1 // SPDX-License-Identifier: MIT 2 /* 3 * Copyright © 2020 Intel Corporation 4 */ 5 6 #include <linux/log2.h> 7 8 #include "gem/i915_gem_lmem.h" 9 10 #include "gen8_ppgtt.h" 11 #include "i915_scatterlist.h" 12 #include "i915_trace.h" 13 #include "i915_pvinfo.h" 14 #include "i915_vgpu.h" 15 #include "intel_gt.h" 16 #include "intel_gtt.h" 17 18 static u64 gen8_pde_encode(const dma_addr_t addr, 19 const enum i915_cache_level level) 20 { 21 u64 pde = addr | GEN8_PAGE_PRESENT | GEN8_PAGE_RW; 22 23 if (level != I915_CACHE_NONE) 24 pde |= PPAT_CACHED_PDE; 25 else 26 pde |= PPAT_UNCACHED; 27 28 return pde; 29 } 30 31 static u64 gen8_pte_encode(dma_addr_t addr, 32 enum i915_cache_level level, 33 u32 flags) 34 { 35 gen8_pte_t pte = addr | GEN8_PAGE_PRESENT | GEN8_PAGE_RW; 36 37 if (unlikely(flags & PTE_READ_ONLY)) 38 pte &= ~GEN8_PAGE_RW; 39 40 if (flags & PTE_LM) 41 pte |= GEN12_PPGTT_PTE_LM; 42 43 switch (level) { 44 case I915_CACHE_NONE: 45 pte |= PPAT_UNCACHED; 46 break; 47 case I915_CACHE_WT: 48 pte |= PPAT_DISPLAY_ELLC; 49 break; 50 default: 51 pte |= PPAT_CACHED; 52 break; 53 } 54 55 return pte; 56 } 57 58 static void gen8_ppgtt_notify_vgt(struct i915_ppgtt *ppgtt, bool create) 59 { 60 struct drm_i915_private *i915 = ppgtt->vm.i915; 61 struct intel_uncore *uncore = ppgtt->vm.gt->uncore; 62 enum vgt_g2v_type msg; 63 int i; 64 65 if (create) 66 atomic_inc(px_used(ppgtt->pd)); /* never remove */ 67 else 68 atomic_dec(px_used(ppgtt->pd)); 69 70 mutex_lock(&i915->vgpu.lock); 71 72 if (i915_vm_is_4lvl(&ppgtt->vm)) { 73 const u64 daddr = px_dma(ppgtt->pd); 74 75 intel_uncore_write(uncore, 76 vgtif_reg(pdp[0].lo), lower_32_bits(daddr)); 77 intel_uncore_write(uncore, 78 vgtif_reg(pdp[0].hi), upper_32_bits(daddr)); 79 80 msg = create ? 81 VGT_G2V_PPGTT_L4_PAGE_TABLE_CREATE : 82 VGT_G2V_PPGTT_L4_PAGE_TABLE_DESTROY; 83 } else { 84 for (i = 0; i < GEN8_3LVL_PDPES; i++) { 85 const u64 daddr = i915_page_dir_dma_addr(ppgtt, i); 86 87 intel_uncore_write(uncore, 88 vgtif_reg(pdp[i].lo), 89 lower_32_bits(daddr)); 90 intel_uncore_write(uncore, 91 vgtif_reg(pdp[i].hi), 92 upper_32_bits(daddr)); 93 } 94 95 msg = create ? 96 VGT_G2V_PPGTT_L3_PAGE_TABLE_CREATE : 97 VGT_G2V_PPGTT_L3_PAGE_TABLE_DESTROY; 98 } 99 100 /* g2v_notify atomically (via hv trap) consumes the message packet. */ 101 intel_uncore_write(uncore, vgtif_reg(g2v_notify), msg); 102 103 mutex_unlock(&i915->vgpu.lock); 104 } 105 106 /* Index shifts into the pagetable are offset by GEN8_PTE_SHIFT [12] */ 107 #define GEN8_PAGE_SIZE (SZ_4K) /* page and page-directory sizes are the same */ 108 #define GEN8_PTE_SHIFT (ilog2(GEN8_PAGE_SIZE)) 109 #define GEN8_PDES (GEN8_PAGE_SIZE / sizeof(u64)) 110 #define gen8_pd_shift(lvl) ((lvl) * ilog2(GEN8_PDES)) 111 #define gen8_pd_index(i, lvl) i915_pde_index((i), gen8_pd_shift(lvl)) 112 #define __gen8_pte_shift(lvl) (GEN8_PTE_SHIFT + gen8_pd_shift(lvl)) 113 #define __gen8_pte_index(a, lvl) i915_pde_index((a), __gen8_pte_shift(lvl)) 114 115 #define as_pd(x) container_of((x), typeof(struct i915_page_directory), pt) 116 117 static unsigned int 118 gen8_pd_range(u64 start, u64 end, int lvl, unsigned int *idx) 119 { 120 const int shift = gen8_pd_shift(lvl); 121 const u64 mask = ~0ull << gen8_pd_shift(lvl + 1); 122 123 GEM_BUG_ON(start >= end); 124 end += ~mask >> gen8_pd_shift(1); 125 126 *idx = i915_pde_index(start, shift); 127 if ((start ^ end) & mask) 128 return GEN8_PDES - *idx; 129 else 130 return i915_pde_index(end, shift) - *idx; 131 } 132 133 static bool gen8_pd_contains(u64 start, u64 end, int lvl) 134 { 135 const u64 mask = ~0ull << gen8_pd_shift(lvl + 1); 136 137 GEM_BUG_ON(start >= end); 138 return (start ^ end) & mask && (start & ~mask) == 0; 139 } 140 141 static unsigned int gen8_pt_count(u64 start, u64 end) 142 { 143 GEM_BUG_ON(start >= end); 144 if ((start ^ end) >> gen8_pd_shift(1)) 145 return GEN8_PDES - (start & (GEN8_PDES - 1)); 146 else 147 return end - start; 148 } 149 150 static unsigned int gen8_pd_top_count(const struct i915_address_space *vm) 151 { 152 unsigned int shift = __gen8_pte_shift(vm->top); 153 154 return (vm->total + (1ull << shift) - 1) >> shift; 155 } 156 157 static struct i915_page_directory * 158 gen8_pdp_for_page_index(struct i915_address_space * const vm, const u64 idx) 159 { 160 struct i915_ppgtt * const ppgtt = i915_vm_to_ppgtt(vm); 161 162 if (vm->top == 2) 163 return ppgtt->pd; 164 else 165 return i915_pd_entry(ppgtt->pd, gen8_pd_index(idx, vm->top)); 166 } 167 168 static struct i915_page_directory * 169 gen8_pdp_for_page_address(struct i915_address_space * const vm, const u64 addr) 170 { 171 return gen8_pdp_for_page_index(vm, addr >> GEN8_PTE_SHIFT); 172 } 173 174 static void __gen8_ppgtt_cleanup(struct i915_address_space *vm, 175 struct i915_page_directory *pd, 176 int count, int lvl) 177 { 178 if (lvl) { 179 void **pde = pd->entry; 180 181 do { 182 if (!*pde) 183 continue; 184 185 __gen8_ppgtt_cleanup(vm, *pde, GEN8_PDES, lvl - 1); 186 } while (pde++, --count); 187 } 188 189 free_px(vm, &pd->pt, lvl); 190 } 191 192 static void gen8_ppgtt_cleanup(struct i915_address_space *vm) 193 { 194 struct i915_ppgtt *ppgtt = i915_vm_to_ppgtt(vm); 195 196 if (intel_vgpu_active(vm->i915)) 197 gen8_ppgtt_notify_vgt(ppgtt, false); 198 199 if (ppgtt->pd) 200 __gen8_ppgtt_cleanup(vm, ppgtt->pd, 201 gen8_pd_top_count(vm), vm->top); 202 203 free_scratch(vm); 204 } 205 206 static u64 __gen8_ppgtt_clear(struct i915_address_space * const vm, 207 struct i915_page_directory * const pd, 208 u64 start, const u64 end, int lvl) 209 { 210 const struct drm_i915_gem_object * const scratch = vm->scratch[lvl]; 211 unsigned int idx, len; 212 213 GEM_BUG_ON(end > vm->total >> GEN8_PTE_SHIFT); 214 215 len = gen8_pd_range(start, end, lvl--, &idx); 216 DBG("%s(%p):{ lvl:%d, start:%llx, end:%llx, idx:%d, len:%d, used:%d }\n", 217 __func__, vm, lvl + 1, start, end, 218 idx, len, atomic_read(px_used(pd))); 219 GEM_BUG_ON(!len || len >= atomic_read(px_used(pd))); 220 221 do { 222 struct i915_page_table *pt = pd->entry[idx]; 223 224 if (atomic_fetch_inc(&pt->used) >> gen8_pd_shift(1) && 225 gen8_pd_contains(start, end, lvl)) { 226 DBG("%s(%p):{ lvl:%d, idx:%d, start:%llx, end:%llx } removing pd\n", 227 __func__, vm, lvl + 1, idx, start, end); 228 clear_pd_entry(pd, idx, scratch); 229 __gen8_ppgtt_cleanup(vm, as_pd(pt), I915_PDES, lvl); 230 start += (u64)I915_PDES << gen8_pd_shift(lvl); 231 continue; 232 } 233 234 if (lvl) { 235 start = __gen8_ppgtt_clear(vm, as_pd(pt), 236 start, end, lvl); 237 } else { 238 unsigned int count; 239 unsigned int pte = gen8_pd_index(start, 0); 240 unsigned int num_ptes; 241 u64 *vaddr; 242 243 count = gen8_pt_count(start, end); 244 DBG("%s(%p):{ lvl:%d, start:%llx, end:%llx, idx:%d, len:%d, used:%d } removing pte\n", 245 __func__, vm, lvl, start, end, 246 gen8_pd_index(start, 0), count, 247 atomic_read(&pt->used)); 248 GEM_BUG_ON(!count || count >= atomic_read(&pt->used)); 249 250 num_ptes = count; 251 if (pt->is_compact) { 252 GEM_BUG_ON(num_ptes % 16); 253 GEM_BUG_ON(pte % 16); 254 num_ptes /= 16; 255 pte /= 16; 256 } 257 258 vaddr = px_vaddr(pt); 259 memset64(vaddr + pte, 260 vm->scratch[0]->encode, 261 num_ptes); 262 263 atomic_sub(count, &pt->used); 264 start += count; 265 } 266 267 if (release_pd_entry(pd, idx, pt, scratch)) 268 free_px(vm, pt, lvl); 269 } while (idx++, --len); 270 271 return start; 272 } 273 274 static void gen8_ppgtt_clear(struct i915_address_space *vm, 275 u64 start, u64 length) 276 { 277 GEM_BUG_ON(!IS_ALIGNED(start, BIT_ULL(GEN8_PTE_SHIFT))); 278 GEM_BUG_ON(!IS_ALIGNED(length, BIT_ULL(GEN8_PTE_SHIFT))); 279 GEM_BUG_ON(range_overflows(start, length, vm->total)); 280 281 start >>= GEN8_PTE_SHIFT; 282 length >>= GEN8_PTE_SHIFT; 283 GEM_BUG_ON(length == 0); 284 285 __gen8_ppgtt_clear(vm, i915_vm_to_ppgtt(vm)->pd, 286 start, start + length, vm->top); 287 } 288 289 static void __gen8_ppgtt_alloc(struct i915_address_space * const vm, 290 struct i915_vm_pt_stash *stash, 291 struct i915_page_directory * const pd, 292 u64 * const start, const u64 end, int lvl) 293 { 294 unsigned int idx, len; 295 296 GEM_BUG_ON(end > vm->total >> GEN8_PTE_SHIFT); 297 298 len = gen8_pd_range(*start, end, lvl--, &idx); 299 DBG("%s(%p):{ lvl:%d, start:%llx, end:%llx, idx:%d, len:%d, used:%d }\n", 300 __func__, vm, lvl + 1, *start, end, 301 idx, len, atomic_read(px_used(pd))); 302 GEM_BUG_ON(!len || (idx + len - 1) >> gen8_pd_shift(1)); 303 304 spin_lock(&pd->lock); 305 GEM_BUG_ON(!atomic_read(px_used(pd))); /* Must be pinned! */ 306 do { 307 struct i915_page_table *pt = pd->entry[idx]; 308 309 if (!pt) { 310 spin_unlock(&pd->lock); 311 312 DBG("%s(%p):{ lvl:%d, idx:%d } allocating new tree\n", 313 __func__, vm, lvl + 1, idx); 314 315 pt = stash->pt[!!lvl]; 316 __i915_gem_object_pin_pages(pt->base); 317 318 fill_px(pt, vm->scratch[lvl]->encode); 319 320 spin_lock(&pd->lock); 321 if (likely(!pd->entry[idx])) { 322 stash->pt[!!lvl] = pt->stash; 323 atomic_set(&pt->used, 0); 324 set_pd_entry(pd, idx, pt); 325 } else { 326 pt = pd->entry[idx]; 327 } 328 } 329 330 if (lvl) { 331 atomic_inc(&pt->used); 332 spin_unlock(&pd->lock); 333 334 __gen8_ppgtt_alloc(vm, stash, 335 as_pd(pt), start, end, lvl); 336 337 spin_lock(&pd->lock); 338 atomic_dec(&pt->used); 339 GEM_BUG_ON(!atomic_read(&pt->used)); 340 } else { 341 unsigned int count = gen8_pt_count(*start, end); 342 343 DBG("%s(%p):{ lvl:%d, start:%llx, end:%llx, idx:%d, len:%d, used:%d } inserting pte\n", 344 __func__, vm, lvl, *start, end, 345 gen8_pd_index(*start, 0), count, 346 atomic_read(&pt->used)); 347 348 atomic_add(count, &pt->used); 349 /* All other pdes may be simultaneously removed */ 350 GEM_BUG_ON(atomic_read(&pt->used) > NALLOC * I915_PDES); 351 *start += count; 352 } 353 } while (idx++, --len); 354 spin_unlock(&pd->lock); 355 } 356 357 static void gen8_ppgtt_alloc(struct i915_address_space *vm, 358 struct i915_vm_pt_stash *stash, 359 u64 start, u64 length) 360 { 361 GEM_BUG_ON(!IS_ALIGNED(start, BIT_ULL(GEN8_PTE_SHIFT))); 362 GEM_BUG_ON(!IS_ALIGNED(length, BIT_ULL(GEN8_PTE_SHIFT))); 363 GEM_BUG_ON(range_overflows(start, length, vm->total)); 364 365 start >>= GEN8_PTE_SHIFT; 366 length >>= GEN8_PTE_SHIFT; 367 GEM_BUG_ON(length == 0); 368 369 __gen8_ppgtt_alloc(vm, stash, i915_vm_to_ppgtt(vm)->pd, 370 &start, start + length, vm->top); 371 } 372 373 static void __gen8_ppgtt_foreach(struct i915_address_space *vm, 374 struct i915_page_directory *pd, 375 u64 *start, u64 end, int lvl, 376 void (*fn)(struct i915_address_space *vm, 377 struct i915_page_table *pt, 378 void *data), 379 void *data) 380 { 381 unsigned int idx, len; 382 383 len = gen8_pd_range(*start, end, lvl--, &idx); 384 385 spin_lock(&pd->lock); 386 do { 387 struct i915_page_table *pt = pd->entry[idx]; 388 389 atomic_inc(&pt->used); 390 spin_unlock(&pd->lock); 391 392 if (lvl) { 393 __gen8_ppgtt_foreach(vm, as_pd(pt), start, end, lvl, 394 fn, data); 395 } else { 396 fn(vm, pt, data); 397 *start += gen8_pt_count(*start, end); 398 } 399 400 spin_lock(&pd->lock); 401 atomic_dec(&pt->used); 402 } while (idx++, --len); 403 spin_unlock(&pd->lock); 404 } 405 406 static void gen8_ppgtt_foreach(struct i915_address_space *vm, 407 u64 start, u64 length, 408 void (*fn)(struct i915_address_space *vm, 409 struct i915_page_table *pt, 410 void *data), 411 void *data) 412 { 413 start >>= GEN8_PTE_SHIFT; 414 length >>= GEN8_PTE_SHIFT; 415 416 __gen8_ppgtt_foreach(vm, i915_vm_to_ppgtt(vm)->pd, 417 &start, start + length, vm->top, 418 fn, data); 419 } 420 421 static __always_inline u64 422 gen8_ppgtt_insert_pte(struct i915_ppgtt *ppgtt, 423 struct i915_page_directory *pdp, 424 struct sgt_dma *iter, 425 u64 idx, 426 enum i915_cache_level cache_level, 427 u32 flags) 428 { 429 struct i915_page_directory *pd; 430 const gen8_pte_t pte_encode = gen8_pte_encode(0, cache_level, flags); 431 gen8_pte_t *vaddr; 432 433 pd = i915_pd_entry(pdp, gen8_pd_index(idx, 2)); 434 vaddr = px_vaddr(i915_pt_entry(pd, gen8_pd_index(idx, 1))); 435 do { 436 GEM_BUG_ON(sg_dma_len(iter->sg) < I915_GTT_PAGE_SIZE); 437 vaddr[gen8_pd_index(idx, 0)] = pte_encode | iter->dma; 438 439 iter->dma += I915_GTT_PAGE_SIZE; 440 if (iter->dma >= iter->max) { 441 iter->sg = __sg_next(iter->sg); 442 if (!iter->sg || sg_dma_len(iter->sg) == 0) { 443 idx = 0; 444 break; 445 } 446 447 iter->dma = sg_dma_address(iter->sg); 448 iter->max = iter->dma + sg_dma_len(iter->sg); 449 } 450 451 if (gen8_pd_index(++idx, 0) == 0) { 452 if (gen8_pd_index(idx, 1) == 0) { 453 /* Limited by sg length for 3lvl */ 454 if (gen8_pd_index(idx, 2) == 0) 455 break; 456 457 pd = pdp->entry[gen8_pd_index(idx, 2)]; 458 } 459 460 drm_clflush_virt_range(vaddr, PAGE_SIZE); 461 vaddr = px_vaddr(i915_pt_entry(pd, gen8_pd_index(idx, 1))); 462 } 463 } while (1); 464 drm_clflush_virt_range(vaddr, PAGE_SIZE); 465 466 return idx; 467 } 468 469 static void 470 xehpsdv_ppgtt_insert_huge(struct i915_address_space *vm, 471 struct i915_vma_resource *vma_res, 472 struct sgt_dma *iter, 473 enum i915_cache_level cache_level, 474 u32 flags) 475 { 476 const gen8_pte_t pte_encode = vm->pte_encode(0, cache_level, flags); 477 unsigned int rem = sg_dma_len(iter->sg); 478 u64 start = vma_res->start; 479 u64 end = start + vma_res->vma_size; 480 481 GEM_BUG_ON(!i915_vm_is_4lvl(vm)); 482 483 do { 484 struct i915_page_directory * const pdp = 485 gen8_pdp_for_page_address(vm, start); 486 struct i915_page_directory * const pd = 487 i915_pd_entry(pdp, __gen8_pte_index(start, 2)); 488 struct i915_page_table *pt = 489 i915_pt_entry(pd, __gen8_pte_index(start, 1)); 490 gen8_pte_t encode = pte_encode; 491 unsigned int page_size; 492 gen8_pte_t *vaddr; 493 u16 index, max, nent, i; 494 495 max = I915_PDES; 496 nent = 1; 497 498 if (vma_res->bi.page_sizes.sg & I915_GTT_PAGE_SIZE_2M && 499 IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_2M) && 500 rem >= I915_GTT_PAGE_SIZE_2M && 501 !__gen8_pte_index(start, 0)) { 502 index = __gen8_pte_index(start, 1); 503 encode |= GEN8_PDE_PS_2M; 504 page_size = I915_GTT_PAGE_SIZE_2M; 505 506 vaddr = px_vaddr(pd); 507 } else { 508 index = __gen8_pte_index(start, 0); 509 page_size = I915_GTT_PAGE_SIZE; 510 511 if (vma_res->bi.page_sizes.sg & I915_GTT_PAGE_SIZE_64K) { 512 /* 513 * Device local-memory on these platforms should 514 * always use 64K pages or larger (including GTT 515 * alignment), therefore if we know the whole 516 * page-table needs to be filled we can always 517 * safely use the compact-layout. Otherwise fall 518 * back to the TLB hint with PS64. If this is 519 * system memory we only bother with PS64. 520 */ 521 if ((encode & GEN12_PPGTT_PTE_LM) && 522 end - start >= SZ_2M && !index) { 523 index = __gen8_pte_index(start, 0) / 16; 524 page_size = I915_GTT_PAGE_SIZE_64K; 525 526 max /= 16; 527 528 vaddr = px_vaddr(pd); 529 vaddr[__gen8_pte_index(start, 1)] |= GEN12_PDE_64K; 530 531 pt->is_compact = true; 532 } else if (IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_64K) && 533 rem >= I915_GTT_PAGE_SIZE_64K && 534 !(index % 16)) { 535 encode |= GEN12_PTE_PS64; 536 page_size = I915_GTT_PAGE_SIZE_64K; 537 nent = 16; 538 } 539 } 540 541 vaddr = px_vaddr(pt); 542 } 543 544 do { 545 GEM_BUG_ON(rem < page_size); 546 547 for (i = 0; i < nent; i++) { 548 vaddr[index++] = 549 encode | (iter->dma + i * 550 I915_GTT_PAGE_SIZE); 551 } 552 553 start += page_size; 554 iter->dma += page_size; 555 rem -= page_size; 556 if (iter->dma >= iter->max) { 557 iter->sg = __sg_next(iter->sg); 558 if (!iter->sg) 559 break; 560 561 rem = sg_dma_len(iter->sg); 562 if (!rem) 563 break; 564 565 iter->dma = sg_dma_address(iter->sg); 566 iter->max = iter->dma + rem; 567 568 if (unlikely(!IS_ALIGNED(iter->dma, page_size))) 569 break; 570 } 571 } while (rem >= page_size && index < max); 572 573 vma_res->page_sizes_gtt |= page_size; 574 } while (iter->sg && sg_dma_len(iter->sg)); 575 } 576 577 static void gen8_ppgtt_insert_huge(struct i915_address_space *vm, 578 struct i915_vma_resource *vma_res, 579 struct sgt_dma *iter, 580 enum i915_cache_level cache_level, 581 u32 flags) 582 { 583 const gen8_pte_t pte_encode = gen8_pte_encode(0, cache_level, flags); 584 unsigned int rem = sg_dma_len(iter->sg); 585 u64 start = vma_res->start; 586 587 GEM_BUG_ON(!i915_vm_is_4lvl(vm)); 588 589 do { 590 struct i915_page_directory * const pdp = 591 gen8_pdp_for_page_address(vm, start); 592 struct i915_page_directory * const pd = 593 i915_pd_entry(pdp, __gen8_pte_index(start, 2)); 594 gen8_pte_t encode = pte_encode; 595 unsigned int maybe_64K = -1; 596 unsigned int page_size; 597 gen8_pte_t *vaddr; 598 u16 index; 599 600 if (vma_res->bi.page_sizes.sg & I915_GTT_PAGE_SIZE_2M && 601 IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_2M) && 602 rem >= I915_GTT_PAGE_SIZE_2M && 603 !__gen8_pte_index(start, 0)) { 604 index = __gen8_pte_index(start, 1); 605 encode |= GEN8_PDE_PS_2M; 606 page_size = I915_GTT_PAGE_SIZE_2M; 607 608 vaddr = px_vaddr(pd); 609 } else { 610 struct i915_page_table *pt = 611 i915_pt_entry(pd, __gen8_pte_index(start, 1)); 612 613 index = __gen8_pte_index(start, 0); 614 page_size = I915_GTT_PAGE_SIZE; 615 616 if (!index && 617 vma_res->bi.page_sizes.sg & I915_GTT_PAGE_SIZE_64K && 618 IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_64K) && 619 (IS_ALIGNED(rem, I915_GTT_PAGE_SIZE_64K) || 620 rem >= (I915_PDES - index) * I915_GTT_PAGE_SIZE)) 621 maybe_64K = __gen8_pte_index(start, 1); 622 623 vaddr = px_vaddr(pt); 624 } 625 626 do { 627 GEM_BUG_ON(sg_dma_len(iter->sg) < page_size); 628 vaddr[index++] = encode | iter->dma; 629 630 start += page_size; 631 iter->dma += page_size; 632 rem -= page_size; 633 if (iter->dma >= iter->max) { 634 iter->sg = __sg_next(iter->sg); 635 if (!iter->sg) 636 break; 637 638 rem = sg_dma_len(iter->sg); 639 if (!rem) 640 break; 641 642 iter->dma = sg_dma_address(iter->sg); 643 iter->max = iter->dma + rem; 644 645 if (maybe_64K != -1 && index < I915_PDES && 646 !(IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_64K) && 647 (IS_ALIGNED(rem, I915_GTT_PAGE_SIZE_64K) || 648 rem >= (I915_PDES - index) * I915_GTT_PAGE_SIZE))) 649 maybe_64K = -1; 650 651 if (unlikely(!IS_ALIGNED(iter->dma, page_size))) 652 break; 653 } 654 } while (rem >= page_size && index < I915_PDES); 655 656 drm_clflush_virt_range(vaddr, PAGE_SIZE); 657 658 /* 659 * Is it safe to mark the 2M block as 64K? -- Either we have 660 * filled whole page-table with 64K entries, or filled part of 661 * it and have reached the end of the sg table and we have 662 * enough padding. 663 */ 664 if (maybe_64K != -1 && 665 (index == I915_PDES || 666 (i915_vm_has_scratch_64K(vm) && 667 !iter->sg && IS_ALIGNED(vma_res->start + 668 vma_res->node_size, 669 I915_GTT_PAGE_SIZE_2M)))) { 670 vaddr = px_vaddr(pd); 671 vaddr[maybe_64K] |= GEN8_PDE_IPS_64K; 672 drm_clflush_virt_range(vaddr, PAGE_SIZE); 673 page_size = I915_GTT_PAGE_SIZE_64K; 674 675 /* 676 * We write all 4K page entries, even when using 64K 677 * pages. In order to verify that the HW isn't cheating 678 * by using the 4K PTE instead of the 64K PTE, we want 679 * to remove all the surplus entries. If the HW skipped 680 * the 64K PTE, it will read/write into the scratch page 681 * instead - which we detect as missing results during 682 * selftests. 683 */ 684 if (I915_SELFTEST_ONLY(vm->scrub_64K)) { 685 u16 i; 686 687 encode = vm->scratch[0]->encode; 688 vaddr = px_vaddr(i915_pt_entry(pd, maybe_64K)); 689 690 for (i = 1; i < index; i += 16) 691 memset64(vaddr + i, encode, 15); 692 693 drm_clflush_virt_range(vaddr, PAGE_SIZE); 694 } 695 } 696 697 vma_res->page_sizes_gtt |= page_size; 698 } while (iter->sg && sg_dma_len(iter->sg)); 699 } 700 701 static void gen8_ppgtt_insert(struct i915_address_space *vm, 702 struct i915_vma_resource *vma_res, 703 enum i915_cache_level cache_level, 704 u32 flags) 705 { 706 struct i915_ppgtt * const ppgtt = i915_vm_to_ppgtt(vm); 707 struct sgt_dma iter = sgt_dma(vma_res); 708 709 if (vma_res->bi.page_sizes.sg > I915_GTT_PAGE_SIZE) { 710 if (HAS_64K_PAGES(vm->i915)) 711 xehpsdv_ppgtt_insert_huge(vm, vma_res, &iter, cache_level, flags); 712 else 713 gen8_ppgtt_insert_huge(vm, vma_res, &iter, cache_level, flags); 714 } else { 715 u64 idx = vma_res->start >> GEN8_PTE_SHIFT; 716 717 do { 718 struct i915_page_directory * const pdp = 719 gen8_pdp_for_page_index(vm, idx); 720 721 idx = gen8_ppgtt_insert_pte(ppgtt, pdp, &iter, idx, 722 cache_level, flags); 723 } while (idx); 724 725 vma_res->page_sizes_gtt = I915_GTT_PAGE_SIZE; 726 } 727 } 728 729 static void gen8_ppgtt_insert_entry(struct i915_address_space *vm, 730 dma_addr_t addr, 731 u64 offset, 732 enum i915_cache_level level, 733 u32 flags) 734 { 735 u64 idx = offset >> GEN8_PTE_SHIFT; 736 struct i915_page_directory * const pdp = 737 gen8_pdp_for_page_index(vm, idx); 738 struct i915_page_directory *pd = 739 i915_pd_entry(pdp, gen8_pd_index(idx, 2)); 740 struct i915_page_table *pt = i915_pt_entry(pd, gen8_pd_index(idx, 1)); 741 gen8_pte_t *vaddr; 742 743 GEM_BUG_ON(pt->is_compact); 744 745 vaddr = px_vaddr(pt); 746 vaddr[gen8_pd_index(idx, 0)] = gen8_pte_encode(addr, level, flags); 747 drm_clflush_virt_range(&vaddr[gen8_pd_index(idx, 0)], sizeof(*vaddr)); 748 } 749 750 static void __xehpsdv_ppgtt_insert_entry_lm(struct i915_address_space *vm, 751 dma_addr_t addr, 752 u64 offset, 753 enum i915_cache_level level, 754 u32 flags) 755 { 756 u64 idx = offset >> GEN8_PTE_SHIFT; 757 struct i915_page_directory * const pdp = 758 gen8_pdp_for_page_index(vm, idx); 759 struct i915_page_directory *pd = 760 i915_pd_entry(pdp, gen8_pd_index(idx, 2)); 761 struct i915_page_table *pt = i915_pt_entry(pd, gen8_pd_index(idx, 1)); 762 gen8_pte_t *vaddr; 763 764 GEM_BUG_ON(!IS_ALIGNED(addr, SZ_64K)); 765 GEM_BUG_ON(!IS_ALIGNED(offset, SZ_64K)); 766 767 /* XXX: we don't strictly need to use this layout */ 768 769 if (!pt->is_compact) { 770 vaddr = px_vaddr(pd); 771 vaddr[gen8_pd_index(idx, 1)] |= GEN12_PDE_64K; 772 pt->is_compact = true; 773 } 774 775 vaddr = px_vaddr(pt); 776 vaddr[gen8_pd_index(idx, 0) / 16] = gen8_pte_encode(addr, level, flags); 777 } 778 779 static void xehpsdv_ppgtt_insert_entry(struct i915_address_space *vm, 780 dma_addr_t addr, 781 u64 offset, 782 enum i915_cache_level level, 783 u32 flags) 784 { 785 if (flags & PTE_LM) 786 return __xehpsdv_ppgtt_insert_entry_lm(vm, addr, offset, 787 level, flags); 788 789 return gen8_ppgtt_insert_entry(vm, addr, offset, level, flags); 790 } 791 792 static int gen8_init_scratch(struct i915_address_space *vm) 793 { 794 u32 pte_flags; 795 int ret; 796 int i; 797 798 /* 799 * If everybody agrees to not to write into the scratch page, 800 * we can reuse it for all vm, keeping contexts and processes separate. 801 */ 802 if (vm->has_read_only && vm->gt->vm && !i915_is_ggtt(vm->gt->vm)) { 803 struct i915_address_space *clone = vm->gt->vm; 804 805 GEM_BUG_ON(!clone->has_read_only); 806 807 vm->scratch_order = clone->scratch_order; 808 for (i = 0; i <= vm->top; i++) 809 vm->scratch[i] = i915_gem_object_get(clone->scratch[i]); 810 811 return 0; 812 } 813 814 ret = setup_scratch_page(vm); 815 if (ret) 816 return ret; 817 818 pte_flags = vm->has_read_only; 819 if (i915_gem_object_is_lmem(vm->scratch[0])) 820 pte_flags |= PTE_LM; 821 822 vm->scratch[0]->encode = 823 gen8_pte_encode(px_dma(vm->scratch[0]), 824 I915_CACHE_NONE, pte_flags); 825 826 for (i = 1; i <= vm->top; i++) { 827 struct drm_i915_gem_object *obj; 828 829 obj = vm->alloc_pt_dma(vm, I915_GTT_PAGE_SIZE_4K); 830 if (IS_ERR(obj)) { 831 ret = PTR_ERR(obj); 832 goto free_scratch; 833 } 834 835 ret = map_pt_dma(vm, obj); 836 if (ret) { 837 i915_gem_object_put(obj); 838 goto free_scratch; 839 } 840 841 fill_px(obj, vm->scratch[i - 1]->encode); 842 obj->encode = gen8_pde_encode(px_dma(obj), I915_CACHE_NONE); 843 844 vm->scratch[i] = obj; 845 } 846 847 return 0; 848 849 free_scratch: 850 while (i--) 851 i915_gem_object_put(vm->scratch[i]); 852 vm->scratch[0] = NULL; 853 return ret; 854 } 855 856 static int gen8_preallocate_top_level_pdp(struct i915_ppgtt *ppgtt) 857 { 858 struct i915_address_space *vm = &ppgtt->vm; 859 struct i915_page_directory *pd = ppgtt->pd; 860 unsigned int idx; 861 862 GEM_BUG_ON(vm->top != 2); 863 GEM_BUG_ON(gen8_pd_top_count(vm) != GEN8_3LVL_PDPES); 864 865 for (idx = 0; idx < GEN8_3LVL_PDPES; idx++) { 866 struct i915_page_directory *pde; 867 int err; 868 869 pde = alloc_pd(vm); 870 if (IS_ERR(pde)) 871 return PTR_ERR(pde); 872 873 err = map_pt_dma(vm, pde->pt.base); 874 if (err) { 875 free_pd(vm, pde); 876 return err; 877 } 878 879 fill_px(pde, vm->scratch[1]->encode); 880 set_pd_entry(pd, idx, pde); 881 atomic_inc(px_used(pde)); /* keep pinned */ 882 } 883 wmb(); 884 885 return 0; 886 } 887 888 static struct i915_page_directory * 889 gen8_alloc_top_pd(struct i915_address_space *vm) 890 { 891 const unsigned int count = gen8_pd_top_count(vm); 892 struct i915_page_directory *pd; 893 int err; 894 895 GEM_BUG_ON(count > I915_PDES); 896 897 pd = __alloc_pd(count); 898 if (unlikely(!pd)) 899 return ERR_PTR(-ENOMEM); 900 901 pd->pt.base = vm->alloc_pt_dma(vm, I915_GTT_PAGE_SIZE_4K); 902 if (IS_ERR(pd->pt.base)) { 903 err = PTR_ERR(pd->pt.base); 904 pd->pt.base = NULL; 905 goto err_pd; 906 } 907 908 err = map_pt_dma(vm, pd->pt.base); 909 if (err) 910 goto err_pd; 911 912 fill_page_dma(px_base(pd), vm->scratch[vm->top]->encode, count); 913 atomic_inc(px_used(pd)); /* mark as pinned */ 914 return pd; 915 916 err_pd: 917 free_pd(vm, pd); 918 return ERR_PTR(err); 919 } 920 921 /* 922 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers 923 * with a net effect resembling a 2-level page table in normal x86 terms. Each 924 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address 925 * space. 926 * 927 */ 928 struct i915_ppgtt *gen8_ppgtt_create(struct intel_gt *gt, 929 unsigned long lmem_pt_obj_flags) 930 { 931 struct i915_page_directory *pd; 932 struct i915_ppgtt *ppgtt; 933 int err; 934 935 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL); 936 if (!ppgtt) 937 return ERR_PTR(-ENOMEM); 938 939 ppgtt_init(ppgtt, gt, lmem_pt_obj_flags); 940 ppgtt->vm.top = i915_vm_is_4lvl(&ppgtt->vm) ? 3 : 2; 941 ppgtt->vm.pd_shift = ilog2(SZ_4K * SZ_4K / sizeof(gen8_pte_t)); 942 943 /* 944 * From bdw, there is hw support for read-only pages in the PPGTT. 945 * 946 * Gen11 has HSDES#:1807136187 unresolved. Disable ro support 947 * for now. 948 * 949 * Gen12 has inherited the same read-only fault issue from gen11. 950 */ 951 ppgtt->vm.has_read_only = !IS_GRAPHICS_VER(gt->i915, 11, 12); 952 953 if (HAS_LMEM(gt->i915)) 954 ppgtt->vm.alloc_pt_dma = alloc_pt_lmem; 955 else 956 ppgtt->vm.alloc_pt_dma = alloc_pt_dma; 957 958 /* 959 * Using SMEM here instead of LMEM has the advantage of not reserving 960 * high performance memory for a "never" used filler page. It also 961 * removes the device access that would be required to initialise the 962 * scratch page, reducing pressure on an even scarcer resource. 963 */ 964 ppgtt->vm.alloc_scratch_dma = alloc_pt_dma; 965 966 ppgtt->vm.pte_encode = gen8_pte_encode; 967 968 ppgtt->vm.bind_async_flags = I915_VMA_LOCAL_BIND; 969 ppgtt->vm.insert_entries = gen8_ppgtt_insert; 970 if (HAS_64K_PAGES(gt->i915)) 971 ppgtt->vm.insert_page = xehpsdv_ppgtt_insert_entry; 972 else 973 ppgtt->vm.insert_page = gen8_ppgtt_insert_entry; 974 ppgtt->vm.allocate_va_range = gen8_ppgtt_alloc; 975 ppgtt->vm.clear_range = gen8_ppgtt_clear; 976 ppgtt->vm.foreach = gen8_ppgtt_foreach; 977 ppgtt->vm.cleanup = gen8_ppgtt_cleanup; 978 979 err = gen8_init_scratch(&ppgtt->vm); 980 if (err) 981 goto err_put; 982 983 pd = gen8_alloc_top_pd(&ppgtt->vm); 984 if (IS_ERR(pd)) { 985 err = PTR_ERR(pd); 986 goto err_put; 987 } 988 ppgtt->pd = pd; 989 990 if (!i915_vm_is_4lvl(&ppgtt->vm)) { 991 err = gen8_preallocate_top_level_pdp(ppgtt); 992 if (err) 993 goto err_put; 994 } 995 996 if (intel_vgpu_active(gt->i915)) 997 gen8_ppgtt_notify_vgt(ppgtt, true); 998 999 return ppgtt; 1000 1001 err_put: 1002 i915_vm_put(&ppgtt->vm); 1003 return ERR_PTR(err); 1004 } 1005