13 files changed, 353 insertions, 137 deletions

diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/Kbuild b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/Kbuild
index 7ba35ea59c06..ea4848931540 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/Kbuild
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/Kbuild
@@ -15,8 +15,7 @@ nvkm-y += nvkm/subdev/mmu/gp100.o
 nvkm-y += nvkm/subdev/mmu/gp10b.o
 nvkm-y += nvkm/subdev/mmu/gv100.o
 nvkm-y += nvkm/subdev/mmu/tu102.o
-
-nvkm-y += nvkm/subdev/mmu/r535.o
+nvkm-y += nvkm/subdev/mmu/gh100.o
 
 nvkm-y += nvkm/subdev/mmu/mem.o
 nvkm-y += nvkm/subdev/mmu/memnv04.o
@@ -38,6 +37,7 @@ nvkm-y += nvkm/subdev/mmu/vmmgp100.o
 nvkm-y += nvkm/subdev/mmu/vmmgp10b.o
 nvkm-y += nvkm/subdev/mmu/vmmgv100.o
 nvkm-y += nvkm/subdev/mmu/vmmtu102.o
+nvkm-y += nvkm/subdev/mmu/vmmgh100.o
 
 nvkm-y += nvkm/subdev/mmu/umem.o
 nvkm-y += nvkm/subdev/mmu/ummu.o
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/gh100.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/gh100.c
new file mode 100644
index 000000000000..2918fb32cc91
--- /dev/null
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/gh100.c
@@ -0,0 +1,25 @@
+/* SPDX-License-Identifier: MIT
+ *
+ * Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
+ */
+#include "mem.h"
+#include "vmm.h"
+
+#include <nvif/class.h>
+
+static const struct nvkm_mmu_func
+gh100_mmu = {
+	.dma_bits = 52,
+	.mmu = {{ -1, -1, NVIF_CLASS_MMU_GF100}},
+	.mem = {{ -1,  0, NVIF_CLASS_MEM_GF100}, gf100_mem_new, gf100_mem_map },
+	.vmm = {{ -1,  0, NVIF_CLASS_VMM_GP100}, gh100_vmm_new },
+	.kind = tu102_mmu_kind,
+	.kind_sys = true,
+};
+
+int
+gh100_mmu_new(struct nvkm_device *device, enum nvkm_subdev_type type, int inst,
+	      struct nvkm_mmu **pmmu)
+{
+	return r535_mmu_new(&gh100_mmu, device, type, inst, pmmu);
+}
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/memgf100.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/memgf100.c
index d9c9bee45222..160a5749a29f 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/memgf100.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/memgf100.c
@@ -60,7 +60,7 @@ gf100_mem_map(struct nvkm_mmu *mmu, struct nvkm_memory *memory, void *argv,
 	if (ret)
 		return ret;
 
-	*paddr = device->func->resource_addr(device, 1) + (*pvma)->addr;
+	*paddr = device->func->resource_addr(device, NVKM_BAR1_FB) + (*pvma)->addr;
 	*psize = (*pvma)->size;
 	return 0;
 }
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/memnv04.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/memnv04.c
index 79a3b0cc9f5b..1e3db52de6cb 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/memnv04.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/memnv04.c
@@ -41,7 +41,7 @@ nv04_mem_map(struct nvkm_mmu *mmu, struct nvkm_memory *memory, void *argv,
 	if ((ret = nvif_unvers(ret, &argv, &argc, args->vn)))
 		return ret;
 
-	*paddr = device->func->resource_addr(device, 1) + addr;
+	*paddr = device->func->resource_addr(device, NVKM_BAR1_FB) + addr;
 	*psize = nvkm_memory_size(memory);
 	*pvma = ERR_PTR(-ENODEV);
 	return 0;
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/memnv50.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/memnv50.c
index 46759b89fc1f..33b2321e9d87 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/memnv50.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/memnv50.c
@@ -57,7 +57,7 @@ nv50_mem_map(struct nvkm_mmu *mmu, struct nvkm_memory *memory, void *argv,
 	if (ret)
 		return ret;
 
-	*paddr = device->func->resource_addr(device, 1) + (*pvma)->addr;
+	*paddr = device->func->resource_addr(device, NVKM_BAR1_FB) + (*pvma)->addr;
 	*psize = (*pvma)->size;
 	return nvkm_memory_map(memory, 0, bar, *pvma, &uvmm, sizeof(uvmm));
 }
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/priv.h b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/priv.h
index e9ca6537778c..90efef8f0b54 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/priv.h
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/priv.h
@@ -53,6 +53,8 @@ const u8 *gf100_mmu_kind(struct nvkm_mmu *, int *count, u8 *invalid);
 
 const u8 *gm200_mmu_kind(struct nvkm_mmu *, int *, u8 *);
 
+const u8 *tu102_mmu_kind(struct nvkm_mmu *, int *, u8 *);
+
 struct nvkm_mmu_pt {
 	union {
 		struct nvkm_mmu_ptc *ptc;
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/r535.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/r535.c
deleted file mode 100644
index d3e95453f25d..000000000000
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/r535.c
+++ /dev/null
@@ -1,123 +0,0 @@
-/*
- * Copyright 2023 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.
- */
-#include "vmm.h"
-
-#include <nvrm/nvtypes.h>
-#include <nvrm/535.113.01/common/sdk/nvidia/inc/class/cl90f1.h>
-#include <nvrm/535.113.01/common/sdk/nvidia/inc/ctrl/ctrl90f1.h>
-#include <nvrm/535.113.01/common/sdk/nvidia/inc/nvos.h>
-
-static int
-r535_mmu_promote_vmm(struct nvkm_vmm *vmm)
-{
-	NV_VASPACE_ALLOCATION_PARAMETERS *args;
-	int ret;
-
-	ret = nvkm_gsp_client_device_ctor(vmm->mmu->subdev.device->gsp,
-					  &vmm->rm.client, &vmm->rm.device);
-	if (ret)
-		return ret;
-
-	args = nvkm_gsp_rm_alloc_get(&vmm->rm.device.object, 0x90f10000, FERMI_VASPACE_A,
-				     sizeof(*args), &vmm->rm.object);
-	if (IS_ERR(args))
-		return PTR_ERR(args);
-
-	args->index = NV_VASPACE_ALLOCATION_INDEX_GPU_NEW;
-
-	ret = nvkm_gsp_rm_alloc_wr(&vmm->rm.object, args);
-	if (ret)
-		return ret;
-
-	{
-		NV90F1_CTRL_VASPACE_COPY_SERVER_RESERVED_PDES_PARAMS *ctrl;
-
-		mutex_lock(&vmm->mutex.vmm);
-		ret = nvkm_vmm_get_locked(vmm, true, false, false, 0x1d, 32, 0x20000000,
-					  &vmm->rm.rsvd);
-		mutex_unlock(&vmm->mutex.vmm);
-		if (ret)
-			return ret;
-
-		ctrl = nvkm_gsp_rm_ctrl_get(&vmm->rm.object,
-					    NV90F1_CTRL_CMD_VASPACE_COPY_SERVER_RESERVED_PDES,
-					    sizeof(*ctrl));
-		if (IS_ERR(ctrl))
-			return PTR_ERR(ctrl);
-
-		ctrl->pageSize = 0x20000000;
-		ctrl->virtAddrLo = vmm->rm.rsvd->addr;
-		ctrl->virtAddrHi = vmm->rm.rsvd->addr + vmm->rm.rsvd->size - 1;
-		ctrl->numLevelsToCopy = vmm->pd->pde[0]->pde[0] ? 3 : 2;
-		ctrl->levels[0].physAddress = vmm->pd->pt[0]->addr;
-		ctrl->levels[0].size = 0x20;
-		ctrl->levels[0].aperture = 1;
-		ctrl->levels[0].pageShift = 0x2f;
-		ctrl->levels[1].physAddress = vmm->pd->pde[0]->pt[0]->addr;
-		ctrl->levels[1].size = 0x1000;
-		ctrl->levels[1].aperture = 1;
-		ctrl->levels[1].pageShift = 0x26;
-		if (vmm->pd->pde[0]->pde[0]) {
-			ctrl->levels[2].physAddress = vmm->pd->pde[0]->pde[0]->pt[0]->addr;
-			ctrl->levels[2].size = 0x1000;
-			ctrl->levels[2].aperture = 1;
-			ctrl->levels[2].pageShift = 0x1d;
-		}
-
-		ret = nvkm_gsp_rm_ctrl_wr(&vmm->rm.object, ctrl);
-	}
-
-	return ret;
-}
-
-static void
-r535_mmu_dtor(struct nvkm_mmu *mmu)
-{
-	kfree(mmu->func);
-}
-
-int
-r535_mmu_new(const struct nvkm_mmu_func *hw,
-	     struct nvkm_device *device, enum nvkm_subdev_type type, int inst,
-	     struct nvkm_mmu **pmmu)
-{
-	struct nvkm_mmu_func *rm;
-	int ret;
-
-	if (!(rm = kzalloc(sizeof(*rm), GFP_KERNEL)))
-		return -ENOMEM;
-
-	rm->dtor = r535_mmu_dtor;
-	rm->dma_bits = hw->dma_bits;
-	rm->mmu = hw->mmu;
-	rm->mem = hw->mem;
-	rm->vmm = hw->vmm;
-	rm->kind = hw->kind;
-	rm->kind_sys = hw->kind_sys;
-	rm->promote_vmm = r535_mmu_promote_vmm;
-
-	ret = nvkm_mmu_new_(rm, device, type, inst, pmmu);
-	if (ret)
-		kfree(rm);
-
-	return ret;
-}
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/tu102.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/tu102.c
index df662ce4a4b0..7acff3642e20 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/tu102.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/tu102.c
@@ -28,7 +28,7 @@
 
 #include <nvif/class.h>
 
-static const u8 *
+const u8 *
 tu102_mmu_kind(struct nvkm_mmu *mmu, int *count, u8 *invalid)
 {
 	static const u8
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c
index 9c97800fe037..f95c58b67633 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.c
@@ -19,7 +19,7 @@
  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  * OTHER DEALINGS IN THE SOFTWARE.
  */
-#define NVKM_VMM_LEVELS_MAX 5
+#define NVKM_VMM_LEVELS_MAX 6
 #include "vmm.h"
 
 #include <subdev/fb.h>
@@ -1030,12 +1030,8 @@ nvkm_vmm_dtor(struct nvkm_vmm *vmm)
 	struct nvkm_vma *vma;
 	struct rb_node *node;
 
-	if (vmm->rm.client.gsp) {
-		nvkm_gsp_rm_free(&vmm->rm.object);
-		nvkm_gsp_device_dtor(&vmm->rm.device);
-		nvkm_gsp_client_dtor(&vmm->rm.client);
-		nvkm_vmm_put(vmm, &vmm->rm.rsvd);
-	}
+	if (vmm->rm.client.gsp)
+		r535_mmu_vaspace_del(vmm);
 
 	if (0)
 		nvkm_vmm_dump(vmm);
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h
index f9bc30cdb2b3..4586a425dbe4 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h
@@ -143,6 +143,8 @@ struct nvkm_vmm_func {
 	int (*aper)(enum nvkm_memory_target);
 	int (*valid)(struct nvkm_vmm *, void *argv, u32 argc,
 		     struct nvkm_vmm_map *);
+	int (*valid2)(struct nvkm_vmm *, bool ro, bool priv, u8 kind, u8 comp,
+		      struct nvkm_vmm_map *);
 	void (*flush)(struct nvkm_vmm *, int depth);
 
 	int (*mthd)(struct nvkm_vmm *, struct nvkm_client *,
@@ -254,6 +256,8 @@ void gp100_vmm_invalidate_pdb(struct nvkm_vmm *, u64 addr);
 
 int gv100_vmm_join(struct nvkm_vmm *, struct nvkm_memory *);
 
+void tu102_vmm_flush(struct nvkm_vmm *, int depth);
+
 int nv04_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
 		 struct lock_class_key *, const char *, struct nvkm_vmm **);
 int nv41_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
@@ -296,6 +300,9 @@ int gv100_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
 int tu102_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
 		  struct lock_class_key *, const char *,
 		  struct nvkm_vmm **);
+int gh100_vmm_new(struct nvkm_mmu *, bool, u64, u64, void *, u32,
+		  struct lock_class_key *, const char *,
+		  struct nvkm_vmm **);
 
 #define VMM_PRINT(l,v,p,f,a...) do {                                           \
 	struct nvkm_vmm *_vmm = (v);                                           \
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgh100.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgh100.c
new file mode 100644
index 000000000000..5614df3432da
--- /dev/null
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgh100.c
@@ -0,0 +1,306 @@
+/* SPDX-License-Identifier: MIT
+ *
+ * Copyright (c) 2025, NVIDIA CORPORATION. All rights reserved.
+ */
+#include "vmm.h"
+
+#include <subdev/fb.h>
+
+#include <nvhw/drf.h>
+#include <nvhw/ref/gh100/dev_mmu.h>
+
+static inline void
+gh100_vmm_pgt_pte(struct nvkm_vmm *vmm, struct nvkm_mmu_pt *pt, u32 ptei, u32 ptes,
+		  struct nvkm_vmm_map *map, u64 addr)
+{
+	u64 data = addr | map->type;
+
+	while (ptes--) {
+		VMM_WO064(pt, vmm, ptei++ * NV_MMU_VER3_PTE__SIZE, data);
+		data += map->next;
+	}
+}
+
+static void
+gh100_vmm_pgt_sgl(struct nvkm_vmm *vmm, struct nvkm_mmu_pt *pt, u32 ptei, u32 ptes,
+		  struct nvkm_vmm_map *map)
+{
+	VMM_MAP_ITER_SGL(vmm, pt, ptei, ptes, map, gh100_vmm_pgt_pte);
+}
+
+static void
+gh100_vmm_pgt_dma(struct nvkm_vmm *vmm, struct nvkm_mmu_pt *pt, u32 ptei, u32 ptes,
+		  struct nvkm_vmm_map *map)
+{
+	if (map->page->shift == PAGE_SHIFT) {
+		VMM_SPAM(vmm, "DMAA %08x %08x PTE(s)", ptei, ptes);
+
+		nvkm_kmap(pt->memory);
+		while (ptes--) {
+			const u64 data = *map->dma++ | map->type;
+
+			VMM_WO064(pt, vmm, ptei++ * NV_MMU_VER3_PTE__SIZE, data);
+		}
+		nvkm_done(pt->memory);
+		return;
+	}
+
+	VMM_MAP_ITER_DMA(vmm, pt, ptei, ptes, map, gh100_vmm_pgt_pte);
+}
+
+static void
+gh100_vmm_pgt_mem(struct nvkm_vmm *vmm, struct nvkm_mmu_pt *pt, u32 ptei, u32 ptes,
+		  struct nvkm_vmm_map *map)
+{
+	VMM_MAP_ITER_MEM(vmm, pt, ptei, ptes, map, gh100_vmm_pgt_pte);
+}
+
+static void
+gh100_vmm_pgt_sparse(struct nvkm_vmm *vmm,
+		     struct nvkm_mmu_pt *pt, u32 ptei, u32 ptes)
+{
+	const u64 data = NVDEF(NV_MMU, VER3_PTE, PCF, SPARSE);
+
+	VMM_FO064(pt, vmm, ptei * NV_MMU_VER3_PTE__SIZE, data, ptes);
+}
+
+static const struct nvkm_vmm_desc_func
+gh100_vmm_desc_spt = {
+	.unmap = gf100_vmm_pgt_unmap,
+	.sparse = gh100_vmm_pgt_sparse,
+	.mem = gh100_vmm_pgt_mem,
+	.dma = gh100_vmm_pgt_dma,
+	.sgl = gh100_vmm_pgt_sgl,
+};
+
+static void
+gh100_vmm_lpt_invalid(struct nvkm_vmm *vmm,
+		      struct nvkm_mmu_pt *pt, u32 ptei, u32 ptes)
+{
+	const u64 data = NVDEF(NV_MMU, VER3_PTE, PCF, NO_VALID_4KB_PAGE);
+
+	VMM_FO064(pt, vmm, ptei * NV_MMU_VER3_PTE__SIZE, data, ptes);
+}
+
+static const struct nvkm_vmm_desc_func
+gh100_vmm_desc_lpt = {
+	.invalid = gh100_vmm_lpt_invalid,
+	.unmap = gf100_vmm_pgt_unmap,
+	.sparse = gh100_vmm_pgt_sparse,
+	.mem = gh100_vmm_pgt_mem,
+};
+
+static inline void
+gh100_vmm_pd0_pte(struct nvkm_vmm *vmm, struct nvkm_mmu_pt *pt,
+		  u32 ptei, u32 ptes, struct nvkm_vmm_map *map, u64 addr)
+{
+	u64 data = addr | map->type;
+
+	while (ptes--) {
+		VMM_WO128(pt, vmm, ptei++ * NV_MMU_VER3_DUAL_PDE__SIZE, data, 0ULL);
+		data += map->next;
+	}
+}
+
+static void
+gh100_vmm_pd0_mem(struct nvkm_vmm *vmm, struct nvkm_mmu_pt *pt,
+		  u32 ptei, u32 ptes, struct nvkm_vmm_map *map)
+{
+	VMM_MAP_ITER_MEM(vmm, pt, ptei, ptes, map, gh100_vmm_pd0_pte);
+}
+
+static inline bool
+gh100_vmm_pde(struct nvkm_mmu_pt *pt, u64 *data)
+{
+	switch (nvkm_memory_target(pt->memory)) {
+	case NVKM_MEM_TARGET_VRAM:
+		*data |= NVDEF(NV_MMU, VER3_PDE, APERTURE, VIDEO_MEMORY);
+		*data |= NVDEF(NV_MMU, VER3_PDE, PCF, VALID_CACHED_ATS_NOT_ALLOWED);
+		break;
+	case NVKM_MEM_TARGET_HOST:
+		*data |= NVDEF(NV_MMU, VER3_PDE, APERTURE, SYSTEM_COHERENT_MEMORY);
+		*data |= NVDEF(NV_MMU, VER3_PDE, PCF, VALID_UNCACHED_ATS_ALLOWED);
+		break;
+	case NVKM_MEM_TARGET_NCOH:
+		*data |= NVDEF(NV_MMU, VER3_PDE, APERTURE, SYSTEM_NON_COHERENT_MEMORY);
+		*data |= NVDEF(NV_MMU, VER3_PDE, PCF, VALID_CACHED_ATS_ALLOWED);
+		break;
+	default:
+		WARN_ON(1);
+		return false;
+	}
+
+	*data |= pt->addr;
+	return true;
+}
+
+static void
+gh100_vmm_pd0_pde(struct nvkm_vmm *vmm, struct nvkm_vmm_pt *pgd, u32 pdei)
+{
+	struct nvkm_vmm_pt *pgt = pgd->pde[pdei];
+	struct nvkm_mmu_pt *pd = pgd->pt[0];
+	u64 data[2] = {};
+
+	if (pgt->pt[0] && !gh100_vmm_pde(pgt->pt[0], &data[0]))
+		return;
+	if (pgt->pt[1] && !gh100_vmm_pde(pgt->pt[1], &data[1]))
+		return;
+
+	nvkm_kmap(pd->memory);
+	VMM_WO128(pd, vmm, pdei * NV_MMU_VER3_DUAL_PDE__SIZE, data[0], data[1]);
+	nvkm_done(pd->memory);
+}
+
+static void
+gh100_vmm_pd0_sparse(struct nvkm_vmm *vmm,
+		     struct nvkm_mmu_pt *pt, u32 pdei, u32 pdes)
+{
+	const u64 data = NVDEF(NV_MMU, VER3_DUAL_PDE, PCF_BIG, SPARSE_ATS_ALLOWED);
+
+	VMM_FO128(pt, vmm, pdei * NV_MMU_VER3_DUAL_PDE__SIZE, data, 0ULL, pdes);
+}
+
+static void
+gh100_vmm_pd0_unmap(struct nvkm_vmm *vmm,
+		    struct nvkm_mmu_pt *pt, u32 pdei, u32 pdes)
+{
+	VMM_FO128(pt, vmm, pdei * NV_MMU_VER3_DUAL_PDE__SIZE, 0ULL, 0ULL, pdes);
+}
+
+static const struct nvkm_vmm_desc_func
+gh100_vmm_desc_pd0 = {
+	.unmap = gh100_vmm_pd0_unmap,
+	.sparse = gh100_vmm_pd0_sparse,
+	.pde = gh100_vmm_pd0_pde,
+	.mem = gh100_vmm_pd0_mem,
+};
+
+static void
+gh100_vmm_pd1_pde(struct nvkm_vmm *vmm, struct nvkm_vmm_pt *pgd, u32 pdei)
+{
+	struct nvkm_vmm_pt *pgt = pgd->pde[pdei];
+	struct nvkm_mmu_pt *pd = pgd->pt[0];
+	u64 data = 0;
+
+	if (!gh100_vmm_pde(pgt->pt[0], &data))
+		return;
+
+	nvkm_kmap(pd->memory);
+	VMM_WO064(pd, vmm, pdei * NV_MMU_VER3_PDE__SIZE, data);
+	nvkm_done(pd->memory);
+}
+
+static const struct nvkm_vmm_desc_func
+gh100_vmm_desc_pd1 = {
+	.unmap = gf100_vmm_pgt_unmap,
+	.sparse = gh100_vmm_pgt_sparse,
+	.pde = gh100_vmm_pd1_pde,
+};
+
+static const struct nvkm_vmm_desc
+gh100_vmm_desc_16[] = {
+	{ LPT, 5,  8, 0x0100, &gh100_vmm_desc_lpt },
+	{ PGD, 8, 16, 0x1000, &gh100_vmm_desc_pd0 },
+	{ PGD, 9,  8, 0x1000, &gh100_vmm_desc_pd1 },
+	{ PGD, 9,  8, 0x1000, &gh100_vmm_desc_pd1 },
+	{ PGD, 9,  8, 0x1000, &gh100_vmm_desc_pd1 },
+	{ PGD, 1,  8, 0x1000, &gh100_vmm_desc_pd1 },
+	{}
+};
+
+static const struct nvkm_vmm_desc
+gh100_vmm_desc_12[] = {
+	{ SPT, 9,  8, 0x1000, &gh100_vmm_desc_spt },
+	{ PGD, 8, 16, 0x1000, &gh100_vmm_desc_pd0 },
+	{ PGD, 9,  8, 0x1000, &gh100_vmm_desc_pd1 },
+	{ PGD, 9,  8, 0x1000, &gh100_vmm_desc_pd1 },
+	{ PGD, 9,  8, 0x1000, &gh100_vmm_desc_pd1 },
+	{ PGD, 1,  8, 0x1000, &gh100_vmm_desc_pd1 },
+	{}
+};
+
+static int
+gh100_vmm_valid(struct nvkm_vmm *vmm, bool ro, bool priv, u8 kind, u8 comp,
+		struct nvkm_vmm_map *map)
+{
+	const enum nvkm_memory_target target = nvkm_memory_target(map->memory);
+	const bool vol = target == NVKM_MEM_TARGET_HOST;
+	const struct nvkm_vmm_page *page = map->page;
+	u8 kind_inv, pcf;
+	int kindn, aper;
+	const u8 *kindm;
+
+	map->next = 1ULL << page->shift;
+	map->type = 0;
+
+	aper = vmm->func->aper(target);
+	if (WARN_ON(aper < 0))
+		return aper;
+
+	kindm = vmm->mmu->func->kind(vmm->mmu, &kindn, &kind_inv);
+	if (kind >= kindn || kindm[kind] == kind_inv) {
+		VMM_DEBUG(vmm, "kind %02x", kind);
+		return -EINVAL;
+	}
+
+	if (priv) {
+		if (ro) {
+			if (vol)
+				pcf = NV_MMU_VER3_PTE_PCF_PRIVILEGE_RO_ATOMIC_UNCACHED_ACD;
+			else
+				pcf = NV_MMU_VER3_PTE_PCF_PRIVILEGE_RO_ATOMIC_CACHED_ACD;
+		} else {
+			if (vol)
+				pcf = NV_MMU_VER3_PTE_PCF_PRIVILEGE_RW_ATOMIC_UNCACHED_ACD;
+			else
+				pcf = NV_MMU_VER3_PTE_PCF_PRIVILEGE_RW_ATOMIC_CACHED_ACD;
+		}
+	} else {
+		if (ro) {
+			if (vol)
+				pcf = NV_MMU_VER3_PTE_PCF_REGULAR_RO_ATOMIC_UNCACHED_ACD;
+			else
+				pcf = NV_MMU_VER3_PTE_PCF_REGULAR_RO_ATOMIC_CACHED_ACD;
+		} else {
+			if (vol)
+				pcf = NV_MMU_VER3_PTE_PCF_REGULAR_RW_ATOMIC_UNCACHED_ACD;
+			else
+				pcf = NV_MMU_VER3_PTE_PCF_REGULAR_RW_ATOMIC_CACHED_ACD;
+		}
+	}
+
+	map->type |= NVDEF(NV_MMU, VER3_PTE, VALID, TRUE);
+	map->type |= NVVAL(NV_MMU, VER3_PTE, APERTURE, aper);
+	map->type |= NVVAL(NV_MMU, VER3_PTE, PCF, pcf);
+	map->type |= NVVAL(NV_MMU, VER3_PTE, KIND, kind);
+	return 0;
+}
+
+static const struct nvkm_vmm_func
+gh100_vmm = {
+	.join = gv100_vmm_join,
+	.part = gf100_vmm_part,
+	.aper = gf100_vmm_aper,
+	.valid = gp100_vmm_valid,
+	.valid2 = gh100_vmm_valid,
+	.flush = tu102_vmm_flush,
+	.page = {
+		{ 56, &gh100_vmm_desc_16[5], NVKM_VMM_PAGE_Sxxx },
+		{ 47, &gh100_vmm_desc_16[4], NVKM_VMM_PAGE_Sxxx },
+		{ 38, &gh100_vmm_desc_16[3], NVKM_VMM_PAGE_Sxxx },
+		{ 29, &gh100_vmm_desc_16[2], NVKM_VMM_PAGE_SVxC },
+		{ 21, &gh100_vmm_desc_16[1], NVKM_VMM_PAGE_SVxC },
+		{ 16, &gh100_vmm_desc_16[0], NVKM_VMM_PAGE_SVxC },
+		{ 12, &gh100_vmm_desc_12[0], NVKM_VMM_PAGE_SVHx },
+		{}
+	}
+};
+
+int
+gh100_vmm_new(struct nvkm_mmu *mmu, bool managed, u64 addr, u64 size,
+	      void *argv, u32 argc, struct lock_class_key *key,
+	      const char *name, struct nvkm_vmm **pvmm)
+{
+	return gp100_vmm_new_(&gh100_vmm, mmu, managed, addr, size,
+			      argv, argc, key, name, pvmm);
+}
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgp100.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgp100.c
index bddac77f48f0..851fd847a2a9 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgp100.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgp100.c
@@ -436,6 +436,9 @@ gp100_vmm_valid(struct nvkm_vmm *vmm, void *argv, u32 argc,
 		return ret;
 	}
 
+	if (vmm->func->valid2)
+		return vmm->func->valid2(vmm, ro, priv, kind, 0, map);
+
 	aper = vmm->func->aper(target);
 	if (WARN_ON(aper < 0))
 		return aper;
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmtu102.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmtu102.c
index 8379e72d77ab..4b30eab40bba 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmtu102.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmtu102.c
@@ -23,7 +23,7 @@
 
 #include <subdev/timer.h>
 
-static void
+void
 tu102_vmm_flush(struct nvkm_vmm *vmm, int depth)
 {
 	struct nvkm_device *device = vmm->mmu->subdev.device;