct nvkm_memory *pgt[2]);
 	void (*map)(struct nvkm_vma *, struct nvkm_memory *,
 		    struct nvkm_mem *, u32 pte, u32 cnt,
@@ -41,7 +41,7 @@ struct nvkm_mmu_func {
 
 extern const struct nvkm_mmu_func nv04_mmu;
 
-void nv50_vm_map_pgt(struct nvkm_gpuobj *, u32, struct nvkm_memory **);
+void nv50_vm_map_pgt(struct nvkm_vmm *, u32, struct nvkm_memory **);
 void nv50_vm_map(struct nvkm_vma *, struct nvkm_memory *, struct nvkm_mem *,
 		 u32, u32, u64, u64);
 void nv50_vm_map_sg(struct nvkm_vma *, struct nvkm_memory *, struct nvkm_mem *,
@@ -49,7 +49,7 @@ void nv50_vm_map_sg(struct nvkm_vma *, struct nvkm_memory *, struct nvkm_mem *,
 void nv50_vm_unmap(struct nvkm_vma *, struct nvkm_memory *, u32, u32);
 void nv50_vm_flush(struct nvkm_vm *);
 
-void gf100_vm_map_pgt(struct nvkm_gpuobj *, u32, struct nvkm_memory **);
+void gf100_vm_map_pgt(struct nvkm_vmm *, u32, struct nvkm_memory **);
 void gf100_vm_map(struct nvkm_vma *, struct nvkm_memory *, struct nvkm_mem *,
 		  u32, u32, u64, u64);
 void gf100_vm_map_sg(struct nvkm_vma *, struct nvkm_memory *, struct nvkm_mem *,
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c
index 25e86ce3d29c..7e00b9adca05 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c
@@ -131,7 +131,7 @@ nvkm_vmm_ctor(const struct nvkm_vmm_func *func, struct nvkm_mmu *mmu,
 	/* ... and the GPU storage for it, except on Tesla-class GPUs that
 	 * have the PD embedded in the instance structure.
 	 */
-	if (desc->size && mmu->func->vmm.global) {
+	if (desc->size) {
 		const u32 size = pd_header + desc->size * (1 << desc->bits);
 		vmm->pd->pt[0] = nvkm_mmu_ptc_get(mmu, size, desc->align, true);
 		if (!vmm->pd->pt[0])
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/secboot/gm200.c b/drivers/gpu/drm/nouveau/nvkm/subdev/secboot/gm200.c
index d35c041fdcec..6c068c5bd3c2 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/secboot/gm200.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/secboot/gm200.c
@@ -112,11 +112,8 @@ gm200_secboot_oneinit(struct nvkm_secboot *sb)
 	int ret;
 
 	/* Allocate instance block and VM */
-	ret = nvkm_gpuobj_new(device, 0x1000, 0, true, NULL, &gsb->inst);
-	if (ret)
-		return ret;
-
-	ret = nvkm_gpuobj_new(device, 0x8000, 0, true, NULL, &gsb->pgd);
+	ret = nvkm_memory_new(device, NVKM_MEM_TARGET_INST, 0x1000, 0, true,
+			      &gsb->inst);
 	if (ret)
 		return ret;
 
@@ -126,18 +123,11 @@ gm200_secboot_oneinit(struct nvkm_secboot *sb)
 
 	atomic_inc(&vm->engref[NVKM_SUBDEV_PMU]);
 
-	ret = nvkm_vm_ref(vm, &gsb->vm, gsb->pgd);
+	ret = nvkm_vm_ref(vm, &gsb->vm, gsb->inst);
 	nvkm_vm_ref(NULL, &vm, NULL);
 	if (ret)
 		return ret;
 
-	nvkm_kmap(gsb->inst);
-	nvkm_wo32(gsb->inst, 0x200, lower_32_bits(gsb->pgd->addr));
-	nvkm_wo32(gsb->inst, 0x204, upper_32_bits(gsb->pgd->addr));
-	nvkm_wo32(gsb->inst, 0x208, lower_32_bits(vm_area_len - 1));
-	nvkm_wo32(gsb->inst, 0x20c, upper_32_bits(vm_area_len - 1));
-	nvkm_done(gsb->inst);
-
 	if (sb->acr->func->oneinit) {
 		ret = sb->acr->func->oneinit(sb->acr, sb);
 		if (ret)
@@ -165,9 +155,8 @@ gm200_secboot_dtor(struct nvkm_secboot *sb)
 
 	sb->acr->func->dtor(sb->acr);
 
-	nvkm_vm_ref(NULL, &gsb->vm, gsb->pgd);
-	nvkm_gpuobj_del(&gsb->pgd);
-	nvkm_gpuobj_del(&gsb->inst);
+	nvkm_vm_ref(NULL, &gsb->vm, gsb->inst);
+	nvkm_memory_unref(&gsb->inst);
 
 	return gsb;
 }
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/secboot/gm200.h b/drivers/gpu/drm/nouveau/nvkm/subdev/secboot/gm200.h
index c8ab3d76bdef..a2b60fab42e8 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/secboot/gm200.h
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/secboot/gm200.h
@@ -29,8 +29,7 @@ struct gm200_secboot {
 	struct nvkm_secboot base;
 
 	/* Instance block & address space used for HS FW execution */
-	struct nvkm_gpuobj *inst;
-	struct nvkm_gpuobj *pgd;
+	struct nvkm_memory *inst;
 	struct nvkm_vm *vm;
 };
 #define gm200_secboot(sb) container_of(sb, struct gm200_secboot, base)