drivers/gpu/drm/nouveau/nouveau_sgdma.c

   1 #include "drmP.h"
   2 #include "nouveau_drv.h"
   3 #include <linux/pagemap.h>
   4 #include <linux/slab.h>
   5
   6 #define NV_CTXDMA_PAGE_SHIFT 12
   7 #define NV_CTXDMA_PAGE_SIZE  (1 << NV_CTXDMA_PAGE_SHIFT)
   8 #define NV_CTXDMA_PAGE_MASK  (NV_CTXDMA_PAGE_SIZE - 1)
   9
  10 struct nouveau_sgdma_be {
  11         struct ttm_backend backend;
  12         struct drm_device *dev;
  13
  14         dma_addr_t *pages;
  15         unsigned nr_pages;
  16
  17         unsigned pte_start;
  18         bool bound;
  19 };
  20
  21 static int
  22 nouveau_sgdma_populate(struct ttm_backend *be, unsigned long num_pages,
  23                        struct page **pages, struct page *dummy_read_page)
  24 {
  25         struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
  26         struct drm_device *dev = nvbe->dev;
  27
  28         NV_DEBUG(nvbe->dev, "num_pages = %ld\n", num_pages);
  29
  30         if (nvbe->pages)
  31                 return -EINVAL;
  32
  33         nvbe->pages = kmalloc(sizeof(dma_addr_t) * num_pages, GFP_KERNEL);
  34         if (!nvbe->pages)
  35                 return -ENOMEM;
  36
  37         nvbe->nr_pages = 0;
  38         while (num_pages--) {
  39                 nvbe->pages[nvbe->nr_pages] =
  40                         pci_map_page(dev->pdev, pages[nvbe->nr_pages], 0,
  41                                      PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
  42                 if (pci_dma_mapping_error(dev->pdev,
  43                                           nvbe->pages[nvbe->nr_pages])) {
  44                         be->func->clear(be);
  45                         return -EFAULT;
  46                 }
  47
  48                 nvbe->nr_pages++;
  49         }
  50
  51         return 0;
  52 }
  53
  54 static void
  55 nouveau_sgdma_clear(struct ttm_backend *be)
  56 {
  57         struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
  58         struct drm_device *dev;
  59
  60         if (nvbe && nvbe->pages) {
  61                 dev = nvbe->dev;
  62                 NV_DEBUG(dev, "\n");
  63
  64                 if (nvbe->bound)
  65                         be->func->unbind(be);
  66
  67                 while (nvbe->nr_pages--) {
  68                         pci_unmap_page(dev->pdev, nvbe->pages[nvbe->nr_pages],
  69                                        PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
  70                 }
  71                 kfree(nvbe->pages);
  72                 nvbe->pages = NULL;
  73                 nvbe->nr_pages = 0;
  74         }
  75 }
  76
  77 static inline unsigned
  78 nouveau_sgdma_pte(struct drm_device *dev, uint64_t offset)
  79 {
  80         struct drm_nouveau_private *dev_priv = dev->dev_private;
  81         unsigned pte = (offset >> NV_CTXDMA_PAGE_SHIFT);
  82
  83         if (dev_priv->card_type < NV_50)
  84                 return pte + 2;
  85
  86         return pte << 1;
  87 }
  88
  89 static int
  90 nouveau_sgdma_bind(struct ttm_backend *be, struct ttm_mem_reg *mem)
  91 {
  92         struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
  93         struct drm_device *dev = nvbe->dev;
  94         struct drm_nouveau_private *dev_priv = dev->dev_private;
  95         struct nouveau_gpuobj *gpuobj = dev_priv->gart_info.sg_ctxdma;
  96         unsigned i, j, pte;
  97
  98         NV_DEBUG(dev, "pg=0x%lx\n", mem->mm_node->start);
  99
 100         pte = nouveau_sgdma_pte(nvbe->dev, mem->mm_node->start << PAGE_SHIFT);
 101         nvbe->pte_start = pte;
 102         for (i = 0; i < nvbe->nr_pages; i++) {
 103                 dma_addr_t dma_offset = nvbe->pages[i];
 104                 uint32_t offset_l = lower_32_bits(dma_offset);
 105                 uint32_t offset_h = upper_32_bits(dma_offset);
 106
 107                 for (j = 0; j < PAGE_SIZE / NV_CTXDMA_PAGE_SIZE; j++) {
 108                         if (dev_priv->card_type < NV_50)
 109                                 nv_wo32(dev, gpuobj, pte++, offset_l | 3);
 110                         else {
 111                                 nv_wo32(dev, gpuobj, pte++, offset_l | 0x21);
 112                                 nv_wo32(dev, gpuobj, pte++, offset_h & 0xff);
 113                         }
 114
 115                         dma_offset += NV_CTXDMA_PAGE_SIZE;
 116                 }
 117         }
 118         dev_priv->engine.instmem.flush(nvbe->dev);
 119
 120         if (dev_priv->card_type == NV_50) {
 121                 nv50_vm_flush(dev, 5); /* PGRAPH */
 122                 nv50_vm_flush(dev, 0); /* PFIFO */
 123         }
 124
 125         nvbe->bound = true;
 126         return 0;
 127 }
 128
 129 static int
 130 nouveau_sgdma_unbind(struct ttm_backend *be)
 131 {
 132         struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
 133         struct drm_device *dev = nvbe->dev;
 134         struct drm_nouveau_private *dev_priv = dev->dev_private;
 135         struct nouveau_gpuobj *gpuobj = dev_priv->gart_info.sg_ctxdma;
 136         unsigned i, j, pte;
 137
 138         NV_DEBUG(dev, "\n");
 139
 140         if (!nvbe->bound)
 141                 return 0;
 142
 143         pte = nvbe->pte_start;
 144         for (i = 0; i < nvbe->nr_pages; i++) {
 145                 dma_addr_t dma_offset = dev_priv->gart_info.sg_dummy_bus;
 146
 147                 for (j = 0; j < PAGE_SIZE / NV_CTXDMA_PAGE_SIZE; j++) {
 148                         if (dev_priv->card_type < NV_50)
 149                                 nv_wo32(dev, gpuobj, pte++, dma_offset | 3);
 150                         else {
 151                                 nv_wo32(dev, gpuobj, pte++, dma_offset | 0x21);
 152                                 nv_wo32(dev, gpuobj, pte++, 0x00000000);
 153                         }
 154
 155                         dma_offset += NV_CTXDMA_PAGE_SIZE;
 156                 }
 157         }
 158         dev_priv->engine.instmem.flush(nvbe->dev);
 159
 160         if (dev_priv->card_type == NV_50) {
 161                 nv50_vm_flush(dev, 5);
 162                 nv50_vm_flush(dev, 0);
 163         }
 164
 165         nvbe->bound = false;
 166         return 0;
 167 }
 168
 169 static void
 170 nouveau_sgdma_destroy(struct ttm_backend *be)
 171 {
 172         struct nouveau_sgdma_be *nvbe = (struct nouveau_sgdma_be *)be;
 173
 174         if (be) {
 175                 NV_DEBUG(nvbe->dev, "\n");
 176
 177                 if (nvbe) {
 178                         if (nvbe->pages)
 179                                 be->func->clear(be);
 180                         kfree(nvbe);
 181                 }
 182         }
 183 }
 184
 185 static struct ttm_backend_func nouveau_sgdma_backend = {
 186         .populate               = nouveau_sgdma_populate,
 187         .clear                  = nouveau_sgdma_clear,
 188         .bind                   = nouveau_sgdma_bind,
 189         .unbind                 = nouveau_sgdma_unbind,
 190         .destroy                = nouveau_sgdma_destroy
 191 };
 192
 193 struct ttm_backend *
 194 nouveau_sgdma_init_ttm(struct drm_device *dev)
 195 {
 196         struct drm_nouveau_private *dev_priv = dev->dev_private;
 197         struct nouveau_sgdma_be *nvbe;
 198
 199         if (!dev_priv->gart_info.sg_ctxdma)
 200                 return NULL;
 201
 202         nvbe = kzalloc(sizeof(*nvbe), GFP_KERNEL);
 203         if (!nvbe)
 204                 return NULL;
 205
 206         nvbe->dev = dev;
 207
 208         nvbe->backend.func      = &nouveau_sgdma_backend;
 209
 210         return &nvbe->backend;
 211 }
 212
 213 int
 214 nouveau_sgdma_init(struct drm_device *dev)
 215 {
 216         struct drm_nouveau_private *dev_priv = dev->dev_private;
 217         struct pci_dev *pdev = dev->pdev;
 218         struct nouveau_gpuobj *gpuobj = NULL;
 219         uint32_t aper_size, obj_size;
 220         int i, ret;
 221
 222         if (dev_priv->card_type < NV_50) {
 223                 aper_size = (64 * 1024 * 1024);
 224                 obj_size  = (aper_size >> NV_CTXDMA_PAGE_SHIFT) * 4;
 225                 obj_size += 8; /* ctxdma header */
 226         } else {
 227                 /* 1 entire VM page table */
 228                 aper_size = (512 * 1024 * 1024);
 229                 obj_size  = (aper_size >> NV_CTXDMA_PAGE_SHIFT) * 8;
 230         }
 231
 232         ret = nouveau_gpuobj_new(dev, NULL, obj_size, 16,
 233                                       NVOBJ_FLAG_ALLOW_NO_REFS |
 234                                       NVOBJ_FLAG_ZERO_ALLOC |
 235                                       NVOBJ_FLAG_ZERO_FREE, &gpuobj);
 236         if (ret) {
 237                 NV_ERROR(dev, "Error creating sgdma object: %d\n", ret);
 238                 return ret;
 239         }
 240
 241         dev_priv->gart_info.sg_dummy_page =
 242                 alloc_page(GFP_KERNEL|__GFP_DMA32);
 243         if (!dev_priv->gart_info.sg_dummy_page) {
 244                 nouveau_gpuobj_del(dev, &gpuobj);
 245                 return -ENOMEM;
 246         }
 247
 248         set_bit(PG_locked, &dev_priv->gart_info.sg_dummy_page->flags);
 249         dev_priv->gart_info.sg_dummy_bus =
 250                 pci_map_page(pdev, dev_priv->gart_info.sg_dummy_page, 0,
 251                              PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 252         if (pci_dma_mapping_error(pdev, dev_priv->gart_info.sg_dummy_bus)) {
 253                 nouveau_gpuobj_del(dev, &gpuobj);
 254                 return -EFAULT;
 255         }
 256
 257         if (dev_priv->card_type < NV_50) {
 258                 /* Maybe use NV_DMA_TARGET_AGP for PCIE? NVIDIA do this, and
 259                  * confirmed to work on c51.  Perhaps means NV_DMA_TARGET_PCIE
 260                  * on those cards? */
 261                 nv_wo32(dev, gpuobj, 0, NV_CLASS_DMA_IN_MEMORY |
 262                                        (1 << 12) /* PT present */ |
 263                                        (0 << 13) /* PT *not* linear */ |
 264                                        (NV_DMA_ACCESS_RW  << 14) |
 265                                        (NV_DMA_TARGET_PCI << 16));
 266                 nv_wo32(dev, gpuobj, 1, aper_size - 1);
 267                 for (i = 2; i < 2 + (aper_size >> 12); i++) {
 268                         nv_wo32(dev, gpuobj, i,
 269                                     dev_priv->gart_info.sg_dummy_bus | 3);
 270                 }
 271         } else {
 272                 for (i = 0; i < obj_size; i += 8) {
 273                         nv_wo32(dev, gpuobj, (i+0)/4,
 274                                     dev_priv->gart_info.sg_dummy_bus | 0x21);
 275                         nv_wo32(dev, gpuobj, (i+4)/4, 0);
 276                 }
 277         }
 278         dev_priv->engine.instmem.flush(dev);
 279
 280         dev_priv->gart_info.type      = NOUVEAU_GART_SGDMA;
 281         dev_priv->gart_info.aper_base = 0;
 282         dev_priv->gart_info.aper_size = aper_size;
 283         dev_priv->gart_info.sg_ctxdma = gpuobj;
 284         return 0;
 285 }
 286
 287 void
 288 nouveau_sgdma_takedown(struct drm_device *dev)
 289 {
 290         struct drm_nouveau_private *dev_priv = dev->dev_private;
 291
 292         if (dev_priv->gart_info.sg_dummy_page) {
 293                 pci_unmap_page(dev->pdev, dev_priv->gart_info.sg_dummy_bus,
 294                                NV_CTXDMA_PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 295                 unlock_page(dev_priv->gart_info.sg_dummy_page);
 296                 __free_page(dev_priv->gart_info.sg_dummy_page);
 297                 dev_priv->gart_info.sg_dummy_page = NULL;
 298                 dev_priv->gart_info.sg_dummy_bus = 0;
 299         }
 300
 301         nouveau_gpuobj_del(dev, &dev_priv->gart_info.sg_ctxdma);
 302 }
 303
 304 int
 305 nouveau_sgdma_get_page(struct drm_device *dev, uint32_t offset, uint32_t *page)
 306 {
 307         struct drm_nouveau_private *dev_priv = dev->dev_private;
 308         struct nouveau_gpuobj *gpuobj = dev_priv->gart_info.sg_ctxdma;
 309         int pte;
 310
 311         pte = (offset >> NV_CTXDMA_PAGE_SHIFT);
 312         if (dev_priv->card_type < NV_50) {
 313                 *page = nv_ro32(dev, gpuobj, (pte + 2)) & ~NV_CTXDMA_PAGE_MASK;
 314                 return 0;
 315         }
 316
 317         NV_ERROR(dev, "Unimplemented on NV50\n");
 318         return -EINVAL;
 319 }