drm/nouveau/mmu/gf100: implement new vmm backend

Adds support for:
- 64KiB big page size.
- System-memory PTs.
Signed-off-by: Ben Skeggs <bskeggs@redhat.com>

drivers/gpu/drm/nouveau/include/nvif/if900d.h

@@ -5,4 +5,17 @@
 struct gf100_vmm_vn {
 	/* nvif_vmm_vX ... */
 };
+
+struct gf100_vmm_map_vn {
+	/* nvif_vmm_map_vX ... */
+};
+
+struct gf100_vmm_map_v0 {
+	/* nvif_vmm_map_vX ... */
+	__u8  version;
+	__u8  vol;
+	__u8  ro;
+	__u8  priv;
+	__u8  kind;
+};
 #endif

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/base.c

@@ -28,6 +28,7 @@
 #include <subdev/fb.h>
 
 #include <nvif/if500d.h>
+#include <nvif/if900d.h>
 
 struct nvkm_mmu_ptp {
 	struct nvkm_mmu_pt *pt;
@@ -222,6 +223,7 @@ nvkm_vm_map_(const struct nvkm_vmm_page *page, struct nvkm_vma *vma, u64 delta,
 {
 	union {
 		struct nv50_vmm_map_v0 nv50;
+		struct gf100_vmm_map_v0 gf100;
 	} args;
 	struct nvkm_vmm *vmm = vma->vm;
 	void *argv = NULL;
@@ -242,6 +244,19 @@ nvkm_vm_map_(const struct nvkm_vmm_page *page, struct nvkm_vma *vma, u64 delta,
 			argv = &args.nv50;
 			argc = sizeof(args.nv50);
 			break;
+		case NV_C0:
+		case NV_E0:
+		case GM100:
+		case GP100: {
+			args.gf100.version = 0;
+			args.gf100.vol = (nvkm_memory_target(map->memory) != NVKM_MEM_TARGET_VRAM);
+			args.gf100.ro = !(vma->access & NV_MEM_ACCESS_WO);
+			args.gf100.priv = !!(vma->access & NV_MEM_ACCESS_SYS);
+			args.gf100.kind = (mem->memtype & 0x0ff);
+			argv = &args.gf100;
+			argc = sizeof(args.gf100);
+		}
+			break;
 		default:
 			break;
 		}

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/gf100.c

@@ -68,6 +68,12 @@ const u8 gf100_pte_storage_type_map[256] =
 	0xfe, 0xfe, 0xfe, 0xfe, 0xff, 0xfd, 0xfe, 0xff
 };
 
+const u8 *
+gf100_mmu_kind(struct nvkm_mmu *mmu, int *count)
+{
+	*count = ARRAY_SIZE(gf100_pte_storage_type_map);
+	return gf100_pte_storage_type_map;
+}
 
 void
 gf100_vm_map_pgt(struct nvkm_vmm *vmm, u32 index, struct nvkm_memory *pgt[2])
@@ -195,12 +201,8 @@ gf100_mmu = {
 	.pgt_bits  = 27 - 12,
 	.spg_shift = 12,
 	.lpg_shift = 17,
-	.map_pgt = gf100_vm_map_pgt,
-	.map = gf100_vm_map,
-	.map_sg = gf100_vm_map_sg,
-	.unmap = gf100_vm_unmap,
-	.flush = gf100_vm_flush,
 	.vmm = {{ -1, -1, NVIF_CLASS_VMM_GF100}, gf100_vmm_new },
+	.kind = gf100_mmu_kind,
 };
 
 int

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/priv.h

@@ -52,6 +52,7 @@ void gf100_vm_map_sg(struct nvkm_vma *, struct nvkm_memory *, struct nvkm_mem *,
 		     u32, u32, dma_addr_t *);
 void gf100_vm_unmap(struct nvkm_vma *, struct nvkm_memory *, u32, u32);
 void gf100_vm_flush(struct nvkm_vm *);
+const u8 *gf100_mmu_kind(struct nvkm_mmu *, int *count);
 
 struct nvkm_mmu_pt {
 	union {

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h

@@ -70,7 +70,15 @@ struct nvkm_vmm_desc_func {
 };
 
 extern const struct nvkm_vmm_desc_func gf100_vmm_pgd;
+void gf100_vmm_pgd_pde(struct nvkm_vmm *, struct nvkm_vmm_pt *, u32);
 extern const struct nvkm_vmm_desc_func gf100_vmm_pgt;
+void gf100_vmm_pgt_unmap(struct nvkm_vmm *, struct nvkm_mmu_pt *, u32, u32);
+void gf100_vmm_pgt_mem(struct nvkm_vmm *, struct nvkm_mmu_pt *, u32, u32,
+		       struct nvkm_vmm_map *);
+void gf100_vmm_pgt_dma(struct nvkm_vmm *, struct nvkm_mmu_pt *, u32, u32,
+		       struct nvkm_vmm_map *);
+void gf100_vmm_pgt_sgl(struct nvkm_vmm *, struct nvkm_mmu_pt *, u32, u32,
+		       struct nvkm_vmm_map *);
 
 struct nvkm_vmm_desc {
 	enum {
@@ -165,6 +173,10 @@ int gf100_vmm_new_(const struct nvkm_vmm_func *, const struct nvkm_vmm_func *,
 int gf100_vmm_join_(struct nvkm_vmm *, struct nvkm_memory *, u64 base);
 int gf100_vmm_join(struct nvkm_vmm *, struct nvkm_memory *);
 void gf100_vmm_part(struct nvkm_vmm *, struct nvkm_memory *);
+int gf100_vmm_aper(enum nvkm_memory_target);
+int gf100_vmm_valid(struct nvkm_vmm *, void *, u32, struct nvkm_vmm_map *);
+void gf100_vmm_flush_(struct nvkm_vmm *, int);
+void gf100_vmm_flush(struct nvkm_vmm *, int);
 
 int gm200_vmm_new_(const struct nvkm_vmm_func *, const struct nvkm_vmm_func *,
 		   struct nvkm_mmu *, u64, u64, void *, u32,

drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgf100.c

@@ -22,13 +22,131 @@
 #include "vmm.h"
 
 #include <subdev/fb.h>
+#include <subdev/ltc.h>
+#include <subdev/timer.h>
+
+#include <nvif/if900d.h>
+#include <nvif/unpack.h>
+
+static inline void
+gf100_vmm_pgt_pte(struct nvkm_vmm *vmm, struct nvkm_mmu_pt *pt,
+		  u32 ptei, u32 ptes, struct nvkm_vmm_map *map, u64 addr)
+{
+	u64 base = (addr >> 8) | map->type;
+	u64 data = base;
+
+	if (map->ctag && !(map->next & (1ULL << 44))) {
+		while (ptes--) {
+			data = base | ((map->ctag >> 1) << 44);
+			if (!(map->ctag++ & 1))
+				data |= BIT_ULL(60);
+
+			VMM_WO064(pt, vmm, ptei++ * 8, data);
+			base += map->next;
+		}
+	} else {
+		map->type += ptes * map->ctag;
+
+		while (ptes--) {
+			VMM_WO064(pt, vmm, ptei++ * 8, data);
+			data += map->next;
+		}
+	}
+}
+
+void
+gf100_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, gf100_vmm_pgt_pte);
+}
+
+void
+gf100_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++ >> 8) | 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, gf100_vmm_pgt_pte);
+}
+
+void
+gf100_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, gf100_vmm_pgt_pte);
+}
+
+void
+gf100_vmm_pgt_unmap(struct nvkm_vmm *vmm,
+		    struct nvkm_mmu_pt *pt, u32 ptei, u32 ptes)
+{
+	VMM_FO064(pt, vmm, ptei * 8, 0ULL, ptes);
+}
 
 const struct nvkm_vmm_desc_func
 gf100_vmm_pgt = {
+	.unmap = gf100_vmm_pgt_unmap,
+	.mem = gf100_vmm_pgt_mem,
+	.dma = gf100_vmm_pgt_dma,
+	.sgl = gf100_vmm_pgt_sgl,
 };
 
+void
+gf100_vmm_pgd_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];
+	struct nvkm_mmu_pt *pt;
+	u64 data = 0;
+
+	if ((pt = pgt->pt[0])) {
+		switch (nvkm_memory_target(pt->memory)) {
+		case NVKM_MEM_TARGET_VRAM: data |= 1ULL << 0; break;
+		case NVKM_MEM_TARGET_HOST: data |= 2ULL << 0;
+			data |= BIT_ULL(35); /* VOL */
+			break;
+		case NVKM_MEM_TARGET_NCOH: data |= 3ULL << 0; break;
+		default:
+			WARN_ON(1);
+			return;
+		}
+		data |= pt->addr >> 8;
+	}
+
+	if ((pt = pgt->pt[1])) {
+		switch (nvkm_memory_target(pt->memory)) {
+		case NVKM_MEM_TARGET_VRAM: data |= 1ULL << 32; break;
+		case NVKM_MEM_TARGET_HOST: data |= 2ULL << 32;
+			data |= BIT_ULL(34); /* VOL */
+			break;
+		case NVKM_MEM_TARGET_NCOH: data |= 3ULL << 32; break;
+		default:
+			WARN_ON(1);
+			return;
+		}
+		data |= pt->addr << 24;
+	}
+
+	nvkm_kmap(pd->memory);
+	VMM_WO064(pd, vmm, pdei * 8, data);
+	nvkm_done(pd->memory);
+}
+
 const struct nvkm_vmm_desc_func
 gf100_vmm_pgd = {
+	.unmap = gf100_vmm_pgt_unmap,
+	.pde = gf100_vmm_pgd_pde,
 };
 
 static const struct nvkm_vmm_desc
@@ -59,6 +177,140 @@ gf100_vmm_desc_16_16[] = {
 	{}
 };
 
+void
+gf100_vmm_flush_(struct nvkm_vmm *vmm, int depth)
+{
+	struct nvkm_subdev *subdev = &vmm->mmu->subdev;
+	struct nvkm_device *device = subdev->device;
+	u32 type;
+
+	type = 0x00000001; /* PAGE_ALL */
+	if (atomic_read(&vmm->engref[NVKM_SUBDEV_BAR]))
+		type |= 0x00000004; /* HUB_ONLY */
+
+	mutex_lock(&subdev->mutex);
+	/* 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, vmm->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(&subdev->mutex);
+}
+
+void
+gf100_vmm_flush(struct nvkm_vmm *vmm, int depth)
+{
+	gf100_vmm_flush_(vmm, 0);
+}
+
+int
+gf100_vmm_valid(struct nvkm_vmm *vmm, void *argv, u32 argc,
+		struct nvkm_vmm_map *map)
+{
+	const enum nvkm_memory_target target = nvkm_memory_target(map->memory);
+	const struct nvkm_vmm_page *page = map->page;
+	const bool gm20x = page->desc->func->sparse != NULL;
+	union {
+		struct gf100_vmm_map_vn vn;
+		struct gf100_vmm_map_v0 v0;
+	} *args = argv;
+	struct nvkm_device *device = vmm->mmu->subdev.device;
+	struct nvkm_memory *memory = map->memory;
+	u8  kind, priv, ro, vol;
+	int kindn, aper, ret = -ENOSYS;
+	const u8 *kindm;
+
+	map->next = (1 << page->shift) >> 8;
+	map->type = map->ctag = 0;
+
+	if (!(ret = nvif_unpack(ret, &argv, &argc, args->v0, 0, 0, false))) {
+		vol  = !!args->v0.vol;
+		ro   = !!args->v0.ro;
+		priv = !!args->v0.priv;
+		kind =   args->v0.kind;
+	} else
+	if (!(ret = nvif_unvers(ret, &argv, &argc, args->vn))) {
+		vol  = target == NVKM_MEM_TARGET_HOST;
+		ro   = 0;
+		priv = 0;
+		kind = 0x00;
+	} else {
+		VMM_DEBUG(vmm, "args");
+		return ret;
+	}
+
+	aper = vmm->func->aper(target);
+	if (WARN_ON(aper < 0))
+		return aper;
+
+	kindm = vmm->mmu->func->kind(vmm->mmu, &kindn);
+	if (kind >= kindn || kindm[kind] == 0xff) {
+		VMM_DEBUG(vmm, "kind %02x", kind);
+		return -EINVAL;
+	}
+
+	if (kindm[kind] != kind) {
+		u32 comp = (page->shift == 16 && !gm20x) ? 16 : 17;
+		u32 tags = ALIGN(nvkm_memory_size(memory), 1 << 17) >> 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,
+					   nvkm_ltc_tags_clear,
+					   &map->tags);
+		if (ret) {
+			VMM_DEBUG(vmm, "comp %d", ret);
+			return ret;
+		}
+
+		if (map->tags->mn) {
+			u64 tags = map->tags->mn->offset + (map->offset >> 17);
+			if (page->shift == 17 || !gm20x) {
+				map->type |= tags << 44;
+				map->ctag |= 1ULL << 44;
+				map->next |= 1ULL << 44;
+			} else {
+				map->ctag |= tags << 1 | 1;
+			}
+		} else {
+			kind = kindm[kind];
+		}
+	}
+
+	map->type |= BIT(0);
+	map->type |= (u64)priv << 1;
+	map->type |= (u64)  ro << 2;
+	map->type |= (u64) vol << 32;
+	map->type |= (u64)aper << 33;
+	map->type |= (u64)kind << 36;
+	return 0;
+}
+
+int
+gf100_vmm_aper(enum nvkm_memory_target target)
+{
+	switch (target) {
+	case NVKM_MEM_TARGET_VRAM: return 0;
+	case NVKM_MEM_TARGET_HOST: return 2;
+	case NVKM_MEM_TARGET_NCOH: return 3;
+	default:
+		return -EINVAL;
+	}
+}
+
 void
 gf100_vmm_part(struct nvkm_vmm *vmm, struct nvkm_memory *inst)
 {
@@ -99,6 +351,9 @@ static const struct nvkm_vmm_func
 gf100_vmm_17 = {
 	.join = gf100_vmm_join,
 	.part = gf100_vmm_part,
+	.aper = gf100_vmm_aper,
+	.valid = gf100_vmm_valid,
+	.flush = gf100_vmm_flush,
 	.page = {
 		{ 17, &gf100_vmm_desc_17_17[0], NVKM_VMM_PAGE_xVxC },
 		{ 12, &gf100_vmm_desc_17_12[0], NVKM_VMM_PAGE_xVHx },
@@ -110,6 +365,9 @@ static const struct nvkm_vmm_func
 gf100_vmm_16 = {
 	.join = gf100_vmm_join,
 	.part = gf100_vmm_part,
+	.aper = gf100_vmm_aper,
+	.valid = gf100_vmm_valid,
+	.flush = gf100_vmm_flush,
 	.page = {
 		{ 16, &gf100_vmm_desc_16_16[0], NVKM_VMM_PAGE_xVxC },
 		{ 12, &gf100_vmm_desc_16_12[0], NVKM_VMM_PAGE_xVHx },