drm/nouveau/mmu/gf100: implement new vmm backend
authorBen Skeggs <bskeggs@redhat.com>
Tue, 31 Oct 2017 17:56:19 +0000 (03:56 +1000)
committerBen Skeggs <bskeggs@redhat.com>
Thu, 2 Nov 2017 03:32:29 +0000 (13:32 +1000)
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
drivers/gpu/drm/nouveau/nvkm/subdev/mmu/base.c
drivers/gpu/drm/nouveau/nvkm/subdev/mmu/gf100.c
drivers/gpu/drm/nouveau/nvkm/subdev/mmu/priv.h
drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmm.h
drivers/gpu/drm/nouveau/nvkm/subdev/mmu/vmmgf100.c

index 112716f889c68baed3fd8c487b5bf3e891a0efc5..49aa50583c3dc5c96d3fbfff77aba3a0d2bd41a0 100644 (file)
@@ -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
index 4d75dcb4146891c94551e14792d571660d97bcee..741021ff8c27ecd3f68d44ef9f4b71c8cfe73365 100644 (file)
@@ -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;
                }
index 167c201d29e3594197bb53eda97b0a3dfe368404..39bf26684b4112c30749c6c57fb8524c02d98fb7 100644 (file)
@@ -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
index 326781246714fae83686a5780f83253677df25ef..64f765ec3e6e47245a8e3ec4080c43f5d5f00dc5 100644 (file)
@@ -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 {
index cbb8d47fbd11919c56f1329a1e5a2a7e5ec6ee21..dfc4b46129bdc564e04ece424878ab1854aadd6b 100644 (file)
@@ -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,
index f8234af0120cda58d197f1cd66c75edf1d9a46b7..ddd44ad797b78988fbfdf0ba676e98f9106629dc 100644 (file)
 #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 },