drivers/gpu/drm/nouveau/nouveau_vm.c

   1 /*
   2  * Copyright 2010 Red Hat Inc.
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice shall be included in
  12  * all copies or substantial portions of the Software.
  13  *
  14  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  15  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  16  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  17  * THE COPYRIGHT HOLDER(S) OR AUTHOR(S) BE LIABLE FOR ANY CLAIM, DAMAGES OR
  18  * OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE,
  19  * ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR
  20  * OTHER DEALINGS IN THE SOFTWARE.
  21  *
  22  * Authors: Ben Skeggs
  23  */
  24
  25 #include "drmP.h"
  26 #include "nouveau_drv.h"
  27 #include "nouveau_mm.h"
  28 #include "nouveau_vm.h"
  29
  30 void
  31 nouveau_vm_map_at(struct nouveau_vma *vma, u64 delta, struct nouveau_mem *node)
  32 {
  33         struct nouveau_vm *vm = vma->vm;
  34         struct nouveau_mm_node *r;
  35         int big = vma->node->type != vm->spg_shift;
  36         u32 offset = vma->node->offset + (delta >> 12);
  37         u32 bits = vma->node->type - 12;
  38         u32 pde  = (offset >> vm->pgt_bits) - vm->fpde;
  39         u32 pte  = (offset & ((1 << vm->pgt_bits) - 1)) >> bits;
  40         u32 max  = 1 << (vm->pgt_bits - bits);
  41         u32 end, len;
  42
  43         delta = 0;
  44         list_for_each_entry(r, &node->regions, rl_entry) {
  45                 u64 phys = (u64)r->offset << 12;
  46                 u32 num  = r->length >> bits;
  47
  48                 while (num) {
  49                         struct nouveau_gpuobj *pgt = vm->pgt[pde].obj[big];
  50
  51                         end = (pte + num);
  52                         if (unlikely(end >= max))
  53                                 end = max;
  54                         len = end - pte;
  55
  56                         vm->map(vma, pgt, node, pte, len, phys, delta);
  57
  58                         num -= len;
  59                         pte += len;
  60                         if (unlikely(end >= max)) {
  61                                 pde++;
  62                                 pte = 0;
  63                         }
  64
  65                         delta += (u64)len << vma->node->type;
  66                 }
  67         }
  68
  69         vm->flush(vm);
  70 }
  71
  72 void
  73 nouveau_vm_map(struct nouveau_vma *vma, struct nouveau_mem *node)
  74 {
  75         nouveau_vm_map_at(vma, 0, node);
  76 }
  77
  78 void
  79 nouveau_vm_map_sg(struct nouveau_vma *vma, u64 delta, u64 length,
  80                   struct nouveau_mem *mem, dma_addr_t *list)
  81 {
  82         struct nouveau_vm *vm = vma->vm;
  83         int big = vma->node->type != vm->spg_shift;
  84         u32 offset = vma->node->offset + (delta >> 12);
  85         u32 bits = vma->node->type - 12;
  86         u32 num  = length >> vma->node->type;
  87         u32 pde  = (offset >> vm->pgt_bits) - vm->fpde;
  88         u32 pte  = (offset & ((1 << vm->pgt_bits) - 1)) >> bits;
  89         u32 max  = 1 << (vm->pgt_bits - bits);
  90         u32 end, len;
  91
  92         while (num) {
  93                 struct nouveau_gpuobj *pgt = vm->pgt[pde].obj[big];
  94
  95                 end = (pte + num);
  96                 if (unlikely(end >= max))
  97                         end = max;
  98                 len = end - pte;
  99
 100                 vm->map_sg(vma, pgt, mem, pte, len, list);
 101
 102                 num  -= len;
 103                 pte  += len;
 104                 list += len;
 105                 if (unlikely(end >= max)) {
 106                         pde++;
 107                         pte = 0;
 108                 }
 109         }
 110
 111         vm->flush(vm);
 112 }
 113
 114 void
 115 nouveau_vm_unmap_at(struct nouveau_vma *vma, u64 delta, u64 length)
 116 {
 117         struct nouveau_vm *vm = vma->vm;
 118         int big = vma->node->type != vm->spg_shift;
 119         u32 offset = vma->node->offset + (delta >> 12);
 120         u32 bits = vma->node->type - 12;
 121         u32 num  = length >> vma->node->type;
 122         u32 pde  = (offset >> vm->pgt_bits) - vm->fpde;
 123         u32 pte  = (offset & ((1 << vm->pgt_bits) - 1)) >> bits;
 124         u32 max  = 1 << (vm->pgt_bits - bits);
 125         u32 end, len;
 126
 127         while (num) {
 128                 struct nouveau_gpuobj *pgt = vm->pgt[pde].obj[big];
 129
 130                 end = (pte + num);
 131                 if (unlikely(end >= max))
 132                         end = max;
 133                 len = end - pte;
 134
 135                 vm->unmap(pgt, pte, len);
 136
 137                 num -= len;
 138                 pte += len;
 139                 if (unlikely(end >= max)) {
 140                         pde++;
 141                         pte = 0;
 142                 }
 143         }
 144
 145         vm->flush(vm);
 146 }
 147
 148 void
 149 nouveau_vm_unmap(struct nouveau_vma *vma)
 150 {
 151         nouveau_vm_unmap_at(vma, 0, (u64)vma->node->length << 12);
 152 }
 153
 154 static void
 155 nouveau_vm_unmap_pgt(struct nouveau_vm *vm, int big, u32 fpde, u32 lpde)
 156 {
 157         struct nouveau_vm_pgd *vpgd;
 158         struct nouveau_vm_pgt *vpgt;
 159         struct nouveau_gpuobj *pgt;
 160         u32 pde;
 161
 162         for (pde = fpde; pde <= lpde; pde++) {
 163                 vpgt = &vm->pgt[pde - vm->fpde];
 164                 if (--vpgt->refcount[big])
 165                         continue;
 166
 167                 pgt = vpgt->obj[big];
 168                 vpgt->obj[big] = NULL;
 169
 170                 list_for_each_entry(vpgd, &vm->pgd_list, head) {
 171                         vm->map_pgt(vpgd->obj, pde, vpgt->obj);
 172                 }
 173
 174                 mutex_unlock(&vm->mm->mutex);
 175                 nouveau_gpuobj_ref(NULL, &pgt);
 176                 mutex_lock(&vm->mm->mutex);
 177         }
 178 }
 179
 180 static int
 181 nouveau_vm_map_pgt(struct nouveau_vm *vm, u32 pde, u32 type)
 182 {
 183         struct nouveau_vm_pgt *vpgt = &vm->pgt[pde - vm->fpde];
 184         struct nouveau_vm_pgd *vpgd;
 185         struct nouveau_gpuobj *pgt;
 186         int big = (type != vm->spg_shift);
 187         u32 pgt_size;
 188         int ret;
 189
 190         pgt_size  = (1 << (vm->pgt_bits + 12)) >> type;
 191         pgt_size *= 8;
 192
 193         mutex_unlock(&vm->mm->mutex);
 194         ret = nouveau_gpuobj_new(vm->dev, NULL, pgt_size, 0x1000,
 195                                  NVOBJ_FLAG_ZERO_ALLOC, &pgt);
 196         mutex_lock(&vm->mm->mutex);
 197         if (unlikely(ret))
 198                 return ret;
 199
 200         /* someone beat us to filling the PDE while we didn't have the lock */
 201         if (unlikely(vpgt->refcount[big]++)) {
 202                 mutex_unlock(&vm->mm->mutex);
 203                 nouveau_gpuobj_ref(NULL, &pgt);
 204                 mutex_lock(&vm->mm->mutex);
 205                 return 0;
 206         }
 207
 208         vpgt->obj[big] = pgt;
 209         list_for_each_entry(vpgd, &vm->pgd_list, head) {
 210                 vm->map_pgt(vpgd->obj, pde, vpgt->obj);
 211         }
 212
 213         return 0;
 214 }
 215
 216 int
 217 nouveau_vm_get(struct nouveau_vm *vm, u64 size, u32 page_shift,
 218                u32 access, struct nouveau_vma *vma)
 219 {
 220         u32 align = (1 << page_shift) >> 12;
 221         u32 msize = size >> 12;
 222         u32 fpde, lpde, pde;
 223         int ret;
 224
 225         mutex_lock(&vm->mm->mutex);
 226         ret = nouveau_mm_get(vm->mm, page_shift, msize, 0, align, &vma->node);
 227         if (unlikely(ret != 0)) {
 228                 mutex_unlock(&vm->mm->mutex);
 229                 return ret;
 230         }
 231
 232         fpde = (vma->node->offset >> vm->pgt_bits);
 233         lpde = (vma->node->offset + vma->node->length - 1) >> vm->pgt_bits;
 234         for (pde = fpde; pde <= lpde; pde++) {
 235                 struct nouveau_vm_pgt *vpgt = &vm->pgt[pde - vm->fpde];
 236                 int big = (vma->node->type != vm->spg_shift);
 237
 238                 if (likely(vpgt->refcount[big])) {
 239                         vpgt->refcount[big]++;
 240                         continue;
 241                 }
 242
 243                 ret = nouveau_vm_map_pgt(vm, pde, vma->node->type);
 244                 if (ret) {
 245                         if (pde != fpde)
 246                                 nouveau_vm_unmap_pgt(vm, big, fpde, pde - 1);
 247                         nouveau_mm_put(vm->mm, vma->node);
 248                         mutex_unlock(&vm->mm->mutex);
 249                         vma->node = NULL;
 250                         return ret;
 251                 }
 252         }
 253         mutex_unlock(&vm->mm->mutex);
 254
 255         vma->vm     = vm;
 256         vma->offset = (u64)vma->node->offset << 12;
 257         vma->access = access;
 258         return 0;
 259 }
 260
 261 void
 262 nouveau_vm_put(struct nouveau_vma *vma)
 263 {
 264         struct nouveau_vm *vm = vma->vm;
 265         u32 fpde, lpde;
 266
 267         if (unlikely(vma->node == NULL))
 268                 return;
 269         fpde = (vma->node->offset >> vm->pgt_bits);
 270         lpde = (vma->node->offset + vma->node->length - 1) >> vm->pgt_bits;
 271
 272         mutex_lock(&vm->mm->mutex);
 273         nouveau_vm_unmap_pgt(vm, vma->node->type != vm->spg_shift, fpde, lpde);
 274         nouveau_mm_put(vm->mm, vma->node);
 275         vma->node = NULL;
 276         mutex_unlock(&vm->mm->mutex);
 277 }
 278
 279 int
 280 nouveau_vm_new(struct drm_device *dev, u64 offset, u64 length, u64 mm_offset,
 281                struct nouveau_vm **pvm)
 282 {
 283         struct drm_nouveau_private *dev_priv = dev->dev_private;
 284         struct nouveau_vm *vm;
 285         u64 mm_length = (offset + length) - mm_offset;
 286         u32 block, pgt_bits;
 287         int ret;
 288
 289         vm = kzalloc(sizeof(*vm), GFP_KERNEL);
 290         if (!vm)
 291                 return -ENOMEM;
 292
 293         if (dev_priv->card_type == NV_50) {
 294                 vm->map_pgt = nv50_vm_map_pgt;
 295                 vm->map = nv50_vm_map;
 296                 vm->map_sg = nv50_vm_map_sg;
 297                 vm->unmap = nv50_vm_unmap;
 298                 vm->flush = nv50_vm_flush;
 299                 vm->spg_shift = 12;
 300                 vm->lpg_shift = 16;
 301
 302                 pgt_bits = 29;
 303                 block = (1 << pgt_bits);
 304                 if (length < block)
 305                         block = length;
 306
 307         } else
 308         if (dev_priv->card_type == NV_C0) {
 309                 vm->map_pgt = nvc0_vm_map_pgt;
 310                 vm->map = nvc0_vm_map;
 311                 vm->map_sg = nvc0_vm_map_sg;
 312                 vm->unmap = nvc0_vm_unmap;
 313                 vm->flush = nvc0_vm_flush;
 314                 vm->spg_shift = 12;
 315                 vm->lpg_shift = 17;
 316                 pgt_bits = 27;
 317                 block = 4096;
 318         } else {
 319                 kfree(vm);
 320                 return -ENOSYS;
 321         }
 322
 323         vm->fpde   = offset >> pgt_bits;
 324         vm->lpde   = (offset + length - 1) >> pgt_bits;
 325         vm->pgt = kcalloc(vm->lpde - vm->fpde + 1, sizeof(*vm->pgt), GFP_KERNEL);
 326         if (!vm->pgt) {
 327                 kfree(vm);
 328                 return -ENOMEM;
 329         }
 330
 331         INIT_LIST_HEAD(&vm->pgd_list);
 332         vm->dev = dev;
 333         vm->refcount = 1;
 334         vm->pgt_bits = pgt_bits - 12;
 335
 336         ret = nouveau_mm_init(&vm->mm, mm_offset >> 12, mm_length >> 12,
 337                               block >> 12);
 338         if (ret) {
 339                 kfree(vm);
 340                 return ret;
 341         }
 342
 343         *pvm = vm;
 344         return 0;
 345 }
 346
 347 static int
 348 nouveau_vm_link(struct nouveau_vm *vm, struct nouveau_gpuobj *pgd)
 349 {
 350         struct nouveau_vm_pgd *vpgd;
 351         int i;
 352
 353         if (!pgd)
 354                 return 0;
 355
 356         vpgd = kzalloc(sizeof(*vpgd), GFP_KERNEL);
 357         if (!vpgd)
 358                 return -ENOMEM;
 359
 360         nouveau_gpuobj_ref(pgd, &vpgd->obj);
 361
 362         mutex_lock(&vm->mm->mutex);
 363         for (i = vm->fpde; i <= vm->lpde; i++)
 364                 vm->map_pgt(pgd, i, vm->pgt[i - vm->fpde].obj);
 365         list_add(&vpgd->head, &vm->pgd_list);
 366         mutex_unlock(&vm->mm->mutex);
 367         return 0;
 368 }
 369
 370 static void
 371 nouveau_vm_unlink(struct nouveau_vm *vm, struct nouveau_gpuobj *pgd)
 372 {
 373         struct nouveau_vm_pgd *vpgd, *tmp;
 374
 375         if (!pgd)
 376                 return;
 377
 378         mutex_lock(&vm->mm->mutex);
 379         list_for_each_entry_safe(vpgd, tmp, &vm->pgd_list, head) {
 380                 if (vpgd->obj != pgd)
 381                         continue;
 382
 383                 list_del(&vpgd->head);
 384                 nouveau_gpuobj_ref(NULL, &vpgd->obj);
 385                 kfree(vpgd);
 386         }
 387         mutex_unlock(&vm->mm->mutex);
 388 }
 389
 390 static void
 391 nouveau_vm_del(struct nouveau_vm *vm)
 392 {
 393         struct nouveau_vm_pgd *vpgd, *tmp;
 394
 395         list_for_each_entry_safe(vpgd, tmp, &vm->pgd_list, head) {
 396                 nouveau_vm_unlink(vm, vpgd->obj);
 397         }
 398         WARN_ON(nouveau_mm_fini(&vm->mm) != 0);
 399
 400         kfree(vm->pgt);
 401         kfree(vm);
 402 }
 403
 404 int
 405 nouveau_vm_ref(struct nouveau_vm *ref, struct nouveau_vm **ptr,
 406                struct nouveau_gpuobj *pgd)
 407 {
 408         struct nouveau_vm *vm;
 409         int ret;
 410
 411         vm = ref;
 412         if (vm) {
 413                 ret = nouveau_vm_link(vm, pgd);
 414                 if (ret)
 415                         return ret;
 416
 417                 vm->refcount++;
 418         }
 419
 420         vm = *ptr;
 421         *ptr = ref;
 422
 423         if (vm) {
 424                 nouveau_vm_unlink(vm, pgd);
 425
 426                 if (--vm->refcount == 0)
 427                         nouveau_vm_del(vm);
 428         }
 429
 430         return 0;
 431 }