1 /* 2 * Copyright © 2010 Daniel Vetter 3 * Copyright © 2011-2014 Intel Corporation 4 * 5 * Permission is hereby granted, free of charge, to any person obtaining a 6 * copy of this software and associated documentation files (the "Software"), 7 * to deal in the Software without restriction, including without limitation 8 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 9 * and/or sell copies of the Software, and to permit persons to whom the 10 * Software is furnished to do so, subject to the following conditions: 11 * 12 * The above copyright notice and this permission notice (including the next 13 * paragraph) shall be included in all copies or substantial portions of the 14 * Software. 15 * 16 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 17 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 18 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 19 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 20 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING 21 * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS 22 * IN THE SOFTWARE. 23 * 24 */ 25 26 #include <linux/seq_file.h> 27 #include <drm/drmP.h> 28 #include <drm/i915_drm.h> 29 #include "i915_drv.h" 30 #include "i915_vgpu.h" 31 #include "i915_trace.h" 32 #include "intel_drv.h" 33 34 /** 35 * DOC: Global GTT views 36 * 37 * Background and previous state 38 * 39 * Historically objects could exists (be bound) in global GTT space only as 40 * singular instances with a view representing all of the object's backing pages 41 * in a linear fashion. This view will be called a normal view. 42 * 43 * To support multiple views of the same object, where the number of mapped 44 * pages is not equal to the backing store, or where the layout of the pages 45 * is not linear, concept of a GGTT view was added. 46 * 47 * One example of an alternative view is a stereo display driven by a single 48 * image. In this case we would have a framebuffer looking like this 49 * (2x2 pages): 50 * 51 * 12 52 * 34 53 * 54 * Above would represent a normal GGTT view as normally mapped for GPU or CPU 55 * rendering. In contrast, fed to the display engine would be an alternative 56 * view which could look something like this: 57 * 58 * 1212 59 * 3434 60 * 61 * In this example both the size and layout of pages in the alternative view is 62 * different from the normal view. 63 * 64 * Implementation and usage 65 * 66 * GGTT views are implemented using VMAs and are distinguished via enum 67 * i915_ggtt_view_type and struct i915_ggtt_view. 68 * 69 * A new flavour of core GEM functions which work with GGTT bound objects were 70 * added with the _ggtt_ infix, and sometimes with _view postfix to avoid 71 * renaming in large amounts of code. They take the struct i915_ggtt_view 72 * parameter encapsulating all metadata required to implement a view. 73 * 74 * As a helper for callers which are only interested in the normal view, 75 * globally const i915_ggtt_view_normal singleton instance exists. All old core 76 * GEM API functions, the ones not taking the view parameter, are operating on, 77 * or with the normal GGTT view. 78 * 79 * Code wanting to add or use a new GGTT view needs to: 80 * 81 * 1. Add a new enum with a suitable name. 82 * 2. Extend the metadata in the i915_ggtt_view structure if required. 83 * 3. Add support to i915_get_vma_pages(). 84 * 85 * New views are required to build a scatter-gather table from within the 86 * i915_get_vma_pages function. This table is stored in the vma.ggtt_view and 87 * exists for the lifetime of an VMA. 88 * 89 * Core API is designed to have copy semantics which means that passed in 90 * struct i915_ggtt_view does not need to be persistent (left around after 91 * calling the core API functions). 92 * 93 */ 94 95 static int 96 i915_get_ggtt_vma_pages(struct i915_vma *vma); 97 98 const struct i915_ggtt_view i915_ggtt_view_normal; 99 const struct i915_ggtt_view i915_ggtt_view_rotated = { 100 .type = I915_GGTT_VIEW_ROTATED 101 }; 102 103 static int sanitize_enable_ppgtt(struct drm_device *dev, int enable_ppgtt) 104 { 105 bool has_aliasing_ppgtt; 106 bool has_full_ppgtt; 107 108 has_aliasing_ppgtt = INTEL_INFO(dev)->gen >= 6; 109 has_full_ppgtt = INTEL_INFO(dev)->gen >= 7; 110 111 if (intel_vgpu_active(dev)) 112 has_full_ppgtt = false; /* emulation is too hard */ 113 114 /* 115 * We don't allow disabling PPGTT for gen9+ as it's a requirement for 116 * execlists, the sole mechanism available to submit work. 117 */ 118 if (INTEL_INFO(dev)->gen < 9 && 119 (enable_ppgtt == 0 || !has_aliasing_ppgtt)) 120 return 0; 121 122 if (enable_ppgtt == 1) 123 return 1; 124 125 if (enable_ppgtt == 2 && has_full_ppgtt) 126 return 2; 127 128 #ifdef CONFIG_INTEL_IOMMU 129 /* Disable ppgtt on SNB if VT-d is on. */ 130 if (INTEL_INFO(dev)->gen == 6 && intel_iommu_gfx_mapped) { 131 DRM_INFO("Disabling PPGTT because VT-d is on\n"); 132 return 0; 133 } 134 #endif 135 136 /* Early VLV doesn't have this */ 137 if (IS_VALLEYVIEW(dev) && !IS_CHERRYVIEW(dev) && 138 dev->pdev->revision < 0xb) { 139 DRM_DEBUG_DRIVER("disabling PPGTT on pre-B3 step VLV\n"); 140 return 0; 141 } 142 143 if (INTEL_INFO(dev)->gen >= 8 && i915.enable_execlists) 144 return 2; 145 else 146 return has_aliasing_ppgtt ? 1 : 0; 147 } 148 149 static int ppgtt_bind_vma(struct i915_vma *vma, 150 enum i915_cache_level cache_level, 151 u32 unused) 152 { 153 u32 pte_flags = 0; 154 155 /* Currently applicable only to VLV */ 156 if (vma->obj->gt_ro) 157 pte_flags |= PTE_READ_ONLY; 158 159 vma->vm->insert_entries(vma->vm, vma->obj->pages, vma->node.start, 160 cache_level, pte_flags); 161 162 return 0; 163 } 164 165 static void ppgtt_unbind_vma(struct i915_vma *vma) 166 { 167 vma->vm->clear_range(vma->vm, 168 vma->node.start, 169 vma->obj->base.size, 170 true); 171 } 172 173 static gen8_pte_t gen8_pte_encode(dma_addr_t addr, 174 enum i915_cache_level level, 175 bool valid) 176 { 177 gen8_pte_t pte = valid ? _PAGE_PRESENT | _PAGE_RW : 0; 178 pte |= addr; 179 180 switch (level) { 181 case I915_CACHE_NONE: 182 pte |= PPAT_UNCACHED_INDEX; 183 break; 184 case I915_CACHE_WT: 185 pte |= PPAT_DISPLAY_ELLC_INDEX; 186 break; 187 default: 188 pte |= PPAT_CACHED_INDEX; 189 break; 190 } 191 192 return pte; 193 } 194 195 static gen8_pde_t gen8_pde_encode(struct drm_device *dev, 196 dma_addr_t addr, 197 enum i915_cache_level level) 198 { 199 gen8_pde_t pde = _PAGE_PRESENT | _PAGE_RW; 200 pde |= addr; 201 if (level != I915_CACHE_NONE) 202 pde |= PPAT_CACHED_PDE_INDEX; 203 else 204 pde |= PPAT_UNCACHED_INDEX; 205 return pde; 206 } 207 208 static gen6_pte_t snb_pte_encode(dma_addr_t addr, 209 enum i915_cache_level level, 210 bool valid, u32 unused) 211 { 212 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0; 213 pte |= GEN6_PTE_ADDR_ENCODE(addr); 214 215 switch (level) { 216 case I915_CACHE_L3_LLC: 217 case I915_CACHE_LLC: 218 pte |= GEN6_PTE_CACHE_LLC; 219 break; 220 case I915_CACHE_NONE: 221 pte |= GEN6_PTE_UNCACHED; 222 break; 223 default: 224 MISSING_CASE(level); 225 } 226 227 return pte; 228 } 229 230 static gen6_pte_t ivb_pte_encode(dma_addr_t addr, 231 enum i915_cache_level level, 232 bool valid, u32 unused) 233 { 234 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0; 235 pte |= GEN6_PTE_ADDR_ENCODE(addr); 236 237 switch (level) { 238 case I915_CACHE_L3_LLC: 239 pte |= GEN7_PTE_CACHE_L3_LLC; 240 break; 241 case I915_CACHE_LLC: 242 pte |= GEN6_PTE_CACHE_LLC; 243 break; 244 case I915_CACHE_NONE: 245 pte |= GEN6_PTE_UNCACHED; 246 break; 247 default: 248 MISSING_CASE(level); 249 } 250 251 return pte; 252 } 253 254 static gen6_pte_t byt_pte_encode(dma_addr_t addr, 255 enum i915_cache_level level, 256 bool valid, u32 flags) 257 { 258 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0; 259 pte |= GEN6_PTE_ADDR_ENCODE(addr); 260 261 if (!(flags & PTE_READ_ONLY)) 262 pte |= BYT_PTE_WRITEABLE; 263 264 if (level != I915_CACHE_NONE) 265 pte |= BYT_PTE_SNOOPED_BY_CPU_CACHES; 266 267 return pte; 268 } 269 270 static gen6_pte_t hsw_pte_encode(dma_addr_t addr, 271 enum i915_cache_level level, 272 bool valid, u32 unused) 273 { 274 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0; 275 pte |= HSW_PTE_ADDR_ENCODE(addr); 276 277 if (level != I915_CACHE_NONE) 278 pte |= HSW_WB_LLC_AGE3; 279 280 return pte; 281 } 282 283 static gen6_pte_t iris_pte_encode(dma_addr_t addr, 284 enum i915_cache_level level, 285 bool valid, u32 unused) 286 { 287 gen6_pte_t pte = valid ? GEN6_PTE_VALID : 0; 288 pte |= HSW_PTE_ADDR_ENCODE(addr); 289 290 switch (level) { 291 case I915_CACHE_NONE: 292 break; 293 case I915_CACHE_WT: 294 pte |= HSW_WT_ELLC_LLC_AGE3; 295 break; 296 default: 297 pte |= HSW_WB_ELLC_LLC_AGE3; 298 break; 299 } 300 301 return pte; 302 } 303 304 #define i915_dma_unmap_single(px, dev) \ 305 __i915_dma_unmap_single((px)->daddr, dev) 306 307 static void __i915_dma_unmap_single(dma_addr_t daddr, 308 struct drm_device *dev) 309 { 310 struct device *device = &dev->pdev->dev; 311 312 dma_unmap_page(device, daddr, 4096, PCI_DMA_BIDIRECTIONAL); 313 } 314 315 /** 316 * i915_dma_map_single() - Create a dma mapping for a page table/dir/etc. 317 * @px: Page table/dir/etc to get a DMA map for 318 * @dev: drm device 319 * 320 * Page table allocations are unified across all gens. They always require a 321 * single 4k allocation, as well as a DMA mapping. If we keep the structs 322 * symmetric here, the simple macro covers us for every page table type. 323 * 324 * Return: 0 if success. 325 */ 326 #define i915_dma_map_single(px, dev) \ 327 i915_dma_map_page_single((px)->page, (dev), &(px)->daddr) 328 329 static int i915_dma_map_page_single(struct page *page, 330 struct drm_device *dev, 331 dma_addr_t *daddr) 332 { 333 struct device *device = &dev->pdev->dev; 334 335 *daddr = dma_map_page(device, page, 0, 4096, PCI_DMA_BIDIRECTIONAL); 336 if (dma_mapping_error(device, *daddr)) 337 return -ENOMEM; 338 339 return 0; 340 } 341 342 static void unmap_and_free_pt(struct i915_page_table *pt, 343 struct drm_device *dev) 344 { 345 if (WARN_ON(!pt->page)) 346 return; 347 348 i915_dma_unmap_single(pt, dev); 349 __free_page(pt->page); 350 kfree(pt->used_ptes); 351 kfree(pt); 352 } 353 354 static void gen8_initialize_pt(struct i915_address_space *vm, 355 struct i915_page_table *pt) 356 { 357 gen8_pte_t *pt_vaddr, scratch_pte; 358 int i; 359 360 pt_vaddr = kmap_atomic(pt->page); 361 scratch_pte = gen8_pte_encode(vm->scratch.addr, 362 I915_CACHE_LLC, true); 363 364 for (i = 0; i < GEN8_PTES; i++) 365 pt_vaddr[i] = scratch_pte; 366 367 if (!HAS_LLC(vm->dev)) 368 drm_clflush_virt_range(pt_vaddr, PAGE_SIZE); 369 kunmap_atomic(pt_vaddr); 370 } 371 372 static struct i915_page_table *alloc_pt_single(struct drm_device *dev) 373 { 374 struct i915_page_table *pt; 375 const size_t count = INTEL_INFO(dev)->gen >= 8 ? 376 GEN8_PTES : GEN6_PTES; 377 int ret = -ENOMEM; 378 379 pt = kzalloc(sizeof(*pt), GFP_KERNEL); 380 if (!pt) 381 return ERR_PTR(-ENOMEM); 382 383 pt->used_ptes = kcalloc(BITS_TO_LONGS(count), sizeof(*pt->used_ptes), 384 GFP_KERNEL); 385 386 if (!pt->used_ptes) 387 goto fail_bitmap; 388 389 pt->page = alloc_page(GFP_KERNEL); 390 if (!pt->page) 391 goto fail_page; 392 393 ret = i915_dma_map_single(pt, dev); 394 if (ret) 395 goto fail_dma; 396 397 return pt; 398 399 fail_dma: 400 __free_page(pt->page); 401 fail_page: 402 kfree(pt->used_ptes); 403 fail_bitmap: 404 kfree(pt); 405 406 return ERR_PTR(ret); 407 } 408 409 static void unmap_and_free_pd(struct i915_page_directory *pd, 410 struct drm_device *dev) 411 { 412 if (pd->page) { 413 i915_dma_unmap_single(pd, dev); 414 __free_page(pd->page); 415 kfree(pd->used_pdes); 416 kfree(pd); 417 } 418 } 419 420 static struct i915_page_directory *alloc_pd_single(struct drm_device *dev) 421 { 422 struct i915_page_directory *pd; 423 int ret = -ENOMEM; 424 425 pd = kzalloc(sizeof(*pd), GFP_KERNEL); 426 if (!pd) 427 return ERR_PTR(-ENOMEM); 428 429 pd->used_pdes = kcalloc(BITS_TO_LONGS(I915_PDES), 430 sizeof(*pd->used_pdes), GFP_KERNEL); 431 if (!pd->used_pdes) 432 goto free_pd; 433 434 pd->page = alloc_page(GFP_KERNEL); 435 if (!pd->page) 436 goto free_bitmap; 437 438 ret = i915_dma_map_single(pd, dev); 439 if (ret) 440 goto free_page; 441 442 return pd; 443 444 free_page: 445 __free_page(pd->page); 446 free_bitmap: 447 kfree(pd->used_pdes); 448 free_pd: 449 kfree(pd); 450 451 return ERR_PTR(ret); 452 } 453 454 /* Broadwell Page Directory Pointer Descriptors */ 455 static int gen8_write_pdp(struct intel_engine_cs *ring, 456 unsigned entry, 457 dma_addr_t addr) 458 { 459 int ret; 460 461 BUG_ON(entry >= 4); 462 463 ret = intel_ring_begin(ring, 6); 464 if (ret) 465 return ret; 466 467 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1)); 468 intel_ring_emit(ring, GEN8_RING_PDP_UDW(ring, entry)); 469 intel_ring_emit(ring, upper_32_bits(addr)); 470 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1)); 471 intel_ring_emit(ring, GEN8_RING_PDP_LDW(ring, entry)); 472 intel_ring_emit(ring, lower_32_bits(addr)); 473 intel_ring_advance(ring); 474 475 return 0; 476 } 477 478 static int gen8_mm_switch(struct i915_hw_ppgtt *ppgtt, 479 struct intel_engine_cs *ring) 480 { 481 int i, ret; 482 483 for (i = GEN8_LEGACY_PDPES - 1; i >= 0; i--) { 484 struct i915_page_directory *pd = ppgtt->pdp.page_directory[i]; 485 dma_addr_t pd_daddr = pd ? pd->daddr : ppgtt->scratch_pd->daddr; 486 /* The page directory might be NULL, but we need to clear out 487 * whatever the previous context might have used. */ 488 ret = gen8_write_pdp(ring, i, pd_daddr); 489 if (ret) 490 return ret; 491 } 492 493 return 0; 494 } 495 496 static void gen8_ppgtt_clear_range(struct i915_address_space *vm, 497 uint64_t start, 498 uint64_t length, 499 bool use_scratch) 500 { 501 struct i915_hw_ppgtt *ppgtt = 502 container_of(vm, struct i915_hw_ppgtt, base); 503 gen8_pte_t *pt_vaddr, scratch_pte; 504 unsigned pdpe = start >> GEN8_PDPE_SHIFT & GEN8_PDPE_MASK; 505 unsigned pde = start >> GEN8_PDE_SHIFT & GEN8_PDE_MASK; 506 unsigned pte = start >> GEN8_PTE_SHIFT & GEN8_PTE_MASK; 507 unsigned num_entries = length >> PAGE_SHIFT; 508 unsigned last_pte, i; 509 510 scratch_pte = gen8_pte_encode(ppgtt->base.scratch.addr, 511 I915_CACHE_LLC, use_scratch); 512 513 while (num_entries) { 514 struct i915_page_directory *pd; 515 struct i915_page_table *pt; 516 struct page *page_table; 517 518 if (WARN_ON(!ppgtt->pdp.page_directory[pdpe])) 519 break; 520 521 pd = ppgtt->pdp.page_directory[pdpe]; 522 523 if (WARN_ON(!pd->page_table[pde])) 524 break; 525 526 pt = pd->page_table[pde]; 527 528 if (WARN_ON(!pt->page)) 529 break; 530 531 page_table = pt->page; 532 533 last_pte = pte + num_entries; 534 if (last_pte > GEN8_PTES) 535 last_pte = GEN8_PTES; 536 537 pt_vaddr = kmap_atomic(page_table); 538 539 for (i = pte; i < last_pte; i++) { 540 pt_vaddr[i] = scratch_pte; 541 num_entries--; 542 } 543 544 if (!HAS_LLC(ppgtt->base.dev)) 545 drm_clflush_virt_range(pt_vaddr, PAGE_SIZE); 546 kunmap_atomic(pt_vaddr); 547 548 pte = 0; 549 if (++pde == I915_PDES) { 550 pdpe++; 551 pde = 0; 552 } 553 } 554 } 555 556 static void gen8_ppgtt_insert_entries(struct i915_address_space *vm, 557 struct sg_table *pages, 558 uint64_t start, 559 enum i915_cache_level cache_level, u32 unused) 560 { 561 struct i915_hw_ppgtt *ppgtt = 562 container_of(vm, struct i915_hw_ppgtt, base); 563 gen8_pte_t *pt_vaddr; 564 unsigned pdpe = start >> GEN8_PDPE_SHIFT & GEN8_PDPE_MASK; 565 unsigned pde = start >> GEN8_PDE_SHIFT & GEN8_PDE_MASK; 566 unsigned pte = start >> GEN8_PTE_SHIFT & GEN8_PTE_MASK; 567 struct sg_page_iter sg_iter; 568 569 pt_vaddr = NULL; 570 571 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) { 572 if (WARN_ON(pdpe >= GEN8_LEGACY_PDPES)) 573 break; 574 575 if (pt_vaddr == NULL) { 576 struct i915_page_directory *pd = ppgtt->pdp.page_directory[pdpe]; 577 struct i915_page_table *pt = pd->page_table[pde]; 578 struct page *page_table = pt->page; 579 580 pt_vaddr = kmap_atomic(page_table); 581 } 582 583 pt_vaddr[pte] = 584 gen8_pte_encode(sg_page_iter_dma_address(&sg_iter), 585 cache_level, true); 586 if (++pte == GEN8_PTES) { 587 if (!HAS_LLC(ppgtt->base.dev)) 588 drm_clflush_virt_range(pt_vaddr, PAGE_SIZE); 589 kunmap_atomic(pt_vaddr); 590 pt_vaddr = NULL; 591 if (++pde == I915_PDES) { 592 pdpe++; 593 pde = 0; 594 } 595 pte = 0; 596 } 597 } 598 if (pt_vaddr) { 599 if (!HAS_LLC(ppgtt->base.dev)) 600 drm_clflush_virt_range(pt_vaddr, PAGE_SIZE); 601 kunmap_atomic(pt_vaddr); 602 } 603 } 604 605 static void __gen8_do_map_pt(gen8_pde_t * const pde, 606 struct i915_page_table *pt, 607 struct drm_device *dev) 608 { 609 gen8_pde_t entry = 610 gen8_pde_encode(dev, pt->daddr, I915_CACHE_LLC); 611 *pde = entry; 612 } 613 614 static void gen8_initialize_pd(struct i915_address_space *vm, 615 struct i915_page_directory *pd) 616 { 617 struct i915_hw_ppgtt *ppgtt = 618 container_of(vm, struct i915_hw_ppgtt, base); 619 gen8_pde_t *page_directory; 620 struct i915_page_table *pt; 621 int i; 622 623 page_directory = kmap_atomic(pd->page); 624 pt = ppgtt->scratch_pt; 625 for (i = 0; i < I915_PDES; i++) 626 /* Map the PDE to the page table */ 627 __gen8_do_map_pt(page_directory + i, pt, vm->dev); 628 629 if (!HAS_LLC(vm->dev)) 630 drm_clflush_virt_range(page_directory, PAGE_SIZE); 631 kunmap_atomic(page_directory); 632 } 633 634 static void gen8_free_page_tables(struct i915_page_directory *pd, struct drm_device *dev) 635 { 636 int i; 637 638 if (!pd->page) 639 return; 640 641 for_each_set_bit(i, pd->used_pdes, I915_PDES) { 642 if (WARN_ON(!pd->page_table[i])) 643 continue; 644 645 unmap_and_free_pt(pd->page_table[i], dev); 646 pd->page_table[i] = NULL; 647 } 648 } 649 650 static void gen8_ppgtt_cleanup(struct i915_address_space *vm) 651 { 652 struct i915_hw_ppgtt *ppgtt = 653 container_of(vm, struct i915_hw_ppgtt, base); 654 int i; 655 656 for_each_set_bit(i, ppgtt->pdp.used_pdpes, GEN8_LEGACY_PDPES) { 657 if (WARN_ON(!ppgtt->pdp.page_directory[i])) 658 continue; 659 660 gen8_free_page_tables(ppgtt->pdp.page_directory[i], ppgtt->base.dev); 661 unmap_and_free_pd(ppgtt->pdp.page_directory[i], ppgtt->base.dev); 662 } 663 664 unmap_and_free_pd(ppgtt->scratch_pd, ppgtt->base.dev); 665 unmap_and_free_pt(ppgtt->scratch_pt, ppgtt->base.dev); 666 } 667 668 /** 669 * gen8_ppgtt_alloc_pagetabs() - Allocate page tables for VA range. 670 * @ppgtt: Master ppgtt structure. 671 * @pd: Page directory for this address range. 672 * @start: Starting virtual address to begin allocations. 673 * @length Size of the allocations. 674 * @new_pts: Bitmap set by function with new allocations. Likely used by the 675 * caller to free on error. 676 * 677 * Allocate the required number of page tables. Extremely similar to 678 * gen8_ppgtt_alloc_page_directories(). The main difference is here we are limited by 679 * the page directory boundary (instead of the page directory pointer). That 680 * boundary is 1GB virtual. Therefore, unlike gen8_ppgtt_alloc_page_directories(), it is 681 * possible, and likely that the caller will need to use multiple calls of this 682 * function to achieve the appropriate allocation. 683 * 684 * Return: 0 if success; negative error code otherwise. 685 */ 686 static int gen8_ppgtt_alloc_pagetabs(struct i915_hw_ppgtt *ppgtt, 687 struct i915_page_directory *pd, 688 uint64_t start, 689 uint64_t length, 690 unsigned long *new_pts) 691 { 692 struct drm_device *dev = ppgtt->base.dev; 693 struct i915_page_table *pt; 694 uint64_t temp; 695 uint32_t pde; 696 697 gen8_for_each_pde(pt, pd, start, length, temp, pde) { 698 /* Don't reallocate page tables */ 699 if (pt) { 700 /* Scratch is never allocated this way */ 701 WARN_ON(pt == ppgtt->scratch_pt); 702 continue; 703 } 704 705 pt = alloc_pt_single(dev); 706 if (IS_ERR(pt)) 707 goto unwind_out; 708 709 gen8_initialize_pt(&ppgtt->base, pt); 710 pd->page_table[pde] = pt; 711 set_bit(pde, new_pts); 712 } 713 714 return 0; 715 716 unwind_out: 717 for_each_set_bit(pde, new_pts, I915_PDES) 718 unmap_and_free_pt(pd->page_table[pde], dev); 719 720 return -ENOMEM; 721 } 722 723 /** 724 * gen8_ppgtt_alloc_page_directories() - Allocate page directories for VA range. 725 * @ppgtt: Master ppgtt structure. 726 * @pdp: Page directory pointer for this address range. 727 * @start: Starting virtual address to begin allocations. 728 * @length Size of the allocations. 729 * @new_pds Bitmap set by function with new allocations. Likely used by the 730 * caller to free on error. 731 * 732 * Allocate the required number of page directories starting at the pde index of 733 * @start, and ending at the pde index @start + @length. This function will skip 734 * over already allocated page directories within the range, and only allocate 735 * new ones, setting the appropriate pointer within the pdp as well as the 736 * correct position in the bitmap @new_pds. 737 * 738 * The function will only allocate the pages within the range for a give page 739 * directory pointer. In other words, if @start + @length straddles a virtually 740 * addressed PDP boundary (512GB for 4k pages), there will be more allocations 741 * required by the caller, This is not currently possible, and the BUG in the 742 * code will prevent it. 743 * 744 * Return: 0 if success; negative error code otherwise. 745 */ 746 static int gen8_ppgtt_alloc_page_directories(struct i915_hw_ppgtt *ppgtt, 747 struct i915_page_directory_pointer *pdp, 748 uint64_t start, 749 uint64_t length, 750 unsigned long *new_pds) 751 { 752 struct drm_device *dev = ppgtt->base.dev; 753 struct i915_page_directory *pd; 754 uint64_t temp; 755 uint32_t pdpe; 756 757 WARN_ON(!bitmap_empty(new_pds, GEN8_LEGACY_PDPES)); 758 759 /* FIXME: upper bound must not overflow 32 bits */ 760 WARN_ON((start + length) > (1ULL << 32)); 761 762 gen8_for_each_pdpe(pd, pdp, start, length, temp, pdpe) { 763 if (pd) 764 continue; 765 766 pd = alloc_pd_single(dev); 767 if (IS_ERR(pd)) 768 goto unwind_out; 769 770 gen8_initialize_pd(&ppgtt->base, pd); 771 pdp->page_directory[pdpe] = pd; 772 set_bit(pdpe, new_pds); 773 } 774 775 return 0; 776 777 unwind_out: 778 for_each_set_bit(pdpe, new_pds, GEN8_LEGACY_PDPES) 779 unmap_and_free_pd(pdp->page_directory[pdpe], dev); 780 781 return -ENOMEM; 782 } 783 784 static void 785 free_gen8_temp_bitmaps(unsigned long *new_pds, unsigned long **new_pts) 786 { 787 int i; 788 789 for (i = 0; i < GEN8_LEGACY_PDPES; i++) 790 kfree(new_pts[i]); 791 kfree(new_pts); 792 kfree(new_pds); 793 } 794 795 /* Fills in the page directory bitmap, and the array of page tables bitmap. Both 796 * of these are based on the number of PDPEs in the system. 797 */ 798 static 799 int __must_check alloc_gen8_temp_bitmaps(unsigned long **new_pds, 800 unsigned long ***new_pts) 801 { 802 int i; 803 unsigned long *pds; 804 unsigned long **pts; 805 806 pds = kcalloc(BITS_TO_LONGS(GEN8_LEGACY_PDPES), sizeof(unsigned long), GFP_KERNEL); 807 if (!pds) 808 return -ENOMEM; 809 810 pts = kcalloc(GEN8_LEGACY_PDPES, sizeof(unsigned long *), GFP_KERNEL); 811 if (!pts) { 812 kfree(pds); 813 return -ENOMEM; 814 } 815 816 for (i = 0; i < GEN8_LEGACY_PDPES; i++) { 817 pts[i] = kcalloc(BITS_TO_LONGS(I915_PDES), 818 sizeof(unsigned long), GFP_KERNEL); 819 if (!pts[i]) 820 goto err_out; 821 } 822 823 *new_pds = pds; 824 *new_pts = pts; 825 826 return 0; 827 828 err_out: 829 free_gen8_temp_bitmaps(pds, pts); 830 return -ENOMEM; 831 } 832 833 static int gen8_alloc_va_range(struct i915_address_space *vm, 834 uint64_t start, 835 uint64_t length) 836 { 837 struct i915_hw_ppgtt *ppgtt = 838 container_of(vm, struct i915_hw_ppgtt, base); 839 unsigned long *new_page_dirs, **new_page_tables; 840 struct i915_page_directory *pd; 841 const uint64_t orig_start = start; 842 const uint64_t orig_length = length; 843 uint64_t temp; 844 uint32_t pdpe; 845 int ret; 846 847 /* Wrap is never okay since we can only represent 48b, and we don't 848 * actually use the other side of the canonical address space. 849 */ 850 if (WARN_ON(start + length < start)) 851 return -ERANGE; 852 853 ret = alloc_gen8_temp_bitmaps(&new_page_dirs, &new_page_tables); 854 if (ret) 855 return ret; 856 857 /* Do the allocations first so we can easily bail out */ 858 ret = gen8_ppgtt_alloc_page_directories(ppgtt, &ppgtt->pdp, start, length, 859 new_page_dirs); 860 if (ret) { 861 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables); 862 return ret; 863 } 864 865 /* For every page directory referenced, allocate page tables */ 866 gen8_for_each_pdpe(pd, &ppgtt->pdp, start, length, temp, pdpe) { 867 ret = gen8_ppgtt_alloc_pagetabs(ppgtt, pd, start, length, 868 new_page_tables[pdpe]); 869 if (ret) 870 goto err_out; 871 } 872 873 start = orig_start; 874 length = orig_length; 875 876 /* Allocations have completed successfully, so set the bitmaps, and do 877 * the mappings. */ 878 gen8_for_each_pdpe(pd, &ppgtt->pdp, start, length, temp, pdpe) { 879 gen8_pde_t *const page_directory = kmap_atomic(pd->page); 880 struct i915_page_table *pt; 881 uint64_t pd_len = gen8_clamp_pd(start, length); 882 uint64_t pd_start = start; 883 uint32_t pde; 884 885 /* Every pd should be allocated, we just did that above. */ 886 WARN_ON(!pd); 887 888 gen8_for_each_pde(pt, pd, pd_start, pd_len, temp, pde) { 889 /* Same reasoning as pd */ 890 WARN_ON(!pt); 891 WARN_ON(!pd_len); 892 WARN_ON(!gen8_pte_count(pd_start, pd_len)); 893 894 /* Set our used ptes within the page table */ 895 bitmap_set(pt->used_ptes, 896 gen8_pte_index(pd_start), 897 gen8_pte_count(pd_start, pd_len)); 898 899 /* Our pde is now pointing to the pagetable, pt */ 900 set_bit(pde, pd->used_pdes); 901 902 /* Map the PDE to the page table */ 903 __gen8_do_map_pt(page_directory + pde, pt, vm->dev); 904 905 /* NB: We haven't yet mapped ptes to pages. At this 906 * point we're still relying on insert_entries() */ 907 } 908 909 if (!HAS_LLC(vm->dev)) 910 drm_clflush_virt_range(page_directory, PAGE_SIZE); 911 912 kunmap_atomic(page_directory); 913 914 set_bit(pdpe, ppgtt->pdp.used_pdpes); 915 } 916 917 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables); 918 return 0; 919 920 err_out: 921 while (pdpe--) { 922 for_each_set_bit(temp, new_page_tables[pdpe], I915_PDES) 923 unmap_and_free_pt(ppgtt->pdp.page_directory[pdpe]->page_table[temp], vm->dev); 924 } 925 926 for_each_set_bit(pdpe, new_page_dirs, GEN8_LEGACY_PDPES) 927 unmap_and_free_pd(ppgtt->pdp.page_directory[pdpe], vm->dev); 928 929 free_gen8_temp_bitmaps(new_page_dirs, new_page_tables); 930 return ret; 931 } 932 933 /* 934 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers 935 * with a net effect resembling a 2-level page table in normal x86 terms. Each 936 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address 937 * space. 938 * 939 */ 940 static int gen8_ppgtt_init(struct i915_hw_ppgtt *ppgtt) 941 { 942 ppgtt->scratch_pt = alloc_pt_single(ppgtt->base.dev); 943 if (IS_ERR(ppgtt->scratch_pt)) 944 return PTR_ERR(ppgtt->scratch_pt); 945 946 ppgtt->scratch_pd = alloc_pd_single(ppgtt->base.dev); 947 if (IS_ERR(ppgtt->scratch_pd)) 948 return PTR_ERR(ppgtt->scratch_pd); 949 950 gen8_initialize_pt(&ppgtt->base, ppgtt->scratch_pt); 951 gen8_initialize_pd(&ppgtt->base, ppgtt->scratch_pd); 952 953 ppgtt->base.start = 0; 954 ppgtt->base.total = 1ULL << 32; 955 if (IS_ENABLED(CONFIG_X86_32)) 956 /* While we have a proliferation of size_t variables 957 * we cannot represent the full ppgtt size on 32bit, 958 * so limit it to the same size as the GGTT (currently 959 * 2GiB). 960 */ 961 ppgtt->base.total = to_i915(ppgtt->base.dev)->gtt.base.total; 962 ppgtt->base.cleanup = gen8_ppgtt_cleanup; 963 ppgtt->base.allocate_va_range = gen8_alloc_va_range; 964 ppgtt->base.insert_entries = gen8_ppgtt_insert_entries; 965 ppgtt->base.clear_range = gen8_ppgtt_clear_range; 966 ppgtt->base.unbind_vma = ppgtt_unbind_vma; 967 ppgtt->base.bind_vma = ppgtt_bind_vma; 968 969 ppgtt->switch_mm = gen8_mm_switch; 970 971 return 0; 972 } 973 974 static void gen6_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m) 975 { 976 struct i915_address_space *vm = &ppgtt->base; 977 struct i915_page_table *unused; 978 gen6_pte_t scratch_pte; 979 uint32_t pd_entry; 980 uint32_t pte, pde, temp; 981 uint32_t start = ppgtt->base.start, length = ppgtt->base.total; 982 983 scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC, true, 0); 984 985 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde) { 986 u32 expected; 987 gen6_pte_t *pt_vaddr; 988 dma_addr_t pt_addr = ppgtt->pd.page_table[pde]->daddr; 989 pd_entry = readl(ppgtt->pd_addr + pde); 990 expected = (GEN6_PDE_ADDR_ENCODE(pt_addr) | GEN6_PDE_VALID); 991 992 if (pd_entry != expected) 993 seq_printf(m, "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n", 994 pde, 995 pd_entry, 996 expected); 997 seq_printf(m, "\tPDE: %x\n", pd_entry); 998 999 pt_vaddr = kmap_atomic(ppgtt->pd.page_table[pde]->page); 1000 for (pte = 0; pte < GEN6_PTES; pte+=4) { 1001 unsigned long va = 1002 (pde * PAGE_SIZE * GEN6_PTES) + 1003 (pte * PAGE_SIZE); 1004 int i; 1005 bool found = false; 1006 for (i = 0; i < 4; i++) 1007 if (pt_vaddr[pte + i] != scratch_pte) 1008 found = true; 1009 if (!found) 1010 continue; 1011 1012 seq_printf(m, "\t\t0x%lx [%03d,%04d]: =", va, pde, pte); 1013 for (i = 0; i < 4; i++) { 1014 if (pt_vaddr[pte + i] != scratch_pte) 1015 seq_printf(m, " %08x", pt_vaddr[pte + i]); 1016 else 1017 seq_puts(m, " SCRATCH "); 1018 } 1019 seq_puts(m, "\n"); 1020 } 1021 kunmap_atomic(pt_vaddr); 1022 } 1023 } 1024 1025 /* Write pde (index) from the page directory @pd to the page table @pt */ 1026 static void gen6_write_pde(struct i915_page_directory *pd, 1027 const int pde, struct i915_page_table *pt) 1028 { 1029 /* Caller needs to make sure the write completes if necessary */ 1030 struct i915_hw_ppgtt *ppgtt = 1031 container_of(pd, struct i915_hw_ppgtt, pd); 1032 u32 pd_entry; 1033 1034 pd_entry = GEN6_PDE_ADDR_ENCODE(pt->daddr); 1035 pd_entry |= GEN6_PDE_VALID; 1036 1037 writel(pd_entry, ppgtt->pd_addr + pde); 1038 } 1039 1040 /* Write all the page tables found in the ppgtt structure to incrementing page 1041 * directories. */ 1042 static void gen6_write_page_range(struct drm_i915_private *dev_priv, 1043 struct i915_page_directory *pd, 1044 uint32_t start, uint32_t length) 1045 { 1046 struct i915_page_table *pt; 1047 uint32_t pde, temp; 1048 1049 gen6_for_each_pde(pt, pd, start, length, temp, pde) 1050 gen6_write_pde(pd, pde, pt); 1051 1052 /* Make sure write is complete before other code can use this page 1053 * table. Also require for WC mapped PTEs */ 1054 readl(dev_priv->gtt.gsm); 1055 } 1056 1057 static uint32_t get_pd_offset(struct i915_hw_ppgtt *ppgtt) 1058 { 1059 BUG_ON(ppgtt->pd.pd_offset & 0x3f); 1060 1061 return (ppgtt->pd.pd_offset / 64) << 16; 1062 } 1063 1064 static int hsw_mm_switch(struct i915_hw_ppgtt *ppgtt, 1065 struct intel_engine_cs *ring) 1066 { 1067 int ret; 1068 1069 /* NB: TLBs must be flushed and invalidated before a switch */ 1070 ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS); 1071 if (ret) 1072 return ret; 1073 1074 ret = intel_ring_begin(ring, 6); 1075 if (ret) 1076 return ret; 1077 1078 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2)); 1079 intel_ring_emit(ring, RING_PP_DIR_DCLV(ring)); 1080 intel_ring_emit(ring, PP_DIR_DCLV_2G); 1081 intel_ring_emit(ring, RING_PP_DIR_BASE(ring)); 1082 intel_ring_emit(ring, get_pd_offset(ppgtt)); 1083 intel_ring_emit(ring, MI_NOOP); 1084 intel_ring_advance(ring); 1085 1086 return 0; 1087 } 1088 1089 static int vgpu_mm_switch(struct i915_hw_ppgtt *ppgtt, 1090 struct intel_engine_cs *ring) 1091 { 1092 struct drm_i915_private *dev_priv = to_i915(ppgtt->base.dev); 1093 1094 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G); 1095 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt)); 1096 return 0; 1097 } 1098 1099 static int gen7_mm_switch(struct i915_hw_ppgtt *ppgtt, 1100 struct intel_engine_cs *ring) 1101 { 1102 int ret; 1103 1104 /* NB: TLBs must be flushed and invalidated before a switch */ 1105 ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS); 1106 if (ret) 1107 return ret; 1108 1109 ret = intel_ring_begin(ring, 6); 1110 if (ret) 1111 return ret; 1112 1113 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(2)); 1114 intel_ring_emit(ring, RING_PP_DIR_DCLV(ring)); 1115 intel_ring_emit(ring, PP_DIR_DCLV_2G); 1116 intel_ring_emit(ring, RING_PP_DIR_BASE(ring)); 1117 intel_ring_emit(ring, get_pd_offset(ppgtt)); 1118 intel_ring_emit(ring, MI_NOOP); 1119 intel_ring_advance(ring); 1120 1121 /* XXX: RCS is the only one to auto invalidate the TLBs? */ 1122 if (ring->id != RCS) { 1123 ret = ring->flush(ring, I915_GEM_GPU_DOMAINS, I915_GEM_GPU_DOMAINS); 1124 if (ret) 1125 return ret; 1126 } 1127 1128 return 0; 1129 } 1130 1131 static int gen6_mm_switch(struct i915_hw_ppgtt *ppgtt, 1132 struct intel_engine_cs *ring) 1133 { 1134 struct drm_device *dev = ppgtt->base.dev; 1135 struct drm_i915_private *dev_priv = dev->dev_private; 1136 1137 1138 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G); 1139 I915_WRITE(RING_PP_DIR_BASE(ring), get_pd_offset(ppgtt)); 1140 1141 POSTING_READ(RING_PP_DIR_DCLV(ring)); 1142 1143 return 0; 1144 } 1145 1146 static void gen8_ppgtt_enable(struct drm_device *dev) 1147 { 1148 struct drm_i915_private *dev_priv = dev->dev_private; 1149 struct intel_engine_cs *ring; 1150 int j; 1151 1152 for_each_ring(ring, dev_priv, j) { 1153 I915_WRITE(RING_MODE_GEN7(ring), 1154 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE)); 1155 } 1156 } 1157 1158 static void gen7_ppgtt_enable(struct drm_device *dev) 1159 { 1160 struct drm_i915_private *dev_priv = dev->dev_private; 1161 struct intel_engine_cs *ring; 1162 uint32_t ecochk, ecobits; 1163 int i; 1164 1165 ecobits = I915_READ(GAC_ECO_BITS); 1166 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B); 1167 1168 ecochk = I915_READ(GAM_ECOCHK); 1169 if (IS_HASWELL(dev)) { 1170 ecochk |= ECOCHK_PPGTT_WB_HSW; 1171 } else { 1172 ecochk |= ECOCHK_PPGTT_LLC_IVB; 1173 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB; 1174 } 1175 I915_WRITE(GAM_ECOCHK, ecochk); 1176 1177 for_each_ring(ring, dev_priv, i) { 1178 /* GFX_MODE is per-ring on gen7+ */ 1179 I915_WRITE(RING_MODE_GEN7(ring), 1180 _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE)); 1181 } 1182 } 1183 1184 static void gen6_ppgtt_enable(struct drm_device *dev) 1185 { 1186 struct drm_i915_private *dev_priv = dev->dev_private; 1187 uint32_t ecochk, gab_ctl, ecobits; 1188 1189 ecobits = I915_READ(GAC_ECO_BITS); 1190 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT | 1191 ECOBITS_PPGTT_CACHE64B); 1192 1193 gab_ctl = I915_READ(GAB_CTL); 1194 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT); 1195 1196 ecochk = I915_READ(GAM_ECOCHK); 1197 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B); 1198 1199 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE)); 1200 } 1201 1202 /* PPGTT support for Sandybdrige/Gen6 and later */ 1203 static void gen6_ppgtt_clear_range(struct i915_address_space *vm, 1204 uint64_t start, 1205 uint64_t length, 1206 bool use_scratch) 1207 { 1208 struct i915_hw_ppgtt *ppgtt = 1209 container_of(vm, struct i915_hw_ppgtt, base); 1210 gen6_pte_t *pt_vaddr, scratch_pte; 1211 unsigned first_entry = start >> PAGE_SHIFT; 1212 unsigned num_entries = length >> PAGE_SHIFT; 1213 unsigned act_pt = first_entry / GEN6_PTES; 1214 unsigned first_pte = first_entry % GEN6_PTES; 1215 unsigned last_pte, i; 1216 1217 scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC, true, 0); 1218 1219 while (num_entries) { 1220 last_pte = first_pte + num_entries; 1221 if (last_pte > GEN6_PTES) 1222 last_pte = GEN6_PTES; 1223 1224 pt_vaddr = kmap_atomic(ppgtt->pd.page_table[act_pt]->page); 1225 1226 for (i = first_pte; i < last_pte; i++) 1227 pt_vaddr[i] = scratch_pte; 1228 1229 kunmap_atomic(pt_vaddr); 1230 1231 num_entries -= last_pte - first_pte; 1232 first_pte = 0; 1233 act_pt++; 1234 } 1235 } 1236 1237 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm, 1238 struct sg_table *pages, 1239 uint64_t start, 1240 enum i915_cache_level cache_level, u32 flags) 1241 { 1242 struct i915_hw_ppgtt *ppgtt = 1243 container_of(vm, struct i915_hw_ppgtt, base); 1244 gen6_pte_t *pt_vaddr; 1245 unsigned first_entry = start >> PAGE_SHIFT; 1246 unsigned act_pt = first_entry / GEN6_PTES; 1247 unsigned act_pte = first_entry % GEN6_PTES; 1248 struct sg_page_iter sg_iter; 1249 1250 pt_vaddr = NULL; 1251 for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) { 1252 if (pt_vaddr == NULL) 1253 pt_vaddr = kmap_atomic(ppgtt->pd.page_table[act_pt]->page); 1254 1255 pt_vaddr[act_pte] = 1256 vm->pte_encode(sg_page_iter_dma_address(&sg_iter), 1257 cache_level, true, flags); 1258 1259 if (++act_pte == GEN6_PTES) { 1260 kunmap_atomic(pt_vaddr); 1261 pt_vaddr = NULL; 1262 act_pt++; 1263 act_pte = 0; 1264 } 1265 } 1266 if (pt_vaddr) 1267 kunmap_atomic(pt_vaddr); 1268 } 1269 1270 /* PDE TLBs are a pain invalidate pre GEN8. It requires a context reload. If we 1271 * are switching between contexts with the same LRCA, we also must do a force 1272 * restore. 1273 */ 1274 static void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt) 1275 { 1276 /* If current vm != vm, */ 1277 ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->base.dev)->ring_mask; 1278 } 1279 1280 static void gen6_initialize_pt(struct i915_address_space *vm, 1281 struct i915_page_table *pt) 1282 { 1283 gen6_pte_t *pt_vaddr, scratch_pte; 1284 int i; 1285 1286 WARN_ON(vm->scratch.addr == 0); 1287 1288 scratch_pte = vm->pte_encode(vm->scratch.addr, 1289 I915_CACHE_LLC, true, 0); 1290 1291 pt_vaddr = kmap_atomic(pt->page); 1292 1293 for (i = 0; i < GEN6_PTES; i++) 1294 pt_vaddr[i] = scratch_pte; 1295 1296 kunmap_atomic(pt_vaddr); 1297 } 1298 1299 static int gen6_alloc_va_range(struct i915_address_space *vm, 1300 uint64_t start, uint64_t length) 1301 { 1302 DECLARE_BITMAP(new_page_tables, I915_PDES); 1303 struct drm_device *dev = vm->dev; 1304 struct drm_i915_private *dev_priv = dev->dev_private; 1305 struct i915_hw_ppgtt *ppgtt = 1306 container_of(vm, struct i915_hw_ppgtt, base); 1307 struct i915_page_table *pt; 1308 const uint32_t start_save = start, length_save = length; 1309 uint32_t pde, temp; 1310 int ret; 1311 1312 WARN_ON(upper_32_bits(start)); 1313 1314 bitmap_zero(new_page_tables, I915_PDES); 1315 1316 /* The allocation is done in two stages so that we can bail out with 1317 * minimal amount of pain. The first stage finds new page tables that 1318 * need allocation. The second stage marks use ptes within the page 1319 * tables. 1320 */ 1321 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) { 1322 if (pt != ppgtt->scratch_pt) { 1323 WARN_ON(bitmap_empty(pt->used_ptes, GEN6_PTES)); 1324 continue; 1325 } 1326 1327 /* We've already allocated a page table */ 1328 WARN_ON(!bitmap_empty(pt->used_ptes, GEN6_PTES)); 1329 1330 pt = alloc_pt_single(dev); 1331 if (IS_ERR(pt)) { 1332 ret = PTR_ERR(pt); 1333 goto unwind_out; 1334 } 1335 1336 gen6_initialize_pt(vm, pt); 1337 1338 ppgtt->pd.page_table[pde] = pt; 1339 set_bit(pde, new_page_tables); 1340 trace_i915_page_table_entry_alloc(vm, pde, start, GEN6_PDE_SHIFT); 1341 } 1342 1343 start = start_save; 1344 length = length_save; 1345 1346 gen6_for_each_pde(pt, &ppgtt->pd, start, length, temp, pde) { 1347 DECLARE_BITMAP(tmp_bitmap, GEN6_PTES); 1348 1349 bitmap_zero(tmp_bitmap, GEN6_PTES); 1350 bitmap_set(tmp_bitmap, gen6_pte_index(start), 1351 gen6_pte_count(start, length)); 1352 1353 if (test_and_clear_bit(pde, new_page_tables)) 1354 gen6_write_pde(&ppgtt->pd, pde, pt); 1355 1356 trace_i915_page_table_entry_map(vm, pde, pt, 1357 gen6_pte_index(start), 1358 gen6_pte_count(start, length), 1359 GEN6_PTES); 1360 bitmap_or(pt->used_ptes, tmp_bitmap, pt->used_ptes, 1361 GEN6_PTES); 1362 } 1363 1364 WARN_ON(!bitmap_empty(new_page_tables, I915_PDES)); 1365 1366 /* Make sure write is complete before other code can use this page 1367 * table. Also require for WC mapped PTEs */ 1368 readl(dev_priv->gtt.gsm); 1369 1370 mark_tlbs_dirty(ppgtt); 1371 return 0; 1372 1373 unwind_out: 1374 for_each_set_bit(pde, new_page_tables, I915_PDES) { 1375 struct i915_page_table *pt = ppgtt->pd.page_table[pde]; 1376 1377 ppgtt->pd.page_table[pde] = ppgtt->scratch_pt; 1378 unmap_and_free_pt(pt, vm->dev); 1379 } 1380 1381 mark_tlbs_dirty(ppgtt); 1382 return ret; 1383 } 1384 1385 static void gen6_ppgtt_cleanup(struct i915_address_space *vm) 1386 { 1387 struct i915_hw_ppgtt *ppgtt = 1388 container_of(vm, struct i915_hw_ppgtt, base); 1389 struct i915_page_table *pt; 1390 uint32_t pde; 1391 1392 1393 drm_mm_remove_node(&ppgtt->node); 1394 1395 gen6_for_all_pdes(pt, ppgtt, pde) { 1396 if (pt != ppgtt->scratch_pt) 1397 unmap_and_free_pt(pt, ppgtt->base.dev); 1398 } 1399 1400 unmap_and_free_pt(ppgtt->scratch_pt, ppgtt->base.dev); 1401 unmap_and_free_pd(&ppgtt->pd, ppgtt->base.dev); 1402 } 1403 1404 static int gen6_ppgtt_allocate_page_directories(struct i915_hw_ppgtt *ppgtt) 1405 { 1406 struct drm_device *dev = ppgtt->base.dev; 1407 struct drm_i915_private *dev_priv = dev->dev_private; 1408 bool retried = false; 1409 int ret; 1410 1411 /* PPGTT PDEs reside in the GGTT and consists of 512 entries. The 1412 * allocator works in address space sizes, so it's multiplied by page 1413 * size. We allocate at the top of the GTT to avoid fragmentation. 1414 */ 1415 BUG_ON(!drm_mm_initialized(&dev_priv->gtt.base.mm)); 1416 ppgtt->scratch_pt = alloc_pt_single(ppgtt->base.dev); 1417 if (IS_ERR(ppgtt->scratch_pt)) 1418 return PTR_ERR(ppgtt->scratch_pt); 1419 1420 gen6_initialize_pt(&ppgtt->base, ppgtt->scratch_pt); 1421 1422 alloc: 1423 ret = drm_mm_insert_node_in_range_generic(&dev_priv->gtt.base.mm, 1424 &ppgtt->node, GEN6_PD_SIZE, 1425 GEN6_PD_ALIGN, 0, 1426 0, dev_priv->gtt.base.total, 1427 DRM_MM_TOPDOWN); 1428 if (ret == -ENOSPC && !retried) { 1429 ret = i915_gem_evict_something(dev, &dev_priv->gtt.base, 1430 GEN6_PD_SIZE, GEN6_PD_ALIGN, 1431 I915_CACHE_NONE, 1432 0, dev_priv->gtt.base.total, 1433 0); 1434 if (ret) 1435 goto err_out; 1436 1437 retried = true; 1438 goto alloc; 1439 } 1440 1441 if (ret) 1442 goto err_out; 1443 1444 1445 if (ppgtt->node.start < dev_priv->gtt.mappable_end) 1446 DRM_DEBUG("Forced to use aperture for PDEs\n"); 1447 1448 return 0; 1449 1450 err_out: 1451 unmap_and_free_pt(ppgtt->scratch_pt, ppgtt->base.dev); 1452 return ret; 1453 } 1454 1455 static int gen6_ppgtt_alloc(struct i915_hw_ppgtt *ppgtt) 1456 { 1457 return gen6_ppgtt_allocate_page_directories(ppgtt); 1458 } 1459 1460 static void gen6_scratch_va_range(struct i915_hw_ppgtt *ppgtt, 1461 uint64_t start, uint64_t length) 1462 { 1463 struct i915_page_table *unused; 1464 uint32_t pde, temp; 1465 1466 gen6_for_each_pde(unused, &ppgtt->pd, start, length, temp, pde) 1467 ppgtt->pd.page_table[pde] = ppgtt->scratch_pt; 1468 } 1469 1470 static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt) 1471 { 1472 struct drm_device *dev = ppgtt->base.dev; 1473 struct drm_i915_private *dev_priv = dev->dev_private; 1474 int ret; 1475 1476 ppgtt->base.pte_encode = dev_priv->gtt.base.pte_encode; 1477 if (IS_GEN6(dev)) { 1478 ppgtt->switch_mm = gen6_mm_switch; 1479 } else if (IS_HASWELL(dev)) { 1480 ppgtt->switch_mm = hsw_mm_switch; 1481 } else if (IS_GEN7(dev)) { 1482 ppgtt->switch_mm = gen7_mm_switch; 1483 } else 1484 BUG(); 1485 1486 if (intel_vgpu_active(dev)) 1487 ppgtt->switch_mm = vgpu_mm_switch; 1488 1489 ret = gen6_ppgtt_alloc(ppgtt); 1490 if (ret) 1491 return ret; 1492 1493 ppgtt->base.allocate_va_range = gen6_alloc_va_range; 1494 ppgtt->base.clear_range = gen6_ppgtt_clear_range; 1495 ppgtt->base.insert_entries = gen6_ppgtt_insert_entries; 1496 ppgtt->base.unbind_vma = ppgtt_unbind_vma; 1497 ppgtt->base.bind_vma = ppgtt_bind_vma; 1498 ppgtt->base.cleanup = gen6_ppgtt_cleanup; 1499 ppgtt->base.start = 0; 1500 ppgtt->base.total = I915_PDES * GEN6_PTES * PAGE_SIZE; 1501 ppgtt->debug_dump = gen6_dump_ppgtt; 1502 1503 ppgtt->pd.pd_offset = 1504 ppgtt->node.start / PAGE_SIZE * sizeof(gen6_pte_t); 1505 1506 ppgtt->pd_addr = (gen6_pte_t __iomem *)dev_priv->gtt.gsm + 1507 ppgtt->pd.pd_offset / sizeof(gen6_pte_t); 1508 1509 gen6_scratch_va_range(ppgtt, 0, ppgtt->base.total); 1510 1511 gen6_write_page_range(dev_priv, &ppgtt->pd, 0, ppgtt->base.total); 1512 1513 DRM_DEBUG_DRIVER("Allocated pde space (%lldM) at GTT entry: %llx\n", 1514 ppgtt->node.size >> 20, 1515 ppgtt->node.start / PAGE_SIZE); 1516 1517 DRM_DEBUG("Adding PPGTT at offset %x\n", 1518 ppgtt->pd.pd_offset << 10); 1519 1520 return 0; 1521 } 1522 1523 static int __hw_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt) 1524 { 1525 struct drm_i915_private *dev_priv = dev->dev_private; 1526 1527 ppgtt->base.dev = dev; 1528 ppgtt->base.scratch = dev_priv->gtt.base.scratch; 1529 1530 if (INTEL_INFO(dev)->gen < 8) 1531 return gen6_ppgtt_init(ppgtt); 1532 else 1533 return gen8_ppgtt_init(ppgtt); 1534 } 1535 int i915_ppgtt_init(struct drm_device *dev, struct i915_hw_ppgtt *ppgtt) 1536 { 1537 struct drm_i915_private *dev_priv = dev->dev_private; 1538 int ret = 0; 1539 1540 ret = __hw_ppgtt_init(dev, ppgtt); 1541 if (ret == 0) { 1542 kref_init(&ppgtt->ref); 1543 drm_mm_init(&ppgtt->base.mm, ppgtt->base.start, 1544 ppgtt->base.total); 1545 i915_init_vm(dev_priv, &ppgtt->base); 1546 } 1547 1548 return ret; 1549 } 1550 1551 int i915_ppgtt_init_hw(struct drm_device *dev) 1552 { 1553 struct drm_i915_private *dev_priv = dev->dev_private; 1554 struct intel_engine_cs *ring; 1555 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt; 1556 int i, ret = 0; 1557 1558 /* In the case of execlists, PPGTT is enabled by the context descriptor 1559 * and the PDPs are contained within the context itself. We don't 1560 * need to do anything here. */ 1561 if (i915.enable_execlists) 1562 return 0; 1563 1564 if (!USES_PPGTT(dev)) 1565 return 0; 1566 1567 if (IS_GEN6(dev)) 1568 gen6_ppgtt_enable(dev); 1569 else if (IS_GEN7(dev)) 1570 gen7_ppgtt_enable(dev); 1571 else if (INTEL_INFO(dev)->gen >= 8) 1572 gen8_ppgtt_enable(dev); 1573 else 1574 MISSING_CASE(INTEL_INFO(dev)->gen); 1575 1576 if (ppgtt) { 1577 for_each_ring(ring, dev_priv, i) { 1578 ret = ppgtt->switch_mm(ppgtt, ring); 1579 if (ret != 0) 1580 return ret; 1581 } 1582 } 1583 1584 return ret; 1585 } 1586 struct i915_hw_ppgtt * 1587 i915_ppgtt_create(struct drm_device *dev, struct drm_i915_file_private *fpriv) 1588 { 1589 struct i915_hw_ppgtt *ppgtt; 1590 int ret; 1591 1592 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL); 1593 if (!ppgtt) 1594 return ERR_PTR(-ENOMEM); 1595 1596 ret = i915_ppgtt_init(dev, ppgtt); 1597 if (ret) { 1598 kfree(ppgtt); 1599 return ERR_PTR(ret); 1600 } 1601 1602 ppgtt->file_priv = fpriv; 1603 1604 trace_i915_ppgtt_create(&ppgtt->base); 1605 1606 return ppgtt; 1607 } 1608 1609 void i915_ppgtt_release(struct kref *kref) 1610 { 1611 struct i915_hw_ppgtt *ppgtt = 1612 container_of(kref, struct i915_hw_ppgtt, ref); 1613 1614 trace_i915_ppgtt_release(&ppgtt->base); 1615 1616 /* vmas should already be unbound */ 1617 WARN_ON(!list_empty(&ppgtt->base.active_list)); 1618 WARN_ON(!list_empty(&ppgtt->base.inactive_list)); 1619 1620 list_del(&ppgtt->base.global_link); 1621 drm_mm_takedown(&ppgtt->base.mm); 1622 1623 ppgtt->base.cleanup(&ppgtt->base); 1624 kfree(ppgtt); 1625 } 1626 1627 extern int intel_iommu_gfx_mapped; 1628 /* Certain Gen5 chipsets require require idling the GPU before 1629 * unmapping anything from the GTT when VT-d is enabled. 1630 */ 1631 static bool needs_idle_maps(struct drm_device *dev) 1632 { 1633 #ifdef CONFIG_INTEL_IOMMU 1634 /* Query intel_iommu to see if we need the workaround. Presumably that 1635 * was loaded first. 1636 */ 1637 if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped) 1638 return true; 1639 #endif 1640 return false; 1641 } 1642 1643 static bool do_idling(struct drm_i915_private *dev_priv) 1644 { 1645 bool ret = dev_priv->mm.interruptible; 1646 1647 if (unlikely(dev_priv->gtt.do_idle_maps)) { 1648 dev_priv->mm.interruptible = false; 1649 if (i915_gpu_idle(dev_priv->dev)) { 1650 DRM_ERROR("Couldn't idle GPU\n"); 1651 /* Wait a bit, in hopes it avoids the hang */ 1652 udelay(10); 1653 } 1654 } 1655 1656 return ret; 1657 } 1658 1659 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible) 1660 { 1661 if (unlikely(dev_priv->gtt.do_idle_maps)) 1662 dev_priv->mm.interruptible = interruptible; 1663 } 1664 1665 void i915_check_and_clear_faults(struct drm_device *dev) 1666 { 1667 struct drm_i915_private *dev_priv = dev->dev_private; 1668 struct intel_engine_cs *ring; 1669 int i; 1670 1671 if (INTEL_INFO(dev)->gen < 6) 1672 return; 1673 1674 for_each_ring(ring, dev_priv, i) { 1675 u32 fault_reg; 1676 fault_reg = I915_READ(RING_FAULT_REG(ring)); 1677 if (fault_reg & RING_FAULT_VALID) { 1678 DRM_DEBUG_DRIVER("Unexpected fault\n" 1679 "\tAddr: 0x%08lx\n" 1680 "\tAddress space: %s\n" 1681 "\tSource ID: %d\n" 1682 "\tType: %d\n", 1683 fault_reg & PAGE_MASK, 1684 fault_reg & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT", 1685 RING_FAULT_SRCID(fault_reg), 1686 RING_FAULT_FAULT_TYPE(fault_reg)); 1687 I915_WRITE(RING_FAULT_REG(ring), 1688 fault_reg & ~RING_FAULT_VALID); 1689 } 1690 } 1691 POSTING_READ(RING_FAULT_REG(&dev_priv->ring[RCS])); 1692 } 1693 1694 static void i915_ggtt_flush(struct drm_i915_private *dev_priv) 1695 { 1696 if (INTEL_INFO(dev_priv->dev)->gen < 6) { 1697 intel_gtt_chipset_flush(); 1698 } else { 1699 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN); 1700 POSTING_READ(GFX_FLSH_CNTL_GEN6); 1701 } 1702 } 1703 1704 void i915_gem_suspend_gtt_mappings(struct drm_device *dev) 1705 { 1706 struct drm_i915_private *dev_priv = dev->dev_private; 1707 1708 /* Don't bother messing with faults pre GEN6 as we have little 1709 * documentation supporting that it's a good idea. 1710 */ 1711 if (INTEL_INFO(dev)->gen < 6) 1712 return; 1713 1714 i915_check_and_clear_faults(dev); 1715 1716 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base, 1717 dev_priv->gtt.base.start, 1718 dev_priv->gtt.base.total, 1719 true); 1720 1721 i915_ggtt_flush(dev_priv); 1722 } 1723 1724 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj) 1725 { 1726 if (obj->has_dma_mapping) 1727 return 0; 1728 1729 if (!dma_map_sg(&obj->base.dev->pdev->dev, 1730 obj->pages->sgl, obj->pages->nents, 1731 PCI_DMA_BIDIRECTIONAL)) 1732 return -ENOSPC; 1733 1734 return 0; 1735 } 1736 1737 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte) 1738 { 1739 #ifdef writeq 1740 writeq(pte, addr); 1741 #else 1742 iowrite32((u32)pte, addr); 1743 iowrite32(pte >> 32, addr + 4); 1744 #endif 1745 } 1746 1747 static void gen8_ggtt_insert_entries(struct i915_address_space *vm, 1748 struct sg_table *st, 1749 uint64_t start, 1750 enum i915_cache_level level, u32 unused) 1751 { 1752 struct drm_i915_private *dev_priv = vm->dev->dev_private; 1753 unsigned first_entry = start >> PAGE_SHIFT; 1754 gen8_pte_t __iomem *gtt_entries = 1755 (gen8_pte_t __iomem *)dev_priv->gtt.gsm + first_entry; 1756 int i = 0; 1757 struct sg_page_iter sg_iter; 1758 dma_addr_t addr = 0; /* shut up gcc */ 1759 1760 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) { 1761 addr = sg_dma_address(sg_iter.sg) + 1762 (sg_iter.sg_pgoffset << PAGE_SHIFT); 1763 gen8_set_pte(>t_entries[i], 1764 gen8_pte_encode(addr, level, true)); 1765 i++; 1766 } 1767 1768 /* 1769 * XXX: This serves as a posting read to make sure that the PTE has 1770 * actually been updated. There is some concern that even though 1771 * registers and PTEs are within the same BAR that they are potentially 1772 * of NUMA access patterns. Therefore, even with the way we assume 1773 * hardware should work, we must keep this posting read for paranoia. 1774 */ 1775 if (i != 0) 1776 WARN_ON(readq(>t_entries[i-1]) 1777 != gen8_pte_encode(addr, level, true)); 1778 1779 /* This next bit makes the above posting read even more important. We 1780 * want to flush the TLBs only after we're certain all the PTE updates 1781 * have finished. 1782 */ 1783 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN); 1784 POSTING_READ(GFX_FLSH_CNTL_GEN6); 1785 } 1786 1787 /* 1788 * Binds an object into the global gtt with the specified cache level. The object 1789 * will be accessible to the GPU via commands whose operands reference offsets 1790 * within the global GTT as well as accessible by the GPU through the GMADR 1791 * mapped BAR (dev_priv->mm.gtt->gtt). 1792 */ 1793 static void gen6_ggtt_insert_entries(struct i915_address_space *vm, 1794 struct sg_table *st, 1795 uint64_t start, 1796 enum i915_cache_level level, u32 flags) 1797 { 1798 struct drm_i915_private *dev_priv = vm->dev->dev_private; 1799 unsigned first_entry = start >> PAGE_SHIFT; 1800 gen6_pte_t __iomem *gtt_entries = 1801 (gen6_pte_t __iomem *)dev_priv->gtt.gsm + first_entry; 1802 int i = 0; 1803 struct sg_page_iter sg_iter; 1804 dma_addr_t addr = 0; 1805 1806 for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) { 1807 addr = sg_page_iter_dma_address(&sg_iter); 1808 iowrite32(vm->pte_encode(addr, level, true, flags), >t_entries[i]); 1809 i++; 1810 } 1811 1812 /* XXX: This serves as a posting read to make sure that the PTE has 1813 * actually been updated. There is some concern that even though 1814 * registers and PTEs are within the same BAR that they are potentially 1815 * of NUMA access patterns. Therefore, even with the way we assume 1816 * hardware should work, we must keep this posting read for paranoia. 1817 */ 1818 if (i != 0) { 1819 unsigned long gtt = readl(>t_entries[i-1]); 1820 WARN_ON(gtt != vm->pte_encode(addr, level, true, flags)); 1821 } 1822 1823 /* This next bit makes the above posting read even more important. We 1824 * want to flush the TLBs only after we're certain all the PTE updates 1825 * have finished. 1826 */ 1827 I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN); 1828 POSTING_READ(GFX_FLSH_CNTL_GEN6); 1829 } 1830 1831 static void gen8_ggtt_clear_range(struct i915_address_space *vm, 1832 uint64_t start, 1833 uint64_t length, 1834 bool use_scratch) 1835 { 1836 struct drm_i915_private *dev_priv = vm->dev->dev_private; 1837 unsigned first_entry = start >> PAGE_SHIFT; 1838 unsigned num_entries = length >> PAGE_SHIFT; 1839 gen8_pte_t scratch_pte, __iomem *gtt_base = 1840 (gen8_pte_t __iomem *) dev_priv->gtt.gsm + first_entry; 1841 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry; 1842 int i; 1843 1844 if (WARN(num_entries > max_entries, 1845 "First entry = %d; Num entries = %d (max=%d)\n", 1846 first_entry, num_entries, max_entries)) 1847 num_entries = max_entries; 1848 1849 scratch_pte = gen8_pte_encode(vm->scratch.addr, 1850 I915_CACHE_LLC, 1851 use_scratch); 1852 for (i = 0; i < num_entries; i++) 1853 gen8_set_pte(>t_base[i], scratch_pte); 1854 readl(gtt_base); 1855 } 1856 1857 static void gen6_ggtt_clear_range(struct i915_address_space *vm, 1858 uint64_t start, 1859 uint64_t length, 1860 bool use_scratch) 1861 { 1862 struct drm_i915_private *dev_priv = vm->dev->dev_private; 1863 unsigned first_entry = start >> PAGE_SHIFT; 1864 unsigned num_entries = length >> PAGE_SHIFT; 1865 gen6_pte_t scratch_pte, __iomem *gtt_base = 1866 (gen6_pte_t __iomem *) dev_priv->gtt.gsm + first_entry; 1867 const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry; 1868 int i; 1869 1870 if (WARN(num_entries > max_entries, 1871 "First entry = %d; Num entries = %d (max=%d)\n", 1872 first_entry, num_entries, max_entries)) 1873 num_entries = max_entries; 1874 1875 scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC, use_scratch, 0); 1876 1877 for (i = 0; i < num_entries; i++) 1878 iowrite32(scratch_pte, >t_base[i]); 1879 readl(gtt_base); 1880 } 1881 1882 static void i915_ggtt_insert_entries(struct i915_address_space *vm, 1883 struct sg_table *pages, 1884 uint64_t start, 1885 enum i915_cache_level cache_level, u32 unused) 1886 { 1887 unsigned int flags = (cache_level == I915_CACHE_NONE) ? 1888 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY; 1889 1890 intel_gtt_insert_sg_entries(pages, start >> PAGE_SHIFT, flags); 1891 1892 } 1893 1894 static void i915_ggtt_clear_range(struct i915_address_space *vm, 1895 uint64_t start, 1896 uint64_t length, 1897 bool unused) 1898 { 1899 unsigned first_entry = start >> PAGE_SHIFT; 1900 unsigned num_entries = length >> PAGE_SHIFT; 1901 intel_gtt_clear_range(first_entry, num_entries); 1902 } 1903 1904 static int ggtt_bind_vma(struct i915_vma *vma, 1905 enum i915_cache_level cache_level, 1906 u32 flags) 1907 { 1908 struct drm_device *dev = vma->vm->dev; 1909 struct drm_i915_private *dev_priv = dev->dev_private; 1910 struct drm_i915_gem_object *obj = vma->obj; 1911 struct sg_table *pages = obj->pages; 1912 u32 pte_flags = 0; 1913 int ret; 1914 1915 ret = i915_get_ggtt_vma_pages(vma); 1916 if (ret) 1917 return ret; 1918 pages = vma->ggtt_view.pages; 1919 1920 /* Currently applicable only to VLV */ 1921 if (obj->gt_ro) 1922 pte_flags |= PTE_READ_ONLY; 1923 1924 1925 if (!dev_priv->mm.aliasing_ppgtt || flags & GLOBAL_BIND) { 1926 vma->vm->insert_entries(vma->vm, pages, 1927 vma->node.start, 1928 cache_level, pte_flags); 1929 } 1930 1931 if (dev_priv->mm.aliasing_ppgtt && flags & LOCAL_BIND) { 1932 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt; 1933 appgtt->base.insert_entries(&appgtt->base, pages, 1934 vma->node.start, 1935 cache_level, pte_flags); 1936 } 1937 1938 return 0; 1939 } 1940 1941 static void ggtt_unbind_vma(struct i915_vma *vma) 1942 { 1943 struct drm_device *dev = vma->vm->dev; 1944 struct drm_i915_private *dev_priv = dev->dev_private; 1945 struct drm_i915_gem_object *obj = vma->obj; 1946 const uint64_t size = min_t(uint64_t, 1947 obj->base.size, 1948 vma->node.size); 1949 1950 if (vma->bound & GLOBAL_BIND) { 1951 vma->vm->clear_range(vma->vm, 1952 vma->node.start, 1953 size, 1954 true); 1955 } 1956 1957 if (dev_priv->mm.aliasing_ppgtt && vma->bound & LOCAL_BIND) { 1958 struct i915_hw_ppgtt *appgtt = dev_priv->mm.aliasing_ppgtt; 1959 1960 appgtt->base.clear_range(&appgtt->base, 1961 vma->node.start, 1962 size, 1963 true); 1964 } 1965 } 1966 1967 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj) 1968 { 1969 struct drm_device *dev = obj->base.dev; 1970 struct drm_i915_private *dev_priv = dev->dev_private; 1971 bool interruptible; 1972 1973 interruptible = do_idling(dev_priv); 1974 1975 if (!obj->has_dma_mapping) 1976 dma_unmap_sg(&dev->pdev->dev, 1977 obj->pages->sgl, obj->pages->nents, 1978 PCI_DMA_BIDIRECTIONAL); 1979 1980 undo_idling(dev_priv, interruptible); 1981 } 1982 1983 static void i915_gtt_color_adjust(struct drm_mm_node *node, 1984 unsigned long color, 1985 u64 *start, 1986 u64 *end) 1987 { 1988 if (node->color != color) 1989 *start += 4096; 1990 1991 if (!list_empty(&node->node_list)) { 1992 node = list_entry(node->node_list.next, 1993 struct drm_mm_node, 1994 node_list); 1995 if (node->allocated && node->color != color) 1996 *end -= 4096; 1997 } 1998 } 1999 2000 static int i915_gem_setup_global_gtt(struct drm_device *dev, 2001 unsigned long start, 2002 unsigned long mappable_end, 2003 unsigned long end) 2004 { 2005 /* Let GEM Manage all of the aperture. 2006 * 2007 * However, leave one page at the end still bound to the scratch page. 2008 * There are a number of places where the hardware apparently prefetches 2009 * past the end of the object, and we've seen multiple hangs with the 2010 * GPU head pointer stuck in a batchbuffer bound at the last page of the 2011 * aperture. One page should be enough to keep any prefetching inside 2012 * of the aperture. 2013 */ 2014 struct drm_i915_private *dev_priv = dev->dev_private; 2015 struct i915_address_space *ggtt_vm = &dev_priv->gtt.base; 2016 struct drm_mm_node *entry; 2017 struct drm_i915_gem_object *obj; 2018 unsigned long hole_start, hole_end; 2019 int ret; 2020 2021 BUG_ON(mappable_end > end); 2022 2023 /* Subtract the guard page ... */ 2024 drm_mm_init(&ggtt_vm->mm, start, end - start - PAGE_SIZE); 2025 2026 dev_priv->gtt.base.start = start; 2027 dev_priv->gtt.base.total = end - start; 2028 2029 if (intel_vgpu_active(dev)) { 2030 ret = intel_vgt_balloon(dev); 2031 if (ret) 2032 return ret; 2033 } 2034 2035 if (!HAS_LLC(dev)) 2036 dev_priv->gtt.base.mm.color_adjust = i915_gtt_color_adjust; 2037 2038 /* Mark any preallocated objects as occupied */ 2039 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) { 2040 struct i915_vma *vma = i915_gem_obj_to_vma(obj, ggtt_vm); 2041 2042 DRM_DEBUG_KMS("reserving preallocated space: %lx + %zx\n", 2043 i915_gem_obj_ggtt_offset(obj), obj->base.size); 2044 2045 WARN_ON(i915_gem_obj_ggtt_bound(obj)); 2046 ret = drm_mm_reserve_node(&ggtt_vm->mm, &vma->node); 2047 if (ret) { 2048 DRM_DEBUG_KMS("Reservation failed: %i\n", ret); 2049 return ret; 2050 } 2051 vma->bound |= GLOBAL_BIND; 2052 } 2053 2054 /* Clear any non-preallocated blocks */ 2055 drm_mm_for_each_hole(entry, &ggtt_vm->mm, hole_start, hole_end) { 2056 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n", 2057 hole_start, hole_end); 2058 ggtt_vm->clear_range(ggtt_vm, hole_start, 2059 hole_end - hole_start, true); 2060 } 2061 2062 /* And finally clear the reserved guard page */ 2063 ggtt_vm->clear_range(ggtt_vm, end - PAGE_SIZE, PAGE_SIZE, true); 2064 2065 if (USES_PPGTT(dev) && !USES_FULL_PPGTT(dev)) { 2066 struct i915_hw_ppgtt *ppgtt; 2067 2068 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL); 2069 if (!ppgtt) 2070 return -ENOMEM; 2071 2072 ret = __hw_ppgtt_init(dev, ppgtt); 2073 if (ret) { 2074 ppgtt->base.cleanup(&ppgtt->base); 2075 kfree(ppgtt); 2076 return ret; 2077 } 2078 2079 if (ppgtt->base.allocate_va_range) 2080 ret = ppgtt->base.allocate_va_range(&ppgtt->base, 0, 2081 ppgtt->base.total); 2082 if (ret) { 2083 ppgtt->base.cleanup(&ppgtt->base); 2084 kfree(ppgtt); 2085 return ret; 2086 } 2087 2088 ppgtt->base.clear_range(&ppgtt->base, 2089 ppgtt->base.start, 2090 ppgtt->base.total, 2091 true); 2092 2093 dev_priv->mm.aliasing_ppgtt = ppgtt; 2094 } 2095 2096 return 0; 2097 } 2098 2099 void i915_gem_init_global_gtt(struct drm_device *dev) 2100 { 2101 struct drm_i915_private *dev_priv = dev->dev_private; 2102 unsigned long gtt_size, mappable_size; 2103 2104 gtt_size = dev_priv->gtt.base.total; 2105 mappable_size = dev_priv->gtt.mappable_end; 2106 2107 i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size); 2108 } 2109 2110 void i915_global_gtt_cleanup(struct drm_device *dev) 2111 { 2112 struct drm_i915_private *dev_priv = dev->dev_private; 2113 struct i915_address_space *vm = &dev_priv->gtt.base; 2114 2115 if (dev_priv->mm.aliasing_ppgtt) { 2116 struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt; 2117 2118 ppgtt->base.cleanup(&ppgtt->base); 2119 } 2120 2121 if (drm_mm_initialized(&vm->mm)) { 2122 if (intel_vgpu_active(dev)) 2123 intel_vgt_deballoon(); 2124 2125 drm_mm_takedown(&vm->mm); 2126 list_del(&vm->global_link); 2127 } 2128 2129 vm->cleanup(vm); 2130 } 2131 2132 static int setup_scratch_page(struct drm_device *dev) 2133 { 2134 struct drm_i915_private *dev_priv = dev->dev_private; 2135 struct page *page; 2136 dma_addr_t dma_addr; 2137 2138 page = alloc_page(GFP_KERNEL | GFP_DMA32 | __GFP_ZERO); 2139 if (page == NULL) 2140 return -ENOMEM; 2141 set_pages_uc(page, 1); 2142 2143 #ifdef CONFIG_INTEL_IOMMU 2144 dma_addr = pci_map_page(dev->pdev, page, 0, PAGE_SIZE, 2145 PCI_DMA_BIDIRECTIONAL); 2146 if (pci_dma_mapping_error(dev->pdev, dma_addr)) 2147 return -EINVAL; 2148 #else 2149 dma_addr = page_to_phys(page); 2150 #endif 2151 dev_priv->gtt.base.scratch.page = page; 2152 dev_priv->gtt.base.scratch.addr = dma_addr; 2153 2154 return 0; 2155 } 2156 2157 static void teardown_scratch_page(struct drm_device *dev) 2158 { 2159 struct drm_i915_private *dev_priv = dev->dev_private; 2160 struct page *page = dev_priv->gtt.base.scratch.page; 2161 2162 set_pages_wb(page, 1); 2163 pci_unmap_page(dev->pdev, dev_priv->gtt.base.scratch.addr, 2164 PAGE_SIZE, PCI_DMA_BIDIRECTIONAL); 2165 __free_page(page); 2166 } 2167 2168 static unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl) 2169 { 2170 snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT; 2171 snb_gmch_ctl &= SNB_GMCH_GGMS_MASK; 2172 return snb_gmch_ctl << 20; 2173 } 2174 2175 static unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl) 2176 { 2177 bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT; 2178 bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK; 2179 if (bdw_gmch_ctl) 2180 bdw_gmch_ctl = 1 << bdw_gmch_ctl; 2181 2182 #ifdef CONFIG_X86_32 2183 /* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * PAGE_SIZE */ 2184 if (bdw_gmch_ctl > 4) 2185 bdw_gmch_ctl = 4; 2186 #endif 2187 2188 return bdw_gmch_ctl << 20; 2189 } 2190 2191 static unsigned int chv_get_total_gtt_size(u16 gmch_ctrl) 2192 { 2193 gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT; 2194 gmch_ctrl &= SNB_GMCH_GGMS_MASK; 2195 2196 if (gmch_ctrl) 2197 return 1 << (20 + gmch_ctrl); 2198 2199 return 0; 2200 } 2201 2202 static size_t gen6_get_stolen_size(u16 snb_gmch_ctl) 2203 { 2204 snb_gmch_ctl >>= SNB_GMCH_GMS_SHIFT; 2205 snb_gmch_ctl &= SNB_GMCH_GMS_MASK; 2206 return snb_gmch_ctl << 25; /* 32 MB units */ 2207 } 2208 2209 static size_t gen8_get_stolen_size(u16 bdw_gmch_ctl) 2210 { 2211 bdw_gmch_ctl >>= BDW_GMCH_GMS_SHIFT; 2212 bdw_gmch_ctl &= BDW_GMCH_GMS_MASK; 2213 return bdw_gmch_ctl << 25; /* 32 MB units */ 2214 } 2215 2216 static size_t chv_get_stolen_size(u16 gmch_ctrl) 2217 { 2218 gmch_ctrl >>= SNB_GMCH_GMS_SHIFT; 2219 gmch_ctrl &= SNB_GMCH_GMS_MASK; 2220 2221 /* 2222 * 0x0 to 0x10: 32MB increments starting at 0MB 2223 * 0x11 to 0x16: 4MB increments starting at 8MB 2224 * 0x17 to 0x1d: 4MB increments start at 36MB 2225 */ 2226 if (gmch_ctrl < 0x11) 2227 return gmch_ctrl << 25; 2228 else if (gmch_ctrl < 0x17) 2229 return (gmch_ctrl - 0x11 + 2) << 22; 2230 else 2231 return (gmch_ctrl - 0x17 + 9) << 22; 2232 } 2233 2234 static size_t gen9_get_stolen_size(u16 gen9_gmch_ctl) 2235 { 2236 gen9_gmch_ctl >>= BDW_GMCH_GMS_SHIFT; 2237 gen9_gmch_ctl &= BDW_GMCH_GMS_MASK; 2238 2239 if (gen9_gmch_ctl < 0xf0) 2240 return gen9_gmch_ctl << 25; /* 32 MB units */ 2241 else 2242 /* 4MB increments starting at 0xf0 for 4MB */ 2243 return (gen9_gmch_ctl - 0xf0 + 1) << 22; 2244 } 2245 2246 static int ggtt_probe_common(struct drm_device *dev, 2247 size_t gtt_size) 2248 { 2249 struct drm_i915_private *dev_priv = dev->dev_private; 2250 phys_addr_t gtt_phys_addr; 2251 int ret; 2252 2253 /* For Modern GENs the PTEs and register space are split in the BAR */ 2254 gtt_phys_addr = pci_resource_start(dev->pdev, 0) + 2255 (pci_resource_len(dev->pdev, 0) / 2); 2256 2257 /* 2258 * On BXT writes larger than 64 bit to the GTT pagetable range will be 2259 * dropped. For WC mappings in general we have 64 byte burst writes 2260 * when the WC buffer is flushed, so we can't use it, but have to 2261 * resort to an uncached mapping. The WC issue is easily caught by the 2262 * readback check when writing GTT PTE entries. 2263 */ 2264 if (IS_BROXTON(dev)) 2265 dev_priv->gtt.gsm = ioremap_nocache(gtt_phys_addr, gtt_size); 2266 else 2267 dev_priv->gtt.gsm = ioremap_wc(gtt_phys_addr, gtt_size); 2268 if (!dev_priv->gtt.gsm) { 2269 DRM_ERROR("Failed to map the gtt page table\n"); 2270 return -ENOMEM; 2271 } 2272 2273 ret = setup_scratch_page(dev); 2274 if (ret) { 2275 DRM_ERROR("Scratch setup failed\n"); 2276 /* iounmap will also get called at remove, but meh */ 2277 iounmap(dev_priv->gtt.gsm); 2278 } 2279 2280 return ret; 2281 } 2282 2283 /* The GGTT and PPGTT need a private PPAT setup in order to handle cacheability 2284 * bits. When using advanced contexts each context stores its own PAT, but 2285 * writing this data shouldn't be harmful even in those cases. */ 2286 static void bdw_setup_private_ppat(struct drm_i915_private *dev_priv) 2287 { 2288 uint64_t pat; 2289 2290 pat = GEN8_PPAT(0, GEN8_PPAT_WB | GEN8_PPAT_LLC) | /* for normal objects, no eLLC */ 2291 GEN8_PPAT(1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC) | /* for something pointing to ptes? */ 2292 GEN8_PPAT(2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC) | /* for scanout with eLLC */ 2293 GEN8_PPAT(3, GEN8_PPAT_UC) | /* Uncached objects, mostly for scanout */ 2294 GEN8_PPAT(4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0)) | 2295 GEN8_PPAT(5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1)) | 2296 GEN8_PPAT(6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2)) | 2297 GEN8_PPAT(7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3)); 2298 2299 if (!USES_PPGTT(dev_priv->dev)) 2300 /* Spec: "For GGTT, there is NO pat_sel[2:0] from the entry, 2301 * so RTL will always use the value corresponding to 2302 * pat_sel = 000". 2303 * So let's disable cache for GGTT to avoid screen corruptions. 2304 * MOCS still can be used though. 2305 * - System agent ggtt writes (i.e. cpu gtt mmaps) already work 2306 * before this patch, i.e. the same uncached + snooping access 2307 * like on gen6/7 seems to be in effect. 2308 * - So this just fixes blitter/render access. Again it looks 2309 * like it's not just uncached access, but uncached + snooping. 2310 * So we can still hold onto all our assumptions wrt cpu 2311 * clflushing on LLC machines. 2312 */ 2313 pat = GEN8_PPAT(0, GEN8_PPAT_UC); 2314 2315 /* XXX: spec defines this as 2 distinct registers. It's unclear if a 64b 2316 * write would work. */ 2317 I915_WRITE(GEN8_PRIVATE_PAT, pat); 2318 I915_WRITE(GEN8_PRIVATE_PAT + 4, pat >> 32); 2319 } 2320 2321 static void chv_setup_private_ppat(struct drm_i915_private *dev_priv) 2322 { 2323 uint64_t pat; 2324 2325 /* 2326 * Map WB on BDW to snooped on CHV. 2327 * 2328 * Only the snoop bit has meaning for CHV, the rest is 2329 * ignored. 2330 * 2331 * The hardware will never snoop for certain types of accesses: 2332 * - CPU GTT (GMADR->GGTT->no snoop->memory) 2333 * - PPGTT page tables 2334 * - some other special cycles 2335 * 2336 * As with BDW, we also need to consider the following for GT accesses: 2337 * "For GGTT, there is NO pat_sel[2:0] from the entry, 2338 * so RTL will always use the value corresponding to 2339 * pat_sel = 000". 2340 * Which means we must set the snoop bit in PAT entry 0 2341 * in order to keep the global status page working. 2342 */ 2343 pat = GEN8_PPAT(0, CHV_PPAT_SNOOP) | 2344 GEN8_PPAT(1, 0) | 2345 GEN8_PPAT(2, 0) | 2346 GEN8_PPAT(3, 0) | 2347 GEN8_PPAT(4, CHV_PPAT_SNOOP) | 2348 GEN8_PPAT(5, CHV_PPAT_SNOOP) | 2349 GEN8_PPAT(6, CHV_PPAT_SNOOP) | 2350 GEN8_PPAT(7, CHV_PPAT_SNOOP); 2351 2352 I915_WRITE(GEN8_PRIVATE_PAT, pat); 2353 I915_WRITE(GEN8_PRIVATE_PAT + 4, pat >> 32); 2354 } 2355 2356 static int gen8_gmch_probe(struct drm_device *dev, 2357 size_t *gtt_total, 2358 size_t *stolen, 2359 phys_addr_t *mappable_base, 2360 unsigned long *mappable_end) 2361 { 2362 struct drm_i915_private *dev_priv = dev->dev_private; 2363 unsigned int gtt_size; 2364 u16 snb_gmch_ctl; 2365 int ret; 2366 2367 /* TODO: We're not aware of mappable constraints on gen8 yet */ 2368 *mappable_base = pci_resource_start(dev->pdev, 2); 2369 *mappable_end = pci_resource_len(dev->pdev, 2); 2370 2371 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(39))) 2372 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(39)); 2373 2374 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl); 2375 2376 if (INTEL_INFO(dev)->gen >= 9) { 2377 *stolen = gen9_get_stolen_size(snb_gmch_ctl); 2378 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl); 2379 } else if (IS_CHERRYVIEW(dev)) { 2380 *stolen = chv_get_stolen_size(snb_gmch_ctl); 2381 gtt_size = chv_get_total_gtt_size(snb_gmch_ctl); 2382 } else { 2383 *stolen = gen8_get_stolen_size(snb_gmch_ctl); 2384 gtt_size = gen8_get_total_gtt_size(snb_gmch_ctl); 2385 } 2386 2387 *gtt_total = (gtt_size / sizeof(gen8_pte_t)) << PAGE_SHIFT; 2388 2389 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev)) 2390 chv_setup_private_ppat(dev_priv); 2391 else 2392 bdw_setup_private_ppat(dev_priv); 2393 2394 ret = ggtt_probe_common(dev, gtt_size); 2395 2396 dev_priv->gtt.base.clear_range = gen8_ggtt_clear_range; 2397 dev_priv->gtt.base.insert_entries = gen8_ggtt_insert_entries; 2398 dev_priv->gtt.base.bind_vma = ggtt_bind_vma; 2399 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma; 2400 2401 return ret; 2402 } 2403 2404 static int gen6_gmch_probe(struct drm_device *dev, 2405 size_t *gtt_total, 2406 size_t *stolen, 2407 phys_addr_t *mappable_base, 2408 unsigned long *mappable_end) 2409 { 2410 struct drm_i915_private *dev_priv = dev->dev_private; 2411 unsigned int gtt_size; 2412 u16 snb_gmch_ctl; 2413 int ret; 2414 2415 *mappable_base = pci_resource_start(dev->pdev, 2); 2416 *mappable_end = pci_resource_len(dev->pdev, 2); 2417 2418 /* 64/512MB is the current min/max we actually know of, but this is just 2419 * a coarse sanity check. 2420 */ 2421 if ((*mappable_end < (64<<20) || (*mappable_end > (512<<20)))) { 2422 DRM_ERROR("Unknown GMADR size (%lx)\n", 2423 dev_priv->gtt.mappable_end); 2424 return -ENXIO; 2425 } 2426 2427 if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(40))) 2428 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(40)); 2429 pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl); 2430 2431 *stolen = gen6_get_stolen_size(snb_gmch_ctl); 2432 2433 gtt_size = gen6_get_total_gtt_size(snb_gmch_ctl); 2434 *gtt_total = (gtt_size / sizeof(gen6_pte_t)) << PAGE_SHIFT; 2435 2436 ret = ggtt_probe_common(dev, gtt_size); 2437 2438 dev_priv->gtt.base.clear_range = gen6_ggtt_clear_range; 2439 dev_priv->gtt.base.insert_entries = gen6_ggtt_insert_entries; 2440 dev_priv->gtt.base.bind_vma = ggtt_bind_vma; 2441 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma; 2442 2443 return ret; 2444 } 2445 2446 static void gen6_gmch_remove(struct i915_address_space *vm) 2447 { 2448 2449 struct i915_gtt *gtt = container_of(vm, struct i915_gtt, base); 2450 2451 iounmap(gtt->gsm); 2452 teardown_scratch_page(vm->dev); 2453 } 2454 2455 static int i915_gmch_probe(struct drm_device *dev, 2456 size_t *gtt_total, 2457 size_t *stolen, 2458 phys_addr_t *mappable_base, 2459 unsigned long *mappable_end) 2460 { 2461 struct drm_i915_private *dev_priv = dev->dev_private; 2462 int ret; 2463 2464 ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->dev->pdev, NULL); 2465 if (!ret) { 2466 DRM_ERROR("failed to set up gmch\n"); 2467 return -EIO; 2468 } 2469 2470 intel_gtt_get(gtt_total, stolen, mappable_base, mappable_end); 2471 2472 dev_priv->gtt.do_idle_maps = needs_idle_maps(dev_priv->dev); 2473 dev_priv->gtt.base.insert_entries = i915_ggtt_insert_entries; 2474 dev_priv->gtt.base.clear_range = i915_ggtt_clear_range; 2475 dev_priv->gtt.base.bind_vma = ggtt_bind_vma; 2476 dev_priv->gtt.base.unbind_vma = ggtt_unbind_vma; 2477 2478 if (unlikely(dev_priv->gtt.do_idle_maps)) 2479 DRM_INFO("applying Ironlake quirks for intel_iommu\n"); 2480 2481 return 0; 2482 } 2483 2484 static void i915_gmch_remove(struct i915_address_space *vm) 2485 { 2486 intel_gmch_remove(); 2487 } 2488 2489 int i915_gem_gtt_init(struct drm_device *dev) 2490 { 2491 struct drm_i915_private *dev_priv = dev->dev_private; 2492 struct i915_gtt *gtt = &dev_priv->gtt; 2493 int ret; 2494 2495 if (INTEL_INFO(dev)->gen <= 5) { 2496 gtt->gtt_probe = i915_gmch_probe; 2497 gtt->base.cleanup = i915_gmch_remove; 2498 } else if (INTEL_INFO(dev)->gen < 8) { 2499 gtt->gtt_probe = gen6_gmch_probe; 2500 gtt->base.cleanup = gen6_gmch_remove; 2501 if (IS_HASWELL(dev) && dev_priv->ellc_size) 2502 gtt->base.pte_encode = iris_pte_encode; 2503 else if (IS_HASWELL(dev)) 2504 gtt->base.pte_encode = hsw_pte_encode; 2505 else if (IS_VALLEYVIEW(dev)) 2506 gtt->base.pte_encode = byt_pte_encode; 2507 else if (INTEL_INFO(dev)->gen >= 7) 2508 gtt->base.pte_encode = ivb_pte_encode; 2509 else 2510 gtt->base.pte_encode = snb_pte_encode; 2511 } else { 2512 dev_priv->gtt.gtt_probe = gen8_gmch_probe; 2513 dev_priv->gtt.base.cleanup = gen6_gmch_remove; 2514 } 2515 2516 ret = gtt->gtt_probe(dev, >t->base.total, >t->stolen_size, 2517 >t->mappable_base, >t->mappable_end); 2518 if (ret) 2519 return ret; 2520 2521 gtt->base.dev = dev; 2522 2523 /* GMADR is the PCI mmio aperture into the global GTT. */ 2524 DRM_INFO("Memory usable by graphics device = %zdM\n", 2525 gtt->base.total >> 20); 2526 DRM_DEBUG_DRIVER("GMADR size = %ldM\n", gtt->mappable_end >> 20); 2527 DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n", gtt->stolen_size >> 20); 2528 #ifdef CONFIG_INTEL_IOMMU 2529 if (intel_iommu_gfx_mapped) 2530 DRM_INFO("VT-d active for gfx access\n"); 2531 #endif 2532 /* 2533 * i915.enable_ppgtt is read-only, so do an early pass to validate the 2534 * user's requested state against the hardware/driver capabilities. We 2535 * do this now so that we can print out any log messages once rather 2536 * than every time we check intel_enable_ppgtt(). 2537 */ 2538 i915.enable_ppgtt = sanitize_enable_ppgtt(dev, i915.enable_ppgtt); 2539 DRM_DEBUG_DRIVER("ppgtt mode: %i\n", i915.enable_ppgtt); 2540 2541 return 0; 2542 } 2543 2544 void i915_gem_restore_gtt_mappings(struct drm_device *dev) 2545 { 2546 struct drm_i915_private *dev_priv = dev->dev_private; 2547 struct drm_i915_gem_object *obj; 2548 struct i915_address_space *vm; 2549 2550 i915_check_and_clear_faults(dev); 2551 2552 /* First fill our portion of the GTT with scratch pages */ 2553 dev_priv->gtt.base.clear_range(&dev_priv->gtt.base, 2554 dev_priv->gtt.base.start, 2555 dev_priv->gtt.base.total, 2556 true); 2557 2558 list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) { 2559 struct i915_vma *vma = i915_gem_obj_to_vma(obj, 2560 &dev_priv->gtt.base); 2561 if (!vma) 2562 continue; 2563 2564 i915_gem_clflush_object(obj, obj->pin_display); 2565 WARN_ON(i915_vma_bind(vma, obj->cache_level, PIN_UPDATE)); 2566 } 2567 2568 2569 if (INTEL_INFO(dev)->gen >= 8) { 2570 if (IS_CHERRYVIEW(dev) || IS_BROXTON(dev)) 2571 chv_setup_private_ppat(dev_priv); 2572 else 2573 bdw_setup_private_ppat(dev_priv); 2574 2575 return; 2576 } 2577 2578 if (USES_PPGTT(dev)) { 2579 list_for_each_entry(vm, &dev_priv->vm_list, global_link) { 2580 /* TODO: Perhaps it shouldn't be gen6 specific */ 2581 2582 struct i915_hw_ppgtt *ppgtt = 2583 container_of(vm, struct i915_hw_ppgtt, 2584 base); 2585 2586 if (i915_is_ggtt(vm)) 2587 ppgtt = dev_priv->mm.aliasing_ppgtt; 2588 2589 gen6_write_page_range(dev_priv, &ppgtt->pd, 2590 0, ppgtt->base.total); 2591 } 2592 } 2593 2594 i915_ggtt_flush(dev_priv); 2595 } 2596 2597 static struct i915_vma * 2598 __i915_gem_vma_create(struct drm_i915_gem_object *obj, 2599 struct i915_address_space *vm, 2600 const struct i915_ggtt_view *ggtt_view) 2601 { 2602 struct i915_vma *vma; 2603 2604 if (WARN_ON(i915_is_ggtt(vm) != !!ggtt_view)) 2605 return ERR_PTR(-EINVAL); 2606 2607 vma = kmem_cache_zalloc(to_i915(obj->base.dev)->vmas, GFP_KERNEL); 2608 if (vma == NULL) 2609 return ERR_PTR(-ENOMEM); 2610 2611 INIT_LIST_HEAD(&vma->vma_link); 2612 INIT_LIST_HEAD(&vma->mm_list); 2613 INIT_LIST_HEAD(&vma->exec_list); 2614 vma->vm = vm; 2615 vma->obj = obj; 2616 2617 if (i915_is_ggtt(vm)) 2618 vma->ggtt_view = *ggtt_view; 2619 2620 list_add_tail(&vma->vma_link, &obj->vma_list); 2621 if (!i915_is_ggtt(vm)) 2622 i915_ppgtt_get(i915_vm_to_ppgtt(vm)); 2623 2624 return vma; 2625 } 2626 2627 struct i915_vma * 2628 i915_gem_obj_lookup_or_create_vma(struct drm_i915_gem_object *obj, 2629 struct i915_address_space *vm) 2630 { 2631 struct i915_vma *vma; 2632 2633 vma = i915_gem_obj_to_vma(obj, vm); 2634 if (!vma) 2635 vma = __i915_gem_vma_create(obj, vm, 2636 i915_is_ggtt(vm) ? &i915_ggtt_view_normal : NULL); 2637 2638 return vma; 2639 } 2640 2641 struct i915_vma * 2642 i915_gem_obj_lookup_or_create_ggtt_vma(struct drm_i915_gem_object *obj, 2643 const struct i915_ggtt_view *view) 2644 { 2645 struct i915_address_space *ggtt = i915_obj_to_ggtt(obj); 2646 struct i915_vma *vma; 2647 2648 if (WARN_ON(!view)) 2649 return ERR_PTR(-EINVAL); 2650 2651 vma = i915_gem_obj_to_ggtt_view(obj, view); 2652 2653 if (IS_ERR(vma)) 2654 return vma; 2655 2656 if (!vma) 2657 vma = __i915_gem_vma_create(obj, ggtt, view); 2658 2659 return vma; 2660 2661 } 2662 2663 static void 2664 rotate_pages(dma_addr_t *in, unsigned int width, unsigned int height, 2665 struct sg_table *st) 2666 { 2667 unsigned int column, row; 2668 unsigned int src_idx; 2669 struct scatterlist *sg = st->sgl; 2670 2671 st->nents = 0; 2672 2673 for (column = 0; column < width; column++) { 2674 src_idx = width * (height - 1) + column; 2675 for (row = 0; row < height; row++) { 2676 st->nents++; 2677 /* We don't need the pages, but need to initialize 2678 * the entries so the sg list can be happily traversed. 2679 * The only thing we need are DMA addresses. 2680 */ 2681 sg_set_page(sg, NULL, PAGE_SIZE, 0); 2682 sg_dma_address(sg) = in[src_idx]; 2683 sg_dma_len(sg) = PAGE_SIZE; 2684 sg = sg_next(sg); 2685 src_idx -= width; 2686 } 2687 } 2688 } 2689 2690 static struct sg_table * 2691 intel_rotate_fb_obj_pages(struct i915_ggtt_view *ggtt_view, 2692 struct drm_i915_gem_object *obj) 2693 { 2694 struct drm_device *dev = obj->base.dev; 2695 struct intel_rotation_info *rot_info = &ggtt_view->rotation_info; 2696 unsigned long size, pages, rot_pages; 2697 struct sg_page_iter sg_iter; 2698 unsigned long i; 2699 dma_addr_t *page_addr_list; 2700 struct sg_table *st; 2701 unsigned int tile_pitch, tile_height; 2702 unsigned int width_pages, height_pages; 2703 int ret = -ENOMEM; 2704 2705 pages = obj->base.size / PAGE_SIZE; 2706 2707 /* Calculate tiling geometry. */ 2708 tile_height = intel_tile_height(dev, rot_info->pixel_format, 2709 rot_info->fb_modifier); 2710 tile_pitch = PAGE_SIZE / tile_height; 2711 width_pages = DIV_ROUND_UP(rot_info->pitch, tile_pitch); 2712 height_pages = DIV_ROUND_UP(rot_info->height, tile_height); 2713 rot_pages = width_pages * height_pages; 2714 size = rot_pages * PAGE_SIZE; 2715 2716 /* Allocate a temporary list of source pages for random access. */ 2717 page_addr_list = drm_malloc_ab(pages, sizeof(dma_addr_t)); 2718 if (!page_addr_list) 2719 return ERR_PTR(ret); 2720 2721 /* Allocate target SG list. */ 2722 st = kmalloc(sizeof(*st), GFP_KERNEL); 2723 if (!st) 2724 goto err_st_alloc; 2725 2726 ret = sg_alloc_table(st, rot_pages, GFP_KERNEL); 2727 if (ret) 2728 goto err_sg_alloc; 2729 2730 /* Populate source page list from the object. */ 2731 i = 0; 2732 for_each_sg_page(obj->pages->sgl, &sg_iter, obj->pages->nents, 0) { 2733 page_addr_list[i] = sg_page_iter_dma_address(&sg_iter); 2734 i++; 2735 } 2736 2737 /* Rotate the pages. */ 2738 rotate_pages(page_addr_list, width_pages, height_pages, st); 2739 2740 DRM_DEBUG_KMS( 2741 "Created rotated page mapping for object size %lu (pitch=%u, height=%u, pixel_format=0x%x, %ux%u tiles, %lu pages).\n", 2742 size, rot_info->pitch, rot_info->height, 2743 rot_info->pixel_format, width_pages, height_pages, 2744 rot_pages); 2745 2746 drm_free_large(page_addr_list); 2747 2748 return st; 2749 2750 err_sg_alloc: 2751 kfree(st); 2752 err_st_alloc: 2753 drm_free_large(page_addr_list); 2754 2755 DRM_DEBUG_KMS( 2756 "Failed to create rotated mapping for object size %lu! (%d) (pitch=%u, height=%u, pixel_format=0x%x, %ux%u tiles, %lu pages)\n", 2757 size, ret, rot_info->pitch, rot_info->height, 2758 rot_info->pixel_format, width_pages, height_pages, 2759 rot_pages); 2760 return ERR_PTR(ret); 2761 } 2762 2763 static struct sg_table * 2764 intel_partial_pages(const struct i915_ggtt_view *view, 2765 struct drm_i915_gem_object *obj) 2766 { 2767 struct sg_table *st; 2768 struct scatterlist *sg; 2769 struct sg_page_iter obj_sg_iter; 2770 int ret = -ENOMEM; 2771 2772 st = kmalloc(sizeof(*st), GFP_KERNEL); 2773 if (!st) 2774 goto err_st_alloc; 2775 2776 ret = sg_alloc_table(st, view->params.partial.size, GFP_KERNEL); 2777 if (ret) 2778 goto err_sg_alloc; 2779 2780 sg = st->sgl; 2781 st->nents = 0; 2782 for_each_sg_page(obj->pages->sgl, &obj_sg_iter, obj->pages->nents, 2783 view->params.partial.offset) 2784 { 2785 if (st->nents >= view->params.partial.size) 2786 break; 2787 2788 sg_set_page(sg, NULL, PAGE_SIZE, 0); 2789 sg_dma_address(sg) = sg_page_iter_dma_address(&obj_sg_iter); 2790 sg_dma_len(sg) = PAGE_SIZE; 2791 2792 sg = sg_next(sg); 2793 st->nents++; 2794 } 2795 2796 return st; 2797 2798 err_sg_alloc: 2799 kfree(st); 2800 err_st_alloc: 2801 return ERR_PTR(ret); 2802 } 2803 2804 static int 2805 i915_get_ggtt_vma_pages(struct i915_vma *vma) 2806 { 2807 int ret = 0; 2808 2809 if (vma->ggtt_view.pages) 2810 return 0; 2811 2812 if (vma->ggtt_view.type == I915_GGTT_VIEW_NORMAL) 2813 vma->ggtt_view.pages = vma->obj->pages; 2814 else if (vma->ggtt_view.type == I915_GGTT_VIEW_ROTATED) 2815 vma->ggtt_view.pages = 2816 intel_rotate_fb_obj_pages(&vma->ggtt_view, vma->obj); 2817 else if (vma->ggtt_view.type == I915_GGTT_VIEW_PARTIAL) 2818 vma->ggtt_view.pages = 2819 intel_partial_pages(&vma->ggtt_view, vma->obj); 2820 else 2821 WARN_ONCE(1, "GGTT view %u not implemented!\n", 2822 vma->ggtt_view.type); 2823 2824 if (!vma->ggtt_view.pages) { 2825 DRM_ERROR("Failed to get pages for GGTT view type %u!\n", 2826 vma->ggtt_view.type); 2827 ret = -EINVAL; 2828 } else if (IS_ERR(vma->ggtt_view.pages)) { 2829 ret = PTR_ERR(vma->ggtt_view.pages); 2830 vma->ggtt_view.pages = NULL; 2831 DRM_ERROR("Failed to get pages for VMA view type %u (%d)!\n", 2832 vma->ggtt_view.type, ret); 2833 } 2834 2835 return ret; 2836 } 2837 2838 /** 2839 * i915_vma_bind - Sets up PTEs for an VMA in it's corresponding address space. 2840 * @vma: VMA to map 2841 * @cache_level: mapping cache level 2842 * @flags: flags like global or local mapping 2843 * 2844 * DMA addresses are taken from the scatter-gather table of this object (or of 2845 * this VMA in case of non-default GGTT views) and PTE entries set up. 2846 * Note that DMA addresses are also the only part of the SG table we care about. 2847 */ 2848 int i915_vma_bind(struct i915_vma *vma, enum i915_cache_level cache_level, 2849 u32 flags) 2850 { 2851 int ret; 2852 u32 bind_flags; 2853 2854 if (WARN_ON(flags == 0)) 2855 return -EINVAL; 2856 2857 bind_flags = 0; 2858 if (flags & PIN_GLOBAL) 2859 bind_flags |= GLOBAL_BIND; 2860 if (flags & PIN_USER) 2861 bind_flags |= LOCAL_BIND; 2862 2863 if (flags & PIN_UPDATE) 2864 bind_flags |= vma->bound; 2865 else 2866 bind_flags &= ~vma->bound; 2867 2868 if (bind_flags == 0) 2869 return 0; 2870 2871 if (vma->bound == 0 && vma->vm->allocate_va_range) { 2872 trace_i915_va_alloc(vma->vm, 2873 vma->node.start, 2874 vma->node.size, 2875 VM_TO_TRACE_NAME(vma->vm)); 2876 2877 ret = vma->vm->allocate_va_range(vma->vm, 2878 vma->node.start, 2879 vma->node.size); 2880 if (ret) 2881 return ret; 2882 } 2883 2884 ret = vma->vm->bind_vma(vma, cache_level, bind_flags); 2885 if (ret) 2886 return ret; 2887 2888 vma->bound |= bind_flags; 2889 2890 return 0; 2891 } 2892 2893 /** 2894 * i915_ggtt_view_size - Get the size of a GGTT view. 2895 * @obj: Object the view is of. 2896 * @view: The view in question. 2897 * 2898 * @return The size of the GGTT view in bytes. 2899 */ 2900 size_t 2901 i915_ggtt_view_size(struct drm_i915_gem_object *obj, 2902 const struct i915_ggtt_view *view) 2903 { 2904 if (view->type == I915_GGTT_VIEW_NORMAL || 2905 view->type == I915_GGTT_VIEW_ROTATED) { 2906 return obj->base.size; 2907 } else if (view->type == I915_GGTT_VIEW_PARTIAL) { 2908 return view->params.partial.size << PAGE_SHIFT; 2909 } else { 2910 WARN_ONCE(1, "GGTT view %u not implemented!\n", view->type); 2911 return obj->base.size; 2912 } 2913 } 2914