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