drm/nouveau/mmu/nv50,g84: implement new vmm backend

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

5 files changed, 347 insertions, 194 deletions

diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/base.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/base.c
index 536187952372..4d75dcb41468 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/base.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/base.c
@@ -27,6 +27,8 @@
 #include <core/gpuobj.h>
 #include <subdev/fb.h>
 
+#include <nvif/if500d.h>
+
 struct nvkm_mmu_ptp {
 	struct nvkm_mmu_pt *pt;
 	struct list_head head;
@@ -218,6 +220,9 @@ nvkm_vm_map_(const struct nvkm_vmm_page *page, struct nvkm_vma *vma, u64 delta,
 	     struct nvkm_mem *mem, nvkm_vmm_pte_func fn,
 	     struct nvkm_vmm_map *map)
 {
+	union {
+		struct nv50_vmm_map_v0 nv50;
+	} args;
 	struct nvkm_vmm *vmm = vma->vm;
 	void *argv = NULL;
 	u32 argc = 0;
@@ -227,6 +232,20 @@ nvkm_vm_map_(const struct nvkm_vmm_page *page, struct nvkm_vma *vma, u64 delta,
 	map->page = page;
 
 	if (vmm->func->valid) {
+		switch (vmm->mmu->subdev.device->card_type) {
+		case NV_50:
+			args.nv50.version = 0;
+			args.nv50.ro = !(vma->access & NV_MEM_ACCESS_WO);
+			args.nv50.priv = !!(vma->access & NV_MEM_ACCESS_SYS);
+			args.nv50.kind = (mem->memtype & 0x07f);
+			args.nv50.comp = (mem->memtype & 0x180) >> 7;
+			argv = &args.nv50;
+			argc = sizeof(args.nv50);
+			break;
+		default:
+			break;
+		}
+
 		ret = vmm->func->valid(vmm, argv, argc, map);
 		if (WARN_ON(ret))
 			return;
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/g84.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/g84.c
index 13f8247a2ee4..e436c610b3df 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/g84.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/g84.c
@@ -30,12 +30,8 @@ g84_mmu = {
 	.pgt_bits  = 29 - 12,
 	.spg_shift = 12,
 	.lpg_shift = 16,
-	.map_pgt = nv50_vm_map_pgt,
-	.map = nv50_vm_map,
-	.map_sg = nv50_vm_map_sg,
-	.unmap = nv50_vm_unmap,
-	.flush = nv50_vm_flush,
 	.vmm = {{ -1, -1, NVIF_CLASS_VMM_NV50}, nv50_vmm_new, false, 0x0200 },
+	.kind = nv50_mmu_kind,
 };
 
 int
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/nv50.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/nv50.c
index da75d1da0658..31d6f9626e56 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/nv50.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/nv50.c
@@ -23,185 +23,40 @@
  */
 #include "vmm.h"
 
-#include <core/gpuobj.h>
-#include <subdev/fb.h>
-#include <subdev/timer.h>
-#include <engine/gr.h>
-
 #include <nvif/class.h>
 
-void
-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;
-
-	if (pgt[0]) {
-		/* present, 4KiB pages */
-		phys = 0x00000003 | nvkm_memory_addr(pgt[0]);
-		coverage = (nvkm_memory_size(pgt[0]) >> 3) << 12;
-	} else
-	if (pgt[1]) {
-		/* present, 64KiB pages  */
-		phys = 0x00000001 | nvkm_memory_addr(pgt[1]);
-		coverage = (nvkm_memory_size(pgt[1]) >> 3) << 16;
-	}
-
-	if (phys & 1) {
-		if (coverage <= 32 * 1024 * 1024)
-			phys |= 0x60;
-		else if (coverage <= 64 * 1024 * 1024)
-			phys |= 0x40;
-		else if (coverage <= 128 * 1024 * 1024)
-			phys |= 0x20;
-	}
-
-	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
-vm_addr(struct nvkm_vma *vma, u64 phys, u32 memtype, u32 target)
+const u8 *
+nv50_mmu_kind(struct nvkm_mmu *base, int *count)
 {
-	phys |= 1; /* present */
-	phys |= (u64)memtype << 40;
-	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 nvkm_vma *vma, struct nvkm_memory *pgt,
-	    struct nvkm_mem *mem, u32 pte, u32 cnt, u64 phys, u64 delta)
-{
-	struct nvkm_ram *ram = vma->vm->mmu->subdev.device->fb->ram;
-	u32 comp = (mem->memtype & 0x180) >> 7;
-	u32 block, target;
-	int i;
-
-	/* IGPs don't have real VRAM, re-target to stolen system memory */
-	target = 0;
-	if (ram->stolen) {
-		phys += ram->stolen;
-		target = 3;
-	}
-
-	phys  = vm_addr(vma, phys, mem->memtype, target);
-	pte <<= 3;
-	cnt <<= 3;
-
-	nvkm_kmap(pgt);
-	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;
-		if (comp) {
-			u32 tag = mem->tag->offset + ((delta >> 16) * comp);
-			offset_h |= (tag << 17);
-			delta    += block << (vma->node->type - 3);
-		}
-
-		while (block) {
-			nvkm_wo32(pgt, pte + 0, offset_l);
-			nvkm_wo32(pgt, pte + 4, offset_h);
-			pte += 8;
-			block -= 8;
-		}
-	}
-	nvkm_done(pgt);
-}
-
-void
-nv50_vm_map_sg(struct nvkm_vma *vma, struct nvkm_memory *pgt,
-	       struct nvkm_mem *mem, u32 pte, u32 cnt, dma_addr_t *list)
-{
-	u32 target = (vma->access & NV_MEM_ACCESS_NOSNOOP) ? 3 : 2;
-	pte <<= 3;
-	nvkm_kmap(pgt);
-	while (cnt--) {
-		u64 phys = vm_addr(vma, (u64)*list++, mem->memtype, target);
-		nvkm_wo32(pgt, pte + 0, lower_32_bits(phys));
-		nvkm_wo32(pgt, pte + 4, upper_32_bits(phys));
-		pte += 8;
-	}
-	nvkm_done(pgt);
-}
-
-void
-nv50_vm_unmap(struct nvkm_vma *vma, struct nvkm_memory *pgt, u32 pte, u32 cnt)
-{
-	pte <<= 3;
-	nvkm_kmap(pgt);
-	while (cnt--) {
-		nvkm_wo32(pgt, pte + 0, 0x00000000);
-		nvkm_wo32(pgt, pte + 4, 0x00000000);
-		pte += 8;
-	}
-	nvkm_done(pgt);
-}
-
-void
-nv50_vm_flush(struct nvkm_vm *vm)
-{
-	struct nvkm_mmu *mmu = vm->mmu;
-	struct nvkm_subdev *subdev = &mmu->subdev;
-	struct nvkm_device *device = subdev->device;
-	int i, vme;
-
-	mutex_lock(&subdev->mutex);
-	for (i = 0; i < NVKM_SUBDEV_NR; i++) {
-		if (!atomic_read(&vm->engref[i]))
-			continue;
-
-		/* unfortunate hw bug workaround... */
-		if (i == NVKM_ENGINE_GR && device->gr) {
-			int ret = nvkm_gr_tlb_flush(device->gr);
-			if (ret != -ENODEV)
-				continue;
-		}
-
-		switch (i) {
-		case NVKM_ENGINE_GR    : vme = 0x00; break;
-		case NVKM_ENGINE_VP    :
-		case NVKM_ENGINE_MSPDEC: vme = 0x01; break;
-		case NVKM_SUBDEV_BAR   : vme = 0x06; break;
-		case NVKM_ENGINE_MSPPP :
-		case NVKM_ENGINE_MPEG  : vme = 0x08; break;
-		case NVKM_ENGINE_BSP   :
-		case NVKM_ENGINE_MSVLD : vme = 0x09; break;
-		case NVKM_ENGINE_CIPHER:
-		case NVKM_ENGINE_SEC   : vme = 0x0a; break;
-		case NVKM_ENGINE_CE0   : vme = 0x0d; break;
-		default:
-			continue;
-		}
-
-		nvkm_wr32(device, 0x100c80, (vme << 16) | 1);
-		if (nvkm_msec(device, 2000,
-			if (!(nvkm_rd32(device, 0x100c80) & 0x00000001))
-				break;
-		) < 0)
-			nvkm_error(subdev, "vm flush timeout: engine %d\n", vme);
-	}
-	mutex_unlock(&subdev->mutex);
+	/* 0x01: no bank swizzle
+	 * 0x02: bank swizzled
+	 * 0x7f: invalid
+	 *
+	 * 0x01/0x02 are values understood by the VRAM allocator,
+	 * and are required to avoid mixing the two types within
+	 * a certain range.
+	 */
+	static const u8
+	kind[128] = {
+		0x01, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, /* 0x00 */
+		0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+		0x01, 0x01, 0x01, 0x01, 0x7f, 0x7f, 0x7f, 0x7f, /* 0x10 */
+		0x02, 0x02, 0x02, 0x02, 0x7f, 0x7f, 0x7f, 0x7f,
+		0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x7f, /* 0x20 */
+		0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x7f,
+		0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, /* 0x30 */
+		0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+		0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x02, /* 0x40 */
+		0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x7f, 0x7f,
+		0x7f, 0x7f, 0x7f, 0x7f, 0x01, 0x01, 0x01, 0x7f, /* 0x50 */
+		0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f, 0x7f,
+		0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x7f, /* 0x60 */
+		0x01, 0x01, 0x01, 0x01, 0x02, 0x02, 0x02, 0x02,
+		0x01, 0x7f, 0x02, 0x7f, 0x01, 0x7f, 0x02, 0x7f, /* 0x70 */
+		0x01, 0x01, 0x02, 0x02, 0x01, 0x01, 0x7f, 0x7f
+	};
+	*count = ARRAY_SIZE(kind);
+	return kind;
 }
 
 static const struct nvkm_mmu_func
@@ -211,12 +66,8 @@ nv50_mmu = {
 	.pgt_bits  = 29 - 12,
 	.spg_shift = 12,
 	.lpg_shift = 16,
-	.map_pgt = nv50_vm_map_pgt,
-	.map = nv50_vm_map,
-	.map_sg = nv50_vm_map_sg,
-	.unmap = nv50_vm_unmap,
-	.flush = nv50_vm_flush,
 	.vmm = {{ -1, -1, NVIF_CLASS_VMM_NV50}, nv50_vmm_new, false, 0x1400 },
+	.kind = nv50_mmu_kind,
 };
 
 int
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/priv.h b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/priv.h
index 3271ec1c3c83..326781246714 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/priv.h
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/priv.h
@@ -37,17 +37,13 @@ struct nvkm_mmu_func {
 		bool global;
 		u32 pd_offset;
 	} vmm;
+
+	const u8 *(*kind)(struct nvkm_mmu *, int *count);
 };
 
 extern const struct nvkm_mmu_func nv04_mmu;
 
-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 *,
-		    u32, u32, dma_addr_t *);
-void nv50_vm_unmap(struct nvkm_vma *, struct nvkm_memory *, u32, u32);
-void nv50_vm_flush(struct nvkm_vm *);
+const u8 *nv50_mmu_kind(struct nvkm_mmu *, int *count);
 
 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 *,
diff --git a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmnv50.c b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmnv50.c
index db27381b67c6..863a2edd9861 100644
--- a/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmnv50.c
+++ b/drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmnv50.c
@@ -21,15 +21,146 @@
  */
 #include "vmm.h"
 
+#include <subdev/fb.h>
+#include <subdev/timer.h>
+#include <engine/gr.h>
+
 #include <nvif/if500d.h>
 #include <nvif/unpack.h>
 
+static inline void
+nv50_vmm_pgt_pte(struct nvkm_vmm *vmm, struct nvkm_mmu_pt *pt,
+		 u32 ptei, u32 ptes, struct nvkm_vmm_map *map, u64 addr)
+{
+	u64 next = addr | map->type, data;
+	u32 pten;
+	int log2blk;
+
+	map->type += ptes * map->ctag;
+
+	while (ptes) {
+		for (log2blk = 7; log2blk >= 0; log2blk--) {
+			pten = 1 << log2blk;
+			if (ptes >= pten && IS_ALIGNED(ptei, pten))
+				break;
+		}
+
+		data  = next | (log2blk << 7);
+		next += pten * map->next;
+		ptes -= pten;
+
+		while (pten--)
+			VMM_WO064(pt, vmm, ptei++ * 8, data);
+	}
+}
+
+static void
+nv50_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, nv50_vmm_pgt_pte);
+}
+
+static void
+nv50_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++ * 8, data);
+			map->type += map->ctag;
+		}
+		nvkm_done(pt->memory);
+		return;
+	}
+
+	VMM_MAP_ITER_DMA(vmm, pt, ptei, ptes, map, nv50_vmm_pgt_pte);
+}
+
+static void
+nv50_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, nv50_vmm_pgt_pte);
+}
+
+static void
+nv50_vmm_pgt_unmap(struct nvkm_vmm *vmm,
+		   struct nvkm_mmu_pt *pt, u32 ptei, u32 ptes)
+{
+	VMM_FO064(pt, vmm, ptei * 8, 0ULL, ptes);
+}
+
 static const struct nvkm_vmm_desc_func
 nv50_vmm_pgt = {
+	.unmap = nv50_vmm_pgt_unmap,
+	.mem = nv50_vmm_pgt_mem,
+	.dma = nv50_vmm_pgt_dma,
+	.sgl = nv50_vmm_pgt_sgl,
 };
 
+static bool
+nv50_vmm_pde(struct nvkm_vmm *vmm, struct nvkm_vmm_pt *pgt, u64 *pdata)
+{
+	struct nvkm_mmu_pt *pt;
+	u64 data = 0xdeadcafe00000000ULL;
+	if (pgt && (pt = pgt->pt[0])) {
+		switch (pgt->page) {
+		case 16: data = 0x00000001; break;
+		case 12: data = 0x00000003;
+			switch (nvkm_memory_size(pt->memory)) {
+			case 0x100000: data |= 0x00000000; break;
+			case 0x040000: data |= 0x00000020; break;
+			case 0x020000: data |= 0x00000040; break;
+			case 0x010000: data |= 0x00000060; break;
+			default:
+				WARN_ON(1);
+				return false;
+			}
+			break;
+		default:
+			WARN_ON(1);
+			return false;
+		}
+
+		switch (nvkm_memory_target(pt->memory)) {
+		case NVKM_MEM_TARGET_VRAM: data |= 0x00000000; break;
+		case NVKM_MEM_TARGET_HOST: data |= 0x00000008; break;
+		case NVKM_MEM_TARGET_NCOH: data |= 0x0000000c; break;
+		default:
+			WARN_ON(1);
+			return false;
+		}
+
+		data |= pt->addr;
+	}
+	*pdata = data;
+	return true;
+}
+
+static void
+nv50_vmm_pgd_pde(struct nvkm_vmm *vmm, struct nvkm_vmm_pt *pgd, u32 pdei)
+{
+	struct nvkm_vmm_join *join;
+	u32 pdeo = vmm->mmu->func->vmm.pd_offset + (pdei * 8);
+	u64 data;
+
+	if (!nv50_vmm_pde(vmm, pgd->pde[pdei], &data))
+		return;
+
+	list_for_each_entry(join, &vmm->join, head) {
+		nvkm_kmap(join->inst);
+		nvkm_wo64(join->inst, pdeo, data);
+		nvkm_done(join->inst);
+	}
+}
+
 static const struct nvkm_vmm_desc_func
 nv50_vmm_pgd = {
+	.pde = nv50_vmm_pgd_pde,
 };
 
 static const struct nvkm_vmm_desc
@@ -47,6 +178,150 @@ nv50_vmm_desc_16[] = {
 };
 
 static void
+nv50_vmm_flush(struct nvkm_vmm *vmm, int level)
+{
+	struct nvkm_subdev *subdev = &vmm->mmu->subdev;
+	struct nvkm_device *device = subdev->device;
+	int i, id;
+
+	mutex_lock(&subdev->mutex);
+	for (i = 0; i < NVKM_SUBDEV_NR; i++) {
+		if (!atomic_read(&vmm->engref[i]))
+			continue;
+
+		/* unfortunate hw bug workaround... */
+		if (i == NVKM_ENGINE_GR && device->gr) {
+			int ret = nvkm_gr_tlb_flush(device->gr);
+			if (ret != -ENODEV)
+				continue;
+		}
+
+		switch (i) {
+		case NVKM_ENGINE_GR    : id = 0x00; break;
+		case NVKM_ENGINE_VP    :
+		case NVKM_ENGINE_MSPDEC: id = 0x01; break;
+		case NVKM_SUBDEV_BAR   : id = 0x06; break;
+		case NVKM_ENGINE_MSPPP :
+		case NVKM_ENGINE_MPEG  : id = 0x08; break;
+		case NVKM_ENGINE_BSP   :
+		case NVKM_ENGINE_MSVLD : id = 0x09; break;
+		case NVKM_ENGINE_CIPHER:
+		case NVKM_ENGINE_SEC   : id = 0x0a; break;
+		case NVKM_ENGINE_CE0   : id = 0x0d; break;
+		default:
+			continue;
+		}
+
+		nvkm_wr32(device, 0x100c80, (id << 16) | 1);
+		if (nvkm_msec(device, 2000,
+			if (!(nvkm_rd32(device, 0x100c80) & 0x00000001))
+				break;
+		) < 0)
+			nvkm_error(subdev, "%s mmu invalidate timeout\n",
+				   nvkm_subdev_name[i]);
+	}
+	mutex_unlock(&subdev->mutex);
+}
+
+static int
+nv50_vmm_valid(struct nvkm_vmm *vmm, void *argv, u32 argc,
+	       struct nvkm_vmm_map *map)
+{
+	const struct nvkm_vmm_page *page = map->page;
+	union {
+		struct nv50_vmm_map_vn vn;
+		struct nv50_vmm_map_v0 v0;
+	} *args = argv;
+	struct nvkm_device *device = vmm->mmu->subdev.device;
+	struct nvkm_ram *ram = device->fb->ram;
+	struct nvkm_memory *memory = map->memory;
+	u8  aper, kind, comp, priv, ro;
+	int kindn, ret = -ENOSYS;
+	const u8 *kindm;
+
+	map->type = map->ctag = 0;
+	map->next = 1 << page->shift;
+
+	if (!(ret = nvif_unpack(ret, &argv, &argc, args->v0, 0, 0, false))) {
+		ro   = !!args->v0.ro;
+		priv = !!args->v0.priv;
+		kind = args->v0.kind & 0x7f;
+		comp = args->v0.comp & 0x03;
+	} else
+	if (!(ret = nvif_unvers(ret, &argv, &argc, args->vn))) {
+		ro   = 0;
+		priv = 0;
+		kind = 0x00;
+		comp = 0;
+	} else {
+		VMM_DEBUG(vmm, "args");
+		return ret;
+	}
+
+	switch (nvkm_memory_target(memory)) {
+	case NVKM_MEM_TARGET_VRAM:
+		if (ram->stolen) {
+			map->type |= ram->stolen;
+			aper = 3;
+		} else {
+			aper = 0;
+		}
+		break;
+	case NVKM_MEM_TARGET_HOST:
+		aper = 2;
+		break;
+	case NVKM_MEM_TARGET_NCOH:
+		aper = 3;
+		break;
+	default:
+		WARN_ON(1);
+		return -EINVAL;
+	}
+
+	kindm = vmm->mmu->func->kind(vmm->mmu, &kindn);
+	if (kind >= kindn || kindm[kind] == 0x7f) {
+		VMM_DEBUG(vmm, "kind %02x", kind);
+		return -EINVAL;
+	}
+
+	if (map->mem && map->mem->type != kindm[kind]) {
+		VMM_DEBUG(vmm, "kind %02x bankswz: %d %d", kind,
+			  kindm[kind], map->mem->type);
+		return -EINVAL;
+	}
+
+	if (comp) {
+		u32 tags = (nvkm_memory_size(memory) >> 16) * comp;
+		if (aper != 0 || !(page->type & NVKM_VMM_PAGE_COMP)) {
+			VMM_DEBUG(vmm, "comp %d %02x", aper, page->type);
+			return -EINVAL;
+		}
+
+		ret = nvkm_memory_tags_get(memory, device, tags, NULL,
+					   &map->tags);
+		if (ret) {
+			VMM_DEBUG(vmm, "comp %d", ret);
+			return ret;
+		}
+
+		if (map->tags->mn) {
+			u32 tags = map->tags->mn->offset + (map->offset >> 16);
+			map->ctag |= (u64)comp << 49;
+			map->type |= (u64)comp << 47;
+			map->type |= (u64)tags << 49;
+			map->next |= map->ctag;
+		}
+	}
+
+	map->type |= BIT(0); /* Valid. */
+	map->type |= (u64)ro << 3;
+	map->type |= (u64)aper << 4;
+	map->type |= (u64)priv << 6;
+	map->type |= (u64)kind << 40;
+	return 0;
+}
+
+static void
 nv50_vmm_part(struct nvkm_vmm *vmm, struct nvkm_memory *inst)
 {
 	struct nvkm_vmm_join *join;
@@ -63,19 +338,35 @@ nv50_vmm_part(struct nvkm_vmm *vmm, struct nvkm_memory *inst)
 static int
 nv50_vmm_join(struct nvkm_vmm *vmm, struct nvkm_memory *inst)
 {
+	const u32 pd_offset = vmm->mmu->func->vmm.pd_offset;
 	struct nvkm_vmm_join *join;
+	int ret = 0;
+	u64 data;
+	u32 pdei;
 
 	if (!(join = kmalloc(sizeof(*join), GFP_KERNEL)))
 		return -ENOMEM;
 	join->inst = inst;
 	list_add_tail(&join->head, &vmm->join);
-	return 0;
+
+	nvkm_kmap(join->inst);
+	for (pdei = vmm->start >> 29; pdei <= (vmm->limit - 1) >> 29; pdei++) {
+		if (!nv50_vmm_pde(vmm, vmm->pd->pde[pdei], &data)) {
+			ret = -EINVAL;
+			break;
+		}
+		nvkm_wo64(join->inst, pd_offset + (pdei * 8), data);
+	}
+	nvkm_done(join->inst);
+	return ret;
 }
 
 static const struct nvkm_vmm_func
 nv50_vmm = {
 	.join = nv50_vmm_join,
 	.part = nv50_vmm_part,
+	.valid = nv50_vmm_valid,
+	.flush = nv50_vmm_flush,
 	.page_block = 1 << 29,
 	.page = {
 		{ 16, &nv50_vmm_desc_16[0], NVKM_VMM_PAGE_xVxC },