1 // SPDX-License-Identifier: MIT 2 /* 3 * Copyright © 2020 Intel Corporation 4 */ 5 6 #include <linux/log2.h> 7 8 #include "gen8_ppgtt.h" 9 #include "i915_scatterlist.h" 10 #include "i915_trace.h" 11 #include "i915_pvinfo.h" 12 #include "i915_vgpu.h" 13 #include "intel_gt.h" 14 #include "intel_gtt.h" 15 16 static u64 gen8_pde_encode(const dma_addr_t addr, 17 const enum i915_cache_level level) 18 { 19 u64 pde = addr | _PAGE_PRESENT | _PAGE_RW; 20 21 if (level != I915_CACHE_NONE) 22 pde |= PPAT_CACHED_PDE; 23 else 24 pde |= PPAT_UNCACHED; 25 26 return pde; 27 } 28 29 static u64 gen8_pte_encode(dma_addr_t addr, 30 enum i915_cache_level level, 31 u32 flags) 32 { 33 gen8_pte_t pte = addr | _PAGE_PRESENT | _PAGE_RW; 34 35 if (unlikely(flags & PTE_READ_ONLY)) 36 pte &= ~_PAGE_RW; 37 38 switch (level) { 39 case I915_CACHE_NONE: 40 pte |= PPAT_UNCACHED; 41 break; 42 case I915_CACHE_WT: 43 pte |= PPAT_DISPLAY_ELLC; 44 break; 45 default: 46 pte |= PPAT_CACHED; 47 break; 48 } 49 50 return pte; 51 } 52 53 static void gen8_ppgtt_notify_vgt(struct i915_ppgtt *ppgtt, bool create) 54 { 55 struct drm_i915_private *i915 = ppgtt->vm.i915; 56 struct intel_uncore *uncore = ppgtt->vm.gt->uncore; 57 enum vgt_g2v_type msg; 58 int i; 59 60 if (create) 61 atomic_inc(px_used(ppgtt->pd)); /* never remove */ 62 else 63 atomic_dec(px_used(ppgtt->pd)); 64 65 mutex_lock(&i915->vgpu.lock); 66 67 if (i915_vm_is_4lvl(&ppgtt->vm)) { 68 const u64 daddr = px_dma(ppgtt->pd); 69 70 intel_uncore_write(uncore, 71 vgtif_reg(pdp[0].lo), lower_32_bits(daddr)); 72 intel_uncore_write(uncore, 73 vgtif_reg(pdp[0].hi), upper_32_bits(daddr)); 74 75 msg = create ? 76 VGT_G2V_PPGTT_L4_PAGE_TABLE_CREATE : 77 VGT_G2V_PPGTT_L4_PAGE_TABLE_DESTROY; 78 } else { 79 for (i = 0; i < GEN8_3LVL_PDPES; i++) { 80 const u64 daddr = i915_page_dir_dma_addr(ppgtt, i); 81 82 intel_uncore_write(uncore, 83 vgtif_reg(pdp[i].lo), 84 lower_32_bits(daddr)); 85 intel_uncore_write(uncore, 86 vgtif_reg(pdp[i].hi), 87 upper_32_bits(daddr)); 88 } 89 90 msg = create ? 91 VGT_G2V_PPGTT_L3_PAGE_TABLE_CREATE : 92 VGT_G2V_PPGTT_L3_PAGE_TABLE_DESTROY; 93 } 94 95 /* g2v_notify atomically (via hv trap) consumes the message packet. */ 96 intel_uncore_write(uncore, vgtif_reg(g2v_notify), msg); 97 98 mutex_unlock(&i915->vgpu.lock); 99 } 100 101 /* Index shifts into the pagetable are offset by GEN8_PTE_SHIFT [12] */ 102 #define GEN8_PAGE_SIZE (SZ_4K) /* page and page-directory sizes are the same */ 103 #define GEN8_PTE_SHIFT (ilog2(GEN8_PAGE_SIZE)) 104 #define GEN8_PDES (GEN8_PAGE_SIZE / sizeof(u64)) 105 #define gen8_pd_shift(lvl) ((lvl) * ilog2(GEN8_PDES)) 106 #define gen8_pd_index(i, lvl) i915_pde_index((i), gen8_pd_shift(lvl)) 107 #define __gen8_pte_shift(lvl) (GEN8_PTE_SHIFT + gen8_pd_shift(lvl)) 108 #define __gen8_pte_index(a, lvl) i915_pde_index((a), __gen8_pte_shift(lvl)) 109 110 #define as_pd(x) container_of((x), typeof(struct i915_page_directory), pt) 111 112 static inline unsigned int 113 gen8_pd_range(u64 start, u64 end, int lvl, unsigned int *idx) 114 { 115 const int shift = gen8_pd_shift(lvl); 116 const u64 mask = ~0ull << gen8_pd_shift(lvl + 1); 117 118 GEM_BUG_ON(start >= end); 119 end += ~mask >> gen8_pd_shift(1); 120 121 *idx = i915_pde_index(start, shift); 122 if ((start ^ end) & mask) 123 return GEN8_PDES - *idx; 124 else 125 return i915_pde_index(end, shift) - *idx; 126 } 127 128 static inline bool gen8_pd_contains(u64 start, u64 end, int lvl) 129 { 130 const u64 mask = ~0ull << gen8_pd_shift(lvl + 1); 131 132 GEM_BUG_ON(start >= end); 133 return (start ^ end) & mask && (start & ~mask) == 0; 134 } 135 136 static inline unsigned int gen8_pt_count(u64 start, u64 end) 137 { 138 GEM_BUG_ON(start >= end); 139 if ((start ^ end) >> gen8_pd_shift(1)) 140 return GEN8_PDES - (start & (GEN8_PDES - 1)); 141 else 142 return end - start; 143 } 144 145 static inline unsigned int 146 gen8_pd_top_count(const struct i915_address_space *vm) 147 { 148 unsigned int shift = __gen8_pte_shift(vm->top); 149 return (vm->total + (1ull << shift) - 1) >> shift; 150 } 151 152 static inline struct i915_page_directory * 153 gen8_pdp_for_page_index(struct i915_address_space * const vm, const u64 idx) 154 { 155 struct i915_ppgtt * const ppgtt = i915_vm_to_ppgtt(vm); 156 157 if (vm->top == 2) 158 return ppgtt->pd; 159 else 160 return i915_pd_entry(ppgtt->pd, gen8_pd_index(idx, vm->top)); 161 } 162 163 static inline struct i915_page_directory * 164 gen8_pdp_for_page_address(struct i915_address_space * const vm, const u64 addr) 165 { 166 return gen8_pdp_for_page_index(vm, addr >> GEN8_PTE_SHIFT); 167 } 168 169 static void __gen8_ppgtt_cleanup(struct i915_address_space *vm, 170 struct i915_page_directory *pd, 171 int count, int lvl) 172 { 173 if (lvl) { 174 void **pde = pd->entry; 175 176 do { 177 if (!*pde) 178 continue; 179 180 __gen8_ppgtt_cleanup(vm, *pde, GEN8_PDES, lvl - 1); 181 } while (pde++, --count); 182 } 183 184 free_px(vm, &pd->pt, lvl); 185 } 186 187 static void gen8_ppgtt_cleanup(struct i915_address_space *vm) 188 { 189 struct i915_ppgtt *ppgtt = i915_vm_to_ppgtt(vm); 190 191 if (intel_vgpu_active(vm->i915)) 192 gen8_ppgtt_notify_vgt(ppgtt, false); 193 194 __gen8_ppgtt_cleanup(vm, ppgtt->pd, gen8_pd_top_count(vm), vm->top); 195 free_scratch(vm); 196 } 197 198 static u64 __gen8_ppgtt_clear(struct i915_address_space * const vm, 199 struct i915_page_directory * const pd, 200 u64 start, const u64 end, int lvl) 201 { 202 const struct drm_i915_gem_object * const scratch = vm->scratch[lvl]; 203 unsigned int idx, len; 204 205 GEM_BUG_ON(end > vm->total >> GEN8_PTE_SHIFT); 206 207 len = gen8_pd_range(start, end, lvl--, &idx); 208 DBG("%s(%p):{ lvl:%d, start:%llx, end:%llx, idx:%d, len:%d, used:%d }\n", 209 __func__, vm, lvl + 1, start, end, 210 idx, len, atomic_read(px_used(pd))); 211 GEM_BUG_ON(!len || len >= atomic_read(px_used(pd))); 212 213 do { 214 struct i915_page_table *pt = pd->entry[idx]; 215 216 if (atomic_fetch_inc(&pt->used) >> gen8_pd_shift(1) && 217 gen8_pd_contains(start, end, lvl)) { 218 DBG("%s(%p):{ lvl:%d, idx:%d, start:%llx, end:%llx } removing pd\n", 219 __func__, vm, lvl + 1, idx, start, end); 220 clear_pd_entry(pd, idx, scratch); 221 __gen8_ppgtt_cleanup(vm, as_pd(pt), I915_PDES, lvl); 222 start += (u64)I915_PDES << gen8_pd_shift(lvl); 223 continue; 224 } 225 226 if (lvl) { 227 start = __gen8_ppgtt_clear(vm, as_pd(pt), 228 start, end, lvl); 229 } else { 230 unsigned int count; 231 u64 *vaddr; 232 233 count = gen8_pt_count(start, end); 234 DBG("%s(%p):{ lvl:%d, start:%llx, end:%llx, idx:%d, len:%d, used:%d } removing pte\n", 235 __func__, vm, lvl, start, end, 236 gen8_pd_index(start, 0), count, 237 atomic_read(&pt->used)); 238 GEM_BUG_ON(!count || count >= atomic_read(&pt->used)); 239 240 vaddr = kmap_atomic_px(pt); 241 memset64(vaddr + gen8_pd_index(start, 0), 242 vm->scratch[0]->encode, 243 count); 244 kunmap_atomic(vaddr); 245 246 atomic_sub(count, &pt->used); 247 start += count; 248 } 249 250 if (release_pd_entry(pd, idx, pt, scratch)) 251 free_px(vm, pt, lvl); 252 } while (idx++, --len); 253 254 return start; 255 } 256 257 static void gen8_ppgtt_clear(struct i915_address_space *vm, 258 u64 start, u64 length) 259 { 260 GEM_BUG_ON(!IS_ALIGNED(start, BIT_ULL(GEN8_PTE_SHIFT))); 261 GEM_BUG_ON(!IS_ALIGNED(length, BIT_ULL(GEN8_PTE_SHIFT))); 262 GEM_BUG_ON(range_overflows(start, length, vm->total)); 263 264 start >>= GEN8_PTE_SHIFT; 265 length >>= GEN8_PTE_SHIFT; 266 GEM_BUG_ON(length == 0); 267 268 __gen8_ppgtt_clear(vm, i915_vm_to_ppgtt(vm)->pd, 269 start, start + length, vm->top); 270 } 271 272 static void __gen8_ppgtt_alloc(struct i915_address_space * const vm, 273 struct i915_vm_pt_stash *stash, 274 struct i915_page_directory * const pd, 275 u64 * const start, const u64 end, int lvl) 276 { 277 unsigned int idx, len; 278 279 GEM_BUG_ON(end > vm->total >> GEN8_PTE_SHIFT); 280 281 len = gen8_pd_range(*start, end, lvl--, &idx); 282 DBG("%s(%p):{ lvl:%d, start:%llx, end:%llx, idx:%d, len:%d, used:%d }\n", 283 __func__, vm, lvl + 1, *start, end, 284 idx, len, atomic_read(px_used(pd))); 285 GEM_BUG_ON(!len || (idx + len - 1) >> gen8_pd_shift(1)); 286 287 spin_lock(&pd->lock); 288 GEM_BUG_ON(!atomic_read(px_used(pd))); /* Must be pinned! */ 289 do { 290 struct i915_page_table *pt = pd->entry[idx]; 291 292 if (!pt) { 293 spin_unlock(&pd->lock); 294 295 DBG("%s(%p):{ lvl:%d, idx:%d } allocating new tree\n", 296 __func__, vm, lvl + 1, idx); 297 298 pt = stash->pt[!!lvl]; 299 __i915_gem_object_pin_pages(pt->base); 300 i915_gem_object_make_unshrinkable(pt->base); 301 302 fill_px(pt, vm->scratch[lvl]->encode); 303 304 spin_lock(&pd->lock); 305 if (likely(!pd->entry[idx])) { 306 stash->pt[!!lvl] = pt->stash; 307 atomic_set(&pt->used, 0); 308 set_pd_entry(pd, idx, pt); 309 } else { 310 pt = pd->entry[idx]; 311 } 312 } 313 314 if (lvl) { 315 atomic_inc(&pt->used); 316 spin_unlock(&pd->lock); 317 318 __gen8_ppgtt_alloc(vm, stash, 319 as_pd(pt), start, end, lvl); 320 321 spin_lock(&pd->lock); 322 atomic_dec(&pt->used); 323 GEM_BUG_ON(!atomic_read(&pt->used)); 324 } else { 325 unsigned int count = gen8_pt_count(*start, end); 326 327 DBG("%s(%p):{ lvl:%d, start:%llx, end:%llx, idx:%d, len:%d, used:%d } inserting pte\n", 328 __func__, vm, lvl, *start, end, 329 gen8_pd_index(*start, 0), count, 330 atomic_read(&pt->used)); 331 332 atomic_add(count, &pt->used); 333 /* All other pdes may be simultaneously removed */ 334 GEM_BUG_ON(atomic_read(&pt->used) > NALLOC * I915_PDES); 335 *start += count; 336 } 337 } while (idx++, --len); 338 spin_unlock(&pd->lock); 339 } 340 341 static void gen8_ppgtt_alloc(struct i915_address_space *vm, 342 struct i915_vm_pt_stash *stash, 343 u64 start, u64 length) 344 { 345 GEM_BUG_ON(!IS_ALIGNED(start, BIT_ULL(GEN8_PTE_SHIFT))); 346 GEM_BUG_ON(!IS_ALIGNED(length, BIT_ULL(GEN8_PTE_SHIFT))); 347 GEM_BUG_ON(range_overflows(start, length, vm->total)); 348 349 start >>= GEN8_PTE_SHIFT; 350 length >>= GEN8_PTE_SHIFT; 351 GEM_BUG_ON(length == 0); 352 353 __gen8_ppgtt_alloc(vm, stash, i915_vm_to_ppgtt(vm)->pd, 354 &start, start + length, vm->top); 355 } 356 357 static __always_inline u64 358 gen8_ppgtt_insert_pte(struct i915_ppgtt *ppgtt, 359 struct i915_page_directory *pdp, 360 struct sgt_dma *iter, 361 u64 idx, 362 enum i915_cache_level cache_level, 363 u32 flags) 364 { 365 struct i915_page_directory *pd; 366 const gen8_pte_t pte_encode = gen8_pte_encode(0, cache_level, flags); 367 gen8_pte_t *vaddr; 368 369 pd = i915_pd_entry(pdp, gen8_pd_index(idx, 2)); 370 vaddr = kmap_atomic_px(i915_pt_entry(pd, gen8_pd_index(idx, 1))); 371 do { 372 GEM_BUG_ON(iter->sg->length < I915_GTT_PAGE_SIZE); 373 vaddr[gen8_pd_index(idx, 0)] = pte_encode | iter->dma; 374 375 iter->dma += I915_GTT_PAGE_SIZE; 376 if (iter->dma >= iter->max) { 377 iter->sg = __sg_next(iter->sg); 378 if (!iter->sg) { 379 idx = 0; 380 break; 381 } 382 383 iter->dma = sg_dma_address(iter->sg); 384 iter->max = iter->dma + iter->sg->length; 385 } 386 387 if (gen8_pd_index(++idx, 0) == 0) { 388 if (gen8_pd_index(idx, 1) == 0) { 389 /* Limited by sg length for 3lvl */ 390 if (gen8_pd_index(idx, 2) == 0) 391 break; 392 393 pd = pdp->entry[gen8_pd_index(idx, 2)]; 394 } 395 396 clflush_cache_range(vaddr, PAGE_SIZE); 397 kunmap_atomic(vaddr); 398 vaddr = kmap_atomic_px(i915_pt_entry(pd, gen8_pd_index(idx, 1))); 399 } 400 } while (1); 401 clflush_cache_range(vaddr, PAGE_SIZE); 402 kunmap_atomic(vaddr); 403 404 return idx; 405 } 406 407 static void gen8_ppgtt_insert_huge(struct i915_vma *vma, 408 struct sgt_dma *iter, 409 enum i915_cache_level cache_level, 410 u32 flags) 411 { 412 const gen8_pte_t pte_encode = gen8_pte_encode(0, cache_level, flags); 413 u64 start = vma->node.start; 414 dma_addr_t rem = iter->sg->length; 415 416 GEM_BUG_ON(!i915_vm_is_4lvl(vma->vm)); 417 418 do { 419 struct i915_page_directory * const pdp = 420 gen8_pdp_for_page_address(vma->vm, start); 421 struct i915_page_directory * const pd = 422 i915_pd_entry(pdp, __gen8_pte_index(start, 2)); 423 gen8_pte_t encode = pte_encode; 424 unsigned int maybe_64K = -1; 425 unsigned int page_size; 426 gen8_pte_t *vaddr; 427 u16 index; 428 429 if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_2M && 430 IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_2M) && 431 rem >= I915_GTT_PAGE_SIZE_2M && 432 !__gen8_pte_index(start, 0)) { 433 index = __gen8_pte_index(start, 1); 434 encode |= GEN8_PDE_PS_2M; 435 page_size = I915_GTT_PAGE_SIZE_2M; 436 437 vaddr = kmap_atomic_px(pd); 438 } else { 439 struct i915_page_table *pt = 440 i915_pt_entry(pd, __gen8_pte_index(start, 1)); 441 442 index = __gen8_pte_index(start, 0); 443 page_size = I915_GTT_PAGE_SIZE; 444 445 if (!index && 446 vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K && 447 IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_64K) && 448 (IS_ALIGNED(rem, I915_GTT_PAGE_SIZE_64K) || 449 rem >= (I915_PDES - index) * I915_GTT_PAGE_SIZE)) 450 maybe_64K = __gen8_pte_index(start, 1); 451 452 vaddr = kmap_atomic_px(pt); 453 } 454 455 do { 456 GEM_BUG_ON(iter->sg->length < page_size); 457 vaddr[index++] = encode | iter->dma; 458 459 start += page_size; 460 iter->dma += page_size; 461 rem -= page_size; 462 if (iter->dma >= iter->max) { 463 iter->sg = __sg_next(iter->sg); 464 if (!iter->sg) 465 break; 466 467 rem = iter->sg->length; 468 iter->dma = sg_dma_address(iter->sg); 469 iter->max = iter->dma + rem; 470 471 if (maybe_64K != -1 && index < I915_PDES && 472 !(IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_64K) && 473 (IS_ALIGNED(rem, I915_GTT_PAGE_SIZE_64K) || 474 rem >= (I915_PDES - index) * I915_GTT_PAGE_SIZE))) 475 maybe_64K = -1; 476 477 if (unlikely(!IS_ALIGNED(iter->dma, page_size))) 478 break; 479 } 480 } while (rem >= page_size && index < I915_PDES); 481 482 clflush_cache_range(vaddr, PAGE_SIZE); 483 kunmap_atomic(vaddr); 484 485 /* 486 * Is it safe to mark the 2M block as 64K? -- Either we have 487 * filled whole page-table with 64K entries, or filled part of 488 * it and have reached the end of the sg table and we have 489 * enough padding. 490 */ 491 if (maybe_64K != -1 && 492 (index == I915_PDES || 493 (i915_vm_has_scratch_64K(vma->vm) && 494 !iter->sg && IS_ALIGNED(vma->node.start + 495 vma->node.size, 496 I915_GTT_PAGE_SIZE_2M)))) { 497 vaddr = kmap_atomic_px(pd); 498 vaddr[maybe_64K] |= GEN8_PDE_IPS_64K; 499 kunmap_atomic(vaddr); 500 page_size = I915_GTT_PAGE_SIZE_64K; 501 502 /* 503 * We write all 4K page entries, even when using 64K 504 * pages. In order to verify that the HW isn't cheating 505 * by using the 4K PTE instead of the 64K PTE, we want 506 * to remove all the surplus entries. If the HW skipped 507 * the 64K PTE, it will read/write into the scratch page 508 * instead - which we detect as missing results during 509 * selftests. 510 */ 511 if (I915_SELFTEST_ONLY(vma->vm->scrub_64K)) { 512 u16 i; 513 514 encode = vma->vm->scratch[0]->encode; 515 vaddr = kmap_atomic_px(i915_pt_entry(pd, maybe_64K)); 516 517 for (i = 1; i < index; i += 16) 518 memset64(vaddr + i, encode, 15); 519 520 kunmap_atomic(vaddr); 521 } 522 } 523 524 vma->page_sizes.gtt |= page_size; 525 } while (iter->sg); 526 } 527 528 static void gen8_ppgtt_insert(struct i915_address_space *vm, 529 struct i915_vma *vma, 530 enum i915_cache_level cache_level, 531 u32 flags) 532 { 533 struct i915_ppgtt * const ppgtt = i915_vm_to_ppgtt(vm); 534 struct sgt_dma iter = sgt_dma(vma); 535 536 if (vma->page_sizes.sg > I915_GTT_PAGE_SIZE) { 537 gen8_ppgtt_insert_huge(vma, &iter, cache_level, flags); 538 } else { 539 u64 idx = vma->node.start >> GEN8_PTE_SHIFT; 540 541 do { 542 struct i915_page_directory * const pdp = 543 gen8_pdp_for_page_index(vm, idx); 544 545 idx = gen8_ppgtt_insert_pte(ppgtt, pdp, &iter, idx, 546 cache_level, flags); 547 } while (idx); 548 549 vma->page_sizes.gtt = I915_GTT_PAGE_SIZE; 550 } 551 } 552 553 static int gen8_init_scratch(struct i915_address_space *vm) 554 { 555 int ret; 556 int i; 557 558 /* 559 * If everybody agrees to not to write into the scratch page, 560 * we can reuse it for all vm, keeping contexts and processes separate. 561 */ 562 if (vm->has_read_only && vm->gt->vm && !i915_is_ggtt(vm->gt->vm)) { 563 struct i915_address_space *clone = vm->gt->vm; 564 565 GEM_BUG_ON(!clone->has_read_only); 566 567 vm->scratch_order = clone->scratch_order; 568 for (i = 0; i <= vm->top; i++) 569 vm->scratch[i] = i915_gem_object_get(clone->scratch[i]); 570 571 return 0; 572 } 573 574 ret = setup_scratch_page(vm); 575 if (ret) 576 return ret; 577 578 vm->scratch[0]->encode = 579 gen8_pte_encode(px_dma(vm->scratch[0]), 580 I915_CACHE_LLC, vm->has_read_only); 581 582 for (i = 1; i <= vm->top; i++) { 583 struct drm_i915_gem_object *obj; 584 585 obj = vm->alloc_pt_dma(vm, I915_GTT_PAGE_SIZE_4K); 586 if (IS_ERR(obj)) 587 goto free_scratch; 588 589 ret = pin_pt_dma(vm, obj); 590 if (ret) { 591 i915_gem_object_put(obj); 592 goto free_scratch; 593 } 594 595 fill_px(obj, vm->scratch[i - 1]->encode); 596 obj->encode = gen8_pde_encode(px_dma(obj), I915_CACHE_LLC); 597 598 vm->scratch[i] = obj; 599 } 600 601 return 0; 602 603 free_scratch: 604 while (i--) 605 i915_gem_object_put(vm->scratch[i]); 606 return -ENOMEM; 607 } 608 609 static int gen8_preallocate_top_level_pdp(struct i915_ppgtt *ppgtt) 610 { 611 struct i915_address_space *vm = &ppgtt->vm; 612 struct i915_page_directory *pd = ppgtt->pd; 613 unsigned int idx; 614 615 GEM_BUG_ON(vm->top != 2); 616 GEM_BUG_ON(gen8_pd_top_count(vm) != GEN8_3LVL_PDPES); 617 618 for (idx = 0; idx < GEN8_3LVL_PDPES; idx++) { 619 struct i915_page_directory *pde; 620 int err; 621 622 pde = alloc_pd(vm); 623 if (IS_ERR(pde)) 624 return PTR_ERR(pde); 625 626 err = pin_pt_dma(vm, pde->pt.base); 627 if (err) { 628 free_pd(vm, pde); 629 return err; 630 } 631 632 fill_px(pde, vm->scratch[1]->encode); 633 set_pd_entry(pd, idx, pde); 634 atomic_inc(px_used(pde)); /* keep pinned */ 635 } 636 wmb(); 637 638 return 0; 639 } 640 641 static struct i915_page_directory * 642 gen8_alloc_top_pd(struct i915_address_space *vm) 643 { 644 const unsigned int count = gen8_pd_top_count(vm); 645 struct i915_page_directory *pd; 646 int err; 647 648 GEM_BUG_ON(count > I915_PDES); 649 650 pd = __alloc_pd(count); 651 if (unlikely(!pd)) 652 return ERR_PTR(-ENOMEM); 653 654 pd->pt.base = vm->alloc_pt_dma(vm, I915_GTT_PAGE_SIZE_4K); 655 if (IS_ERR(pd->pt.base)) { 656 err = PTR_ERR(pd->pt.base); 657 pd->pt.base = NULL; 658 goto err_pd; 659 } 660 661 err = pin_pt_dma(vm, pd->pt.base); 662 if (err) 663 goto err_pd; 664 665 fill_page_dma(px_base(pd), vm->scratch[vm->top]->encode, count); 666 atomic_inc(px_used(pd)); /* mark as pinned */ 667 return pd; 668 669 err_pd: 670 free_pd(vm, pd); 671 return ERR_PTR(err); 672 } 673 674 /* 675 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers 676 * with a net effect resembling a 2-level page table in normal x86 terms. Each 677 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address 678 * space. 679 * 680 */ 681 struct i915_ppgtt *gen8_ppgtt_create(struct intel_gt *gt) 682 { 683 struct i915_ppgtt *ppgtt; 684 int err; 685 686 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL); 687 if (!ppgtt) 688 return ERR_PTR(-ENOMEM); 689 690 ppgtt_init(ppgtt, gt); 691 ppgtt->vm.top = i915_vm_is_4lvl(&ppgtt->vm) ? 3 : 2; 692 ppgtt->vm.pd_shift = ilog2(SZ_4K * SZ_4K / sizeof(gen8_pte_t)); 693 694 /* 695 * From bdw, there is hw support for read-only pages in the PPGTT. 696 * 697 * Gen11 has HSDES#:1807136187 unresolved. Disable ro support 698 * for now. 699 * 700 * Gen12 has inherited the same read-only fault issue from gen11. 701 */ 702 ppgtt->vm.has_read_only = !IS_GEN_RANGE(gt->i915, 11, 12); 703 704 ppgtt->vm.alloc_pt_dma = alloc_pt_dma; 705 706 err = gen8_init_scratch(&ppgtt->vm); 707 if (err) 708 goto err_free; 709 710 ppgtt->pd = gen8_alloc_top_pd(&ppgtt->vm); 711 if (IS_ERR(ppgtt->pd)) { 712 err = PTR_ERR(ppgtt->pd); 713 goto err_free_scratch; 714 } 715 716 if (!i915_vm_is_4lvl(&ppgtt->vm)) { 717 err = gen8_preallocate_top_level_pdp(ppgtt); 718 if (err) 719 goto err_free_pd; 720 } 721 722 ppgtt->vm.bind_async_flags = I915_VMA_LOCAL_BIND; 723 ppgtt->vm.insert_entries = gen8_ppgtt_insert; 724 ppgtt->vm.allocate_va_range = gen8_ppgtt_alloc; 725 ppgtt->vm.clear_range = gen8_ppgtt_clear; 726 727 ppgtt->vm.pte_encode = gen8_pte_encode; 728 729 if (intel_vgpu_active(gt->i915)) 730 gen8_ppgtt_notify_vgt(ppgtt, true); 731 732 ppgtt->vm.cleanup = gen8_ppgtt_cleanup; 733 734 return ppgtt; 735 736 err_free_pd: 737 __gen8_ppgtt_cleanup(&ppgtt->vm, ppgtt->pd, 738 gen8_pd_top_count(&ppgtt->vm), ppgtt->vm.top); 739 err_free_scratch: 740 free_scratch(&ppgtt->vm); 741 err_free: 742 kfree(ppgtt); 743 return ERR_PTR(err); 744 } 745