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_vram *vram)
  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         list_for_each_entry(r, &vram->regions, rl_entry) {
  44                 u64 phys = (u64)r->offset << 12;
  45                 u32 num  = r->length >> bits;
  46
  47                 while (num) {
  48                         struct nouveau_gpuobj *pgt = vm->pgt[pde].obj[big];
  49
  50                         end = (pte + num);
  51                         if (unlikely(end >= max))
  52                                 end = max;
  53                         len = end - pte;
  54
  55                         vm->map(vma, pgt, vram, pte, len, phys);
  56
  57                         num -= len;
  58                         pte += len;
  59                         if (unlikely(end >= max)) {
  60                                 pde++;
  61                                 pte = 0;
  62                         }
  63                 }
  64         }
  65
  66         vm->flush(vm);
  67 }
  68
  69 void
  70 nouveau_vm_map(struct nouveau_vma *vma, struct nouveau_vram *vram)
  71 {
  72         nouveau_vm_map_at(vma, 0, vram);
  73 }
  74
  75 void
  76 nouveau_vm_map_sg(struct nouveau_vma *vma, u64 delta, u64 length,
  77                   dma_addr_t *list)
  78 {
  79         struct nouveau_vm *vm = vma->vm;
  80         int big = vma->node->type != vm->spg_shift;
  81         u32 offset = vma->node->offset + (delta >> 12);
  82         u32 bits = vma->node->type - 12;
  83         u32 num  = length >> vma->node->type;
  84         u32 pde  = (offset >> vm->pgt_bits) - vm->fpde;
  85         u32 pte  = (offset & ((1 << vm->pgt_bits) - 1)) >> bits;
  86         u32 max  = 1 << (vm->pgt_bits - bits);
  87         u32 end, len;
  88
  89         while (num) {
  90                 struct nouveau_gpuobj *pgt = vm->pgt[pde].obj[big];
  91
  92                 end = (pte + num);
  93                 if (unlikely(end >= max))
  94                         end = max;
  95                 len = end - pte;
  96
  97                 vm->map_sg(vma, pgt, pte, list, len);
  98
  99                 num  -= len;
 100                 pte  += len;
 101                 list += len;
 102                 if (unlikely(end >= max)) {
 103                         pde++;
 104                         pte = 0;
 105                 }
 106         }
 107
 108         vm->flush(vm);
 109 }
 110
 111 void
 112 nouveau_vm_unmap_at(struct nouveau_vma *vma, u64 delta, u64 length)
 113 {
 114         struct nouveau_vm *vm = vma->vm;
 115         int big = vma->node->type != vm->spg_shift;
 116         u32 offset = vma->node->offset + (delta >> 12);
 117         u32 bits = vma->node->type - 12;
 118         u32 num  = length >> vma->node->type;
 119         u32 pde  = (offset >> vm->pgt_bits) - vm->fpde;
 120         u32 pte  = (offset & ((1 << vm->pgt_bits) - 1)) >> bits;
 121         u32 max  = 1 << (vm->pgt_bits - bits);
 122         u32 end, len;
 123
 124         while (num) {
 125                 struct nouveau_gpuobj *pgt = vm->pgt[pde].obj[big];
 126
 127                 end = (pte + num);
 128                 if (unlikely(end >= max))
 129                         end = max;
 130                 len = end - pte;
 131
 132                 vm->unmap(pgt, pte, len);
 133
 134                 num -= len;
 135                 pte += len;
 136                 if (unlikely(end >= max)) {
 137                         pde++;
 138                         pte = 0;
 139                 }
 140         }
 141
 142         vm->flush(vm);
 143 }
 144
 145 void
 146 nouveau_vm_unmap(struct nouveau_vma *vma)
 147 {
 148         nouveau_vm_unmap_at(vma, 0, (u64)vma->node->length << 12);
 149 }
 150
 151 static void
 152 nouveau_vm_unmap_pgt(struct nouveau_vm *vm, int big, u32 fpde, u32 lpde)
 153 {
 154         struct nouveau_vm_pgd *vpgd;
 155         struct nouveau_vm_pgt *vpgt;
 156         struct nouveau_gpuobj *pgt;
 157         u32 pde;
 158
 159         for (pde = fpde; pde <= lpde; pde++) {
 160                 vpgt = &vm->pgt[pde - vm->fpde];
 161                 if (--vpgt->refcount[big])
 162                         continue;
 163
 164                 pgt = vpgt->obj[big];
 165                 vpgt->obj[big] = NULL;
 166
 167                 list_for_each_entry(vpgd, &vm->pgd_list, head) {
 168                         vm->map_pgt(vpgd->obj, pde, vpgt->obj);
 169                 }
 170
 171                 mutex_unlock(&vm->mm->mutex);
 172                 nouveau_gpuobj_ref(NULL, &pgt);
 173                 mutex_lock(&vm->mm->mutex);
 174         }
 175 }
 176
 177 static int
 178 nouveau_vm_map_pgt(struct nouveau_vm *vm, u32 pde, u32 type)
 179 {
 180         struct nouveau_vm_pgt *vpgt = &vm->pgt[pde - vm->fpde];
 181         struct nouveau_vm_pgd *vpgd;
 182         struct nouveau_gpuobj *pgt;
 183         int big = (type != vm->spg_shift);
 184         u32 pgt_size;
 185         int ret;
 186
 187         pgt_size  = (1 << (vm->pgt_bits + 12)) >> type;
 188         pgt_size *= 8;
 189
 190         mutex_unlock(&vm->mm->mutex);
 191         ret = nouveau_gpuobj_new(vm->dev, NULL, pgt_size, 0x1000,
 192                                  NVOBJ_FLAG_ZERO_ALLOC, &pgt);
 193         mutex_lock(&vm->mm->mutex);
 194         if (unlikely(ret))
 195                 return ret;
 196
 197         /* someone beat us to filling the PDE while we didn't have the lock */
 198         if (unlikely(vpgt->refcount[big]++)) {
 199                 mutex_unlock(&vm->mm->mutex);
 200                 nouveau_gpuobj_ref(NULL, &pgt);
 201                 mutex_lock(&vm->mm->mutex);
 202                 return 0;
 203         }
 204
 205         vpgt->obj[big] = pgt;
 206         list_for_each_entry(vpgd, &vm->pgd_list, head) {
 207                 vm->map_pgt(vpgd->obj, pde, vpgt->obj);
 208         }
 209
 210         return 0;
 211 }
 212
 213 int
 214 nouveau_vm_get(struct nouveau_vm *vm, u64 size, u32 page_shift,
 215                u32 access, struct nouveau_vma *vma)
 216 {
 217         u32 align = (1 << page_shift) >> 12;
 218         u32 msize = size >> 12;
 219         u32 fpde, lpde, pde;
 220         int ret;
 221
 222         mutex_lock(&vm->mm->mutex);
 223         ret = nouveau_mm_get(vm->mm, page_shift, msize, 0, align, &vma->node);
 224         if (unlikely(ret != 0)) {
 225                 mutex_unlock(&vm->mm->mutex);
 226                 return ret;
 227         }
 228
 229         fpde = (vma->node->offset >> vm->pgt_bits);
 230         lpde = (vma->node->offset + vma->node->length - 1) >> vm->pgt_bits;
 231         for (pde = fpde; pde <= lpde; pde++) {
 232                 struct nouveau_vm_pgt *vpgt = &vm->pgt[pde - vm->fpde];
 233                 int big = (vma->node->type != vm->spg_shift);
 234
 235                 if (likely(vpgt->refcount[big])) {
 236                         vpgt->refcount[big]++;
 237                         continue;
 238                 }
 239
 240                 ret = nouveau_vm_map_pgt(vm, pde, vma->node->type);
 241                 if (ret) {
 242                         if (pde != fpde)
 243                                 nouveau_vm_unmap_pgt(vm, big, fpde, pde - 1);
 244                         nouveau_mm_put(vm->mm, vma->node);
 245                         mutex_unlock(&vm->mm->mutex);
 246                         vma->node = NULL;
 247                         return ret;
 248                 }
 249         }
 250         mutex_unlock(&vm->mm->mutex);
 251
 252         vma->vm     = vm;
 253         vma->offset = (u64)vma->node->offset << 12;
 254         vma->access = access;
 255         return 0;
 256 }
 257
 258 void
 259 nouveau_vm_put(struct nouveau_vma *vma)
 260 {
 261         struct nouveau_vm *vm = vma->vm;
 262         u32 fpde, lpde;
 263
 264         if (unlikely(vma->node == NULL))
 265                 return;
 266         fpde = (vma->node->offset >> vm->pgt_bits);
 267         lpde = (vma->node->offset + vma->node->length - 1) >> vm->pgt_bits;
 268
 269         mutex_lock(&vm->mm->mutex);
 270         nouveau_vm_unmap_pgt(vm, vma->node->type != vm->spg_shift, fpde, lpde);
 271         nouveau_mm_put(vm->mm, vma->node);
 272         vma->node = NULL;
 273         mutex_unlock(&vm->mm->mutex);
 274 }
 275
 276 int
 277 nouveau_vm_new(struct drm_device *dev, u64 offset, u64 length, u64 mm_offset,
 278                struct nouveau_vm **pvm)
 279 {
 280         struct drm_nouveau_private *dev_priv = dev->dev_private;
 281         struct nouveau_vm *vm;
 282         u64 mm_length = (offset + length) - mm_offset;
 283         u32 block, pgt_bits;
 284         int ret;
 285
 286         vm = kzalloc(sizeof(*vm), GFP_KERNEL);
 287         if (!vm)
 288                 return -ENOMEM;
 289
 290         if (dev_priv->card_type == NV_50) {
 291                 vm->map_pgt = nv50_vm_map_pgt;
 292                 vm->map = nv50_vm_map;
 293                 vm->map_sg = nv50_vm_map_sg;
 294                 vm->unmap = nv50_vm_unmap;
 295                 vm->flush = nv50_vm_flush;
 296                 vm->spg_shift = 12;
 297                 vm->lpg_shift = 16;
 298                 pgt_bits = 29;
 299         } else {
 300                 kfree(vm);
 301                 return -ENOSYS;
 302         }
 303
 304         vm->fpde   = offset >> pgt_bits;
 305         vm->lpde   = (offset + length - 1) >> pgt_bits;
 306         vm->pgt = kcalloc(vm->lpde - vm->fpde + 1, sizeof(*vm->pgt), GFP_KERNEL);
 307         if (!vm->pgt) {
 308                 kfree(vm);
 309                 return -ENOMEM;
 310         }
 311
 312         INIT_LIST_HEAD(&vm->pgd_list);
 313         vm->dev = dev;
 314         vm->refcount = 1;
 315         vm->pgt_bits = pgt_bits - 12;
 316
 317         block = (1 << pgt_bits);
 318         if (length < block)
 319                 block = length;
 320
 321         ret = nouveau_mm_init(&vm->mm, mm_offset >> 12, mm_length >> 12,
 322                               block >> 12);
 323         if (ret) {
 324                 kfree(vm);
 325                 return ret;
 326         }
 327
 328         *pvm = vm;
 329         return 0;
 330 }
 331
 332 static int
 333 nouveau_vm_link(struct nouveau_vm *vm, struct nouveau_gpuobj *pgd)
 334 {
 335         struct nouveau_vm_pgd *vpgd;
 336         int i;
 337
 338         if (!pgd)
 339                 return 0;
 340
 341         vpgd = kzalloc(sizeof(*vpgd), GFP_KERNEL);
 342         if (!vpgd)
 343                 return -ENOMEM;
 344
 345         nouveau_gpuobj_ref(pgd, &vpgd->obj);
 346
 347         mutex_lock(&vm->mm->mutex);
 348         for (i = vm->fpde; i <= vm->lpde; i++)
 349                 vm->map_pgt(pgd, i, vm->pgt[i - vm->fpde].obj);
 350         list_add(&vpgd->head, &vm->pgd_list);
 351         mutex_unlock(&vm->mm->mutex);
 352         return 0;
 353 }
 354
 355 static void
 356 nouveau_vm_unlink(struct nouveau_vm *vm, struct nouveau_gpuobj *pgd)
 357 {
 358         struct nouveau_vm_pgd *vpgd, *tmp;
 359
 360         if (!pgd)
 361                 return;
 362
 363         mutex_lock(&vm->mm->mutex);
 364         list_for_each_entry_safe(vpgd, tmp, &vm->pgd_list, head) {
 365                 if (vpgd->obj != pgd)
 366                         continue;
 367
 368                 list_del(&vpgd->head);
 369                 nouveau_gpuobj_ref(NULL, &vpgd->obj);
 370                 kfree(vpgd);
 371         }
 372         mutex_unlock(&vm->mm->mutex);
 373 }
 374
 375 static void
 376 nouveau_vm_del(struct nouveau_vm *vm)
 377 {
 378         struct nouveau_vm_pgd *vpgd, *tmp;
 379
 380         list_for_each_entry_safe(vpgd, tmp, &vm->pgd_list, head) {
 381                 nouveau_vm_unlink(vm, vpgd->obj);
 382         }
 383         WARN_ON(nouveau_mm_fini(&vm->mm) != 0);
 384
 385         kfree(vm->pgt);
 386         kfree(vm);
 387 }
 388
 389 int
 390 nouveau_vm_ref(struct nouveau_vm *ref, struct nouveau_vm **ptr,
 391                struct nouveau_gpuobj *pgd)
 392 {
 393         struct nouveau_vm *vm;
 394         int ret;
 395
 396         vm = ref;
 397         if (vm) {
 398                 ret = nouveau_vm_link(vm, pgd);
 399                 if (ret)
 400                         return ret;
 401
 402                 vm->refcount++;
 403         }
 404
 405         vm = *ptr;
 406         *ptr = ref;
 407
 408         if (vm) {
 409                 nouveau_vm_unlink(vm, pgd);
 410
 411                 if (--vm->refcount == 0)
 412                         nouveau_vm_del(vm);
 413         }
 414
 415         return 0;
 416 }