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#include <subdev/mmu.h> 25#include <subdev/fb.h> 26 27#include <core/gpuobj.h> 28 29void 30nvkm_vm_map_at(struct nvkm_vma *vma, u64 delta, struct nvkm_mem *node) 31{ 32 struct nvkm_vm *vm = vma->vm; 33 struct nvkm_mmu *mmu = vm->mmu; 34 struct nvkm_mm_node *r; 35 int big = vma->node->type != mmu->spg_shift; 36 u32 offset = vma->node->offset + (delta >> 12); 37 u32 bits = vma->node->type - 12; 38 u32 pde = (offset >> mmu->pgt_bits) - vm->fpde; 39 u32 pte = (offset & ((1 << mmu->pgt_bits) - 1)) >> bits; 40 u32 max = 1 << (mmu->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 nvkm_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 mmu->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 mmu->flush(vm); 71} 72 73static void 74nvkm_vm_map_sg_table(struct nvkm_vma *vma, u64 delta, u64 length, 75 struct nvkm_mem *mem) 76{ 77 struct nvkm_vm *vm = vma->vm; 78 struct nvkm_mmu *mmu = vm->mmu; 79 int big = vma->node->type != mmu->spg_shift; 80 u32 offset = vma->node->offset + (delta >> 12); 81 u32 bits = vma->node->type - 12; 82 u32 num = length >> vma->node->type; 83 u32 pde = (offset >> mmu->pgt_bits) - vm->fpde; 84 u32 pte = (offset & ((1 << mmu->pgt_bits) - 1)) >> bits; 85 u32 max = 1 << (mmu->pgt_bits - bits); 86 unsigned m, sglen; 87 u32 end, len; 88 int i; 89 struct scatterlist *sg; 90 91 for_each_sg(mem->sg->sgl, sg, mem->sg->nents, i) { 92 struct nvkm_gpuobj *pgt = vm->pgt[pde].obj[big]; 93 sglen = sg_dma_len(sg) >> PAGE_SHIFT; 94 95 end = pte + sglen; 96 if (unlikely(end >= max)) 97 end = max; 98 len = end - pte; 99 100 for (m = 0; m < len; m++) { 101 dma_addr_t addr = sg_dma_address(sg) + (m << PAGE_SHIFT); 102 103 mmu->map_sg(vma, pgt, mem, pte, 1, &addr); 104 num--; 105 pte++; 106 107 if (num == 0) 108 goto finish; 109 } 110 if (unlikely(end >= max)) { 111 pde++; 112 pte = 0; 113 } 114 if (m < sglen) { 115 for (; m < sglen; m++) { 116 dma_addr_t addr = sg_dma_address(sg) + (m << PAGE_SHIFT); 117 118 mmu->map_sg(vma, pgt, mem, pte, 1, &addr); 119 num--; 120 pte++; 121 if (num == 0) 122 goto finish; 123 } 124 } 125 126 } 127finish: 128 mmu->flush(vm); 129} 130 131static void 132nvkm_vm_map_sg(struct nvkm_vma *vma, u64 delta, u64 length, 133 struct nvkm_mem *mem) 134{ 135 struct nvkm_vm *vm = vma->vm; 136 struct nvkm_mmu *mmu = vm->mmu; 137 dma_addr_t *list = mem->pages; 138 int big = vma->node->type != mmu->spg_shift; 139 u32 offset = vma->node->offset + (delta >> 12); 140 u32 bits = vma->node->type - 12; 141 u32 num = length >> vma->node->type; 142 u32 pde = (offset >> mmu->pgt_bits) - vm->fpde; 143 u32 pte = (offset & ((1 << mmu->pgt_bits) - 1)) >> bits; 144 u32 max = 1 << (mmu->pgt_bits - bits); 145 u32 end, len; 146 147 while (num) { 148 struct nvkm_gpuobj *pgt = vm->pgt[pde].obj[big]; 149 150 end = (pte + num); 151 if (unlikely(end >= max)) 152 end = max; 153 len = end - pte; 154 155 mmu->map_sg(vma, pgt, mem, pte, len, list); 156 157 num -= len; 158 pte += len; 159 list += len; 160 if (unlikely(end >= max)) { 161 pde++; 162 pte = 0; 163 } 164 } 165 166 mmu->flush(vm); 167} 168 169void 170nvkm_vm_map(struct nvkm_vma *vma, struct nvkm_mem *node) 171{ 172 if (node->sg) 173 nvkm_vm_map_sg_table(vma, 0, node->size << 12, node); 174 else 175 if (node->pages) 176 nvkm_vm_map_sg(vma, 0, node->size << 12, node); 177 else 178 nvkm_vm_map_at(vma, 0, node); 179} 180 181void 182nvkm_vm_unmap_at(struct nvkm_vma *vma, u64 delta, u64 length) 183{ 184 struct nvkm_vm *vm = vma->vm; 185 struct nvkm_mmu *mmu = vm->mmu; 186 int big = vma->node->type != mmu->spg_shift; 187 u32 offset = vma->node->offset + (delta >> 12); 188 u32 bits = vma->node->type - 12; 189 u32 num = length >> vma->node->type; 190 u32 pde = (offset >> mmu->pgt_bits) - vm->fpde; 191 u32 pte = (offset & ((1 << mmu->pgt_bits) - 1)) >> bits; 192 u32 max = 1 << (mmu->pgt_bits - bits); 193 u32 end, len; 194 195 while (num) { 196 struct nvkm_gpuobj *pgt = vm->pgt[pde].obj[big]; 197 198 end = (pte + num); 199 if (unlikely(end >= max)) 200 end = max; 201 len = end - pte; 202 203 mmu->unmap(pgt, pte, len); 204 205 num -= len; 206 pte += len; 207 if (unlikely(end >= max)) { 208 pde++; 209 pte = 0; 210 } 211 } 212 213 mmu->flush(vm); 214} 215 216void 217nvkm_vm_unmap(struct nvkm_vma *vma) 218{ 219 nvkm_vm_unmap_at(vma, 0, (u64)vma->node->length << 12); 220} 221 222static void 223nvkm_vm_unmap_pgt(struct nvkm_vm *vm, int big, u32 fpde, u32 lpde) 224{ 225 struct nvkm_mmu *mmu = vm->mmu; 226 struct nvkm_vm_pgd *vpgd; 227 struct nvkm_vm_pgt *vpgt; 228 struct nvkm_gpuobj *pgt; 229 u32 pde; 230 231 for (pde = fpde; pde <= lpde; pde++) { 232 vpgt = &vm->pgt[pde - vm->fpde]; 233 if (--vpgt->refcount[big]) 234 continue; 235 236 pgt = vpgt->obj[big]; 237 vpgt->obj[big] = NULL; 238 239 list_for_each_entry(vpgd, &vm->pgd_list, head) { 240 mmu->map_pgt(vpgd->obj, pde, vpgt->obj); 241 } 242 243 mutex_unlock(&nv_subdev(mmu)->mutex); 244 nvkm_gpuobj_ref(NULL, &pgt); 245 mutex_lock(&nv_subdev(mmu)->mutex); 246 } 247} 248 249static int 250nvkm_vm_map_pgt(struct nvkm_vm *vm, u32 pde, u32 type) 251{ 252 struct nvkm_mmu *mmu = vm->mmu; 253 struct nvkm_vm_pgt *vpgt = &vm->pgt[pde - vm->fpde]; 254 struct nvkm_vm_pgd *vpgd; 255 struct nvkm_gpuobj *pgt; 256 int big = (type != mmu->spg_shift); 257 u32 pgt_size; 258 int ret; 259 260 pgt_size = (1 << (mmu->pgt_bits + 12)) >> type; 261 pgt_size *= 8; 262 263 mutex_unlock(&nv_subdev(mmu)->mutex); 264 ret = nvkm_gpuobj_new(nv_object(vm->mmu), NULL, pgt_size, 0x1000, 265 NVOBJ_FLAG_ZERO_ALLOC, &pgt); 266 mutex_lock(&nv_subdev(mmu)->mutex); 267 if (unlikely(ret)) 268 return ret; 269 270 /* someone beat us to filling the PDE while we didn't have the lock */ 271 if (unlikely(vpgt->refcount[big]++)) { 272 mutex_unlock(&nv_subdev(mmu)->mutex); 273 nvkm_gpuobj_ref(NULL, &pgt); 274 mutex_lock(&nv_subdev(mmu)->mutex); 275 return 0; 276 } 277 278 vpgt->obj[big] = pgt; 279 list_for_each_entry(vpgd, &vm->pgd_list, head) { 280 mmu->map_pgt(vpgd->obj, pde, vpgt->obj); 281 } 282 283 return 0; 284} 285 286int 287nvkm_vm_get(struct nvkm_vm *vm, u64 size, u32 page_shift, u32 access, 288 struct nvkm_vma *vma) 289{ 290 struct nvkm_mmu *mmu = vm->mmu; 291 u32 align = (1 << page_shift) >> 12; 292 u32 msize = size >> 12; 293 u32 fpde, lpde, pde; 294 int ret; 295 296 mutex_lock(&nv_subdev(mmu)->mutex); 297 ret = nvkm_mm_head(&vm->mm, 0, page_shift, msize, msize, align, 298 &vma->node); 299 if (unlikely(ret != 0)) { 300 mutex_unlock(&nv_subdev(mmu)->mutex); 301 return ret; 302 } 303 304 fpde = (vma->node->offset >> mmu->pgt_bits); 305 lpde = (vma->node->offset + vma->node->length - 1) >> mmu->pgt_bits; 306 307 for (pde = fpde; pde <= lpde; pde++) { 308 struct nvkm_vm_pgt *vpgt = &vm->pgt[pde - vm->fpde]; 309 int big = (vma->node->type != mmu->spg_shift); 310 311 if (likely(vpgt->refcount[big])) { 312 vpgt->refcount[big]++; 313 continue; 314 } 315 316 ret = nvkm_vm_map_pgt(vm, pde, vma->node->type); 317 if (ret) { 318 if (pde != fpde) 319 nvkm_vm_unmap_pgt(vm, big, fpde, pde - 1); 320 nvkm_mm_free(&vm->mm, &vma->node); 321 mutex_unlock(&nv_subdev(mmu)->mutex); 322 return ret; 323 } 324 } 325 mutex_unlock(&nv_subdev(mmu)->mutex); 326 327 vma->vm = NULL; 328 nvkm_vm_ref(vm, &vma->vm, NULL); 329 vma->offset = (u64)vma->node->offset << 12; 330 vma->access = access; 331 return 0; 332} 333 334void 335nvkm_vm_put(struct nvkm_vma *vma) 336{ 337 struct nvkm_vm *vm = vma->vm; 338 struct nvkm_mmu *mmu = vm->mmu; 339 u32 fpde, lpde; 340 341 if (unlikely(vma->node == NULL)) 342 return; 343 fpde = (vma->node->offset >> mmu->pgt_bits); 344 lpde = (vma->node->offset + vma->node->length - 1) >> mmu->pgt_bits; 345 346 mutex_lock(&nv_subdev(mmu)->mutex); 347 nvkm_vm_unmap_pgt(vm, vma->node->type != mmu->spg_shift, fpde, lpde); 348 nvkm_mm_free(&vm->mm, &vma->node); 349 mutex_unlock(&nv_subdev(mmu)->mutex); 350 351 nvkm_vm_ref(NULL, &vma->vm, NULL); 352} 353 354int 355nvkm_vm_create(struct nvkm_mmu *mmu, u64 offset, u64 length, u64 mm_offset, 356 u32 block, struct nvkm_vm **pvm) 357{ 358 struct nvkm_vm *vm; 359 u64 mm_length = (offset + length) - mm_offset; 360 int ret; 361 362 vm = kzalloc(sizeof(*vm), GFP_KERNEL); 363 if (!vm) 364 return -ENOMEM; 365 366 INIT_LIST_HEAD(&vm->pgd_list); 367 vm->mmu = mmu; 368 kref_init(&vm->refcount); 369 vm->fpde = offset >> (mmu->pgt_bits + 12); 370 vm->lpde = (offset + length - 1) >> (mmu->pgt_bits + 12); 371 372 vm->pgt = vzalloc((vm->lpde - vm->fpde + 1) * sizeof(*vm->pgt)); 373 if (!vm->pgt) { 374 kfree(vm); 375 return -ENOMEM; 376 } 377 378 ret = nvkm_mm_init(&vm->mm, mm_offset >> 12, mm_length >> 12, 379 block >> 12); 380 if (ret) { 381 vfree(vm->pgt); 382 kfree(vm); 383 return ret; 384 } 385 386 *pvm = vm; 387 388 return 0; 389} 390 391int 392nvkm_vm_new(struct nvkm_device *device, u64 offset, u64 length, u64 mm_offset, 393 struct nvkm_vm **pvm) 394{ 395 struct nvkm_mmu *mmu = nvkm_mmu(device); 396 return mmu->create(mmu, offset, length, mm_offset, pvm); 397} 398 399static int 400nvkm_vm_link(struct nvkm_vm *vm, struct nvkm_gpuobj *pgd) 401{ 402 struct nvkm_mmu *mmu = vm->mmu; 403 struct nvkm_vm_pgd *vpgd; 404 int i; 405 406 if (!pgd) 407 return 0; 408 409 vpgd = kzalloc(sizeof(*vpgd), GFP_KERNEL); 410 if (!vpgd) 411 return -ENOMEM; 412 413 nvkm_gpuobj_ref(pgd, &vpgd->obj); 414 415 mutex_lock(&nv_subdev(mmu)->mutex); 416 for (i = vm->fpde; i <= vm->lpde; i++) 417 mmu->map_pgt(pgd, i, vm->pgt[i - vm->fpde].obj); 418 list_add(&vpgd->head, &vm->pgd_list); 419 mutex_unlock(&nv_subdev(mmu)->mutex); 420 return 0; 421} 422 423static void 424nvkm_vm_unlink(struct nvkm_vm *vm, struct nvkm_gpuobj *mpgd) 425{ 426 struct nvkm_mmu *mmu = vm->mmu; 427 struct nvkm_vm_pgd *vpgd, *tmp; 428 struct nvkm_gpuobj *pgd = NULL; 429 430 if (!mpgd) 431 return; 432 433 mutex_lock(&nv_subdev(mmu)->mutex); 434 list_for_each_entry_safe(vpgd, tmp, &vm->pgd_list, head) { 435 if (vpgd->obj == mpgd) { 436 pgd = vpgd->obj; 437 list_del(&vpgd->head); 438 kfree(vpgd); 439 break; 440 } 441 } 442 mutex_unlock(&nv_subdev(mmu)->mutex); 443 444 nvkm_gpuobj_ref(NULL, &pgd); 445} 446 447static void 448nvkm_vm_del(struct kref *kref) 449{ 450 struct nvkm_vm *vm = container_of(kref, typeof(*vm), refcount); 451 struct nvkm_vm_pgd *vpgd, *tmp; 452 453 list_for_each_entry_safe(vpgd, tmp, &vm->pgd_list, head) { 454 nvkm_vm_unlink(vm, vpgd->obj); 455 } 456 457 nvkm_mm_fini(&vm->mm); 458 vfree(vm->pgt); 459 kfree(vm); 460} 461 462int 463nvkm_vm_ref(struct nvkm_vm *ref, struct nvkm_vm **ptr, struct nvkm_gpuobj *pgd) 464{ 465 if (ref) { 466 int ret = nvkm_vm_link(ref, pgd); 467 if (ret) 468 return ret; 469 470 kref_get(&ref->refcount); 471 } 472 473 if (*ptr) { 474 nvkm_vm_unlink(*ptr, pgd); 475 kref_put(&(*ptr)->refcount, nvkm_vm_del); 476 } 477 478 *ptr = ref; 479 return 0; 480} 481