drivers/gpu/drm/i915/i915_gem_gtt.c

   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/slab.h> /* fault-inject.h is not standalone! */
  27
  28 #include <linux/fault-inject.h>
  29 #include <linux/log2.h>
  30 #include <linux/random.h>
  31 #include <linux/seq_file.h>
  32 #include <linux/stop_machine.h>
  33
  34 #include <asm/set_memory.h>
  35
  36 #include <drm/drmP.h>
  37 #include <drm/i915_drm.h>
  38
  39 #include "i915_drv.h"
  40 #include "i915_vgpu.h"
  41 #include "i915_trace.h"
  42 #include "intel_drv.h"
  43 #include "intel_frontbuffer.h"
  44
  45 #define I915_GFP_ALLOW_FAIL (GFP_KERNEL | __GFP_RETRY_MAYFAIL | __GFP_NOWARN)
  46
  47 /**
  48  * DOC: Global GTT views
  49  *
  50  * Background and previous state
  51  *
  52  * Historically objects could exists (be bound) in global GTT space only as
  53  * singular instances with a view representing all of the object's backing pages
  54  * in a linear fashion. This view will be called a normal view.
  55  *
  56  * To support multiple views of the same object, where the number of mapped
  57  * pages is not equal to the backing store, or where the layout of the pages
  58  * is not linear, concept of a GGTT view was added.
  59  *
  60  * One example of an alternative view is a stereo display driven by a single
  61  * image. In this case we would have a framebuffer looking like this
  62  * (2x2 pages):
  63  *
  64  *    12
  65  *    34
  66  *
  67  * Above would represent a normal GGTT view as normally mapped for GPU or CPU
  68  * rendering. In contrast, fed to the display engine would be an alternative
  69  * view which could look something like this:
  70  *
  71  *   1212
  72  *   3434
  73  *
  74  * In this example both the size and layout of pages in the alternative view is
  75  * different from the normal view.
  76  *
  77  * Implementation and usage
  78  *
  79  * GGTT views are implemented using VMAs and are distinguished via enum
  80  * i915_ggtt_view_type and struct i915_ggtt_view.
  81  *
  82  * A new flavour of core GEM functions which work with GGTT bound objects were
  83  * added with the _ggtt_ infix, and sometimes with _view postfix to avoid
  84  * renaming  in large amounts of code. They take the struct i915_ggtt_view
  85  * parameter encapsulating all metadata required to implement a view.
  86  *
  87  * As a helper for callers which are only interested in the normal view,
  88  * globally const i915_ggtt_view_normal singleton instance exists. All old core
  89  * GEM API functions, the ones not taking the view parameter, are operating on,
  90  * or with the normal GGTT view.
  91  *
  92  * Code wanting to add or use a new GGTT view needs to:
  93  *
  94  * 1. Add a new enum with a suitable name.
  95  * 2. Extend the metadata in the i915_ggtt_view structure if required.
  96  * 3. Add support to i915_get_vma_pages().
  97  *
  98  * New views are required to build a scatter-gather table from within the
  99  * i915_get_vma_pages function. This table is stored in the vma.ggtt_view and
 100  * exists for the lifetime of an VMA.
 101  *
 102  * Core API is designed to have copy semantics which means that passed in
 103  * struct i915_ggtt_view does not need to be persistent (left around after
 104  * calling the core API functions).
 105  *
 106  */
 107
 108 static int
 109 i915_get_ggtt_vma_pages(struct i915_vma *vma);
 110
 111 static void gen6_ggtt_invalidate(struct drm_i915_private *dev_priv)
 112 {
 113         /*
 114          * Note that as an uncached mmio write, this will flush the
 115          * WCB of the writes into the GGTT before it triggers the invalidate.
 116          */
 117         I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
 118 }
 119
 120 static void guc_ggtt_invalidate(struct drm_i915_private *dev_priv)
 121 {
 122         gen6_ggtt_invalidate(dev_priv);
 123         I915_WRITE(GEN8_GTCR, GEN8_GTCR_INVALIDATE);
 124 }
 125
 126 static void gmch_ggtt_invalidate(struct drm_i915_private *dev_priv)
 127 {
 128         intel_gtt_chipset_flush();
 129 }
 130
 131 static inline void i915_ggtt_invalidate(struct drm_i915_private *i915)
 132 {
 133         i915->ggtt.invalidate(i915);
 134 }
 135
 136 int intel_sanitize_enable_ppgtt(struct drm_i915_private *dev_priv,
 137                                 int enable_ppgtt)
 138 {
 139         bool has_full_ppgtt;
 140         bool has_full_48bit_ppgtt;
 141
 142         if (!dev_priv->info.has_aliasing_ppgtt)
 143                 return 0;
 144
 145         has_full_ppgtt = dev_priv->info.has_full_ppgtt;
 146         has_full_48bit_ppgtt = dev_priv->info.has_full_48bit_ppgtt;
 147
 148         if (intel_vgpu_active(dev_priv)) {
 149                 /* GVT-g has no support for 32bit ppgtt */
 150                 has_full_ppgtt = false;
 151                 has_full_48bit_ppgtt = intel_vgpu_has_full_48bit_ppgtt(dev_priv);
 152         }
 153
 154         /*
 155          * We don't allow disabling PPGTT for gen9+ as it's a requirement for
 156          * execlists, the sole mechanism available to submit work.
 157          */
 158         if (enable_ppgtt == 0 && INTEL_GEN(dev_priv) < 9)
 159                 return 0;
 160
 161         if (enable_ppgtt == 1)
 162                 return 1;
 163
 164         if (enable_ppgtt == 2 && has_full_ppgtt)
 165                 return 2;
 166
 167         if (enable_ppgtt == 3 && has_full_48bit_ppgtt)
 168                 return 3;
 169
 170         /* Disable ppgtt on SNB if VT-d is on. */
 171         if (IS_GEN6(dev_priv) && intel_vtd_active()) {
 172                 DRM_INFO("Disabling PPGTT because VT-d is on\n");
 173                 return 0;
 174         }
 175
 176         if (has_full_48bit_ppgtt)
 177                 return 3;
 178
 179         if (has_full_ppgtt)
 180                 return 2;
 181
 182         return 1;
 183 }
 184
 185 static int ppgtt_bind_vma(struct i915_vma *vma,
 186                           enum i915_cache_level cache_level,
 187                           u32 unused)
 188 {
 189         u32 pte_flags;
 190         int err;
 191
 192         if (!(vma->flags & I915_VMA_LOCAL_BIND)) {
 193                 err = vma->vm->allocate_va_range(vma->vm,
 194                                                  vma->node.start, vma->size);
 195                 if (err)
 196                         return err;
 197         }
 198
 199         /* Applicable to VLV, and gen8+ */
 200         pte_flags = 0;
 201         if (i915_gem_object_is_readonly(vma->obj))
 202                 pte_flags |= PTE_READ_ONLY;
 203
 204         vma->vm->insert_entries(vma->vm, vma, cache_level, pte_flags);
 205
 206         return 0;
 207 }
 208
 209 static void ppgtt_unbind_vma(struct i915_vma *vma)
 210 {
 211         vma->vm->clear_range(vma->vm, vma->node.start, vma->size);
 212 }
 213
 214 static int ppgtt_set_pages(struct i915_vma *vma)
 215 {
 216         GEM_BUG_ON(vma->pages);
 217
 218         vma->pages = vma->obj->mm.pages;
 219
 220         vma->page_sizes = vma->obj->mm.page_sizes;
 221
 222         return 0;
 223 }
 224
 225 static void clear_pages(struct i915_vma *vma)
 226 {
 227         GEM_BUG_ON(!vma->pages);
 228
 229         if (vma->pages != vma->obj->mm.pages) {
 230                 sg_free_table(vma->pages);
 231                 kfree(vma->pages);
 232         }
 233         vma->pages = NULL;
 234
 235         memset(&vma->page_sizes, 0, sizeof(vma->page_sizes));
 236 }
 237
 238 static gen8_pte_t gen8_pte_encode(dma_addr_t addr,
 239                                   enum i915_cache_level level,
 240                                   u32 flags)
 241 {
 242         gen8_pte_t pte = addr | _PAGE_PRESENT | _PAGE_RW;
 243
 244         if (unlikely(flags & PTE_READ_ONLY))
 245                 pte &= ~_PAGE_RW;
 246
 247         switch (level) {
 248         case I915_CACHE_NONE:
 249                 pte |= PPAT_UNCACHED;
 250                 break;
 251         case I915_CACHE_WT:
 252                 pte |= PPAT_DISPLAY_ELLC;
 253                 break;
 254         default:
 255                 pte |= PPAT_CACHED;
 256                 break;
 257         }
 258
 259         return pte;
 260 }
 261
 262 static gen8_pde_t gen8_pde_encode(const dma_addr_t addr,
 263                                   const enum i915_cache_level level)
 264 {
 265         gen8_pde_t pde = _PAGE_PRESENT | _PAGE_RW;
 266         pde |= addr;
 267         if (level != I915_CACHE_NONE)
 268                 pde |= PPAT_CACHED_PDE;
 269         else
 270                 pde |= PPAT_UNCACHED;
 271         return pde;
 272 }
 273
 274 #define gen8_pdpe_encode gen8_pde_encode
 275 #define gen8_pml4e_encode gen8_pde_encode
 276
 277 static gen6_pte_t snb_pte_encode(dma_addr_t addr,
 278                                  enum i915_cache_level level,
 279                                  u32 unused)
 280 {
 281         gen6_pte_t pte = GEN6_PTE_VALID;
 282         pte |= GEN6_PTE_ADDR_ENCODE(addr);
 283
 284         switch (level) {
 285         case I915_CACHE_L3_LLC:
 286         case I915_CACHE_LLC:
 287                 pte |= GEN6_PTE_CACHE_LLC;
 288                 break;
 289         case I915_CACHE_NONE:
 290                 pte |= GEN6_PTE_UNCACHED;
 291                 break;
 292         default:
 293                 MISSING_CASE(level);
 294         }
 295
 296         return pte;
 297 }
 298
 299 static gen6_pte_t ivb_pte_encode(dma_addr_t addr,
 300                                  enum i915_cache_level level,
 301                                  u32 unused)
 302 {
 303         gen6_pte_t pte = GEN6_PTE_VALID;
 304         pte |= GEN6_PTE_ADDR_ENCODE(addr);
 305
 306         switch (level) {
 307         case I915_CACHE_L3_LLC:
 308                 pte |= GEN7_PTE_CACHE_L3_LLC;
 309                 break;
 310         case I915_CACHE_LLC:
 311                 pte |= GEN6_PTE_CACHE_LLC;
 312                 break;
 313         case I915_CACHE_NONE:
 314                 pte |= GEN6_PTE_UNCACHED;
 315                 break;
 316         default:
 317                 MISSING_CASE(level);
 318         }
 319
 320         return pte;
 321 }
 322
 323 static gen6_pte_t byt_pte_encode(dma_addr_t addr,
 324                                  enum i915_cache_level level,
 325                                  u32 flags)
 326 {
 327         gen6_pte_t pte = GEN6_PTE_VALID;
 328         pte |= GEN6_PTE_ADDR_ENCODE(addr);
 329
 330         if (!(flags & PTE_READ_ONLY))
 331                 pte |= BYT_PTE_WRITEABLE;
 332
 333         if (level != I915_CACHE_NONE)
 334                 pte |= BYT_PTE_SNOOPED_BY_CPU_CACHES;
 335
 336         return pte;
 337 }
 338
 339 static gen6_pte_t hsw_pte_encode(dma_addr_t addr,
 340                                  enum i915_cache_level level,
 341                                  u32 unused)
 342 {
 343         gen6_pte_t pte = GEN6_PTE_VALID;
 344         pte |= HSW_PTE_ADDR_ENCODE(addr);
 345
 346         if (level != I915_CACHE_NONE)
 347                 pte |= HSW_WB_LLC_AGE3;
 348
 349         return pte;
 350 }
 351
 352 static gen6_pte_t iris_pte_encode(dma_addr_t addr,
 353                                   enum i915_cache_level level,
 354                                   u32 unused)
 355 {
 356         gen6_pte_t pte = GEN6_PTE_VALID;
 357         pte |= HSW_PTE_ADDR_ENCODE(addr);
 358
 359         switch (level) {
 360         case I915_CACHE_NONE:
 361                 break;
 362         case I915_CACHE_WT:
 363                 pte |= HSW_WT_ELLC_LLC_AGE3;
 364                 break;
 365         default:
 366                 pte |= HSW_WB_ELLC_LLC_AGE3;
 367                 break;
 368         }
 369
 370         return pte;
 371 }
 372
 373 static void stash_init(struct pagestash *stash)
 374 {
 375         pagevec_init(&stash->pvec);
 376         spin_lock_init(&stash->lock);
 377 }
 378
 379 static struct page *stash_pop_page(struct pagestash *stash)
 380 {
 381         struct page *page = NULL;
 382
 383         spin_lock(&stash->lock);
 384         if (likely(stash->pvec.nr))
 385                 page = stash->pvec.pages[--stash->pvec.nr];
 386         spin_unlock(&stash->lock);
 387
 388         return page;
 389 }
 390
 391 static void stash_push_pagevec(struct pagestash *stash, struct pagevec *pvec)
 392 {
 393         int nr;
 394
 395         spin_lock_nested(&stash->lock, SINGLE_DEPTH_NESTING);
 396
 397         nr = min_t(int, pvec->nr, pagevec_space(&stash->pvec));
 398         memcpy(stash->pvec.pages + stash->pvec.nr,
 399                pvec->pages + pvec->nr - nr,
 400                sizeof(pvec->pages[0]) * nr);
 401         stash->pvec.nr += nr;
 402
 403         spin_unlock(&stash->lock);
 404
 405         pvec->nr -= nr;
 406 }
 407
 408 static struct page *vm_alloc_page(struct i915_address_space *vm, gfp_t gfp)
 409 {
 410         struct pagevec stack;
 411         struct page *page;
 412
 413         if (I915_SELFTEST_ONLY(should_fail(&vm->fault_attr, 1)))
 414                 i915_gem_shrink_all(vm->i915);
 415
 416         page = stash_pop_page(&vm->free_pages);
 417         if (page)
 418                 return page;
 419
 420         if (!vm->pt_kmap_wc)
 421                 return alloc_page(gfp);
 422
 423         /* Look in our global stash of WC pages... */
 424         page = stash_pop_page(&vm->i915->mm.wc_stash);
 425         if (page)
 426                 return page;
 427
 428         /*
 429          * Otherwise batch allocate pages to amortize cost of set_pages_wc.
 430          *
 431          * We have to be careful as page allocation may trigger the shrinker
 432          * (via direct reclaim) which will fill up the WC stash underneath us.
 433          * So we add our WB pages into a temporary pvec on the stack and merge
 434          * them into the WC stash after all the allocations are complete.
 435          */
 436         pagevec_init(&stack);
 437         do {
 438                 struct page *page;
 439
 440                 page = alloc_page(gfp);
 441                 if (unlikely(!page))
 442                         break;
 443
 444                 stack.pages[stack.nr++] = page;
 445         } while (pagevec_space(&stack));
 446
 447         if (stack.nr && !set_pages_array_wc(stack.pages, stack.nr)) {
 448                 page = stack.pages[--stack.nr];
 449
 450                 /* Merge spare WC pages to the global stash */
 451                 stash_push_pagevec(&vm->i915->mm.wc_stash, &stack);
 452
 453                 /* Push any surplus WC pages onto the local VM stash */
 454                 if (stack.nr)
 455                         stash_push_pagevec(&vm->free_pages, &stack);
 456         }
 457
 458         /* Return unwanted leftovers */
 459         if (unlikely(stack.nr)) {
 460                 WARN_ON_ONCE(set_pages_array_wb(stack.pages, stack.nr));
 461                 __pagevec_release(&stack);
 462         }
 463
 464         return page;
 465 }
 466
 467 static void vm_free_pages_release(struct i915_address_space *vm,
 468                                   bool immediate)
 469 {
 470         struct pagevec *pvec = &vm->free_pages.pvec;
 471         struct pagevec stack;
 472
 473         lockdep_assert_held(&vm->free_pages.lock);
 474         GEM_BUG_ON(!pagevec_count(pvec));
 475
 476         if (vm->pt_kmap_wc) {
 477                 /*
 478                  * When we use WC, first fill up the global stash and then
 479                  * only if full immediately free the overflow.
 480                  */
 481                 stash_push_pagevec(&vm->i915->mm.wc_stash, pvec);
 482
 483                 /*
 484                  * As we have made some room in the VM's free_pages,
 485                  * we can wait for it to fill again. Unless we are
 486                  * inside i915_address_space_fini() and must
 487                  * immediately release the pages!
 488                  */
 489                 if (pvec->nr <= (immediate ? 0 : PAGEVEC_SIZE - 1))
 490                         return;
 491
 492                 /*
 493                  * We have to drop the lock to allow ourselves to sleep,
 494                  * so take a copy of the pvec and clear the stash for
 495                  * others to use it as we sleep.
 496                  */
 497                 stack = *pvec;
 498                 pagevec_reinit(pvec);
 499                 spin_unlock(&vm->free_pages.lock);
 500
 501                 pvec = &stack;
 502                 set_pages_array_wb(pvec->pages, pvec->nr);
 503
 504                 spin_lock(&vm->free_pages.lock);
 505         }
 506
 507         __pagevec_release(pvec);
 508 }
 509
 510 static void vm_free_page(struct i915_address_space *vm, struct page *page)
 511 {
 512         /*
 513          * On !llc, we need to change the pages back to WB. We only do so
 514          * in bulk, so we rarely need to change the page attributes here,
 515          * but doing so requires a stop_machine() from deep inside arch/x86/mm.
 516          * To make detection of the possible sleep more likely, use an
 517          * unconditional might_sleep() for everybody.
 518          */
 519         might_sleep();
 520         spin_lock(&vm->free_pages.lock);
 521         if (!pagevec_add(&vm->free_pages.pvec, page))
 522                 vm_free_pages_release(vm, false);
 523         spin_unlock(&vm->free_pages.lock);
 524 }
 525
 526 static void i915_address_space_init(struct i915_address_space *vm,
 527                                     struct drm_i915_private *dev_priv)
 528 {
 529         /*
 530          * The vm->mutex must be reclaim safe (for use in the shrinker).
 531          * Do a dummy acquire now under fs_reclaim so that any allocation
 532          * attempt holding the lock is immediately reported by lockdep.
 533          */
 534         mutex_init(&vm->mutex);
 535         i915_gem_shrinker_taints_mutex(&vm->mutex);
 536
 537         GEM_BUG_ON(!vm->total);
 538         drm_mm_init(&vm->mm, 0, vm->total);
 539         vm->mm.head_node.color = I915_COLOR_UNEVICTABLE;
 540
 541         stash_init(&vm->free_pages);
 542
 543         INIT_LIST_HEAD(&vm->active_list);
 544         INIT_LIST_HEAD(&vm->inactive_list);
 545         INIT_LIST_HEAD(&vm->unbound_list);
 546 }
 547
 548 static void i915_address_space_fini(struct i915_address_space *vm)
 549 {
 550         spin_lock(&vm->free_pages.lock);
 551         if (pagevec_count(&vm->free_pages.pvec))
 552                 vm_free_pages_release(vm, true);
 553         GEM_BUG_ON(pagevec_count(&vm->free_pages.pvec));
 554         spin_unlock(&vm->free_pages.lock);
 555
 556         drm_mm_takedown(&vm->mm);
 557
 558         mutex_destroy(&vm->mutex);
 559 }
 560
 561 static int __setup_page_dma(struct i915_address_space *vm,
 562                             struct i915_page_dma *p,
 563                             gfp_t gfp)
 564 {
 565         p->page = vm_alloc_page(vm, gfp | I915_GFP_ALLOW_FAIL);
 566         if (unlikely(!p->page))
 567                 return -ENOMEM;
 568
 569         p->daddr = dma_map_page_attrs(vm->dma,
 570                                       p->page, 0, PAGE_SIZE,
 571                                       PCI_DMA_BIDIRECTIONAL,
 572                                       DMA_ATTR_SKIP_CPU_SYNC |
 573                                       DMA_ATTR_NO_WARN);
 574         if (unlikely(dma_mapping_error(vm->dma, p->daddr))) {
 575                 vm_free_page(vm, p->page);
 576                 return -ENOMEM;
 577         }
 578
 579         return 0;
 580 }
 581
 582 static int setup_page_dma(struct i915_address_space *vm,
 583                           struct i915_page_dma *p)
 584 {
 585         return __setup_page_dma(vm, p, __GFP_HIGHMEM);
 586 }
 587
 588 static void cleanup_page_dma(struct i915_address_space *vm,
 589                              struct i915_page_dma *p)
 590 {
 591         dma_unmap_page(vm->dma, p->daddr, PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 592         vm_free_page(vm, p->page);
 593 }
 594
 595 #define kmap_atomic_px(px) kmap_atomic(px_base(px)->page)
 596
 597 #define setup_px(vm, px) setup_page_dma((vm), px_base(px))
 598 #define cleanup_px(vm, px) cleanup_page_dma((vm), px_base(px))
 599 #define fill_px(vm, px, v) fill_page_dma((vm), px_base(px), (v))
 600 #define fill32_px(vm, px, v) fill_page_dma_32((vm), px_base(px), (v))
 601
 602 static void fill_page_dma(struct i915_address_space *vm,
 603                           struct i915_page_dma *p,
 604                           const u64 val)
 605 {
 606         u64 * const vaddr = kmap_atomic(p->page);
 607
 608         memset64(vaddr, val, PAGE_SIZE / sizeof(val));
 609
 610         kunmap_atomic(vaddr);
 611 }
 612
 613 static void fill_page_dma_32(struct i915_address_space *vm,
 614                              struct i915_page_dma *p,
 615                              const u32 v)
 616 {
 617         fill_page_dma(vm, p, (u64)v << 32 | v);
 618 }
 619
 620 static int
 621 setup_scratch_page(struct i915_address_space *vm, gfp_t gfp)
 622 {
 623         unsigned long size;
 624
 625         /*
 626          * In order to utilize 64K pages for an object with a size < 2M, we will
 627          * need to support a 64K scratch page, given that every 16th entry for a
 628          * page-table operating in 64K mode must point to a properly aligned 64K
 629          * region, including any PTEs which happen to point to scratch.
 630          *
 631          * This is only relevant for the 48b PPGTT where we support
 632          * huge-gtt-pages, see also i915_vma_insert().
 633          *
 634          * TODO: we should really consider write-protecting the scratch-page and
 635          * sharing between ppgtt
 636          */
 637         size = I915_GTT_PAGE_SIZE_4K;
 638         if (i915_vm_is_48bit(vm) &&
 639             HAS_PAGE_SIZES(vm->i915, I915_GTT_PAGE_SIZE_64K)) {
 640                 size = I915_GTT_PAGE_SIZE_64K;
 641                 gfp |= __GFP_NOWARN;
 642         }
 643         gfp |= __GFP_ZERO | __GFP_RETRY_MAYFAIL;
 644
 645         do {
 646                 int order = get_order(size);
 647                 struct page *page;
 648                 dma_addr_t addr;
 649
 650                 page = alloc_pages(gfp, order);
 651                 if (unlikely(!page))
 652                         goto skip;
 653
 654                 addr = dma_map_page_attrs(vm->dma,
 655                                           page, 0, size,
 656                                           PCI_DMA_BIDIRECTIONAL,
 657                                           DMA_ATTR_SKIP_CPU_SYNC |
 658                                           DMA_ATTR_NO_WARN);
 659                 if (unlikely(dma_mapping_error(vm->dma, addr)))
 660                         goto free_page;
 661
 662                 if (unlikely(!IS_ALIGNED(addr, size)))
 663                         goto unmap_page;
 664
 665                 vm->scratch_page.page = page;
 666                 vm->scratch_page.daddr = addr;
 667                 vm->scratch_page.order = order;
 668                 return 0;
 669
 670 unmap_page:
 671                 dma_unmap_page(vm->dma, addr, size, PCI_DMA_BIDIRECTIONAL);
 672 free_page:
 673                 __free_pages(page, order);
 674 skip:
 675                 if (size == I915_GTT_PAGE_SIZE_4K)
 676                         return -ENOMEM;
 677
 678                 size = I915_GTT_PAGE_SIZE_4K;
 679                 gfp &= ~__GFP_NOWARN;
 680         } while (1);
 681 }
 682
 683 static void cleanup_scratch_page(struct i915_address_space *vm)
 684 {
 685         struct i915_page_dma *p = &vm->scratch_page;
 686
 687         dma_unmap_page(vm->dma, p->daddr, BIT(p->order) << PAGE_SHIFT,
 688                        PCI_DMA_BIDIRECTIONAL);
 689         __free_pages(p->page, p->order);
 690 }
 691
 692 static struct i915_page_table *alloc_pt(struct i915_address_space *vm)
 693 {
 694         struct i915_page_table *pt;
 695
 696         pt = kmalloc(sizeof(*pt), I915_GFP_ALLOW_FAIL);
 697         if (unlikely(!pt))
 698                 return ERR_PTR(-ENOMEM);
 699
 700         if (unlikely(setup_px(vm, pt))) {
 701                 kfree(pt);
 702                 return ERR_PTR(-ENOMEM);
 703         }
 704
 705         pt->used_ptes = 0;
 706         return pt;
 707 }
 708
 709 static void free_pt(struct i915_address_space *vm, struct i915_page_table *pt)
 710 {
 711         cleanup_px(vm, pt);
 712         kfree(pt);
 713 }
 714
 715 static void gen8_initialize_pt(struct i915_address_space *vm,
 716                                struct i915_page_table *pt)
 717 {
 718         fill_px(vm, pt,
 719                 gen8_pte_encode(vm->scratch_page.daddr, I915_CACHE_LLC, 0));
 720 }
 721
 722 static void gen6_initialize_pt(struct gen6_hw_ppgtt *ppgtt,
 723                                struct i915_page_table *pt)
 724 {
 725         fill32_px(&ppgtt->base.vm, pt, ppgtt->scratch_pte);
 726 }
 727
 728 static struct i915_page_directory *alloc_pd(struct i915_address_space *vm)
 729 {
 730         struct i915_page_directory *pd;
 731
 732         pd = kzalloc(sizeof(*pd), I915_GFP_ALLOW_FAIL);
 733         if (unlikely(!pd))
 734                 return ERR_PTR(-ENOMEM);
 735
 736         if (unlikely(setup_px(vm, pd))) {
 737                 kfree(pd);
 738                 return ERR_PTR(-ENOMEM);
 739         }
 740
 741         pd->used_pdes = 0;
 742         return pd;
 743 }
 744
 745 static void free_pd(struct i915_address_space *vm,
 746                     struct i915_page_directory *pd)
 747 {
 748         cleanup_px(vm, pd);
 749         kfree(pd);
 750 }
 751
 752 static void gen8_initialize_pd(struct i915_address_space *vm,
 753                                struct i915_page_directory *pd)
 754 {
 755         fill_px(vm, pd,
 756                 gen8_pde_encode(px_dma(vm->scratch_pt), I915_CACHE_LLC));
 757         memset_p((void **)pd->page_table, vm->scratch_pt, I915_PDES);
 758 }
 759
 760 static int __pdp_init(struct i915_address_space *vm,
 761                       struct i915_page_directory_pointer *pdp)
 762 {
 763         const unsigned int pdpes = i915_pdpes_per_pdp(vm);
 764
 765         pdp->page_directory = kmalloc_array(pdpes, sizeof(*pdp->page_directory),
 766                                             I915_GFP_ALLOW_FAIL);
 767         if (unlikely(!pdp->page_directory))
 768                 return -ENOMEM;
 769
 770         memset_p((void **)pdp->page_directory, vm->scratch_pd, pdpes);
 771
 772         return 0;
 773 }
 774
 775 static void __pdp_fini(struct i915_page_directory_pointer *pdp)
 776 {
 777         kfree(pdp->page_directory);
 778         pdp->page_directory = NULL;
 779 }
 780
 781 static inline bool use_4lvl(const struct i915_address_space *vm)
 782 {
 783         return i915_vm_is_48bit(vm);
 784 }
 785
 786 static struct i915_page_directory_pointer *
 787 alloc_pdp(struct i915_address_space *vm)
 788 {
 789         struct i915_page_directory_pointer *pdp;
 790         int ret = -ENOMEM;
 791
 792         GEM_BUG_ON(!use_4lvl(vm));
 793
 794         pdp = kzalloc(sizeof(*pdp), GFP_KERNEL);
 795         if (!pdp)
 796                 return ERR_PTR(-ENOMEM);
 797
 798         ret = __pdp_init(vm, pdp);
 799         if (ret)
 800                 goto fail_bitmap;
 801
 802         ret = setup_px(vm, pdp);
 803         if (ret)
 804                 goto fail_page_m;
 805
 806         return pdp;
 807
 808 fail_page_m:
 809         __pdp_fini(pdp);
 810 fail_bitmap:
 811         kfree(pdp);
 812
 813         return ERR_PTR(ret);
 814 }
 815
 816 static void free_pdp(struct i915_address_space *vm,
 817                      struct i915_page_directory_pointer *pdp)
 818 {
 819         __pdp_fini(pdp);
 820
 821         if (!use_4lvl(vm))
 822                 return;
 823
 824         cleanup_px(vm, pdp);
 825         kfree(pdp);
 826 }
 827
 828 static void gen8_initialize_pdp(struct i915_address_space *vm,
 829                                 struct i915_page_directory_pointer *pdp)
 830 {
 831         gen8_ppgtt_pdpe_t scratch_pdpe;
 832
 833         scratch_pdpe = gen8_pdpe_encode(px_dma(vm->scratch_pd), I915_CACHE_LLC);
 834
 835         fill_px(vm, pdp, scratch_pdpe);
 836 }
 837
 838 static void gen8_initialize_pml4(struct i915_address_space *vm,
 839                                  struct i915_pml4 *pml4)
 840 {
 841         fill_px(vm, pml4,
 842                 gen8_pml4e_encode(px_dma(vm->scratch_pdp), I915_CACHE_LLC));
 843         memset_p((void **)pml4->pdps, vm->scratch_pdp, GEN8_PML4ES_PER_PML4);
 844 }
 845
 846 /* PDE TLBs are a pain to invalidate on GEN8+. When we modify
 847  * the page table structures, we mark them dirty so that
 848  * context switching/execlist queuing code takes extra steps
 849  * to ensure that tlbs are flushed.
 850  */
 851 static void mark_tlbs_dirty(struct i915_hw_ppgtt *ppgtt)
 852 {
 853         ppgtt->pd_dirty_rings = INTEL_INFO(ppgtt->vm.i915)->ring_mask;
 854 }
 855
 856 /* Removes entries from a single page table, releasing it if it's empty.
 857  * Caller can use the return value to update higher-level entries.
 858  */
 859 static bool gen8_ppgtt_clear_pt(struct i915_address_space *vm,
 860                                 struct i915_page_table *pt,
 861                                 u64 start, u64 length)
 862 {
 863         unsigned int num_entries = gen8_pte_count(start, length);
 864         unsigned int pte = gen8_pte_index(start);
 865         unsigned int pte_end = pte + num_entries;
 866         const gen8_pte_t scratch_pte =
 867                 gen8_pte_encode(vm->scratch_page.daddr, I915_CACHE_LLC, 0);
 868         gen8_pte_t *vaddr;
 869
 870         GEM_BUG_ON(num_entries > pt->used_ptes);
 871
 872         pt->used_ptes -= num_entries;
 873         if (!pt->used_ptes)
 874                 return true;
 875
 876         vaddr = kmap_atomic_px(pt);
 877         while (pte < pte_end)
 878                 vaddr[pte++] = scratch_pte;
 879         kunmap_atomic(vaddr);
 880
 881         return false;
 882 }
 883
 884 static void gen8_ppgtt_set_pde(struct i915_address_space *vm,
 885                                struct i915_page_directory *pd,
 886                                struct i915_page_table *pt,
 887                                unsigned int pde)
 888 {
 889         gen8_pde_t *vaddr;
 890
 891         pd->page_table[pde] = pt;
 892
 893         vaddr = kmap_atomic_px(pd);
 894         vaddr[pde] = gen8_pde_encode(px_dma(pt), I915_CACHE_LLC);
 895         kunmap_atomic(vaddr);
 896 }
 897
 898 static bool gen8_ppgtt_clear_pd(struct i915_address_space *vm,
 899                                 struct i915_page_directory *pd,
 900                                 u64 start, u64 length)
 901 {
 902         struct i915_page_table *pt;
 903         u32 pde;
 904
 905         gen8_for_each_pde(pt, pd, start, length, pde) {
 906                 GEM_BUG_ON(pt == vm->scratch_pt);
 907
 908                 if (!gen8_ppgtt_clear_pt(vm, pt, start, length))
 909                         continue;
 910
 911                 gen8_ppgtt_set_pde(vm, pd, vm->scratch_pt, pde);
 912                 GEM_BUG_ON(!pd->used_pdes);
 913                 pd->used_pdes--;
 914
 915                 free_pt(vm, pt);
 916         }
 917
 918         return !pd->used_pdes;
 919 }
 920
 921 static void gen8_ppgtt_set_pdpe(struct i915_address_space *vm,
 922                                 struct i915_page_directory_pointer *pdp,
 923                                 struct i915_page_directory *pd,
 924                                 unsigned int pdpe)
 925 {
 926         gen8_ppgtt_pdpe_t *vaddr;
 927
 928         pdp->page_directory[pdpe] = pd;
 929         if (!use_4lvl(vm))
 930                 return;
 931
 932         vaddr = kmap_atomic_px(pdp);
 933         vaddr[pdpe] = gen8_pdpe_encode(px_dma(pd), I915_CACHE_LLC);
 934         kunmap_atomic(vaddr);
 935 }
 936
 937 /* Removes entries from a single page dir pointer, releasing it if it's empty.
 938  * Caller can use the return value to update higher-level entries
 939  */
 940 static bool gen8_ppgtt_clear_pdp(struct i915_address_space *vm,
 941                                  struct i915_page_directory_pointer *pdp,
 942                                  u64 start, u64 length)
 943 {
 944         struct i915_page_directory *pd;
 945         unsigned int pdpe;
 946
 947         gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
 948                 GEM_BUG_ON(pd == vm->scratch_pd);
 949
 950                 if (!gen8_ppgtt_clear_pd(vm, pd, start, length))
 951                         continue;
 952
 953                 gen8_ppgtt_set_pdpe(vm, pdp, vm->scratch_pd, pdpe);
 954                 GEM_BUG_ON(!pdp->used_pdpes);
 955                 pdp->used_pdpes--;
 956
 957                 free_pd(vm, pd);
 958         }
 959
 960         return !pdp->used_pdpes;
 961 }
 962
 963 static void gen8_ppgtt_clear_3lvl(struct i915_address_space *vm,
 964                                   u64 start, u64 length)
 965 {
 966         gen8_ppgtt_clear_pdp(vm, &i915_vm_to_ppgtt(vm)->pdp, start, length);
 967 }
 968
 969 static void gen8_ppgtt_set_pml4e(struct i915_pml4 *pml4,
 970                                  struct i915_page_directory_pointer *pdp,
 971                                  unsigned int pml4e)
 972 {
 973         gen8_ppgtt_pml4e_t *vaddr;
 974
 975         pml4->pdps[pml4e] = pdp;
 976
 977         vaddr = kmap_atomic_px(pml4);
 978         vaddr[pml4e] = gen8_pml4e_encode(px_dma(pdp), I915_CACHE_LLC);
 979         kunmap_atomic(vaddr);
 980 }
 981
 982 /* Removes entries from a single pml4.
 983  * This is the top-level structure in 4-level page tables used on gen8+.
 984  * Empty entries are always scratch pml4e.
 985  */
 986 static void gen8_ppgtt_clear_4lvl(struct i915_address_space *vm,
 987                                   u64 start, u64 length)
 988 {
 989         struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
 990         struct i915_pml4 *pml4 = &ppgtt->pml4;
 991         struct i915_page_directory_pointer *pdp;
 992         unsigned int pml4e;
 993
 994         GEM_BUG_ON(!use_4lvl(vm));
 995
 996         gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
 997                 GEM_BUG_ON(pdp == vm->scratch_pdp);
 998
 999                 if (!gen8_ppgtt_clear_pdp(vm, pdp, start, length))
1000                         continue;
1001
1002                 gen8_ppgtt_set_pml4e(pml4, vm->scratch_pdp, pml4e);
1003
1004                 free_pdp(vm, pdp);
1005         }
1006 }
1007
1008 static inline struct sgt_dma {
1009         struct scatterlist *sg;
1010         dma_addr_t dma, max;
1011 } sgt_dma(struct i915_vma *vma) {
1012         struct scatterlist *sg = vma->pages->sgl;
1013         dma_addr_t addr = sg_dma_address(sg);
1014         return (struct sgt_dma) { sg, addr, addr + sg->length };
1015 }
1016
1017 struct gen8_insert_pte {
1018         u16 pml4e;
1019         u16 pdpe;
1020         u16 pde;
1021         u16 pte;
1022 };
1023
1024 static __always_inline struct gen8_insert_pte gen8_insert_pte(u64 start)
1025 {
1026         return (struct gen8_insert_pte) {
1027                  gen8_pml4e_index(start),
1028                  gen8_pdpe_index(start),
1029                  gen8_pde_index(start),
1030                  gen8_pte_index(start),
1031         };
1032 }
1033
1034 static __always_inline bool
1035 gen8_ppgtt_insert_pte_entries(struct i915_hw_ppgtt *ppgtt,
1036                               struct i915_page_directory_pointer *pdp,
1037                               struct sgt_dma *iter,
1038                               struct gen8_insert_pte *idx,
1039                               enum i915_cache_level cache_level,
1040                               u32 flags)
1041 {
1042         struct i915_page_directory *pd;
1043         const gen8_pte_t pte_encode = gen8_pte_encode(0, cache_level, flags);
1044         gen8_pte_t *vaddr;
1045         bool ret;
1046
1047         GEM_BUG_ON(idx->pdpe >= i915_pdpes_per_pdp(&ppgtt->vm));
1048         pd = pdp->page_directory[idx->pdpe];
1049         vaddr = kmap_atomic_px(pd->page_table[idx->pde]);
1050         do {
1051                 vaddr[idx->pte] = pte_encode | iter->dma;
1052
1053                 iter->dma += I915_GTT_PAGE_SIZE;
1054                 if (iter->dma >= iter->max) {
1055                         iter->sg = __sg_next(iter->sg);
1056                         if (!iter->sg) {
1057                                 ret = false;
1058                                 break;
1059                         }
1060
1061                         iter->dma = sg_dma_address(iter->sg);
1062                         iter->max = iter->dma + iter->sg->length;
1063                 }
1064
1065                 if (++idx->pte == GEN8_PTES) {
1066                         idx->pte = 0;
1067
1068                         if (++idx->pde == I915_PDES) {
1069                                 idx->pde = 0;
1070
1071                                 /* Limited by sg length for 3lvl */
1072                                 if (++idx->pdpe == GEN8_PML4ES_PER_PML4) {
1073                                         idx->pdpe = 0;
1074                                         ret = true;
1075                                         break;
1076                                 }
1077
1078                                 GEM_BUG_ON(idx->pdpe >= i915_pdpes_per_pdp(&ppgtt->vm));
1079                                 pd = pdp->page_directory[idx->pdpe];
1080                         }
1081
1082                         kunmap_atomic(vaddr);
1083                         vaddr = kmap_atomic_px(pd->page_table[idx->pde]);
1084                 }
1085         } while (1);
1086         kunmap_atomic(vaddr);
1087
1088         return ret;
1089 }
1090
1091 static void gen8_ppgtt_insert_3lvl(struct i915_address_space *vm,
1092                                    struct i915_vma *vma,
1093                                    enum i915_cache_level cache_level,
1094                                    u32 flags)
1095 {
1096         struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1097         struct sgt_dma iter = sgt_dma(vma);
1098         struct gen8_insert_pte idx = gen8_insert_pte(vma->node.start);
1099
1100         gen8_ppgtt_insert_pte_entries(ppgtt, &ppgtt->pdp, &iter, &idx,
1101                                       cache_level, flags);
1102
1103         vma->page_sizes.gtt = I915_GTT_PAGE_SIZE;
1104 }
1105
1106 static void gen8_ppgtt_insert_huge_entries(struct i915_vma *vma,
1107                                            struct i915_page_directory_pointer **pdps,
1108                                            struct sgt_dma *iter,
1109                                            enum i915_cache_level cache_level,
1110                                            u32 flags)
1111 {
1112         const gen8_pte_t pte_encode = gen8_pte_encode(0, cache_level, flags);
1113         u64 start = vma->node.start;
1114         dma_addr_t rem = iter->sg->length;
1115
1116         do {
1117                 struct gen8_insert_pte idx = gen8_insert_pte(start);
1118                 struct i915_page_directory_pointer *pdp = pdps[idx.pml4e];
1119                 struct i915_page_directory *pd = pdp->page_directory[idx.pdpe];
1120                 unsigned int page_size;
1121                 bool maybe_64K = false;
1122                 gen8_pte_t encode = pte_encode;
1123                 gen8_pte_t *vaddr;
1124                 u16 index, max;
1125
1126                 if (vma->page_sizes.sg & I915_GTT_PAGE_SIZE_2M &&
1127                     IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_2M) &&
1128                     rem >= I915_GTT_PAGE_SIZE_2M && !idx.pte) {
1129                         index = idx.pde;
1130                         max = I915_PDES;
1131                         page_size = I915_GTT_PAGE_SIZE_2M;
1132
1133                         encode |= GEN8_PDE_PS_2M;
1134
1135                         vaddr = kmap_atomic_px(pd);
1136                 } else {
1137                         struct i915_page_table *pt = pd->page_table[idx.pde];
1138
1139                         index = idx.pte;
1140                         max = GEN8_PTES;
1141                         page_size = I915_GTT_PAGE_SIZE;
1142
1143                         if (!index &&
1144                             vma->page_sizes.sg & I915_GTT_PAGE_SIZE_64K &&
1145                             IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_64K) &&
1146                             (IS_ALIGNED(rem, I915_GTT_PAGE_SIZE_64K) ||
1147                              rem >= (max - index) * I915_GTT_PAGE_SIZE))
1148                                 maybe_64K = true;
1149
1150                         vaddr = kmap_atomic_px(pt);
1151                 }
1152
1153                 do {
1154                         GEM_BUG_ON(iter->sg->length < page_size);
1155                         vaddr[index++] = encode | iter->dma;
1156
1157                         start += page_size;
1158                         iter->dma += page_size;
1159                         rem -= page_size;
1160                         if (iter->dma >= iter->max) {
1161                                 iter->sg = __sg_next(iter->sg);
1162                                 if (!iter->sg)
1163                                         break;
1164
1165                                 rem = iter->sg->length;
1166                                 iter->dma = sg_dma_address(iter->sg);
1167                                 iter->max = iter->dma + rem;
1168
1169                                 if (maybe_64K && index < max &&
1170                                     !(IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_64K) &&
1171                                       (IS_ALIGNED(rem, I915_GTT_PAGE_SIZE_64K) ||
1172                                        rem >= (max - index) * I915_GTT_PAGE_SIZE)))
1173                                         maybe_64K = false;
1174
1175                                 if (unlikely(!IS_ALIGNED(iter->dma, page_size)))
1176                                         break;
1177                         }
1178                 } while (rem >= page_size && index < max);
1179
1180                 kunmap_atomic(vaddr);
1181
1182                 /*
1183                  * Is it safe to mark the 2M block as 64K? -- Either we have
1184                  * filled whole page-table with 64K entries, or filled part of
1185                  * it and have reached the end of the sg table and we have
1186                  * enough padding.
1187                  */
1188                 if (maybe_64K &&
1189                     (index == max ||
1190                      (i915_vm_has_scratch_64K(vma->vm) &&
1191                       !iter->sg && IS_ALIGNED(vma->node.start +
1192                                               vma->node.size,
1193                                               I915_GTT_PAGE_SIZE_2M)))) {
1194                         vaddr = kmap_atomic_px(pd);
1195                         vaddr[idx.pde] |= GEN8_PDE_IPS_64K;
1196                         kunmap_atomic(vaddr);
1197                         page_size = I915_GTT_PAGE_SIZE_64K;
1198
1199                         /*
1200                          * We write all 4K page entries, even when using 64K
1201                          * pages. In order to verify that the HW isn't cheating
1202                          * by using the 4K PTE instead of the 64K PTE, we want
1203                          * to remove all the surplus entries. If the HW skipped
1204                          * the 64K PTE, it will read/write into the scratch page
1205                          * instead - which we detect as missing results during
1206                          * selftests.
1207                          */
1208                         if (I915_SELFTEST_ONLY(vma->vm->scrub_64K)) {
1209                                 u16 i;
1210
1211                                 encode = pte_encode | vma->vm->scratch_page.daddr;
1212                                 vaddr = kmap_atomic_px(pd->page_table[idx.pde]);
1213
1214                                 for (i = 1; i < index; i += 16)
1215                                         memset64(vaddr + i, encode, 15);
1216
1217                                 kunmap_atomic(vaddr);
1218                         }
1219                 }
1220
1221                 vma->page_sizes.gtt |= page_size;
1222         } while (iter->sg);
1223 }
1224
1225 static void gen8_ppgtt_insert_4lvl(struct i915_address_space *vm,
1226                                    struct i915_vma *vma,
1227                                    enum i915_cache_level cache_level,
1228                                    u32 flags)
1229 {
1230         struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1231         struct sgt_dma iter = sgt_dma(vma);
1232         struct i915_page_directory_pointer **pdps = ppgtt->pml4.pdps;
1233
1234         if (vma->page_sizes.sg > I915_GTT_PAGE_SIZE) {
1235                 gen8_ppgtt_insert_huge_entries(vma, pdps, &iter, cache_level,
1236                                                flags);
1237         } else {
1238                 struct gen8_insert_pte idx = gen8_insert_pte(vma->node.start);
1239
1240                 while (gen8_ppgtt_insert_pte_entries(ppgtt, pdps[idx.pml4e++],
1241                                                      &iter, &idx, cache_level,
1242                                                      flags))
1243                         GEM_BUG_ON(idx.pml4e >= GEN8_PML4ES_PER_PML4);
1244
1245                 vma->page_sizes.gtt = I915_GTT_PAGE_SIZE;
1246         }
1247 }
1248
1249 static void gen8_free_page_tables(struct i915_address_space *vm,
1250                                   struct i915_page_directory *pd)
1251 {
1252         int i;
1253
1254         for (i = 0; i < I915_PDES; i++) {
1255                 if (pd->page_table[i] != vm->scratch_pt)
1256                         free_pt(vm, pd->page_table[i]);
1257         }
1258 }
1259
1260 static int gen8_init_scratch(struct i915_address_space *vm)
1261 {
1262         int ret;
1263
1264         ret = setup_scratch_page(vm, __GFP_HIGHMEM);
1265         if (ret)
1266                 return ret;
1267
1268         vm->scratch_pt = alloc_pt(vm);
1269         if (IS_ERR(vm->scratch_pt)) {
1270                 ret = PTR_ERR(vm->scratch_pt);
1271                 goto free_scratch_page;
1272         }
1273
1274         vm->scratch_pd = alloc_pd(vm);
1275         if (IS_ERR(vm->scratch_pd)) {
1276                 ret = PTR_ERR(vm->scratch_pd);
1277                 goto free_pt;
1278         }
1279
1280         if (use_4lvl(vm)) {
1281                 vm->scratch_pdp = alloc_pdp(vm);
1282                 if (IS_ERR(vm->scratch_pdp)) {
1283                         ret = PTR_ERR(vm->scratch_pdp);
1284                         goto free_pd;
1285                 }
1286         }
1287
1288         gen8_initialize_pt(vm, vm->scratch_pt);
1289         gen8_initialize_pd(vm, vm->scratch_pd);
1290         if (use_4lvl(vm))
1291                 gen8_initialize_pdp(vm, vm->scratch_pdp);
1292
1293         return 0;
1294
1295 free_pd:
1296         free_pd(vm, vm->scratch_pd);
1297 free_pt:
1298         free_pt(vm, vm->scratch_pt);
1299 free_scratch_page:
1300         cleanup_scratch_page(vm);
1301
1302         return ret;
1303 }
1304
1305 static int gen8_ppgtt_notify_vgt(struct i915_hw_ppgtt *ppgtt, bool create)
1306 {
1307         struct i915_address_space *vm = &ppgtt->vm;
1308         struct drm_i915_private *dev_priv = vm->i915;
1309         enum vgt_g2v_type msg;
1310         int i;
1311
1312         if (use_4lvl(vm)) {
1313                 const u64 daddr = px_dma(&ppgtt->pml4);
1314
1315                 I915_WRITE(vgtif_reg(pdp[0].lo), lower_32_bits(daddr));
1316                 I915_WRITE(vgtif_reg(pdp[0].hi), upper_32_bits(daddr));
1317
1318                 msg = (create ? VGT_G2V_PPGTT_L4_PAGE_TABLE_CREATE :
1319                                 VGT_G2V_PPGTT_L4_PAGE_TABLE_DESTROY);
1320         } else {
1321                 for (i = 0; i < GEN8_3LVL_PDPES; i++) {
1322                         const u64 daddr = i915_page_dir_dma_addr(ppgtt, i);
1323
1324                         I915_WRITE(vgtif_reg(pdp[i].lo), lower_32_bits(daddr));
1325                         I915_WRITE(vgtif_reg(pdp[i].hi), upper_32_bits(daddr));
1326                 }
1327
1328                 msg = (create ? VGT_G2V_PPGTT_L3_PAGE_TABLE_CREATE :
1329                                 VGT_G2V_PPGTT_L3_PAGE_TABLE_DESTROY);
1330         }
1331
1332         I915_WRITE(vgtif_reg(g2v_notify), msg);
1333
1334         return 0;
1335 }
1336
1337 static void gen8_free_scratch(struct i915_address_space *vm)
1338 {
1339         if (use_4lvl(vm))
1340                 free_pdp(vm, vm->scratch_pdp);
1341         free_pd(vm, vm->scratch_pd);
1342         free_pt(vm, vm->scratch_pt);
1343         cleanup_scratch_page(vm);
1344 }
1345
1346 static void gen8_ppgtt_cleanup_3lvl(struct i915_address_space *vm,
1347                                     struct i915_page_directory_pointer *pdp)
1348 {
1349         const unsigned int pdpes = i915_pdpes_per_pdp(vm);
1350         int i;
1351
1352         for (i = 0; i < pdpes; i++) {
1353                 if (pdp->page_directory[i] == vm->scratch_pd)
1354                         continue;
1355
1356                 gen8_free_page_tables(vm, pdp->page_directory[i]);
1357                 free_pd(vm, pdp->page_directory[i]);
1358         }
1359
1360         free_pdp(vm, pdp);
1361 }
1362
1363 static void gen8_ppgtt_cleanup_4lvl(struct i915_hw_ppgtt *ppgtt)
1364 {
1365         int i;
1366
1367         for (i = 0; i < GEN8_PML4ES_PER_PML4; i++) {
1368                 if (ppgtt->pml4.pdps[i] == ppgtt->vm.scratch_pdp)
1369                         continue;
1370
1371                 gen8_ppgtt_cleanup_3lvl(&ppgtt->vm, ppgtt->pml4.pdps[i]);
1372         }
1373
1374         cleanup_px(&ppgtt->vm, &ppgtt->pml4);
1375 }
1376
1377 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
1378 {
1379         struct drm_i915_private *dev_priv = vm->i915;
1380         struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1381
1382         if (intel_vgpu_active(dev_priv))
1383                 gen8_ppgtt_notify_vgt(ppgtt, false);
1384
1385         if (use_4lvl(vm))
1386                 gen8_ppgtt_cleanup_4lvl(ppgtt);
1387         else
1388                 gen8_ppgtt_cleanup_3lvl(&ppgtt->vm, &ppgtt->pdp);
1389
1390         gen8_free_scratch(vm);
1391 }
1392
1393 static int gen8_ppgtt_alloc_pd(struct i915_address_space *vm,
1394                                struct i915_page_directory *pd,
1395                                u64 start, u64 length)
1396 {
1397         struct i915_page_table *pt;
1398         u64 from = start;
1399         unsigned int pde;
1400
1401         gen8_for_each_pde(pt, pd, start, length, pde) {
1402                 int count = gen8_pte_count(start, length);
1403
1404                 if (pt == vm->scratch_pt) {
1405                         pd->used_pdes++;
1406
1407                         pt = alloc_pt(vm);
1408                         if (IS_ERR(pt)) {
1409                                 pd->used_pdes--;
1410                                 goto unwind;
1411                         }
1412
1413                         if (count < GEN8_PTES || intel_vgpu_active(vm->i915))
1414                                 gen8_initialize_pt(vm, pt);
1415
1416                         gen8_ppgtt_set_pde(vm, pd, pt, pde);
1417                         GEM_BUG_ON(pd->used_pdes > I915_PDES);
1418                 }
1419
1420                 pt->used_ptes += count;
1421         }
1422         return 0;
1423
1424 unwind:
1425         gen8_ppgtt_clear_pd(vm, pd, from, start - from);
1426         return -ENOMEM;
1427 }
1428
1429 static int gen8_ppgtt_alloc_pdp(struct i915_address_space *vm,
1430                                 struct i915_page_directory_pointer *pdp,
1431                                 u64 start, u64 length)
1432 {
1433         struct i915_page_directory *pd;
1434         u64 from = start;
1435         unsigned int pdpe;
1436         int ret;
1437
1438         gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1439                 if (pd == vm->scratch_pd) {
1440                         pdp->used_pdpes++;
1441
1442                         pd = alloc_pd(vm);
1443                         if (IS_ERR(pd)) {
1444                                 pdp->used_pdpes--;
1445                                 goto unwind;
1446                         }
1447
1448                         gen8_initialize_pd(vm, pd);
1449                         gen8_ppgtt_set_pdpe(vm, pdp, pd, pdpe);
1450                         GEM_BUG_ON(pdp->used_pdpes > i915_pdpes_per_pdp(vm));
1451
1452                         mark_tlbs_dirty(i915_vm_to_ppgtt(vm));
1453                 }
1454
1455                 ret = gen8_ppgtt_alloc_pd(vm, pd, start, length);
1456                 if (unlikely(ret))
1457                         goto unwind_pd;
1458         }
1459
1460         return 0;
1461
1462 unwind_pd:
1463         if (!pd->used_pdes) {
1464                 gen8_ppgtt_set_pdpe(vm, pdp, vm->scratch_pd, pdpe);
1465                 GEM_BUG_ON(!pdp->used_pdpes);
1466                 pdp->used_pdpes--;
1467                 free_pd(vm, pd);
1468         }
1469 unwind:
1470         gen8_ppgtt_clear_pdp(vm, pdp, from, start - from);
1471         return -ENOMEM;
1472 }
1473
1474 static int gen8_ppgtt_alloc_3lvl(struct i915_address_space *vm,
1475                                  u64 start, u64 length)
1476 {
1477         return gen8_ppgtt_alloc_pdp(vm,
1478                                     &i915_vm_to_ppgtt(vm)->pdp, start, length);
1479 }
1480
1481 static int gen8_ppgtt_alloc_4lvl(struct i915_address_space *vm,
1482                                  u64 start, u64 length)
1483 {
1484         struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1485         struct i915_pml4 *pml4 = &ppgtt->pml4;
1486         struct i915_page_directory_pointer *pdp;
1487         u64 from = start;
1488         u32 pml4e;
1489         int ret;
1490
1491         gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1492                 if (pml4->pdps[pml4e] == vm->scratch_pdp) {
1493                         pdp = alloc_pdp(vm);
1494                         if (IS_ERR(pdp))
1495                                 goto unwind;
1496
1497                         gen8_initialize_pdp(vm, pdp);
1498                         gen8_ppgtt_set_pml4e(pml4, pdp, pml4e);
1499                 }
1500
1501                 ret = gen8_ppgtt_alloc_pdp(vm, pdp, start, length);
1502                 if (unlikely(ret))
1503                         goto unwind_pdp;
1504         }
1505
1506         return 0;
1507
1508 unwind_pdp:
1509         if (!pdp->used_pdpes) {
1510                 gen8_ppgtt_set_pml4e(pml4, vm->scratch_pdp, pml4e);
1511                 free_pdp(vm, pdp);
1512         }
1513 unwind:
1514         gen8_ppgtt_clear_4lvl(vm, from, start - from);
1515         return -ENOMEM;
1516 }
1517
1518 static void gen8_dump_pdp(struct i915_hw_ppgtt *ppgtt,
1519                           struct i915_page_directory_pointer *pdp,
1520                           u64 start, u64 length,
1521                           gen8_pte_t scratch_pte,
1522                           struct seq_file *m)
1523 {
1524         struct i915_address_space *vm = &ppgtt->vm;
1525         struct i915_page_directory *pd;
1526         u32 pdpe;
1527
1528         gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1529                 struct i915_page_table *pt;
1530                 u64 pd_len = length;
1531                 u64 pd_start = start;
1532                 u32 pde;
1533
1534                 if (pdp->page_directory[pdpe] == ppgtt->vm.scratch_pd)
1535                         continue;
1536
1537                 seq_printf(m, "\tPDPE #%d\n", pdpe);
1538                 gen8_for_each_pde(pt, pd, pd_start, pd_len, pde) {
1539                         u32 pte;
1540                         gen8_pte_t *pt_vaddr;
1541
1542                         if (pd->page_table[pde] == ppgtt->vm.scratch_pt)
1543                                 continue;
1544
1545                         pt_vaddr = kmap_atomic_px(pt);
1546                         for (pte = 0; pte < GEN8_PTES; pte += 4) {
1547                                 u64 va = (pdpe << GEN8_PDPE_SHIFT |
1548                                           pde << GEN8_PDE_SHIFT |
1549                                           pte << GEN8_PTE_SHIFT);
1550                                 int i;
1551                                 bool found = false;
1552
1553                                 for (i = 0; i < 4; i++)
1554                                         if (pt_vaddr[pte + i] != scratch_pte)
1555                                                 found = true;
1556                                 if (!found)
1557                                         continue;
1558
1559                                 seq_printf(m, "\t\t0x%llx [%03d,%03d,%04d]: =", va, pdpe, pde, pte);
1560                                 for (i = 0; i < 4; i++) {
1561                                         if (pt_vaddr[pte + i] != scratch_pte)
1562                                                 seq_printf(m, " %llx", pt_vaddr[pte + i]);
1563                                         else
1564                                                 seq_puts(m, "  SCRATCH ");
1565                                 }
1566                                 seq_puts(m, "\n");
1567                         }
1568                         kunmap_atomic(pt_vaddr);
1569                 }
1570         }
1571 }
1572
1573 static void gen8_dump_ppgtt(struct i915_hw_ppgtt *ppgtt, struct seq_file *m)
1574 {
1575         struct i915_address_space *vm = &ppgtt->vm;
1576         const gen8_pte_t scratch_pte =
1577                 gen8_pte_encode(vm->scratch_page.daddr, I915_CACHE_LLC, 0);
1578         u64 start = 0, length = ppgtt->vm.total;
1579
1580         if (use_4lvl(vm)) {
1581                 u64 pml4e;
1582                 struct i915_pml4 *pml4 = &ppgtt->pml4;
1583                 struct i915_page_directory_pointer *pdp;
1584
1585                 gen8_for_each_pml4e(pdp, pml4, start, length, pml4e) {
1586                         if (pml4->pdps[pml4e] == ppgtt->vm.scratch_pdp)
1587                                 continue;
1588
1589                         seq_printf(m, "    PML4E #%llu\n", pml4e);
1590                         gen8_dump_pdp(ppgtt, pdp, start, length, scratch_pte, m);
1591                 }
1592         } else {
1593                 gen8_dump_pdp(ppgtt, &ppgtt->pdp, start, length, scratch_pte, m);
1594         }
1595 }
1596
1597 static int gen8_preallocate_top_level_pdp(struct i915_hw_ppgtt *ppgtt)
1598 {
1599         struct i915_address_space *vm = &ppgtt->vm;
1600         struct i915_page_directory_pointer *pdp = &ppgtt->pdp;
1601         struct i915_page_directory *pd;
1602         u64 start = 0, length = ppgtt->vm.total;
1603         u64 from = start;
1604         unsigned int pdpe;
1605
1606         gen8_for_each_pdpe(pd, pdp, start, length, pdpe) {
1607                 pd = alloc_pd(vm);
1608                 if (IS_ERR(pd))
1609                         goto unwind;
1610
1611                 gen8_initialize_pd(vm, pd);
1612                 gen8_ppgtt_set_pdpe(vm, pdp, pd, pdpe);
1613                 pdp->used_pdpes++;
1614         }
1615
1616         pdp->used_pdpes++; /* never remove */
1617         return 0;
1618
1619 unwind:
1620         start -= from;
1621         gen8_for_each_pdpe(pd, pdp, from, start, pdpe) {
1622                 gen8_ppgtt_set_pdpe(vm, pdp, vm->scratch_pd, pdpe);
1623                 free_pd(vm, pd);
1624         }
1625         pdp->used_pdpes = 0;
1626         return -ENOMEM;
1627 }
1628
1629 /*
1630  * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
1631  * with a net effect resembling a 2-level page table in normal x86 terms. Each
1632  * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
1633  * space.
1634  *
1635  */
1636 static struct i915_hw_ppgtt *gen8_ppgtt_create(struct drm_i915_private *i915)
1637 {
1638         struct i915_hw_ppgtt *ppgtt;
1639         int err;
1640
1641         ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
1642         if (!ppgtt)
1643                 return ERR_PTR(-ENOMEM);
1644
1645         kref_init(&ppgtt->ref);
1646
1647         ppgtt->vm.i915 = i915;
1648         ppgtt->vm.dma = &i915->drm.pdev->dev;
1649
1650         ppgtt->vm.total = USES_FULL_48BIT_PPGTT(i915) ?
1651                 1ULL << 48 :
1652                 1ULL << 32;
1653
1654         /*
1655          * From bdw, there is support for read-only pages in the PPGTT.
1656          *
1657          * XXX GVT is not honouring the lack of RW in the PTE bits.
1658          */
1659         ppgtt->vm.has_read_only = !intel_vgpu_active(i915);
1660
1661         i915_address_space_init(&ppgtt->vm, i915);
1662
1663         /* There are only few exceptions for gen >=6. chv and bxt.
1664          * And we are not sure about the latter so play safe for now.
1665          */
1666         if (IS_CHERRYVIEW(i915) || IS_BROXTON(i915))
1667                 ppgtt->vm.pt_kmap_wc = true;
1668
1669         err = gen8_init_scratch(&ppgtt->vm);
1670         if (err)
1671                 goto err_free;
1672
1673         if (use_4lvl(&ppgtt->vm)) {
1674                 err = setup_px(&ppgtt->vm, &ppgtt->pml4);
1675                 if (err)
1676                         goto err_scratch;
1677
1678                 gen8_initialize_pml4(&ppgtt->vm, &ppgtt->pml4);
1679
1680                 ppgtt->vm.allocate_va_range = gen8_ppgtt_alloc_4lvl;
1681                 ppgtt->vm.insert_entries = gen8_ppgtt_insert_4lvl;
1682                 ppgtt->vm.clear_range = gen8_ppgtt_clear_4lvl;
1683         } else {
1684                 err = __pdp_init(&ppgtt->vm, &ppgtt->pdp);
1685                 if (err)
1686                         goto err_scratch;
1687
1688                 if (intel_vgpu_active(i915)) {
1689                         err = gen8_preallocate_top_level_pdp(ppgtt);
1690                         if (err) {
1691                                 __pdp_fini(&ppgtt->pdp);
1692                                 goto err_scratch;
1693                         }
1694                 }
1695
1696                 ppgtt->vm.allocate_va_range = gen8_ppgtt_alloc_3lvl;
1697                 ppgtt->vm.insert_entries = gen8_ppgtt_insert_3lvl;
1698                 ppgtt->vm.clear_range = gen8_ppgtt_clear_3lvl;
1699         }
1700
1701         if (intel_vgpu_active(i915))
1702                 gen8_ppgtt_notify_vgt(ppgtt, true);
1703
1704         ppgtt->vm.cleanup = gen8_ppgtt_cleanup;
1705         ppgtt->debug_dump = gen8_dump_ppgtt;
1706
1707         ppgtt->vm.vma_ops.bind_vma    = ppgtt_bind_vma;
1708         ppgtt->vm.vma_ops.unbind_vma  = ppgtt_unbind_vma;
1709         ppgtt->vm.vma_ops.set_pages   = ppgtt_set_pages;
1710         ppgtt->vm.vma_ops.clear_pages = clear_pages;
1711
1712         return ppgtt;
1713
1714 err_scratch:
1715         gen8_free_scratch(&ppgtt->vm);
1716 err_free:
1717         kfree(ppgtt);
1718         return ERR_PTR(err);
1719 }
1720
1721 static void gen6_dump_ppgtt(struct i915_hw_ppgtt *base, struct seq_file *m)
1722 {
1723         struct gen6_hw_ppgtt *ppgtt = to_gen6_ppgtt(base);
1724         const gen6_pte_t scratch_pte = ppgtt->scratch_pte;
1725         struct i915_page_table *pt;
1726         u32 pte, pde;
1727
1728         gen6_for_all_pdes(pt, &base->pd, pde) {
1729                 gen6_pte_t *vaddr;
1730
1731                 if (pt == base->vm.scratch_pt)
1732                         continue;
1733
1734                 if (i915_vma_is_bound(ppgtt->vma, I915_VMA_GLOBAL_BIND)) {
1735                         u32 expected =
1736                                 GEN6_PDE_ADDR_ENCODE(px_dma(pt)) |
1737                                 GEN6_PDE_VALID;
1738                         u32 pd_entry = readl(ppgtt->pd_addr + pde);
1739
1740                         if (pd_entry != expected)
1741                                 seq_printf(m,
1742                                            "\tPDE #%d mismatch: Actual PDE: %x Expected PDE: %x\n",
1743                                            pde,
1744                                            pd_entry,
1745                                            expected);
1746
1747                         seq_printf(m, "\tPDE: %x\n", pd_entry);
1748                 }
1749
1750                 vaddr = kmap_atomic_px(base->pd.page_table[pde]);
1751                 for (pte = 0; pte < GEN6_PTES; pte += 4) {
1752                         int i;
1753
1754                         for (i = 0; i < 4; i++)
1755                                 if (vaddr[pte + i] != scratch_pte)
1756                                         break;
1757                         if (i == 4)
1758                                 continue;
1759
1760                         seq_printf(m, "\t\t(%03d, %04d) %08llx: ",
1761                                    pde, pte,
1762                                    (pde * GEN6_PTES + pte) * I915_GTT_PAGE_SIZE);
1763                         for (i = 0; i < 4; i++) {
1764                                 if (vaddr[pte + i] != scratch_pte)
1765                                         seq_printf(m, " %08x", vaddr[pte + i]);
1766                                 else
1767                                         seq_puts(m, "  SCRATCH");
1768                         }
1769                         seq_puts(m, "\n");
1770                 }
1771                 kunmap_atomic(vaddr);
1772         }
1773 }
1774
1775 /* Write pde (index) from the page directory @pd to the page table @pt */
1776 static inline void gen6_write_pde(const struct gen6_hw_ppgtt *ppgtt,
1777                                   const unsigned int pde,
1778                                   const struct i915_page_table *pt)
1779 {
1780         /* Caller needs to make sure the write completes if necessary */
1781         iowrite32(GEN6_PDE_ADDR_ENCODE(px_dma(pt)) | GEN6_PDE_VALID,
1782                   ppgtt->pd_addr + pde);
1783 }
1784
1785 static void gen8_ppgtt_enable(struct drm_i915_private *dev_priv)
1786 {
1787         struct intel_engine_cs *engine;
1788         enum intel_engine_id id;
1789
1790         for_each_engine(engine, dev_priv, id) {
1791                 u32 four_level = USES_FULL_48BIT_PPGTT(dev_priv) ?
1792                                  GEN8_GFX_PPGTT_48B : 0;
1793                 I915_WRITE(RING_MODE_GEN7(engine),
1794                            _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE | four_level));
1795         }
1796 }
1797
1798 static void gen7_ppgtt_enable(struct drm_i915_private *dev_priv)
1799 {
1800         struct intel_engine_cs *engine;
1801         u32 ecochk, ecobits;
1802         enum intel_engine_id id;
1803
1804         ecobits = I915_READ(GAC_ECO_BITS);
1805         I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
1806
1807         ecochk = I915_READ(GAM_ECOCHK);
1808         if (IS_HASWELL(dev_priv)) {
1809                 ecochk |= ECOCHK_PPGTT_WB_HSW;
1810         } else {
1811                 ecochk |= ECOCHK_PPGTT_LLC_IVB;
1812                 ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
1813         }
1814         I915_WRITE(GAM_ECOCHK, ecochk);
1815
1816         for_each_engine(engine, dev_priv, id) {
1817                 /* GFX_MODE is per-ring on gen7+ */
1818                 I915_WRITE(RING_MODE_GEN7(engine),
1819                            _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1820         }
1821 }
1822
1823 static void gen6_ppgtt_enable(struct drm_i915_private *dev_priv)
1824 {
1825         u32 ecochk, gab_ctl, ecobits;
1826
1827         ecobits = I915_READ(GAC_ECO_BITS);
1828         I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
1829                    ECOBITS_PPGTT_CACHE64B);
1830
1831         gab_ctl = I915_READ(GAB_CTL);
1832         I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
1833
1834         ecochk = I915_READ(GAM_ECOCHK);
1835         I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT | ECOCHK_PPGTT_CACHE64B);
1836
1837         I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
1838 }
1839
1840 /* PPGTT support for Sandybdrige/Gen6 and later */
1841 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
1842                                    u64 start, u64 length)
1843 {
1844         struct gen6_hw_ppgtt *ppgtt = to_gen6_ppgtt(i915_vm_to_ppgtt(vm));
1845         unsigned int first_entry = start / I915_GTT_PAGE_SIZE;
1846         unsigned int pde = first_entry / GEN6_PTES;
1847         unsigned int pte = first_entry % GEN6_PTES;
1848         unsigned int num_entries = length / I915_GTT_PAGE_SIZE;
1849         const gen6_pte_t scratch_pte = ppgtt->scratch_pte;
1850
1851         while (num_entries) {
1852                 struct i915_page_table *pt = ppgtt->base.pd.page_table[pde++];
1853                 const unsigned int end = min(pte + num_entries, GEN6_PTES);
1854                 const unsigned int count = end - pte;
1855                 gen6_pte_t *vaddr;
1856
1857                 GEM_BUG_ON(pt == vm->scratch_pt);
1858
1859                 num_entries -= count;
1860
1861                 GEM_BUG_ON(count > pt->used_ptes);
1862                 pt->used_ptes -= count;
1863                 if (!pt->used_ptes)
1864                         ppgtt->scan_for_unused_pt = true;
1865
1866                 /*
1867                  * Note that the hw doesn't support removing PDE on the fly
1868                  * (they are cached inside the context with no means to
1869                  * invalidate the cache), so we can only reset the PTE
1870                  * entries back to scratch.
1871                  */
1872
1873                 vaddr = kmap_atomic_px(pt);
1874                 do {
1875                         vaddr[pte++] = scratch_pte;
1876                 } while (pte < end);
1877                 kunmap_atomic(vaddr);
1878
1879                 pte = 0;
1880         }
1881 }
1882
1883 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
1884                                       struct i915_vma *vma,
1885                                       enum i915_cache_level cache_level,
1886                                       u32 flags)
1887 {
1888         struct i915_hw_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
1889         unsigned first_entry = vma->node.start / I915_GTT_PAGE_SIZE;
1890         unsigned act_pt = first_entry / GEN6_PTES;
1891         unsigned act_pte = first_entry % GEN6_PTES;
1892         const u32 pte_encode = vm->pte_encode(0, cache_level, flags);
1893         struct sgt_dma iter = sgt_dma(vma);
1894         gen6_pte_t *vaddr;
1895
1896         GEM_BUG_ON(ppgtt->pd.page_table[act_pt] == vm->scratch_pt);
1897
1898         vaddr = kmap_atomic_px(ppgtt->pd.page_table[act_pt]);
1899         do {
1900                 vaddr[act_pte] = pte_encode | GEN6_PTE_ADDR_ENCODE(iter.dma);
1901
1902                 iter.dma += I915_GTT_PAGE_SIZE;
1903                 if (iter.dma == iter.max) {
1904                         iter.sg = __sg_next(iter.sg);
1905                         if (!iter.sg)
1906                                 break;
1907
1908                         iter.dma = sg_dma_address(iter.sg);
1909                         iter.max = iter.dma + iter.sg->length;
1910                 }
1911
1912                 if (++act_pte == GEN6_PTES) {
1913                         kunmap_atomic(vaddr);
1914                         vaddr = kmap_atomic_px(ppgtt->pd.page_table[++act_pt]);
1915                         act_pte = 0;
1916                 }
1917         } while (1);
1918         kunmap_atomic(vaddr);
1919
1920         vma->page_sizes.gtt = I915_GTT_PAGE_SIZE;
1921 }
1922
1923 static int gen6_alloc_va_range(struct i915_address_space *vm,
1924                                u64 start, u64 length)
1925 {
1926         struct gen6_hw_ppgtt *ppgtt = to_gen6_ppgtt(i915_vm_to_ppgtt(vm));
1927         struct i915_page_table *pt;
1928         u64 from = start;
1929         unsigned int pde;
1930         bool flush = false;
1931
1932         gen6_for_each_pde(pt, &ppgtt->base.pd, start, length, pde) {
1933                 const unsigned int count = gen6_pte_count(start, length);
1934
1935                 if (pt == vm->scratch_pt) {
1936                         pt = alloc_pt(vm);
1937                         if (IS_ERR(pt))
1938                                 goto unwind_out;
1939
1940                         gen6_initialize_pt(ppgtt, pt);
1941                         ppgtt->base.pd.page_table[pde] = pt;
1942
1943                         if (i915_vma_is_bound(ppgtt->vma,
1944                                               I915_VMA_GLOBAL_BIND)) {
1945                                 gen6_write_pde(ppgtt, pde, pt);
1946                                 flush = true;
1947                         }
1948
1949                         GEM_BUG_ON(pt->used_ptes);
1950                 }
1951
1952                 pt->used_ptes += count;
1953         }
1954
1955         if (flush) {
1956                 mark_tlbs_dirty(&ppgtt->base);
1957                 gen6_ggtt_invalidate(ppgtt->base.vm.i915);
1958         }
1959
1960         return 0;
1961
1962 unwind_out:
1963         gen6_ppgtt_clear_range(vm, from, start - from);
1964         return -ENOMEM;
1965 }
1966
1967 static int gen6_ppgtt_init_scratch(struct gen6_hw_ppgtt *ppgtt)
1968 {
1969         struct i915_address_space * const vm = &ppgtt->base.vm;
1970         struct i915_page_table *unused;
1971         u32 pde;
1972         int ret;
1973
1974         ret = setup_scratch_page(vm, __GFP_HIGHMEM);
1975         if (ret)
1976                 return ret;
1977
1978         ppgtt->scratch_pte =
1979                 vm->pte_encode(vm->scratch_page.daddr,
1980                                I915_CACHE_NONE, PTE_READ_ONLY);
1981
1982         vm->scratch_pt = alloc_pt(vm);
1983         if (IS_ERR(vm->scratch_pt)) {
1984                 cleanup_scratch_page(vm);
1985                 return PTR_ERR(vm->scratch_pt);
1986         }
1987
1988         gen6_initialize_pt(ppgtt, vm->scratch_pt);
1989         gen6_for_all_pdes(unused, &ppgtt->base.pd, pde)
1990                 ppgtt->base.pd.page_table[pde] = vm->scratch_pt;
1991
1992         return 0;
1993 }
1994
1995 static void gen6_ppgtt_free_scratch(struct i915_address_space *vm)
1996 {
1997         free_pt(vm, vm->scratch_pt);
1998         cleanup_scratch_page(vm);
1999 }
2000
2001 static void gen6_ppgtt_free_pd(struct gen6_hw_ppgtt *ppgtt)
2002 {
2003         struct i915_page_table *pt;
2004         u32 pde;
2005
2006         gen6_for_all_pdes(pt, &ppgtt->base.pd, pde)
2007                 if (pt != ppgtt->base.vm.scratch_pt)
2008                         free_pt(&ppgtt->base.vm, pt);
2009 }
2010
2011 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
2012 {
2013         struct gen6_hw_ppgtt *ppgtt = to_gen6_ppgtt(i915_vm_to_ppgtt(vm));
2014
2015         i915_vma_destroy(ppgtt->vma);
2016
2017         gen6_ppgtt_free_pd(ppgtt);
2018         gen6_ppgtt_free_scratch(vm);
2019 }
2020
2021 static int pd_vma_set_pages(struct i915_vma *vma)
2022 {
2023         vma->pages = ERR_PTR(-ENODEV);
2024         return 0;
2025 }
2026
2027 static void pd_vma_clear_pages(struct i915_vma *vma)
2028 {
2029         GEM_BUG_ON(!vma->pages);
2030
2031         vma->pages = NULL;
2032 }
2033
2034 static int pd_vma_bind(struct i915_vma *vma,
2035                        enum i915_cache_level cache_level,
2036                        u32 unused)
2037 {
2038         struct i915_ggtt *ggtt = i915_vm_to_ggtt(vma->vm);
2039         struct gen6_hw_ppgtt *ppgtt = vma->private;
2040         u32 ggtt_offset = i915_ggtt_offset(vma) / I915_GTT_PAGE_SIZE;
2041         struct i915_page_table *pt;
2042         unsigned int pde;
2043
2044         ppgtt->base.pd.base.ggtt_offset = ggtt_offset * sizeof(gen6_pte_t);
2045         ppgtt->pd_addr = (gen6_pte_t __iomem *)ggtt->gsm + ggtt_offset;
2046
2047         gen6_for_all_pdes(pt, &ppgtt->base.pd, pde)
2048                 gen6_write_pde(ppgtt, pde, pt);
2049
2050         mark_tlbs_dirty(&ppgtt->base);
2051         gen6_ggtt_invalidate(ppgtt->base.vm.i915);
2052
2053         return 0;
2054 }
2055
2056 static void pd_vma_unbind(struct i915_vma *vma)
2057 {
2058         struct gen6_hw_ppgtt *ppgtt = vma->private;
2059         struct i915_page_table * const scratch_pt = ppgtt->base.vm.scratch_pt;
2060         struct i915_page_table *pt;
2061         unsigned int pde;
2062
2063         if (!ppgtt->scan_for_unused_pt)
2064                 return;
2065
2066         /* Free all no longer used page tables */
2067         gen6_for_all_pdes(pt, &ppgtt->base.pd, pde) {
2068                 if (pt->used_ptes || pt == scratch_pt)
2069                         continue;
2070
2071                 free_pt(&ppgtt->base.vm, pt);
2072                 ppgtt->base.pd.page_table[pde] = scratch_pt;
2073         }
2074
2075         ppgtt->scan_for_unused_pt = false;
2076 }
2077
2078 static const struct i915_vma_ops pd_vma_ops = {
2079         .set_pages = pd_vma_set_pages,
2080         .clear_pages = pd_vma_clear_pages,
2081         .bind_vma = pd_vma_bind,
2082         .unbind_vma = pd_vma_unbind,
2083 };
2084
2085 static struct i915_vma *pd_vma_create(struct gen6_hw_ppgtt *ppgtt, int size)
2086 {
2087         struct drm_i915_private *i915 = ppgtt->base.vm.i915;
2088         struct i915_ggtt *ggtt = &i915->ggtt;
2089         struct i915_vma *vma;
2090
2091         GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
2092         GEM_BUG_ON(size > ggtt->vm.total);
2093
2094         vma = kmem_cache_zalloc(i915->vmas, GFP_KERNEL);
2095         if (!vma)
2096                 return ERR_PTR(-ENOMEM);
2097
2098         init_request_active(&vma->last_fence, NULL);
2099
2100         vma->vm = &ggtt->vm;
2101         vma->ops = &pd_vma_ops;
2102         vma->private = ppgtt;
2103
2104         vma->active = RB_ROOT;
2105
2106         vma->size = size;
2107         vma->fence_size = size;
2108         vma->flags = I915_VMA_GGTT;
2109         vma->ggtt_view.type = I915_GGTT_VIEW_ROTATED; /* prevent fencing */
2110
2111         INIT_LIST_HEAD(&vma->obj_link);
2112         list_add(&vma->vm_link, &vma->vm->unbound_list);
2113
2114         return vma;
2115 }
2116
2117 int gen6_ppgtt_pin(struct i915_hw_ppgtt *base)
2118 {
2119         struct gen6_hw_ppgtt *ppgtt = to_gen6_ppgtt(base);
2120
2121         /*
2122          * Workaround the limited maximum vma->pin_count and the aliasing_ppgtt
2123          * which will be pinned into every active context.
2124          * (When vma->pin_count becomes atomic, I expect we will naturally
2125          * need a larger, unpacked, type and kill this redundancy.)
2126          */
2127         if (ppgtt->pin_count++)
2128                 return 0;
2129
2130         /*
2131          * PPGTT PDEs reside in the GGTT and consists of 512 entries. The
2132          * allocator works in address space sizes, so it's multiplied by page
2133          * size. We allocate at the top of the GTT to avoid fragmentation.
2134          */
2135         return i915_vma_pin(ppgtt->vma,
2136                             0, GEN6_PD_ALIGN,
2137                             PIN_GLOBAL | PIN_HIGH);
2138 }
2139
2140 void gen6_ppgtt_unpin(struct i915_hw_ppgtt *base)
2141 {
2142         struct gen6_hw_ppgtt *ppgtt = to_gen6_ppgtt(base);
2143
2144         GEM_BUG_ON(!ppgtt->pin_count);
2145         if (--ppgtt->pin_count)
2146                 return;
2147
2148         i915_vma_unpin(ppgtt->vma);
2149 }
2150
2151 static struct i915_hw_ppgtt *gen6_ppgtt_create(struct drm_i915_private *i915)
2152 {
2153         struct i915_ggtt * const ggtt = &i915->ggtt;
2154         struct gen6_hw_ppgtt *ppgtt;
2155         int err;
2156
2157         ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
2158         if (!ppgtt)
2159                 return ERR_PTR(-ENOMEM);
2160
2161         kref_init(&ppgtt->base.ref);
2162
2163         ppgtt->base.vm.i915 = i915;
2164         ppgtt->base.vm.dma = &i915->drm.pdev->dev;
2165
2166         ppgtt->base.vm.total = I915_PDES * GEN6_PTES * I915_GTT_PAGE_SIZE;
2167
2168         i915_address_space_init(&ppgtt->base.vm, i915);
2169
2170         ppgtt->base.vm.allocate_va_range = gen6_alloc_va_range;
2171         ppgtt->base.vm.clear_range = gen6_ppgtt_clear_range;
2172         ppgtt->base.vm.insert_entries = gen6_ppgtt_insert_entries;
2173         ppgtt->base.vm.cleanup = gen6_ppgtt_cleanup;
2174         ppgtt->base.debug_dump = gen6_dump_ppgtt;
2175
2176         ppgtt->base.vm.vma_ops.bind_vma    = ppgtt_bind_vma;
2177         ppgtt->base.vm.vma_ops.unbind_vma  = ppgtt_unbind_vma;
2178         ppgtt->base.vm.vma_ops.set_pages   = ppgtt_set_pages;
2179         ppgtt->base.vm.vma_ops.clear_pages = clear_pages;
2180
2181         ppgtt->base.vm.pte_encode = ggtt->vm.pte_encode;
2182
2183         err = gen6_ppgtt_init_scratch(ppgtt);
2184         if (err)
2185                 goto err_free;
2186
2187         ppgtt->vma = pd_vma_create(ppgtt, GEN6_PD_SIZE);
2188         if (IS_ERR(ppgtt->vma)) {
2189                 err = PTR_ERR(ppgtt->vma);
2190                 goto err_scratch;
2191         }
2192
2193         return &ppgtt->base;
2194
2195 err_scratch:
2196         gen6_ppgtt_free_scratch(&ppgtt->base.vm);
2197 err_free:
2198         kfree(ppgtt);
2199         return ERR_PTR(err);
2200 }
2201
2202 static void gtt_write_workarounds(struct drm_i915_private *dev_priv)
2203 {
2204         /* This function is for gtt related workarounds. This function is
2205          * called on driver load and after a GPU reset, so you can place
2206          * workarounds here even if they get overwritten by GPU reset.
2207          */
2208         /* WaIncreaseDefaultTLBEntries:chv,bdw,skl,bxt,kbl,glk,cfl,cnl,icl */
2209         if (IS_BROADWELL(dev_priv))
2210                 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN8_L3_LRA_1_GPGPU_DEFAULT_VALUE_BDW);
2211         else if (IS_CHERRYVIEW(dev_priv))
2212                 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN8_L3_LRA_1_GPGPU_DEFAULT_VALUE_CHV);
2213         else if (IS_GEN9_LP(dev_priv))
2214                 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN9_L3_LRA_1_GPGPU_DEFAULT_VALUE_BXT);
2215         else if (INTEL_GEN(dev_priv) >= 9)
2216                 I915_WRITE(GEN8_L3_LRA_1_GPGPU, GEN9_L3_LRA_1_GPGPU_DEFAULT_VALUE_SKL);
2217
2218         /*
2219          * To support 64K PTEs we need to first enable the use of the
2220          * Intermediate-Page-Size(IPS) bit of the PDE field via some magical
2221          * mmio, otherwise the page-walker will simply ignore the IPS bit. This
2222          * shouldn't be needed after GEN10.
2223          *
2224          * 64K pages were first introduced from BDW+, although technically they
2225          * only *work* from gen9+. For pre-BDW we instead have the option for
2226          * 32K pages, but we don't currently have any support for it in our
2227          * driver.
2228          */
2229         if (HAS_PAGE_SIZES(dev_priv, I915_GTT_PAGE_SIZE_64K) &&
2230             INTEL_GEN(dev_priv) <= 10)
2231                 I915_WRITE(GEN8_GAMW_ECO_DEV_RW_IA,
2232                            I915_READ(GEN8_GAMW_ECO_DEV_RW_IA) |
2233                            GAMW_ECO_ENABLE_64K_IPS_FIELD);
2234 }
2235
2236 int i915_ppgtt_init_hw(struct drm_i915_private *dev_priv)
2237 {
2238         gtt_write_workarounds(dev_priv);
2239
2240         /* In the case of execlists, PPGTT is enabled by the context descriptor
2241          * and the PDPs are contained within the context itself.  We don't
2242          * need to do anything here. */
2243         if (HAS_LOGICAL_RING_CONTEXTS(dev_priv))
2244                 return 0;
2245
2246         if (!USES_PPGTT(dev_priv))
2247                 return 0;
2248
2249         if (IS_GEN6(dev_priv))
2250                 gen6_ppgtt_enable(dev_priv);
2251         else if (IS_GEN7(dev_priv))
2252                 gen7_ppgtt_enable(dev_priv);
2253         else if (INTEL_GEN(dev_priv) >= 8)
2254                 gen8_ppgtt_enable(dev_priv);
2255         else
2256                 MISSING_CASE(INTEL_GEN(dev_priv));
2257
2258         return 0;
2259 }
2260
2261 static struct i915_hw_ppgtt *
2262 __hw_ppgtt_create(struct drm_i915_private *i915)
2263 {
2264         if (INTEL_GEN(i915) < 8)
2265                 return gen6_ppgtt_create(i915);
2266         else
2267                 return gen8_ppgtt_create(i915);
2268 }
2269
2270 struct i915_hw_ppgtt *
2271 i915_ppgtt_create(struct drm_i915_private *i915,
2272                   struct drm_i915_file_private *fpriv)
2273 {
2274         struct i915_hw_ppgtt *ppgtt;
2275
2276         ppgtt = __hw_ppgtt_create(i915);
2277         if (IS_ERR(ppgtt))
2278                 return ppgtt;
2279
2280         ppgtt->vm.file = fpriv;
2281
2282         trace_i915_ppgtt_create(&ppgtt->vm);
2283
2284         return ppgtt;
2285 }
2286
2287 void i915_ppgtt_close(struct i915_address_space *vm)
2288 {
2289         GEM_BUG_ON(vm->closed);
2290         vm->closed = true;
2291 }
2292
2293 static void ppgtt_destroy_vma(struct i915_address_space *vm)
2294 {
2295         struct list_head *phases[] = {
2296                 &vm->active_list,
2297                 &vm->inactive_list,
2298                 &vm->unbound_list,
2299                 NULL,
2300         }, **phase;
2301
2302         vm->closed = true;
2303         for (phase = phases; *phase; phase++) {
2304                 struct i915_vma *vma, *vn;
2305
2306                 list_for_each_entry_safe(vma, vn, *phase, vm_link)
2307                         i915_vma_destroy(vma);
2308         }
2309 }
2310
2311 void i915_ppgtt_release(struct kref *kref)
2312 {
2313         struct i915_hw_ppgtt *ppgtt =
2314                 container_of(kref, struct i915_hw_ppgtt, ref);
2315
2316         trace_i915_ppgtt_release(&ppgtt->vm);
2317
2318         ppgtt_destroy_vma(&ppgtt->vm);
2319
2320         GEM_BUG_ON(!list_empty(&ppgtt->vm.active_list));
2321         GEM_BUG_ON(!list_empty(&ppgtt->vm.inactive_list));
2322         GEM_BUG_ON(!list_empty(&ppgtt->vm.unbound_list));
2323
2324         ppgtt->vm.cleanup(&ppgtt->vm);
2325         i915_address_space_fini(&ppgtt->vm);
2326         kfree(ppgtt);
2327 }
2328
2329 /* Certain Gen5 chipsets require require idling the GPU before
2330  * unmapping anything from the GTT when VT-d is enabled.
2331  */
2332 static bool needs_idle_maps(struct drm_i915_private *dev_priv)
2333 {
2334         /* Query intel_iommu to see if we need the workaround. Presumably that
2335          * was loaded first.
2336          */
2337         return IS_GEN5(dev_priv) && IS_MOBILE(dev_priv) && intel_vtd_active();
2338 }
2339
2340 static void gen6_check_faults(struct drm_i915_private *dev_priv)
2341 {
2342         struct intel_engine_cs *engine;
2343         enum intel_engine_id id;
2344         u32 fault;
2345
2346         for_each_engine(engine, dev_priv, id) {
2347                 fault = I915_READ(RING_FAULT_REG(engine));
2348                 if (fault & RING_FAULT_VALID) {
2349                         DRM_DEBUG_DRIVER("Unexpected fault\n"
2350                                          "\tAddr: 0x%08lx\n"
2351                                          "\tAddress space: %s\n"
2352                                          "\tSource ID: %d\n"
2353                                          "\tType: %d\n",
2354                                          fault & PAGE_MASK,
2355                                          fault & RING_FAULT_GTTSEL_MASK ? "GGTT" : "PPGTT",
2356                                          RING_FAULT_SRCID(fault),
2357                                          RING_FAULT_FAULT_TYPE(fault));
2358                 }
2359         }
2360 }
2361
2362 static void gen8_check_faults(struct drm_i915_private *dev_priv)
2363 {
2364         u32 fault = I915_READ(GEN8_RING_FAULT_REG);
2365
2366         if (fault & RING_FAULT_VALID) {
2367                 u32 fault_data0, fault_data1;
2368                 u64 fault_addr;
2369
2370                 fault_data0 = I915_READ(GEN8_FAULT_TLB_DATA0);
2371                 fault_data1 = I915_READ(GEN8_FAULT_TLB_DATA1);
2372                 fault_addr = ((u64)(fault_data1 & FAULT_VA_HIGH_BITS) << 44) |
2373                              ((u64)fault_data0 << 12);
2374
2375                 DRM_DEBUG_DRIVER("Unexpected fault\n"
2376                                  "\tAddr: 0x%08x_%08x\n"
2377                                  "\tAddress space: %s\n"
2378                                  "\tEngine ID: %d\n"
2379                                  "\tSource ID: %d\n"
2380                                  "\tType: %d\n",
2381                                  upper_32_bits(fault_addr),
2382                                  lower_32_bits(fault_addr),
2383                                  fault_data1 & FAULT_GTT_SEL ? "GGTT" : "PPGTT",
2384                                  GEN8_RING_FAULT_ENGINE_ID(fault),
2385                                  RING_FAULT_SRCID(fault),
2386                                  RING_FAULT_FAULT_TYPE(fault));
2387         }
2388 }
2389
2390 void i915_check_and_clear_faults(struct drm_i915_private *dev_priv)
2391 {
2392         /* From GEN8 onwards we only have one 'All Engine Fault Register' */
2393         if (INTEL_GEN(dev_priv) >= 8)
2394                 gen8_check_faults(dev_priv);
2395         else if (INTEL_GEN(dev_priv) >= 6)
2396                 gen6_check_faults(dev_priv);
2397         else
2398                 return;
2399
2400         i915_clear_error_registers(dev_priv);
2401 }
2402
2403 void i915_gem_suspend_gtt_mappings(struct drm_i915_private *dev_priv)
2404 {
2405         struct i915_ggtt *ggtt = &dev_priv->ggtt;
2406
2407         /* Don't bother messing with faults pre GEN6 as we have little
2408          * documentation supporting that it's a good idea.
2409          */
2410         if (INTEL_GEN(dev_priv) < 6)
2411                 return;
2412
2413         i915_check_and_clear_faults(dev_priv);
2414
2415         ggtt->vm.clear_range(&ggtt->vm, 0, ggtt->vm.total);
2416
2417         i915_ggtt_invalidate(dev_priv);
2418 }
2419
2420 int i915_gem_gtt_prepare_pages(struct drm_i915_gem_object *obj,
2421                                struct sg_table *pages)
2422 {
2423         do {
2424                 if (dma_map_sg_attrs(&obj->base.dev->pdev->dev,
2425                                      pages->sgl, pages->nents,
2426                                      PCI_DMA_BIDIRECTIONAL,
2427                                      DMA_ATTR_NO_WARN))
2428                         return 0;
2429
2430                 /* If the DMA remap fails, one cause can be that we have
2431                  * too many objects pinned in a small remapping table,
2432                  * such as swiotlb. Incrementally purge all other objects and
2433                  * try again - if there are no more pages to remove from
2434                  * the DMA remapper, i915_gem_shrink will return 0.
2435                  */
2436                 GEM_BUG_ON(obj->mm.pages == pages);
2437         } while (i915_gem_shrink(to_i915(obj->base.dev),
2438                                  obj->base.size >> PAGE_SHIFT, NULL,
2439                                  I915_SHRINK_BOUND |
2440                                  I915_SHRINK_UNBOUND |
2441                                  I915_SHRINK_ACTIVE));
2442
2443         return -ENOSPC;
2444 }
2445
2446 static void gen8_set_pte(void __iomem *addr, gen8_pte_t pte)
2447 {
2448         writeq(pte, addr);
2449 }
2450
2451 static void gen8_ggtt_insert_page(struct i915_address_space *vm,
2452                                   dma_addr_t addr,
2453                                   u64 offset,
2454                                   enum i915_cache_level level,
2455                                   u32 unused)
2456 {
2457         struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2458         gen8_pte_t __iomem *pte =
2459                 (gen8_pte_t __iomem *)ggtt->gsm + offset / I915_GTT_PAGE_SIZE;
2460
2461         gen8_set_pte(pte, gen8_pte_encode(addr, level, 0));
2462
2463         ggtt->invalidate(vm->i915);
2464 }
2465
2466 static void gen8_ggtt_insert_entries(struct i915_address_space *vm,
2467                                      struct i915_vma *vma,
2468                                      enum i915_cache_level level,
2469                                      u32 flags)
2470 {
2471         struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2472         struct sgt_iter sgt_iter;
2473         gen8_pte_t __iomem *gtt_entries;
2474         const gen8_pte_t pte_encode = gen8_pte_encode(0, level, 0);
2475         dma_addr_t addr;
2476
2477         /*
2478          * Note that we ignore PTE_READ_ONLY here. The caller must be careful
2479          * not to allow the user to override access to a read only page.
2480          */
2481
2482         gtt_entries = (gen8_pte_t __iomem *)ggtt->gsm;
2483         gtt_entries += vma->node.start / I915_GTT_PAGE_SIZE;
2484         for_each_sgt_dma(addr, sgt_iter, vma->pages)
2485                 gen8_set_pte(gtt_entries++, pte_encode | addr);
2486
2487         /*
2488          * We want to flush the TLBs only after we're certain all the PTE
2489          * updates have finished.
2490          */
2491         ggtt->invalidate(vm->i915);
2492 }
2493
2494 static void gen6_ggtt_insert_page(struct i915_address_space *vm,
2495                                   dma_addr_t addr,
2496                                   u64 offset,
2497                                   enum i915_cache_level level,
2498                                   u32 flags)
2499 {
2500         struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2501         gen6_pte_t __iomem *pte =
2502                 (gen6_pte_t __iomem *)ggtt->gsm + offset / I915_GTT_PAGE_SIZE;
2503
2504         iowrite32(vm->pte_encode(addr, level, flags), pte);
2505
2506         ggtt->invalidate(vm->i915);
2507 }
2508
2509 /*
2510  * Binds an object into the global gtt with the specified cache level. The object
2511  * will be accessible to the GPU via commands whose operands reference offsets
2512  * within the global GTT as well as accessible by the GPU through the GMADR
2513  * mapped BAR (dev_priv->mm.gtt->gtt).
2514  */
2515 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
2516                                      struct i915_vma *vma,
2517                                      enum i915_cache_level level,
2518                                      u32 flags)
2519 {
2520         struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2521         gen6_pte_t __iomem *entries = (gen6_pte_t __iomem *)ggtt->gsm;
2522         unsigned int i = vma->node.start / I915_GTT_PAGE_SIZE;
2523         struct sgt_iter iter;
2524         dma_addr_t addr;
2525         for_each_sgt_dma(addr, iter, vma->pages)
2526                 iowrite32(vm->pte_encode(addr, level, flags), &entries[i++]);
2527
2528         /*
2529          * We want to flush the TLBs only after we're certain all the PTE
2530          * updates have finished.
2531          */
2532         ggtt->invalidate(vm->i915);
2533 }
2534
2535 static void nop_clear_range(struct i915_address_space *vm,
2536                             u64 start, u64 length)
2537 {
2538 }
2539
2540 static void gen8_ggtt_clear_range(struct i915_address_space *vm,
2541                                   u64 start, u64 length)
2542 {
2543         struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2544         unsigned first_entry = start / I915_GTT_PAGE_SIZE;
2545         unsigned num_entries = length / I915_GTT_PAGE_SIZE;
2546         const gen8_pte_t scratch_pte =
2547                 gen8_pte_encode(vm->scratch_page.daddr, I915_CACHE_LLC, 0);
2548         gen8_pte_t __iomem *gtt_base =
2549                 (gen8_pte_t __iomem *)ggtt->gsm + first_entry;
2550         const int max_entries = ggtt_total_entries(ggtt) - first_entry;
2551         int i;
2552
2553         if (WARN(num_entries > max_entries,
2554                  "First entry = %d; Num entries = %d (max=%d)\n",
2555                  first_entry, num_entries, max_entries))
2556                 num_entries = max_entries;
2557
2558         for (i = 0; i < num_entries; i++)
2559                 gen8_set_pte(&gtt_base[i], scratch_pte);
2560 }
2561
2562 static void bxt_vtd_ggtt_wa(struct i915_address_space *vm)
2563 {
2564         struct drm_i915_private *dev_priv = vm->i915;
2565
2566         /*
2567          * Make sure the internal GAM fifo has been cleared of all GTT
2568          * writes before exiting stop_machine(). This guarantees that
2569          * any aperture accesses waiting to start in another process
2570          * cannot back up behind the GTT writes causing a hang.
2571          * The register can be any arbitrary GAM register.
2572          */
2573         POSTING_READ(GFX_FLSH_CNTL_GEN6);
2574 }
2575
2576 struct insert_page {
2577         struct i915_address_space *vm;
2578         dma_addr_t addr;
2579         u64 offset;
2580         enum i915_cache_level level;
2581 };
2582
2583 static int bxt_vtd_ggtt_insert_page__cb(void *_arg)
2584 {
2585         struct insert_page *arg = _arg;
2586
2587         gen8_ggtt_insert_page(arg->vm, arg->addr, arg->offset, arg->level, 0);
2588         bxt_vtd_ggtt_wa(arg->vm);
2589
2590         return 0;
2591 }
2592
2593 static void bxt_vtd_ggtt_insert_page__BKL(struct i915_address_space *vm,
2594                                           dma_addr_t addr,
2595                                           u64 offset,
2596                                           enum i915_cache_level level,
2597                                           u32 unused)
2598 {
2599         struct insert_page arg = { vm, addr, offset, level };
2600
2601         stop_machine(bxt_vtd_ggtt_insert_page__cb, &arg, NULL);
2602 }
2603
2604 struct insert_entries {
2605         struct i915_address_space *vm;
2606         struct i915_vma *vma;
2607         enum i915_cache_level level;
2608         u32 flags;
2609 };
2610
2611 static int bxt_vtd_ggtt_insert_entries__cb(void *_arg)
2612 {
2613         struct insert_entries *arg = _arg;
2614
2615         gen8_ggtt_insert_entries(arg->vm, arg->vma, arg->level, arg->flags);
2616         bxt_vtd_ggtt_wa(arg->vm);
2617
2618         return 0;
2619 }
2620
2621 static void bxt_vtd_ggtt_insert_entries__BKL(struct i915_address_space *vm,
2622                                              struct i915_vma *vma,
2623                                              enum i915_cache_level level,
2624                                              u32 flags)
2625 {
2626         struct insert_entries arg = { vm, vma, level, flags };
2627
2628         stop_machine(bxt_vtd_ggtt_insert_entries__cb, &arg, NULL);
2629 }
2630
2631 struct clear_range {
2632         struct i915_address_space *vm;
2633         u64 start;
2634         u64 length;
2635 };
2636
2637 static int bxt_vtd_ggtt_clear_range__cb(void *_arg)
2638 {
2639         struct clear_range *arg = _arg;
2640
2641         gen8_ggtt_clear_range(arg->vm, arg->start, arg->length);
2642         bxt_vtd_ggtt_wa(arg->vm);
2643
2644         return 0;
2645 }
2646
2647 static void bxt_vtd_ggtt_clear_range__BKL(struct i915_address_space *vm,
2648                                           u64 start,
2649                                           u64 length)
2650 {
2651         struct clear_range arg = { vm, start, length };
2652
2653         stop_machine(bxt_vtd_ggtt_clear_range__cb, &arg, NULL);
2654 }
2655
2656 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
2657                                   u64 start, u64 length)
2658 {
2659         struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
2660         unsigned first_entry = start / I915_GTT_PAGE_SIZE;
2661         unsigned num_entries = length / I915_GTT_PAGE_SIZE;
2662         gen6_pte_t scratch_pte, __iomem *gtt_base =
2663                 (gen6_pte_t __iomem *)ggtt->gsm + first_entry;
2664         const int max_entries = ggtt_total_entries(ggtt) - first_entry;
2665         int i;
2666
2667         if (WARN(num_entries > max_entries,
2668                  "First entry = %d; Num entries = %d (max=%d)\n",
2669                  first_entry, num_entries, max_entries))
2670                 num_entries = max_entries;
2671
2672         scratch_pte = vm->pte_encode(vm->scratch_page.daddr,
2673                                      I915_CACHE_LLC, 0);
2674
2675         for (i = 0; i < num_entries; i++)
2676                 iowrite32(scratch_pte, &gtt_base[i]);
2677 }
2678
2679 static void i915_ggtt_insert_page(struct i915_address_space *vm,
2680                                   dma_addr_t addr,
2681                                   u64 offset,
2682                                   enum i915_cache_level cache_level,
2683                                   u32 unused)
2684 {
2685         unsigned int flags = (cache_level == I915_CACHE_NONE) ?
2686                 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
2687
2688         intel_gtt_insert_page(addr, offset >> PAGE_SHIFT, flags);
2689 }
2690
2691 static void i915_ggtt_insert_entries(struct i915_address_space *vm,
2692                                      struct i915_vma *vma,
2693                                      enum i915_cache_level cache_level,
2694                                      u32 unused)
2695 {
2696         unsigned int flags = (cache_level == I915_CACHE_NONE) ?
2697                 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
2698
2699         intel_gtt_insert_sg_entries(vma->pages, vma->node.start >> PAGE_SHIFT,
2700                                     flags);
2701 }
2702
2703 static void i915_ggtt_clear_range(struct i915_address_space *vm,
2704                                   u64 start, u64 length)
2705 {
2706         intel_gtt_clear_range(start >> PAGE_SHIFT, length >> PAGE_SHIFT);
2707 }
2708
2709 static int ggtt_bind_vma(struct i915_vma *vma,
2710                          enum i915_cache_level cache_level,
2711                          u32 flags)
2712 {
2713         struct drm_i915_private *i915 = vma->vm->i915;
2714         struct drm_i915_gem_object *obj = vma->obj;
2715         u32 pte_flags;
2716
2717         /* Applicable to VLV (gen8+ do not support RO in the GGTT) */
2718         pte_flags = 0;
2719         if (i915_gem_object_is_readonly(obj))
2720                 pte_flags |= PTE_READ_ONLY;
2721
2722         intel_runtime_pm_get(i915);
2723         vma->vm->insert_entries(vma->vm, vma, cache_level, pte_flags);
2724         intel_runtime_pm_put(i915);
2725
2726         vma->page_sizes.gtt = I915_GTT_PAGE_SIZE;
2727
2728         /*
2729          * Without aliasing PPGTT there's no difference between
2730          * GLOBAL/LOCAL_BIND, it's all the same ptes. Hence unconditionally
2731          * upgrade to both bound if we bind either to avoid double-binding.
2732          */
2733         vma->flags |= I915_VMA_GLOBAL_BIND | I915_VMA_LOCAL_BIND;
2734
2735         return 0;
2736 }
2737
2738 static void ggtt_unbind_vma(struct i915_vma *vma)
2739 {
2740         struct drm_i915_private *i915 = vma->vm->i915;
2741
2742         intel_runtime_pm_get(i915);
2743         vma->vm->clear_range(vma->vm, vma->node.start, vma->size);
2744         intel_runtime_pm_put(i915);
2745 }
2746
2747 static int aliasing_gtt_bind_vma(struct i915_vma *vma,
2748                                  enum i915_cache_level cache_level,
2749                                  u32 flags)
2750 {
2751         struct drm_i915_private *i915 = vma->vm->i915;
2752         u32 pte_flags;
2753         int ret;
2754
2755         /* Currently applicable only to VLV */
2756         pte_flags = 0;
2757         if (i915_gem_object_is_readonly(vma->obj))
2758                 pte_flags |= PTE_READ_ONLY;
2759
2760         if (flags & I915_VMA_LOCAL_BIND) {
2761                 struct i915_hw_ppgtt *appgtt = i915->mm.aliasing_ppgtt;
2762
2763                 if (!(vma->flags & I915_VMA_LOCAL_BIND)) {
2764                         ret = appgtt->vm.allocate_va_range(&appgtt->vm,
2765                                                            vma->node.start,
2766                                                            vma->size);
2767                         if (ret)
2768                                 return ret;
2769                 }
2770
2771                 appgtt->vm.insert_entries(&appgtt->vm, vma, cache_level,
2772                                           pte_flags);
2773         }
2774
2775         if (flags & I915_VMA_GLOBAL_BIND) {
2776                 intel_runtime_pm_get(i915);
2777                 vma->vm->insert_entries(vma->vm, vma, cache_level, pte_flags);
2778                 intel_runtime_pm_put(i915);
2779         }
2780
2781         return 0;
2782 }
2783
2784 static void aliasing_gtt_unbind_vma(struct i915_vma *vma)
2785 {
2786         struct drm_i915_private *i915 = vma->vm->i915;
2787
2788         if (vma->flags & I915_VMA_GLOBAL_BIND) {
2789                 intel_runtime_pm_get(i915);
2790                 vma->vm->clear_range(vma->vm, vma->node.start, vma->size);
2791                 intel_runtime_pm_put(i915);
2792         }
2793
2794         if (vma->flags & I915_VMA_LOCAL_BIND) {
2795                 struct i915_address_space *vm = &i915->mm.aliasing_ppgtt->vm;
2796
2797                 vm->clear_range(vm, vma->node.start, vma->size);
2798         }
2799 }
2800
2801 void i915_gem_gtt_finish_pages(struct drm_i915_gem_object *obj,
2802                                struct sg_table *pages)
2803 {
2804         struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
2805         struct device *kdev = &dev_priv->drm.pdev->dev;
2806         struct i915_ggtt *ggtt = &dev_priv->ggtt;
2807
2808         if (unlikely(ggtt->do_idle_maps)) {
2809                 if (i915_gem_wait_for_idle(dev_priv, 0, MAX_SCHEDULE_TIMEOUT)) {
2810                         DRM_ERROR("Failed to wait for idle; VT'd may hang.\n");
2811                         /* Wait a bit, in hopes it avoids the hang */
2812                         udelay(10);
2813                 }
2814         }
2815
2816         dma_unmap_sg(kdev, pages->sgl, pages->nents, PCI_DMA_BIDIRECTIONAL);
2817 }
2818
2819 static int ggtt_set_pages(struct i915_vma *vma)
2820 {
2821         int ret;
2822
2823         GEM_BUG_ON(vma->pages);
2824
2825         ret = i915_get_ggtt_vma_pages(vma);
2826         if (ret)
2827                 return ret;
2828
2829         vma->page_sizes = vma->obj->mm.page_sizes;
2830
2831         return 0;
2832 }
2833
2834 static void i915_gtt_color_adjust(const struct drm_mm_node *node,
2835                                   unsigned long color,
2836                                   u64 *start,
2837                                   u64 *end)
2838 {
2839         if (node->allocated && node->color != color)
2840                 *start += I915_GTT_PAGE_SIZE;
2841
2842         /* Also leave a space between the unallocated reserved node after the
2843          * GTT and any objects within the GTT, i.e. we use the color adjustment
2844          * to insert a guard page to prevent prefetches crossing over the
2845          * GTT boundary.
2846          */
2847         node = list_next_entry(node, node_list);
2848         if (node->color != color)
2849                 *end -= I915_GTT_PAGE_SIZE;
2850 }
2851
2852 int i915_gem_init_aliasing_ppgtt(struct drm_i915_private *i915)
2853 {
2854         struct i915_ggtt *ggtt = &i915->ggtt;
2855         struct i915_hw_ppgtt *ppgtt;
2856         int err;
2857
2858         ppgtt = i915_ppgtt_create(i915, ERR_PTR(-EPERM));
2859         if (IS_ERR(ppgtt))
2860                 return PTR_ERR(ppgtt);
2861
2862         if (GEM_WARN_ON(ppgtt->vm.total < ggtt->vm.total)) {
2863                 err = -ENODEV;
2864                 goto err_ppgtt;
2865         }
2866
2867         /*
2868          * Note we only pre-allocate as far as the end of the global
2869          * GTT. On 48b / 4-level page-tables, the difference is very,
2870          * very significant! We have to preallocate as GVT/vgpu does
2871          * not like the page directory disappearing.
2872          */
2873         err = ppgtt->vm.allocate_va_range(&ppgtt->vm, 0, ggtt->vm.total);
2874         if (err)
2875                 goto err_ppgtt;
2876
2877         i915->mm.aliasing_ppgtt = ppgtt;
2878
2879         GEM_BUG_ON(ggtt->vm.vma_ops.bind_vma != ggtt_bind_vma);
2880         ggtt->vm.vma_ops.bind_vma = aliasing_gtt_bind_vma;
2881
2882         GEM_BUG_ON(ggtt->vm.vma_ops.unbind_vma != ggtt_unbind_vma);
2883         ggtt->vm.vma_ops.unbind_vma = aliasing_gtt_unbind_vma;
2884
2885         return 0;
2886
2887 err_ppgtt:
2888         i915_ppgtt_put(ppgtt);
2889         return err;
2890 }
2891
2892 void i915_gem_fini_aliasing_ppgtt(struct drm_i915_private *i915)
2893 {
2894         struct i915_ggtt *ggtt = &i915->ggtt;
2895         struct i915_hw_ppgtt *ppgtt;
2896
2897         ppgtt = fetch_and_zero(&i915->mm.aliasing_ppgtt);
2898         if (!ppgtt)
2899                 return;
2900
2901         i915_ppgtt_put(ppgtt);
2902
2903         ggtt->vm.vma_ops.bind_vma   = ggtt_bind_vma;
2904         ggtt->vm.vma_ops.unbind_vma = ggtt_unbind_vma;
2905 }
2906
2907 int i915_gem_init_ggtt(struct drm_i915_private *dev_priv)
2908 {
2909         /* Let GEM Manage all of the aperture.
2910          *
2911          * However, leave one page at the end still bound to the scratch page.
2912          * There are a number of places where the hardware apparently prefetches
2913          * past the end of the object, and we've seen multiple hangs with the
2914          * GPU head pointer stuck in a batchbuffer bound at the last page of the
2915          * aperture.  One page should be enough to keep any prefetching inside
2916          * of the aperture.
2917          */
2918         struct i915_ggtt *ggtt = &dev_priv->ggtt;
2919         unsigned long hole_start, hole_end;
2920         struct drm_mm_node *entry;
2921         int ret;
2922
2923         /*
2924          * GuC requires all resources that we're sharing with it to be placed in
2925          * non-WOPCM memory. If GuC is not present or not in use we still need a
2926          * small bias as ring wraparound at offset 0 sometimes hangs. No idea
2927          * why.
2928          */
2929         ggtt->pin_bias = max_t(u32, I915_GTT_PAGE_SIZE,
2930                                intel_guc_reserved_gtt_size(&dev_priv->guc));
2931
2932         ret = intel_vgt_balloon(dev_priv);
2933         if (ret)
2934                 return ret;
2935
2936         /* Reserve a mappable slot for our lockless error capture */
2937         ret = drm_mm_insert_node_in_range(&ggtt->vm.mm, &ggtt->error_capture,
2938                                           PAGE_SIZE, 0, I915_COLOR_UNEVICTABLE,
2939                                           0, ggtt->mappable_end,
2940                                           DRM_MM_INSERT_LOW);
2941         if (ret)
2942                 return ret;
2943
2944         /* Clear any non-preallocated blocks */
2945         drm_mm_for_each_hole(entry, &ggtt->vm.mm, hole_start, hole_end) {
2946                 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
2947                               hole_start, hole_end);
2948                 ggtt->vm.clear_range(&ggtt->vm, hole_start,
2949                                      hole_end - hole_start);
2950         }
2951
2952         /* And finally clear the reserved guard page */
2953         ggtt->vm.clear_range(&ggtt->vm, ggtt->vm.total - PAGE_SIZE, PAGE_SIZE);
2954
2955         if (USES_PPGTT(dev_priv) && !USES_FULL_PPGTT(dev_priv)) {
2956                 ret = i915_gem_init_aliasing_ppgtt(dev_priv);
2957                 if (ret)
2958                         goto err;
2959         }
2960
2961         return 0;
2962
2963 err:
2964         drm_mm_remove_node(&ggtt->error_capture);
2965         return ret;
2966 }
2967
2968 /**
2969  * i915_ggtt_cleanup_hw - Clean up GGTT hardware initialization
2970  * @dev_priv: i915 device
2971  */
2972 void i915_ggtt_cleanup_hw(struct drm_i915_private *dev_priv)
2973 {
2974         struct i915_ggtt *ggtt = &dev_priv->ggtt;
2975         struct i915_vma *vma, *vn;
2976         struct pagevec *pvec;
2977
2978         ggtt->vm.closed = true;
2979
2980         mutex_lock(&dev_priv->drm.struct_mutex);
2981         i915_gem_fini_aliasing_ppgtt(dev_priv);
2982
2983         GEM_BUG_ON(!list_empty(&ggtt->vm.active_list));
2984         list_for_each_entry_safe(vma, vn, &ggtt->vm.inactive_list, vm_link)
2985                 WARN_ON(i915_vma_unbind(vma));
2986
2987         if (drm_mm_node_allocated(&ggtt->error_capture))
2988                 drm_mm_remove_node(&ggtt->error_capture);
2989
2990         if (drm_mm_initialized(&ggtt->vm.mm)) {
2991                 intel_vgt_deballoon(dev_priv);
2992                 i915_address_space_fini(&ggtt->vm);
2993         }
2994
2995         ggtt->vm.cleanup(&ggtt->vm);
2996
2997         pvec = &dev_priv->mm.wc_stash.pvec;
2998         if (pvec->nr) {
2999                 set_pages_array_wb(pvec->pages, pvec->nr);
3000                 __pagevec_release(pvec);
3001         }
3002
3003         mutex_unlock(&dev_priv->drm.struct_mutex);
3004
3005         arch_phys_wc_del(ggtt->mtrr);
3006         io_mapping_fini(&ggtt->iomap);
3007
3008         i915_gem_cleanup_stolen(dev_priv);
3009 }
3010
3011 static unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
3012 {
3013         snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
3014         snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
3015         return snb_gmch_ctl << 20;
3016 }
3017
3018 static unsigned int gen8_get_total_gtt_size(u16 bdw_gmch_ctl)
3019 {
3020         bdw_gmch_ctl >>= BDW_GMCH_GGMS_SHIFT;
3021         bdw_gmch_ctl &= BDW_GMCH_GGMS_MASK;
3022         if (bdw_gmch_ctl)
3023                 bdw_gmch_ctl = 1 << bdw_gmch_ctl;
3024
3025 #ifdef CONFIG_X86_32
3026         /* Limit 32b platforms to a 2GB GGTT: 4 << 20 / pte size * I915_GTT_PAGE_SIZE */
3027         if (bdw_gmch_ctl > 4)
3028                 bdw_gmch_ctl = 4;
3029 #endif
3030
3031         return bdw_gmch_ctl << 20;
3032 }
3033
3034 static unsigned int chv_get_total_gtt_size(u16 gmch_ctrl)
3035 {
3036         gmch_ctrl >>= SNB_GMCH_GGMS_SHIFT;
3037         gmch_ctrl &= SNB_GMCH_GGMS_MASK;
3038
3039         if (gmch_ctrl)
3040                 return 1 << (20 + gmch_ctrl);
3041
3042         return 0;
3043 }
3044
3045 static int ggtt_probe_common(struct i915_ggtt *ggtt, u64 size)
3046 {
3047         struct drm_i915_private *dev_priv = ggtt->vm.i915;
3048         struct pci_dev *pdev = dev_priv->drm.pdev;
3049         phys_addr_t phys_addr;
3050         int ret;
3051
3052         /* For Modern GENs the PTEs and register space are split in the BAR */
3053         phys_addr = pci_resource_start(pdev, 0) + pci_resource_len(pdev, 0) / 2;
3054
3055         /*
3056          * On BXT+/CNL+ writes larger than 64 bit to the GTT pagetable range
3057          * will be dropped. For WC mappings in general we have 64 byte burst
3058          * writes when the WC buffer is flushed, so we can't use it, but have to
3059          * resort to an uncached mapping. The WC issue is easily caught by the
3060          * readback check when writing GTT PTE entries.
3061          */
3062         if (IS_GEN9_LP(dev_priv) || INTEL_GEN(dev_priv) >= 10)
3063                 ggtt->gsm = ioremap_nocache(phys_addr, size);
3064         else
3065                 ggtt->gsm = ioremap_wc(phys_addr, size);
3066         if (!ggtt->gsm) {
3067                 DRM_ERROR("Failed to map the ggtt page table\n");
3068                 return -ENOMEM;
3069         }
3070
3071         ret = setup_scratch_page(&ggtt->vm, GFP_DMA32);
3072         if (ret) {
3073                 DRM_ERROR("Scratch setup failed\n");
3074                 /* iounmap will also get called at remove, but meh */
3075                 iounmap(ggtt->gsm);
3076                 return ret;
3077         }
3078
3079         return 0;
3080 }
3081
3082 static struct intel_ppat_entry *
3083 __alloc_ppat_entry(struct intel_ppat *ppat, unsigned int index, u8 value)
3084 {
3085         struct intel_ppat_entry *entry = &ppat->entries[index];
3086
3087         GEM_BUG_ON(index >= ppat->max_entries);
3088         GEM_BUG_ON(test_bit(index, ppat->used));
3089
3090         entry->ppat = ppat;
3091         entry->value = value;
3092         kref_init(&entry->ref);
3093         set_bit(index, ppat->used);
3094         set_bit(index, ppat->dirty);
3095
3096         return entry;
3097 }
3098
3099 static void __free_ppat_entry(struct intel_ppat_entry *entry)
3100 {
3101         struct intel_ppat *ppat = entry->ppat;
3102         unsigned int index = entry - ppat->entries;
3103
3104         GEM_BUG_ON(index >= ppat->max_entries);
3105         GEM_BUG_ON(!test_bit(index, ppat->used));
3106
3107         entry->value = ppat->clear_value;
3108         clear_bit(index, ppat->used);
3109         set_bit(index, ppat->dirty);
3110 }
3111
3112 /**
3113  * intel_ppat_get - get a usable PPAT entry
3114  * @i915: i915 device instance
3115  * @value: the PPAT value required by the caller
3116  *
3117  * The function tries to search if there is an existing PPAT entry which
3118  * matches with the required value. If perfectly matched, the existing PPAT
3119  * entry will be used. If only partially matched, it will try to check if
3120  * there is any available PPAT index. If yes, it will allocate a new PPAT
3121  * index for the required entry and update the HW. If not, the partially
3122  * matched entry will be used.
3123  */
3124 const struct intel_ppat_entry *
3125 intel_ppat_get(struct drm_i915_private *i915, u8 value)
3126 {
3127         struct intel_ppat *ppat = &i915->ppat;
3128         struct intel_ppat_entry *entry = NULL;
3129         unsigned int scanned, best_score;
3130         int i;
3131
3132         GEM_BUG_ON(!ppat->max_entries);
3133
3134         scanned = best_score = 0;
3135         for_each_set_bit(i, ppat->used, ppat->max_entries) {
3136                 unsigned int score;
3137
3138                 score = ppat->match(ppat->entries[i].value, value);
3139                 if (score > best_score) {
3140                         entry = &ppat->entries[i];
3141                         if (score == INTEL_PPAT_PERFECT_MATCH) {
3142                                 kref_get(&entry->ref);
3143                                 return entry;
3144                         }
3145                         best_score = score;
3146                 }
3147                 scanned++;
3148         }
3149
3150         if (scanned == ppat->max_entries) {
3151                 if (!entry)
3152                         return ERR_PTR(-ENOSPC);
3153
3154                 kref_get(&entry->ref);
3155                 return entry;
3156         }
3157
3158         i = find_first_zero_bit(ppat->used, ppat->max_entries);
3159         entry = __alloc_ppat_entry(ppat, i, value);
3160         ppat->update_hw(i915);
3161         return entry;
3162 }
3163
3164 static void release_ppat(struct kref *kref)
3165 {
3166         struct intel_ppat_entry *entry =
3167                 container_of(kref, struct intel_ppat_entry, ref);
3168         struct drm_i915_private *i915 = entry->ppat->i915;
3169
3170         __free_ppat_entry(entry);
3171         entry->ppat->update_hw(i915);
3172 }
3173
3174 /**
3175  * intel_ppat_put - put back the PPAT entry got from intel_ppat_get()
3176  * @entry: an intel PPAT entry
3177  *
3178  * Put back the PPAT entry got from intel_ppat_get(). If the PPAT index of the
3179  * entry is dynamically allocated, its reference count will be decreased. Once
3180  * the reference count becomes into zero, the PPAT index becomes free again.
3181  */
3182 void intel_ppat_put(const struct intel_ppat_entry *entry)
3183 {
3184         struct intel_ppat *ppat = entry->ppat;
3185         unsigned int index = entry - ppat->entries;
3186
3187         GEM_BUG_ON(!ppat->max_entries);
3188
3189         kref_put(&ppat->entries[index].ref, release_ppat);
3190 }
3191
3192 static void cnl_private_pat_update_hw(struct drm_i915_private *dev_priv)
3193 {
3194         struct intel_ppat *ppat = &dev_priv->ppat;
3195         int i;
3196
3197         for_each_set_bit(i, ppat->dirty, ppat->max_entries) {
3198                 I915_WRITE(GEN10_PAT_INDEX(i), ppat->entries[i].value);
3199                 clear_bit(i, ppat->dirty);
3200         }
3201 }
3202
3203 static void bdw_private_pat_update_hw(struct drm_i915_private *dev_priv)
3204 {
3205         struct intel_ppat *ppat = &dev_priv->ppat;
3206         u64 pat = 0;
3207         int i;
3208
3209         for (i = 0; i < ppat->max_entries; i++)
3210                 pat |= GEN8_PPAT(i, ppat->entries[i].value);
3211
3212         bitmap_clear(ppat->dirty, 0, ppat->max_entries);
3213
3214         I915_WRITE(GEN8_PRIVATE_PAT_LO, lower_32_bits(pat));
3215         I915_WRITE(GEN8_PRIVATE_PAT_HI, upper_32_bits(pat));
3216 }
3217
3218 static unsigned int bdw_private_pat_match(u8 src, u8 dst)
3219 {
3220         unsigned int score = 0;
3221         enum {
3222                 AGE_MATCH = BIT(0),
3223                 TC_MATCH = BIT(1),
3224                 CA_MATCH = BIT(2),
3225         };
3226
3227         /* Cache attribute has to be matched. */
3228         if (GEN8_PPAT_GET_CA(src) != GEN8_PPAT_GET_CA(dst))
3229                 return 0;
3230
3231         score |= CA_MATCH;
3232
3233         if (GEN8_PPAT_GET_TC(src) == GEN8_PPAT_GET_TC(dst))
3234                 score |= TC_MATCH;
3235
3236         if (GEN8_PPAT_GET_AGE(src) == GEN8_PPAT_GET_AGE(dst))
3237                 score |= AGE_MATCH;
3238
3239         if (score == (AGE_MATCH | TC_MATCH | CA_MATCH))
3240                 return INTEL_PPAT_PERFECT_MATCH;
3241
3242         return score;
3243 }
3244
3245 static unsigned int chv_private_pat_match(u8 src, u8 dst)
3246 {
3247         return (CHV_PPAT_GET_SNOOP(src) == CHV_PPAT_GET_SNOOP(dst)) ?
3248                 INTEL_PPAT_PERFECT_MATCH : 0;
3249 }
3250
3251 static void cnl_setup_private_ppat(struct intel_ppat *ppat)
3252 {
3253         ppat->max_entries = 8;
3254         ppat->update_hw = cnl_private_pat_update_hw;
3255         ppat->match = bdw_private_pat_match;
3256         ppat->clear_value = GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3);
3257
3258         __alloc_ppat_entry(ppat, 0, GEN8_PPAT_WB | GEN8_PPAT_LLC);
3259         __alloc_ppat_entry(ppat, 1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC);
3260         __alloc_ppat_entry(ppat, 2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC);
3261         __alloc_ppat_entry(ppat, 3, GEN8_PPAT_UC);
3262         __alloc_ppat_entry(ppat, 4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0));
3263         __alloc_ppat_entry(ppat, 5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1));
3264         __alloc_ppat_entry(ppat, 6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2));
3265         __alloc_ppat_entry(ppat, 7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
3266 }
3267
3268 /* The GGTT and PPGTT need a private PPAT setup in order to handle cacheability
3269  * bits. When using advanced contexts each context stores its own PAT, but
3270  * writing this data shouldn't be harmful even in those cases. */
3271 static void bdw_setup_private_ppat(struct intel_ppat *ppat)
3272 {
3273         ppat->max_entries = 8;
3274         ppat->update_hw = bdw_private_pat_update_hw;
3275         ppat->match = bdw_private_pat_match;
3276         ppat->clear_value = GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3);
3277
3278         if (!USES_PPGTT(ppat->i915)) {
3279                 /* Spec: "For GGTT, there is NO pat_sel[2:0] from the entry,
3280                  * so RTL will always use the value corresponding to
3281                  * pat_sel = 000".
3282                  * So let's disable cache for GGTT to avoid screen corruptions.
3283                  * MOCS still can be used though.
3284                  * - System agent ggtt writes (i.e. cpu gtt mmaps) already work
3285                  * before this patch, i.e. the same uncached + snooping access
3286                  * like on gen6/7 seems to be in effect.
3287                  * - So this just fixes blitter/render access. Again it looks
3288                  * like it's not just uncached access, but uncached + snooping.
3289                  * So we can still hold onto all our assumptions wrt cpu
3290                  * clflushing on LLC machines.
3291                  */
3292                 __alloc_ppat_entry(ppat, 0, GEN8_PPAT_UC);
3293                 return;
3294         }
3295
3296         __alloc_ppat_entry(ppat, 0, GEN8_PPAT_WB | GEN8_PPAT_LLC);      /* for normal objects, no eLLC */
3297         __alloc_ppat_entry(ppat, 1, GEN8_PPAT_WC | GEN8_PPAT_LLCELLC);  /* for something pointing to ptes? */
3298         __alloc_ppat_entry(ppat, 2, GEN8_PPAT_WT | GEN8_PPAT_LLCELLC);  /* for scanout with eLLC */
3299         __alloc_ppat_entry(ppat, 3, GEN8_PPAT_UC);                      /* Uncached objects, mostly for scanout */
3300         __alloc_ppat_entry(ppat, 4, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(0));
3301         __alloc_ppat_entry(ppat, 5, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(1));
3302         __alloc_ppat_entry(ppat, 6, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(2));
3303         __alloc_ppat_entry(ppat, 7, GEN8_PPAT_WB | GEN8_PPAT_LLCELLC | GEN8_PPAT_AGE(3));
3304 }
3305
3306 static void chv_setup_private_ppat(struct intel_ppat *ppat)
3307 {
3308         ppat->max_entries = 8;
3309         ppat->update_hw = bdw_private_pat_update_hw;
3310         ppat->match = chv_private_pat_match;
3311         ppat->clear_value = CHV_PPAT_SNOOP;
3312
3313         /*
3314          * Map WB on BDW to snooped on CHV.
3315          *
3316          * Only the snoop bit has meaning for CHV, the rest is
3317          * ignored.
3318          *
3319          * The hardware will never snoop for certain types of accesses:
3320          * - CPU GTT (GMADR->GGTT->no snoop->memory)
3321          * - PPGTT page tables
3322          * - some other special cycles
3323          *
3324          * As with BDW, we also need to consider the following for GT accesses:
3325          * "For GGTT, there is NO pat_sel[2:0] from the entry,
3326          * so RTL will always use the value corresponding to
3327          * pat_sel = 000".
3328          * Which means we must set the snoop bit in PAT entry 0
3329          * in order to keep the global status page working.
3330          */
3331
3332         __alloc_ppat_entry(ppat, 0, CHV_PPAT_SNOOP);
3333         __alloc_ppat_entry(ppat, 1, 0);
3334         __alloc_ppat_entry(ppat, 2, 0);
3335         __alloc_ppat_entry(ppat, 3, 0);
3336         __alloc_ppat_entry(ppat, 4, CHV_PPAT_SNOOP);
3337         __alloc_ppat_entry(ppat, 5, CHV_PPAT_SNOOP);
3338         __alloc_ppat_entry(ppat, 6, CHV_PPAT_SNOOP);
3339         __alloc_ppat_entry(ppat, 7, CHV_PPAT_SNOOP);
3340 }
3341
3342 static void gen6_gmch_remove(struct i915_address_space *vm)
3343 {
3344         struct i915_ggtt *ggtt = i915_vm_to_ggtt(vm);
3345
3346         iounmap(ggtt->gsm);
3347         cleanup_scratch_page(vm);
3348 }
3349
3350 static void setup_private_pat(struct drm_i915_private *dev_priv)
3351 {
3352         struct intel_ppat *ppat = &dev_priv->ppat;
3353         int i;
3354
3355         ppat->i915 = dev_priv;
3356
3357         if (INTEL_GEN(dev_priv) >= 10)
3358                 cnl_setup_private_ppat(ppat);
3359         else if (IS_CHERRYVIEW(dev_priv) || IS_GEN9_LP(dev_priv))
3360                 chv_setup_private_ppat(ppat);
3361         else
3362                 bdw_setup_private_ppat(ppat);
3363
3364         GEM_BUG_ON(ppat->max_entries > INTEL_MAX_PPAT_ENTRIES);
3365
3366         for_each_clear_bit(i, ppat->used, ppat->max_entries) {
3367                 ppat->entries[i].value = ppat->clear_value;
3368                 ppat->entries[i].ppat = ppat;
3369                 set_bit(i, ppat->dirty);
3370         }
3371
3372         ppat->update_hw(dev_priv);
3373 }
3374
3375 static int gen8_gmch_probe(struct i915_ggtt *ggtt)
3376 {
3377         struct drm_i915_private *dev_priv = ggtt->vm.i915;
3378         struct pci_dev *pdev = dev_priv->drm.pdev;
3379         unsigned int size;
3380         u16 snb_gmch_ctl;
3381         int err;
3382
3383         /* TODO: We're not aware of mappable constraints on gen8 yet */
3384         ggtt->gmadr =
3385                 (struct resource) DEFINE_RES_MEM(pci_resource_start(pdev, 2),
3386                                                  pci_resource_len(pdev, 2));
3387         ggtt->mappable_end = resource_size(&ggtt->gmadr);
3388
3389         err = pci_set_dma_mask(pdev, DMA_BIT_MASK(39));
3390         if (!err)
3391                 err = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(39));
3392         if (err)
3393                 DRM_ERROR("Can't set DMA mask/consistent mask (%d)\n", err);
3394
3395         pci_read_config_word(pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3396         if (IS_CHERRYVIEW(dev_priv))
3397                 size = chv_get_total_gtt_size(snb_gmch_ctl);
3398         else
3399                 size = gen8_get_total_gtt_size(snb_gmch_ctl);
3400
3401         ggtt->vm.total = (size / sizeof(gen8_pte_t)) * I915_GTT_PAGE_SIZE;
3402         ggtt->vm.cleanup = gen6_gmch_remove;
3403         ggtt->vm.insert_page = gen8_ggtt_insert_page;
3404         ggtt->vm.clear_range = nop_clear_range;
3405         if (!USES_FULL_PPGTT(dev_priv) || intel_scanout_needs_vtd_wa(dev_priv))
3406                 ggtt->vm.clear_range = gen8_ggtt_clear_range;
3407
3408         ggtt->vm.insert_entries = gen8_ggtt_insert_entries;
3409
3410         /* Serialize GTT updates with aperture access on BXT if VT-d is on. */
3411         if (intel_ggtt_update_needs_vtd_wa(dev_priv)) {
3412                 ggtt->vm.insert_entries = bxt_vtd_ggtt_insert_entries__BKL;
3413                 ggtt->vm.insert_page    = bxt_vtd_ggtt_insert_page__BKL;
3414                 if (ggtt->vm.clear_range != nop_clear_range)
3415                         ggtt->vm.clear_range = bxt_vtd_ggtt_clear_range__BKL;
3416         }
3417
3418         ggtt->invalidate = gen6_ggtt_invalidate;
3419
3420         ggtt->vm.vma_ops.bind_vma    = ggtt_bind_vma;
3421         ggtt->vm.vma_ops.unbind_vma  = ggtt_unbind_vma;
3422         ggtt->vm.vma_ops.set_pages   = ggtt_set_pages;
3423         ggtt->vm.vma_ops.clear_pages = clear_pages;
3424
3425         setup_private_pat(dev_priv);
3426
3427         return ggtt_probe_common(ggtt, size);
3428 }
3429
3430 static int gen6_gmch_probe(struct i915_ggtt *ggtt)
3431 {
3432         struct drm_i915_private *dev_priv = ggtt->vm.i915;
3433         struct pci_dev *pdev = dev_priv->drm.pdev;
3434         unsigned int size;
3435         u16 snb_gmch_ctl;
3436         int err;
3437
3438         ggtt->gmadr =
3439                 (struct resource) DEFINE_RES_MEM(pci_resource_start(pdev, 2),
3440                                                  pci_resource_len(pdev, 2));
3441         ggtt->mappable_end = resource_size(&ggtt->gmadr);
3442
3443         /* 64/512MB is the current min/max we actually know of, but this is just
3444          * a coarse sanity check.
3445          */
3446         if (ggtt->mappable_end < (64<<20) || ggtt->mappable_end > (512<<20)) {
3447                 DRM_ERROR("Unknown GMADR size (%pa)\n", &ggtt->mappable_end);
3448                 return -ENXIO;
3449         }
3450
3451         err = pci_set_dma_mask(pdev, DMA_BIT_MASK(40));
3452         if (!err)
3453                 err = pci_set_consistent_dma_mask(pdev, DMA_BIT_MASK(40));
3454         if (err)
3455                 DRM_ERROR("Can't set DMA mask/consistent mask (%d)\n", err);
3456         pci_read_config_word(pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
3457
3458         size = gen6_get_total_gtt_size(snb_gmch_ctl);
3459         ggtt->vm.total = (size / sizeof(gen6_pte_t)) * I915_GTT_PAGE_SIZE;
3460
3461         ggtt->vm.clear_range = gen6_ggtt_clear_range;
3462         ggtt->vm.insert_page = gen6_ggtt_insert_page;
3463         ggtt->vm.insert_entries = gen6_ggtt_insert_entries;
3464         ggtt->vm.cleanup = gen6_gmch_remove;
3465
3466         ggtt->invalidate = gen6_ggtt_invalidate;
3467
3468         if (HAS_EDRAM(dev_priv))
3469                 ggtt->vm.pte_encode = iris_pte_encode;
3470         else if (IS_HASWELL(dev_priv))
3471                 ggtt->vm.pte_encode = hsw_pte_encode;
3472         else if (IS_VALLEYVIEW(dev_priv))
3473                 ggtt->vm.pte_encode = byt_pte_encode;
3474         else if (INTEL_GEN(dev_priv) >= 7)
3475                 ggtt->vm.pte_encode = ivb_pte_encode;
3476         else
3477                 ggtt->vm.pte_encode = snb_pte_encode;
3478
3479         ggtt->vm.vma_ops.bind_vma    = ggtt_bind_vma;
3480         ggtt->vm.vma_ops.unbind_vma  = ggtt_unbind_vma;
3481         ggtt->vm.vma_ops.set_pages   = ggtt_set_pages;
3482         ggtt->vm.vma_ops.clear_pages = clear_pages;
3483
3484         return ggtt_probe_common(ggtt, size);
3485 }
3486
3487 static void i915_gmch_remove(struct i915_address_space *vm)
3488 {
3489         intel_gmch_remove();
3490 }
3491
3492 static int i915_gmch_probe(struct i915_ggtt *ggtt)
3493 {
3494         struct drm_i915_private *dev_priv = ggtt->vm.i915;
3495         phys_addr_t gmadr_base;
3496         int ret;
3497
3498         ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->drm.pdev, NULL);
3499         if (!ret) {
3500                 DRM_ERROR("failed to set up gmch\n");
3501                 return -EIO;
3502         }
3503
3504         intel_gtt_get(&ggtt->vm.total, &gmadr_base, &ggtt->mappable_end);
3505
3506         ggtt->gmadr =
3507                 (struct resource) DEFINE_RES_MEM(gmadr_base,
3508                                                  ggtt->mappable_end);
3509
3510         ggtt->do_idle_maps = needs_idle_maps(dev_priv);
3511         ggtt->vm.insert_page = i915_ggtt_insert_page;
3512         ggtt->vm.insert_entries = i915_ggtt_insert_entries;
3513         ggtt->vm.clear_range = i915_ggtt_clear_range;
3514         ggtt->vm.cleanup = i915_gmch_remove;
3515
3516         ggtt->invalidate = gmch_ggtt_invalidate;
3517
3518         ggtt->vm.vma_ops.bind_vma    = ggtt_bind_vma;
3519         ggtt->vm.vma_ops.unbind_vma  = ggtt_unbind_vma;
3520         ggtt->vm.vma_ops.set_pages   = ggtt_set_pages;
3521         ggtt->vm.vma_ops.clear_pages = clear_pages;
3522
3523         if (unlikely(ggtt->do_idle_maps))
3524                 DRM_INFO("applying Ironlake quirks for intel_iommu\n");
3525
3526         return 0;
3527 }
3528
3529 /**
3530  * i915_ggtt_probe_hw - Probe GGTT hardware location
3531  * @dev_priv: i915 device
3532  */
3533 int i915_ggtt_probe_hw(struct drm_i915_private *dev_priv)
3534 {
3535         struct i915_ggtt *ggtt = &dev_priv->ggtt;
3536         int ret;
3537
3538         ggtt->vm.i915 = dev_priv;
3539         ggtt->vm.dma = &dev_priv->drm.pdev->dev;
3540
3541         if (INTEL_GEN(dev_priv) <= 5)
3542                 ret = i915_gmch_probe(ggtt);
3543         else if (INTEL_GEN(dev_priv) < 8)
3544                 ret = gen6_gmch_probe(ggtt);
3545         else
3546                 ret = gen8_gmch_probe(ggtt);
3547         if (ret)
3548                 return ret;
3549
3550         /* Trim the GGTT to fit the GuC mappable upper range (when enabled).
3551          * This is easier than doing range restriction on the fly, as we
3552          * currently don't have any bits spare to pass in this upper
3553          * restriction!
3554          */
3555         if (USES_GUC(dev_priv)) {
3556                 ggtt->vm.total = min_t(u64, ggtt->vm.total, GUC_GGTT_TOP);
3557                 ggtt->mappable_end =
3558                         min_t(u64, ggtt->mappable_end, ggtt->vm.total);
3559         }
3560
3561         if ((ggtt->vm.total - 1) >> 32) {
3562                 DRM_ERROR("We never expected a Global GTT with more than 32bits"
3563                           " of address space! Found %lldM!\n",
3564                           ggtt->vm.total >> 20);
3565                 ggtt->vm.total = 1ULL << 32;
3566                 ggtt->mappable_end =
3567                         min_t(u64, ggtt->mappable_end, ggtt->vm.total);
3568         }
3569
3570         if (ggtt->mappable_end > ggtt->vm.total) {
3571                 DRM_ERROR("mappable aperture extends past end of GGTT,"
3572                           " aperture=%pa, total=%llx\n",
3573                           &ggtt->mappable_end, ggtt->vm.total);
3574                 ggtt->mappable_end = ggtt->vm.total;
3575         }
3576
3577         /* GMADR is the PCI mmio aperture into the global GTT. */
3578         DRM_DEBUG_DRIVER("GGTT size = %lluM\n", ggtt->vm.total >> 20);
3579         DRM_DEBUG_DRIVER("GMADR size = %lluM\n", (u64)ggtt->mappable_end >> 20);
3580         DRM_DEBUG_DRIVER("DSM size = %lluM\n",
3581                          (u64)resource_size(&intel_graphics_stolen_res) >> 20);
3582         if (intel_vtd_active())
3583                 DRM_INFO("VT-d active for gfx access\n");
3584
3585         return 0;
3586 }
3587
3588 /**
3589  * i915_ggtt_init_hw - Initialize GGTT hardware
3590  * @dev_priv: i915 device
3591  */
3592 int i915_ggtt_init_hw(struct drm_i915_private *dev_priv)
3593 {
3594         struct i915_ggtt *ggtt = &dev_priv->ggtt;
3595         int ret;
3596
3597         stash_init(&dev_priv->mm.wc_stash);
3598
3599         /* Note that we use page colouring to enforce a guard page at the
3600          * end of the address space. This is required as the CS may prefetch
3601          * beyond the end of the batch buffer, across the page boundary,
3602          * and beyond the end of the GTT if we do not provide a guard.
3603          */
3604         mutex_lock(&dev_priv->drm.struct_mutex);
3605         i915_address_space_init(&ggtt->vm, dev_priv);
3606
3607         ggtt->vm.is_ggtt = true;
3608
3609         /* Only VLV supports read-only GGTT mappings */
3610         ggtt->vm.has_read_only = IS_VALLEYVIEW(dev_priv);
3611
3612         if (!HAS_LLC(dev_priv) && !USES_PPGTT(dev_priv))
3613                 ggtt->vm.mm.color_adjust = i915_gtt_color_adjust;
3614         mutex_unlock(&dev_priv->drm.struct_mutex);
3615
3616         if (!io_mapping_init_wc(&dev_priv->ggtt.iomap,
3617                                 dev_priv->ggtt.gmadr.start,
3618                                 dev_priv->ggtt.mappable_end)) {
3619                 ret = -EIO;
3620                 goto out_gtt_cleanup;
3621         }
3622
3623         ggtt->mtrr = arch_phys_wc_add(ggtt->gmadr.start, ggtt->mappable_end);
3624
3625         /*
3626          * Initialise stolen early so that we may reserve preallocated
3627          * objects for the BIOS to KMS transition.
3628          */
3629         ret = i915_gem_init_stolen(dev_priv);
3630         if (ret)
3631                 goto out_gtt_cleanup;
3632
3633         return 0;
3634
3635 out_gtt_cleanup:
3636         ggtt->vm.cleanup(&ggtt->vm);
3637         return ret;
3638 }
3639
3640 int i915_ggtt_enable_hw(struct drm_i915_private *dev_priv)
3641 {
3642         if (INTEL_GEN(dev_priv) < 6 && !intel_enable_gtt())
3643                 return -EIO;
3644
3645         return 0;
3646 }
3647
3648 void i915_ggtt_enable_guc(struct drm_i915_private *i915)
3649 {
3650         GEM_BUG_ON(i915->ggtt.invalidate != gen6_ggtt_invalidate);
3651
3652         i915->ggtt.invalidate = guc_ggtt_invalidate;
3653
3654         i915_ggtt_invalidate(i915);
3655 }
3656
3657 void i915_ggtt_disable_guc(struct drm_i915_private *i915)
3658 {
3659         /* XXX Temporary pardon for error unload */
3660         if (i915->ggtt.invalidate == gen6_ggtt_invalidate)
3661                 return;
3662
3663         /* We should only be called after i915_ggtt_enable_guc() */
3664         GEM_BUG_ON(i915->ggtt.invalidate != guc_ggtt_invalidate);
3665
3666         i915->ggtt.invalidate = gen6_ggtt_invalidate;
3667
3668         i915_ggtt_invalidate(i915);
3669 }
3670
3671 void i915_gem_restore_gtt_mappings(struct drm_i915_private *dev_priv)
3672 {
3673         struct i915_ggtt *ggtt = &dev_priv->ggtt;
3674         struct i915_vma *vma, *vn;
3675
3676         i915_check_and_clear_faults(dev_priv);
3677
3678         /* First fill our portion of the GTT with scratch pages */
3679         ggtt->vm.clear_range(&ggtt->vm, 0, ggtt->vm.total);
3680
3681         ggtt->vm.closed = true; /* skip rewriting PTE on VMA unbind */
3682
3683         /* clflush objects bound into the GGTT and rebind them. */
3684         GEM_BUG_ON(!list_empty(&ggtt->vm.active_list));
3685         list_for_each_entry_safe(vma, vn, &ggtt->vm.inactive_list, vm_link) {
3686                 struct drm_i915_gem_object *obj = vma->obj;
3687
3688                 if (!(vma->flags & I915_VMA_GLOBAL_BIND))
3689                         continue;
3690
3691                 if (!i915_vma_unbind(vma))
3692                         continue;
3693
3694                 WARN_ON(i915_vma_bind(vma,
3695                                       obj ? obj->cache_level : 0,
3696                                       PIN_UPDATE));
3697                 if (obj)
3698                         WARN_ON(i915_gem_object_set_to_gtt_domain(obj, false));
3699         }
3700
3701         ggtt->vm.closed = false;
3702         i915_ggtt_invalidate(dev_priv);
3703
3704         if (INTEL_GEN(dev_priv) >= 8) {
3705                 struct intel_ppat *ppat = &dev_priv->ppat;
3706
3707                 bitmap_set(ppat->dirty, 0, ppat->max_entries);
3708                 dev_priv->ppat.update_hw(dev_priv);
3709                 return;
3710         }
3711 }
3712
3713 static struct scatterlist *
3714 rotate_pages(const dma_addr_t *in, unsigned int offset,
3715              unsigned int width, unsigned int height,
3716              unsigned int stride,
3717              struct sg_table *st, struct scatterlist *sg)
3718 {
3719         unsigned int column, row;
3720         unsigned int src_idx;
3721
3722         for (column = 0; column < width; column++) {
3723                 src_idx = stride * (height - 1) + column;
3724                 for (row = 0; row < height; row++) {
3725                         st->nents++;
3726                         /* We don't need the pages, but need to initialize
3727                          * the entries so the sg list can be happily traversed.
3728                          * The only thing we need are DMA addresses.
3729                          */
3730                         sg_set_page(sg, NULL, I915_GTT_PAGE_SIZE, 0);
3731                         sg_dma_address(sg) = in[offset + src_idx];
3732                         sg_dma_len(sg) = I915_GTT_PAGE_SIZE;
3733                         sg = sg_next(sg);
3734                         src_idx -= stride;
3735                 }
3736         }
3737
3738         return sg;
3739 }
3740
3741 static noinline struct sg_table *
3742 intel_rotate_pages(struct intel_rotation_info *rot_info,
3743                    struct drm_i915_gem_object *obj)
3744 {
3745         const unsigned long n_pages = obj->base.size / I915_GTT_PAGE_SIZE;
3746         unsigned int size = intel_rotation_info_size(rot_info);
3747         struct sgt_iter sgt_iter;
3748         dma_addr_t dma_addr;
3749         unsigned long i;
3750         dma_addr_t *page_addr_list;
3751         struct sg_table *st;
3752         struct scatterlist *sg;
3753         int ret = -ENOMEM;
3754
3755         /* Allocate a temporary list of source pages for random access. */
3756         page_addr_list = kvmalloc_array(n_pages,
3757                                         sizeof(dma_addr_t),
3758                                         GFP_KERNEL);
3759         if (!page_addr_list)
3760                 return ERR_PTR(ret);
3761
3762         /* Allocate target SG list. */
3763         st = kmalloc(sizeof(*st), GFP_KERNEL);
3764         if (!st)
3765                 goto err_st_alloc;
3766
3767         ret = sg_alloc_table(st, size, GFP_KERNEL);
3768         if (ret)
3769                 goto err_sg_alloc;
3770
3771         /* Populate source page list from the object. */
3772         i = 0;
3773         for_each_sgt_dma(dma_addr, sgt_iter, obj->mm.pages)
3774                 page_addr_list[i++] = dma_addr;
3775
3776         GEM_BUG_ON(i != n_pages);
3777         st->nents = 0;
3778         sg = st->sgl;
3779
3780         for (i = 0 ; i < ARRAY_SIZE(rot_info->plane); i++) {
3781                 sg = rotate_pages(page_addr_list, rot_info->plane[i].offset,
3782                                   rot_info->plane[i].width, rot_info->plane[i].height,
3783                                   rot_info->plane[i].stride, st, sg);
3784         }
3785
3786         kvfree(page_addr_list);
3787
3788         return st;
3789
3790 err_sg_alloc:
3791         kfree(st);
3792 err_st_alloc:
3793         kvfree(page_addr_list);
3794
3795         DRM_DEBUG_DRIVER("Failed to create rotated mapping for object size %zu! (%ux%u tiles, %u pages)\n",
3796                          obj->base.size, rot_info->plane[0].width, rot_info->plane[0].height, size);
3797
3798         return ERR_PTR(ret);
3799 }
3800
3801 static noinline struct sg_table *
3802 intel_partial_pages(const struct i915_ggtt_view *view,
3803                     struct drm_i915_gem_object *obj)
3804 {
3805         struct sg_table *st;
3806         struct scatterlist *sg, *iter;
3807         unsigned int count = view->partial.size;
3808         unsigned int offset;
3809         int ret = -ENOMEM;
3810
3811         st = kmalloc(sizeof(*st), GFP_KERNEL);
3812         if (!st)
3813                 goto err_st_alloc;
3814
3815         ret = sg_alloc_table(st, count, GFP_KERNEL);
3816         if (ret)
3817                 goto err_sg_alloc;
3818
3819         iter = i915_gem_object_get_sg(obj, view->partial.offset, &offset);
3820         GEM_BUG_ON(!iter);
3821
3822         sg = st->sgl;
3823         st->nents = 0;
3824         do {
3825                 unsigned int len;
3826
3827                 len = min(iter->length - (offset << PAGE_SHIFT),
3828                           count << PAGE_SHIFT);
3829                 sg_set_page(sg, NULL, len, 0);
3830                 sg_dma_address(sg) =
3831                         sg_dma_address(iter) + (offset << PAGE_SHIFT);
3832                 sg_dma_len(sg) = len;
3833
3834                 st->nents++;
3835                 count -= len >> PAGE_SHIFT;
3836                 if (count == 0) {
3837                         sg_mark_end(sg);
3838                         return st;
3839                 }
3840
3841                 sg = __sg_next(sg);
3842                 iter = __sg_next(iter);
3843                 offset = 0;
3844         } while (1);
3845
3846 err_sg_alloc:
3847         kfree(st);
3848 err_st_alloc:
3849         return ERR_PTR(ret);
3850 }
3851
3852 static int
3853 i915_get_ggtt_vma_pages(struct i915_vma *vma)
3854 {
3855         int ret;
3856
3857         /* The vma->pages are only valid within the lifespan of the borrowed
3858          * obj->mm.pages. When the obj->mm.pages sg_table is regenerated, so
3859          * must be the vma->pages. A simple rule is that vma->pages must only
3860          * be accessed when the obj->mm.pages are pinned.
3861          */
3862         GEM_BUG_ON(!i915_gem_object_has_pinned_pages(vma->obj));
3863
3864         switch (vma->ggtt_view.type) {
3865         default:
3866                 GEM_BUG_ON(vma->ggtt_view.type);
3867                 /* fall through */
3868         case I915_GGTT_VIEW_NORMAL:
3869                 vma->pages = vma->obj->mm.pages;
3870                 return 0;
3871
3872         case I915_GGTT_VIEW_ROTATED:
3873                 vma->pages =
3874                         intel_rotate_pages(&vma->ggtt_view.rotated, vma->obj);
3875                 break;
3876
3877         case I915_GGTT_VIEW_PARTIAL:
3878                 vma->pages = intel_partial_pages(&vma->ggtt_view, vma->obj);
3879                 break;
3880         }
3881
3882         ret = 0;
3883         if (unlikely(IS_ERR(vma->pages))) {
3884                 ret = PTR_ERR(vma->pages);
3885                 vma->pages = NULL;
3886                 DRM_ERROR("Failed to get pages for VMA view type %u (%d)!\n",
3887                           vma->ggtt_view.type, ret);
3888         }
3889         return ret;
3890 }
3891
3892 /**
3893  * i915_gem_gtt_reserve - reserve a node in an address_space (GTT)
3894  * @vm: the &struct i915_address_space
3895  * @node: the &struct drm_mm_node (typically i915_vma.mode)
3896  * @size: how much space to allocate inside the GTT,
3897  *        must be #I915_GTT_PAGE_SIZE aligned
3898  * @offset: where to insert inside the GTT,
3899  *          must be #I915_GTT_MIN_ALIGNMENT aligned, and the node
3900  *          (@offset + @size) must fit within the address space
3901  * @color: color to apply to node, if this node is not from a VMA,
3902  *         color must be #I915_COLOR_UNEVICTABLE
3903  * @flags: control search and eviction behaviour
3904  *
3905  * i915_gem_gtt_reserve() tries to insert the @node at the exact @offset inside
3906  * the address space (using @size and @color). If the @node does not fit, it
3907  * tries to evict any overlapping nodes from the GTT, including any
3908  * neighbouring nodes if the colors do not match (to ensure guard pages between
3909  * differing domains). See i915_gem_evict_for_node() for the gory details
3910  * on the eviction algorithm. #PIN_NONBLOCK may used to prevent waiting on
3911  * evicting active overlapping objects, and any overlapping node that is pinned
3912  * or marked as unevictable will also result in failure.
3913  *
3914  * Returns: 0 on success, -ENOSPC if no suitable hole is found, -EINTR if
3915  * asked to wait for eviction and interrupted.
3916  */
3917 int i915_gem_gtt_reserve(struct i915_address_space *vm,
3918                          struct drm_mm_node *node,
3919                          u64 size, u64 offset, unsigned long color,
3920                          unsigned int flags)
3921 {
3922         int err;
3923
3924         GEM_BUG_ON(!size);
3925         GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
3926         GEM_BUG_ON(!IS_ALIGNED(offset, I915_GTT_MIN_ALIGNMENT));
3927         GEM_BUG_ON(range_overflows(offset, size, vm->total));
3928         GEM_BUG_ON(vm == &vm->i915->mm.aliasing_ppgtt->vm);
3929         GEM_BUG_ON(drm_mm_node_allocated(node));
3930
3931         node->size = size;
3932         node->start = offset;
3933         node->color = color;
3934
3935         err = drm_mm_reserve_node(&vm->mm, node);
3936         if (err != -ENOSPC)
3937                 return err;
3938
3939         if (flags & PIN_NOEVICT)
3940                 return -ENOSPC;
3941
3942         err = i915_gem_evict_for_node(vm, node, flags);
3943         if (err == 0)
3944                 err = drm_mm_reserve_node(&vm->mm, node);
3945
3946         return err;
3947 }
3948
3949 static u64 random_offset(u64 start, u64 end, u64 len, u64 align)
3950 {
3951         u64 range, addr;
3952
3953         GEM_BUG_ON(range_overflows(start, len, end));
3954         GEM_BUG_ON(round_up(start, align) > round_down(end - len, align));
3955
3956         range = round_down(end - len, align) - round_up(start, align);
3957         if (range) {
3958                 if (sizeof(unsigned long) == sizeof(u64)) {
3959                         addr = get_random_long();
3960                 } else {
3961                         addr = get_random_int();
3962                         if (range > U32_MAX) {
3963                                 addr <<= 32;
3964                                 addr |= get_random_int();
3965                         }
3966                 }
3967                 div64_u64_rem(addr, range, &addr);
3968                 start += addr;
3969         }
3970
3971         return round_up(start, align);
3972 }
3973
3974 /**
3975  * i915_gem_gtt_insert - insert a node into an address_space (GTT)
3976  * @vm: the &struct i915_address_space
3977  * @node: the &struct drm_mm_node (typically i915_vma.node)
3978  * @size: how much space to allocate inside the GTT,
3979  *        must be #I915_GTT_PAGE_SIZE aligned
3980  * @alignment: required alignment of starting offset, may be 0 but
3981  *             if specified, this must be a power-of-two and at least
3982  *             #I915_GTT_MIN_ALIGNMENT
3983  * @color: color to apply to node
3984  * @start: start of any range restriction inside GTT (0 for all),
3985  *         must be #I915_GTT_PAGE_SIZE aligned
3986  * @end: end of any range restriction inside GTT (U64_MAX for all),
3987  *       must be #I915_GTT_PAGE_SIZE aligned if not U64_MAX
3988  * @flags: control search and eviction behaviour
3989  *
3990  * i915_gem_gtt_insert() first searches for an available hole into which
3991  * is can insert the node. The hole address is aligned to @alignment and
3992  * its @size must then fit entirely within the [@start, @end] bounds. The
3993  * nodes on either side of the hole must match @color, or else a guard page
3994  * will be inserted between the two nodes (or the node evicted). If no
3995  * suitable hole is found, first a victim is randomly selected and tested
3996  * for eviction, otherwise then the LRU list of objects within the GTT
3997  * is scanned to find the first set of replacement nodes to create the hole.
3998  * Those old overlapping nodes are evicted from the GTT (and so must be
3999  * rebound before any future use). Any node that is currently pinned cannot
4000  * be evicted (see i915_vma_pin()). Similar if the node's VMA is currently
4001  * active and #PIN_NONBLOCK is specified, that node is also skipped when
4002  * searching for an eviction candidate. See i915_gem_evict_something() for
4003  * the gory details on the eviction algorithm.
4004  *
4005  * Returns: 0 on success, -ENOSPC if no suitable hole is found, -EINTR if
4006  * asked to wait for eviction and interrupted.
4007  */
4008 int i915_gem_gtt_insert(struct i915_address_space *vm,
4009                         struct drm_mm_node *node,
4010                         u64 size, u64 alignment, unsigned long color,
4011                         u64 start, u64 end, unsigned int flags)
4012 {
4013         enum drm_mm_insert_mode mode;
4014         u64 offset;
4015         int err;
4016
4017         lockdep_assert_held(&vm->i915->drm.struct_mutex);
4018         GEM_BUG_ON(!size);
4019         GEM_BUG_ON(!IS_ALIGNED(size, I915_GTT_PAGE_SIZE));
4020         GEM_BUG_ON(alignment && !is_power_of_2(alignment));
4021         GEM_BUG_ON(alignment && !IS_ALIGNED(alignment, I915_GTT_MIN_ALIGNMENT));
4022         GEM_BUG_ON(start >= end);
4023         GEM_BUG_ON(start > 0  && !IS_ALIGNED(start, I915_GTT_PAGE_SIZE));
4024         GEM_BUG_ON(end < U64_MAX && !IS_ALIGNED(end, I915_GTT_PAGE_SIZE));
4025         GEM_BUG_ON(vm == &vm->i915->mm.aliasing_ppgtt->vm);
4026         GEM_BUG_ON(drm_mm_node_allocated(node));
4027
4028         if (unlikely(range_overflows(start, size, end)))
4029                 return -ENOSPC;
4030
4031         if (unlikely(round_up(start, alignment) > round_down(end - size, alignment)))
4032                 return -ENOSPC;
4033
4034         mode = DRM_MM_INSERT_BEST;
4035         if (flags & PIN_HIGH)
4036                 mode = DRM_MM_INSERT_HIGHEST;
4037         if (flags & PIN_MAPPABLE)
4038                 mode = DRM_MM_INSERT_LOW;
4039
4040         /* We only allocate in PAGE_SIZE/GTT_PAGE_SIZE (4096) chunks,
4041          * so we know that we always have a minimum alignment of 4096.
4042          * The drm_mm range manager is optimised to return results
4043          * with zero alignment, so where possible use the optimal
4044          * path.
4045          */
4046         BUILD_BUG_ON(I915_GTT_MIN_ALIGNMENT > I915_GTT_PAGE_SIZE);
4047         if (alignment <= I915_GTT_MIN_ALIGNMENT)
4048                 alignment = 0;
4049
4050         err = drm_mm_insert_node_in_range(&vm->mm, node,
4051                                           size, alignment, color,
4052                                           start, end, mode);
4053         if (err != -ENOSPC)
4054                 return err;
4055
4056         if (mode & DRM_MM_INSERT_ONCE) {
4057                 err = drm_mm_insert_node_in_range(&vm->mm, node,
4058                                                   size, alignment, color,
4059                                                   start, end,
4060                                                   DRM_MM_INSERT_BEST);
4061                 if (err != -ENOSPC)
4062                         return err;
4063         }
4064
4065         if (flags & PIN_NOEVICT)
4066                 return -ENOSPC;
4067
4068         /* No free space, pick a slot at random.
4069          *
4070          * There is a pathological case here using a GTT shared between
4071          * mmap and GPU (i.e. ggtt/aliasing_ppgtt but not full-ppgtt):
4072          *
4073          *    |<-- 256 MiB aperture -->||<-- 1792 MiB unmappable -->|
4074          *         (64k objects)             (448k objects)
4075          *
4076          * Now imagine that the eviction LRU is ordered top-down (just because
4077          * pathology meets real life), and that we need to evict an object to
4078          * make room inside the aperture. The eviction scan then has to walk
4079          * the 448k list before it finds one within range. And now imagine that
4080          * it has to search for a new hole between every byte inside the memcpy,
4081          * for several simultaneous clients.
4082          *
4083          * On a full-ppgtt system, if we have run out of available space, there
4084          * will be lots and lots of objects in the eviction list! Again,
4085          * searching that LRU list may be slow if we are also applying any
4086          * range restrictions (e.g. restriction to low 4GiB) and so, for
4087          * simplicity and similarilty between different GTT, try the single
4088          * random replacement first.
4089          */
4090         offset = random_offset(start, end,
4091                                size, alignment ?: I915_GTT_MIN_ALIGNMENT);
4092         err = i915_gem_gtt_reserve(vm, node, size, offset, color, flags);
4093         if (err != -ENOSPC)
4094                 return err;
4095
4096         /* Randomly selected placement is pinned, do a search */
4097         err = i915_gem_evict_something(vm, size, alignment, color,
4098                                        start, end, flags);
4099         if (err)
4100                 return err;
4101
4102         return drm_mm_insert_node_in_range(&vm->mm, node,
4103                                            size, alignment, color,
4104                                            start, end, DRM_MM_INSERT_EVICT);
4105 }
4106
4107 #if IS_ENABLED(CONFIG_DRM_I915_SELFTEST)
4108 #include "selftests/mock_gtt.c"
4109 #include "selftests/i915_gem_gtt.c"
4110 #endif