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                                 phys += len << (bits + 12);
  62                                 pde++;
  63                                 pte = 0;
  64                         }
  65
  66                         delta += (u64)len << vma->node->type;
  67                 }
  68         }
  69
  70         vm->flush(vm);
  71 }
  72
  73 void
  74 nouveau_vm_map(struct nouveau_vma *vma, struct nouveau_mem *node)
  75 {
  76         nouveau_vm_map_at(vma, 0, node);
  77 }
  78
  79 void
  80 nouveau_vm_map_sg_table(struct nouveau_vma *vma, u64 delta, u64 length,
  81                         struct nouveau_mem *mem)
  82 {
  83         struct nouveau_vm *vm = vma->vm;
  84         int big = vma->node->type != vm->spg_shift;
  85         u32 offset = vma->node->offset + (delta >> 12);
  86         u32 bits = vma->node->type - 12;
  87         u32 num  = length >> vma->node->type;
  88         u32 pde  = (offset >> vm->pgt_bits) - vm->fpde;
  89         u32 pte  = (offset & ((1 << vm->pgt_bits) - 1)) >> bits;
  90         u32 max  = 1 << (vm->pgt_bits - bits);
  91         unsigned m, sglen;
  92         u32 end, len;
  93         int i;
  94         struct scatterlist *sg;
  95
  96         for_each_sg(mem->sg->sgl, sg, mem->sg->nents, i) {
  97                 struct nouveau_gpuobj *pgt = vm->pgt[pde].obj[big];
  98                 sglen = sg_dma_len(sg) >> PAGE_SHIFT;
  99
 100                 end = pte + sglen;
 101                 if (unlikely(end >= max))
 102                         end = max;
 103                 len = end - pte;
 104
 105                 for (m = 0; m < len; m++) {
 106                         dma_addr_t addr = sg_dma_address(sg) + (m << PAGE_SHIFT);
 107
 108                         vm->map_sg(vma, pgt, mem, pte, 1, &addr);
 109                         num--;
 110                         pte++;
 111
 112                         if (num == 0)
 113                                 goto finish;
 114                 }
 115                 if (unlikely(end >= max)) {
 116                         pde++;
 117                         pte = 0;
 118                 }
 119                 if (m < sglen) {
 120                         for (; m < sglen; m++) {
 121                                 dma_addr_t addr = sg_dma_address(sg) + (m << PAGE_SHIFT);
 122
 123                                 vm->map_sg(vma, pgt, mem, pte, 1, &addr);
 124                                 num--;
 125                                 pte++;
 126                                 if (num == 0)
 127                                         goto finish;
 128                         }
 129                 }
 130
 131         }
 132 finish:
 133         vm->flush(vm);
 134 }
 135
 136 void
 137 nouveau_vm_map_sg(struct nouveau_vma *vma, u64 delta, u64 length,
 138                   struct nouveau_mem *mem)
 139 {
 140         struct nouveau_vm *vm = vma->vm;
 141         dma_addr_t *list = mem->pages;
 142         int big = vma->node->type != vm->spg_shift;
 143         u32 offset = vma->node->offset + (delta >> 12);
 144         u32 bits = vma->node->type - 12;
 145         u32 num  = length >> vma->node->type;
 146         u32 pde  = (offset >> vm->pgt_bits) - vm->fpde;
 147         u32 pte  = (offset & ((1 << vm->pgt_bits) - 1)) >> bits;
 148         u32 max  = 1 << (vm->pgt_bits - bits);
 149         u32 end, len;
 150
 151         while (num) {
 152                 struct nouveau_gpuobj *pgt = vm->pgt[pde].obj[big];
 153
 154                 end = (pte + num);
 155                 if (unlikely(end >= max))
 156                         end = max;
 157                 len = end - pte;
 158
 159                 vm->map_sg(vma, pgt, mem, pte, len, list);
 160
 161                 num  -= len;
 162                 pte  += len;
 163                 list += len;
 164                 if (unlikely(end >= max)) {
 165                         pde++;
 166                         pte = 0;
 167                 }
 168         }
 169
 170         vm->flush(vm);
 171 }
 172
 173 void
 174 nouveau_vm_unmap_at(struct nouveau_vma *vma, u64 delta, u64 length)
 175 {
 176         struct nouveau_vm *vm = vma->vm;
 177         int big = vma->node->type != vm->spg_shift;
 178         u32 offset = vma->node->offset + (delta >> 12);
 179         u32 bits = vma->node->type - 12;
 180         u32 num  = length >> vma->node->type;
 181         u32 pde  = (offset >> vm->pgt_bits) - vm->fpde;
 182         u32 pte  = (offset & ((1 << vm->pgt_bits) - 1)) >> bits;
 183         u32 max  = 1 << (vm->pgt_bits - bits);
 184         u32 end, len;
 185
 186         while (num) {
 187                 struct nouveau_gpuobj *pgt = vm->pgt[pde].obj[big];
 188
 189                 end = (pte + num);
 190                 if (unlikely(end >= max))
 191                         end = max;
 192                 len = end - pte;
 193
 194                 vm->unmap(pgt, pte, len);
 195
 196                 num -= len;
 197                 pte += len;
 198                 if (unlikely(end >= max)) {
 199                         pde++;
 200                         pte = 0;
 201                 }
 202         }
 203
 204         vm->flush(vm);
 205 }
 206
 207 void
 208 nouveau_vm_unmap(struct nouveau_vma *vma)
 209 {
 210         nouveau_vm_unmap_at(vma, 0, (u64)vma->node->length << 12);
 211 }
 212
 213 static void
 214 nouveau_vm_unmap_pgt(struct nouveau_vm *vm, int big, u32 fpde, u32 lpde)
 215 {
 216         struct nouveau_vm_pgd *vpgd;
 217         struct nouveau_vm_pgt *vpgt;
 218         struct nouveau_gpuobj *pgt;
 219         u32 pde;
 220
 221         for (pde = fpde; pde <= lpde; pde++) {
 222                 vpgt = &vm->pgt[pde - vm->fpde];
 223                 if (--vpgt->refcount[big])
 224                         continue;
 225
 226                 pgt = vpgt->obj[big];
 227                 vpgt->obj[big] = NULL;
 228
 229                 list_for_each_entry(vpgd, &vm->pgd_list, head) {
 230                         vm->map_pgt(vpgd->obj, pde, vpgt->obj);
 231                 }
 232
 233                 mutex_unlock(&vm->mm.mutex);
 234                 nouveau_gpuobj_ref(NULL, &pgt);
 235                 mutex_lock(&vm->mm.mutex);
 236         }
 237 }
 238
 239 static int
 240 nouveau_vm_map_pgt(struct nouveau_vm *vm, u32 pde, u32 type)
 241 {
 242         struct nouveau_vm_pgt *vpgt = &vm->pgt[pde - vm->fpde];
 243         struct nouveau_vm_pgd *vpgd;
 244         struct nouveau_gpuobj *pgt;
 245         int big = (type != vm->spg_shift);
 246         u32 pgt_size;
 247         int ret;
 248
 249         pgt_size  = (1 << (vm->pgt_bits + 12)) >> type;
 250         pgt_size *= 8;
 251
 252         mutex_unlock(&vm->mm.mutex);
 253         ret = nouveau_gpuobj_new(vm->dev, NULL, pgt_size, 0x1000,
 254                                  NVOBJ_FLAG_ZERO_ALLOC, &pgt);
 255         mutex_lock(&vm->mm.mutex);
 256         if (unlikely(ret))
 257                 return ret;
 258
 259         /* someone beat us to filling the PDE while we didn't have the lock */
 260         if (unlikely(vpgt->refcount[big]++)) {
 261                 mutex_unlock(&vm->mm.mutex);
 262                 nouveau_gpuobj_ref(NULL, &pgt);
 263                 mutex_lock(&vm->mm.mutex);
 264                 return 0;
 265         }
 266
 267         vpgt->obj[big] = pgt;
 268         list_for_each_entry(vpgd, &vm->pgd_list, head) {
 269                 vm->map_pgt(vpgd->obj, pde, vpgt->obj);
 270         }
 271
 272         return 0;
 273 }
 274
 275 int
 276 nouveau_vm_get(struct nouveau_vm *vm, u64 size, u32 page_shift,
 277                u32 access, struct nouveau_vma *vma)
 278 {
 279         u32 align = (1 << page_shift) >> 12;
 280         u32 msize = size >> 12;
 281         u32 fpde, lpde, pde;
 282         int ret;
 283
 284         mutex_lock(&vm->mm.mutex);
 285         ret = nouveau_mm_get(&vm->mm, page_shift, msize, 0, align, &vma->node);
 286         if (unlikely(ret != 0)) {
 287                 mutex_unlock(&vm->mm.mutex);
 288                 return ret;
 289         }
 290
 291         fpde = (vma->node->offset >> vm->pgt_bits);
 292         lpde = (vma->node->offset + vma->node->length - 1) >> vm->pgt_bits;
 293         for (pde = fpde; pde <= lpde; pde++) {
 294                 struct nouveau_vm_pgt *vpgt = &vm->pgt[pde - vm->fpde];
 295                 int big = (vma->node->type != vm->spg_shift);
 296
 297                 if (likely(vpgt->refcount[big])) {
 298                         vpgt->refcount[big]++;
 299                         continue;
 300                 }
 301
 302                 ret = nouveau_vm_map_pgt(vm, pde, vma->node->type);
 303                 if (ret) {
 304                         if (pde != fpde)
 305                                 nouveau_vm_unmap_pgt(vm, big, fpde, pde - 1);
 306                         nouveau_mm_put(&vm->mm, vma->node);
 307                         mutex_unlock(&vm->mm.mutex);
 308                         vma->node = NULL;
 309                         return ret;
 310                 }
 311         }
 312         mutex_unlock(&vm->mm.mutex);
 313
 314         vma->vm     = vm;
 315         vma->offset = (u64)vma->node->offset << 12;
 316         vma->access = access;
 317         return 0;
 318 }
 319
 320 void
 321 nouveau_vm_put(struct nouveau_vma *vma)
 322 {
 323         struct nouveau_vm *vm = vma->vm;
 324         u32 fpde, lpde;
 325
 326         if (unlikely(vma->node == NULL))
 327                 return;
 328         fpde = (vma->node->offset >> vm->pgt_bits);
 329         lpde = (vma->node->offset + vma->node->length - 1) >> vm->pgt_bits;
 330
 331         mutex_lock(&vm->mm.mutex);
 332         nouveau_vm_unmap_pgt(vm, vma->node->type != vm->spg_shift, fpde, lpde);
 333         nouveau_mm_put(&vm->mm, vma->node);
 334         vma->node = NULL;
 335         mutex_unlock(&vm->mm.mutex);
 336 }
 337
 338 int
 339 nouveau_vm_new(struct drm_device *dev, u64 offset, u64 length, u64 mm_offset,
 340                struct nouveau_vm **pvm)
 341 {
 342         struct drm_nouveau_private *dev_priv = dev->dev_private;
 343         struct nouveau_vm *vm;
 344         u64 mm_length = (offset + length) - mm_offset;
 345         u32 block, pgt_bits;
 346         int ret;
 347
 348         vm = kzalloc(sizeof(*vm), GFP_KERNEL);
 349         if (!vm)
 350                 return -ENOMEM;
 351
 352         if (dev_priv->card_type == NV_50) {
 353                 vm->map_pgt = nv50_vm_map_pgt;
 354                 vm->map = nv50_vm_map;
 355                 vm->map_sg = nv50_vm_map_sg;
 356                 vm->unmap = nv50_vm_unmap;
 357                 vm->flush = nv50_vm_flush;
 358                 vm->spg_shift = 12;
 359                 vm->lpg_shift = 16;
 360
 361                 pgt_bits = 29;
 362                 block = (1 << pgt_bits);
 363                 if (length < block)
 364                         block = length;
 365
 366         } else
 367         if (dev_priv->card_type >= NV_C0) {
 368                 vm->map_pgt = nvc0_vm_map_pgt;
 369                 vm->map = nvc0_vm_map;
 370                 vm->map_sg = nvc0_vm_map_sg;
 371                 vm->unmap = nvc0_vm_unmap;
 372                 vm->flush = nvc0_vm_flush;
 373                 vm->spg_shift = 12;
 374                 vm->lpg_shift = 17;
 375                 pgt_bits = 27;
 376                 block = 4096;
 377         } else {
 378                 kfree(vm);
 379                 return -ENOSYS;
 380         }
 381
 382         vm->fpde   = offset >> pgt_bits;
 383         vm->lpde   = (offset + length - 1) >> pgt_bits;
 384         vm->pgt = kcalloc(vm->lpde - vm->fpde + 1, sizeof(*vm->pgt), GFP_KERNEL);
 385         if (!vm->pgt) {
 386                 kfree(vm);
 387                 return -ENOMEM;
 388         }
 389
 390         INIT_LIST_HEAD(&vm->pgd_list);
 391         vm->dev = dev;
 392         vm->refcount = 1;
 393         vm->pgt_bits = pgt_bits - 12;
 394
 395         ret = nouveau_mm_init(&vm->mm, mm_offset >> 12, mm_length >> 12,
 396                               block >> 12);
 397         if (ret) {
 398                 kfree(vm);
 399                 return ret;
 400         }
 401
 402         *pvm = vm;
 403         return 0;
 404 }
 405
 406 static int
 407 nouveau_vm_link(struct nouveau_vm *vm, struct nouveau_gpuobj *pgd)
 408 {
 409         struct nouveau_vm_pgd *vpgd;
 410         int i;
 411
 412         if (!pgd)
 413                 return 0;
 414
 415         vpgd = kzalloc(sizeof(*vpgd), GFP_KERNEL);
 416         if (!vpgd)
 417                 return -ENOMEM;
 418
 419         nouveau_gpuobj_ref(pgd, &vpgd->obj);
 420
 421         mutex_lock(&vm->mm.mutex);
 422         for (i = vm->fpde; i <= vm->lpde; i++)
 423                 vm->map_pgt(pgd, i, vm->pgt[i - vm->fpde].obj);
 424         list_add(&vpgd->head, &vm->pgd_list);
 425         mutex_unlock(&vm->mm.mutex);
 426         return 0;
 427 }
 428
 429 static void
 430 nouveau_vm_unlink(struct nouveau_vm *vm, struct nouveau_gpuobj *mpgd)
 431 {
 432         struct nouveau_vm_pgd *vpgd, *tmp;
 433         struct nouveau_gpuobj *pgd = NULL;
 434
 435         if (!mpgd)
 436                 return;
 437
 438         mutex_lock(&vm->mm.mutex);
 439         list_for_each_entry_safe(vpgd, tmp, &vm->pgd_list, head) {
 440                 if (vpgd->obj == mpgd) {
 441                         pgd = vpgd->obj;
 442                         list_del(&vpgd->head);
 443                         kfree(vpgd);
 444                         break;
 445                 }
 446         }
 447         mutex_unlock(&vm->mm.mutex);
 448
 449         nouveau_gpuobj_ref(NULL, &pgd);
 450 }
 451
 452 static void
 453 nouveau_vm_del(struct nouveau_vm *vm)
 454 {
 455         struct nouveau_vm_pgd *vpgd, *tmp;
 456
 457         list_for_each_entry_safe(vpgd, tmp, &vm->pgd_list, head) {
 458                 nouveau_vm_unlink(vm, vpgd->obj);
 459         }
 460
 461         nouveau_mm_fini(&vm->mm);
 462         kfree(vm->pgt);
 463         kfree(vm);
 464 }
 465
 466 int
 467 nouveau_vm_ref(struct nouveau_vm *ref, struct nouveau_vm **ptr,
 468                struct nouveau_gpuobj *pgd)
 469 {
 470         struct nouveau_vm *vm;
 471         int ret;
 472
 473         vm = ref;
 474         if (vm) {
 475                 ret = nouveau_vm_link(vm, pgd);
 476                 if (ret)
 477                         return ret;
 478
 479                 vm->refcount++;
 480         }
 481
 482         vm = *ptr;
 483         *ptr = ref;
 484
 485         if (vm) {
 486                 nouveau_vm_unlink(vm, pgd);
 487
 488                 if (--vm->refcount == 0)
 489                         nouveau_vm_del(vm);
 490         }
 491
 492         return 0;
 493 }