Pileus Git - ~andy/linux/blob - drivers/gpu/drm/i915/i915_gem_gtt.c

   1 /*
   2  * Copyright © 2010 Daniel Vetter
   3  *
   4  * Permission is hereby granted, free of charge, to any person obtaining a
   5  * copy of this software and associated documentation files (the "Software"),
   6  * to deal in the Software without restriction, including without limitation
   7  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   8  * and/or sell copies of the Software, and to permit persons to whom the
   9  * Software is furnished to do so, subject to the following conditions:
  10  *
  11  * The above copyright notice and this permission notice (including the next
  12  * paragraph) shall be included in all copies or substantial portions of the
  13  * Software.
  14  *
  15  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  16  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  17  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  18  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  19  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
  20  * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS
  21  * IN THE SOFTWARE.
  22  *
  23  */
  24
  25 #include <drm/drmP.h>
  26 #include <drm/i915_drm.h>
  27 #include "i915_drv.h"
  28 #include "i915_trace.h"
  29 #include "intel_drv.h"
  30
  31 #define GEN6_PPGTT_PD_ENTRIES 512
  32 #define I915_PPGTT_PT_ENTRIES (PAGE_SIZE / sizeof(gen6_gtt_pte_t))
  33
  34 /* PPGTT stuff */
  35 #define GEN6_GTT_ADDR_ENCODE(addr)      ((addr) | (((addr) >> 28) & 0xff0))
  36 #define HSW_GTT_ADDR_ENCODE(addr)       ((addr) | (((addr) >> 28) & 0x7f0))
  37
  38 #define GEN6_PDE_VALID                  (1 << 0)
  39 /* gen6+ has bit 11-4 for physical addr bit 39-32 */
  40 #define GEN6_PDE_ADDR_ENCODE(addr)      GEN6_GTT_ADDR_ENCODE(addr)
  41
  42 #define GEN6_PTE_VALID                  (1 << 0)
  43 #define GEN6_PTE_UNCACHED               (1 << 1)
  44 #define HSW_PTE_UNCACHED                (0)
  45 #define GEN6_PTE_CACHE_LLC              (2 << 1)
  46 #define GEN7_PTE_CACHE_L3_LLC           (3 << 1)
  47 #define GEN6_PTE_ADDR_ENCODE(addr)      GEN6_GTT_ADDR_ENCODE(addr)
  48 #define HSW_PTE_ADDR_ENCODE(addr)       HSW_GTT_ADDR_ENCODE(addr)
  49
  50 /* Cacheability Control is a 4-bit value. The low three bits are stored in *
  51  * bits 3:1 of the PTE, while the fourth bit is stored in bit 11 of the PTE.
  52  */
  53 #define HSW_CACHEABILITY_CONTROL(bits)  ((((bits) & 0x7) << 1) | \
  54                                          (((bits) & 0x8) << (11 - 3)))
  55 #define HSW_WB_LLC_AGE3                 HSW_CACHEABILITY_CONTROL(0x2)
  56 #define HSW_WB_LLC_AGE0                 HSW_CACHEABILITY_CONTROL(0x3)
  57 #define HSW_WB_ELLC_LLC_AGE0            HSW_CACHEABILITY_CONTROL(0xb)
  58 #define HSW_WT_ELLC_LLC_AGE0            HSW_CACHEABILITY_CONTROL(0x6)
  59
  60 static gen6_gtt_pte_t snb_pte_encode(dma_addr_t addr,
  61                                      enum i915_cache_level level)
  62 {
  63         gen6_gtt_pte_t pte = GEN6_PTE_VALID;
  64         pte |= GEN6_PTE_ADDR_ENCODE(addr);
  65
  66         switch (level) {
  67         case I915_CACHE_L3_LLC:
  68         case I915_CACHE_LLC:
  69                 pte |= GEN6_PTE_CACHE_LLC;
  70                 break;
  71         case I915_CACHE_NONE:
  72                 pte |= GEN6_PTE_UNCACHED;
  73                 break;
  74         default:
  75                 WARN_ON(1);
  76         }
  77
  78         return pte;
  79 }
  80
  81 static gen6_gtt_pte_t ivb_pte_encode(dma_addr_t addr,
  82                                      enum i915_cache_level level)
  83 {
  84         gen6_gtt_pte_t pte = GEN6_PTE_VALID;
  85         pte |= GEN6_PTE_ADDR_ENCODE(addr);
  86
  87         switch (level) {
  88         case I915_CACHE_L3_LLC:
  89                 pte |= GEN7_PTE_CACHE_L3_LLC;
  90                 break;
  91         case I915_CACHE_LLC:
  92                 pte |= GEN6_PTE_CACHE_LLC;
  93                 break;
  94         case I915_CACHE_NONE:
  95                 pte |= GEN6_PTE_UNCACHED;
  96                 break;
  97         default:
  98                 WARN_ON(1);
  99         }
 100
 101         return pte;
 102 }
 103
 104 #define BYT_PTE_WRITEABLE               (1 << 1)
 105 #define BYT_PTE_SNOOPED_BY_CPU_CACHES   (1 << 2)
 106
 107 static gen6_gtt_pte_t byt_pte_encode(dma_addr_t addr,
 108                                      enum i915_cache_level level)
 109 {
 110         gen6_gtt_pte_t pte = GEN6_PTE_VALID;
 111         pte |= GEN6_PTE_ADDR_ENCODE(addr);
 112
 113         /* Mark the page as writeable.  Other platforms don't have a
 114          * setting for read-only/writable, so this matches that behavior.
 115          */
 116         pte |= BYT_PTE_WRITEABLE;
 117
 118         if (level != I915_CACHE_NONE)
 119                 pte |= BYT_PTE_SNOOPED_BY_CPU_CACHES;
 120
 121         return pte;
 122 }
 123
 124 static gen6_gtt_pte_t hsw_pte_encode(dma_addr_t addr,
 125                                      enum i915_cache_level level)
 126 {
 127         gen6_gtt_pte_t pte = GEN6_PTE_VALID;
 128         pte |= HSW_PTE_ADDR_ENCODE(addr);
 129
 130         if (level != I915_CACHE_NONE)
 131                 pte |= HSW_WB_LLC_AGE3;
 132
 133         return pte;
 134 }
 135
 136 static gen6_gtt_pte_t iris_pte_encode(dma_addr_t addr,
 137                                       enum i915_cache_level level)
 138 {
 139         gen6_gtt_pte_t pte = GEN6_PTE_VALID;
 140         pte |= HSW_PTE_ADDR_ENCODE(addr);
 141
 142         switch (level) {
 143         case I915_CACHE_NONE:
 144                 break;
 145         case I915_CACHE_WT:
 146                 pte |= HSW_WT_ELLC_LLC_AGE0;
 147                 break;
 148         default:
 149                 pte |= HSW_WB_ELLC_LLC_AGE0;
 150                 break;
 151         }
 152
 153         return pte;
 154 }
 155
 156 static void gen6_write_pdes(struct i915_hw_ppgtt *ppgtt)
 157 {
 158         struct drm_i915_private *dev_priv = ppgtt->base.dev->dev_private;
 159         gen6_gtt_pte_t __iomem *pd_addr;
 160         uint32_t pd_entry;
 161         int i;
 162
 163         WARN_ON(ppgtt->pd_offset & 0x3f);
 164         pd_addr = (gen6_gtt_pte_t __iomem*)dev_priv->gtt.gsm +
 165                 ppgtt->pd_offset / sizeof(gen6_gtt_pte_t);
 166         for (i = 0; i < ppgtt->num_pd_entries; i++) {
 167                 dma_addr_t pt_addr;
 168
 169                 pt_addr = ppgtt->pt_dma_addr[i];
 170                 pd_entry = GEN6_PDE_ADDR_ENCODE(pt_addr);
 171                 pd_entry |= GEN6_PDE_VALID;
 172
 173                 writel(pd_entry, pd_addr + i);
 174         }
 175         readl(pd_addr);
 176 }
 177
 178 static int gen6_ppgtt_enable(struct drm_device *dev)
 179 {
 180         drm_i915_private_t *dev_priv = dev->dev_private;
 181         uint32_t pd_offset;
 182         struct intel_ring_buffer *ring;
 183         struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
 184         int i;
 185
 186         BUG_ON(ppgtt->pd_offset & 0x3f);
 187
 188         gen6_write_pdes(ppgtt);
 189
 190         pd_offset = ppgtt->pd_offset;
 191         pd_offset /= 64; /* in cachelines, */
 192         pd_offset <<= 16;
 193
 194         if (INTEL_INFO(dev)->gen == 6) {
 195                 uint32_t ecochk, gab_ctl, ecobits;
 196
 197                 ecobits = I915_READ(GAC_ECO_BITS);
 198                 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_SNB_BIT |
 199                                          ECOBITS_PPGTT_CACHE64B);
 200
 201                 gab_ctl = I915_READ(GAB_CTL);
 202                 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
 203
 204                 ecochk = I915_READ(GAM_ECOCHK);
 205                 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT |
 206                                        ECOCHK_PPGTT_CACHE64B);
 207                 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
 208         } else if (INTEL_INFO(dev)->gen >= 7) {
 209                 uint32_t ecochk, ecobits;
 210
 211                 ecobits = I915_READ(GAC_ECO_BITS);
 212                 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
 213
 214                 ecochk = I915_READ(GAM_ECOCHK);
 215                 if (IS_HASWELL(dev)) {
 216                         ecochk |= ECOCHK_PPGTT_WB_HSW;
 217                 } else {
 218                         ecochk |= ECOCHK_PPGTT_LLC_IVB;
 219                         ecochk &= ~ECOCHK_PPGTT_GFDT_IVB;
 220                 }
 221                 I915_WRITE(GAM_ECOCHK, ecochk);
 222                 /* GFX_MODE is per-ring on gen7+ */
 223         }
 224
 225         for_each_ring(ring, dev_priv, i) {
 226                 if (INTEL_INFO(dev)->gen >= 7)
 227                         I915_WRITE(RING_MODE_GEN7(ring),
 228                                    _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
 229
 230                 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
 231                 I915_WRITE(RING_PP_DIR_BASE(ring), pd_offset);
 232         }
 233         return 0;
 234 }
 235
 236 /* PPGTT support for Sandybdrige/Gen6 and later */
 237 static void gen6_ppgtt_clear_range(struct i915_address_space *vm,
 238                                    unsigned first_entry,
 239                                    unsigned num_entries)
 240 {
 241         struct i915_hw_ppgtt *ppgtt =
 242                 container_of(vm, struct i915_hw_ppgtt, base);
 243         gen6_gtt_pte_t *pt_vaddr, scratch_pte;
 244         unsigned act_pt = first_entry / I915_PPGTT_PT_ENTRIES;
 245         unsigned first_pte = first_entry % I915_PPGTT_PT_ENTRIES;
 246         unsigned last_pte, i;
 247
 248         scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC);
 249
 250         while (num_entries) {
 251                 last_pte = first_pte + num_entries;
 252                 if (last_pte > I915_PPGTT_PT_ENTRIES)
 253                         last_pte = I915_PPGTT_PT_ENTRIES;
 254
 255                 pt_vaddr = kmap_atomic(ppgtt->pt_pages[act_pt]);
 256
 257                 for (i = first_pte; i < last_pte; i++)
 258                         pt_vaddr[i] = scratch_pte;
 259
 260                 kunmap_atomic(pt_vaddr);
 261
 262                 num_entries -= last_pte - first_pte;
 263                 first_pte = 0;
 264                 act_pt++;
 265         }
 266 }
 267
 268 static void gen6_ppgtt_insert_entries(struct i915_address_space *vm,
 269                                       struct sg_table *pages,
 270                                       unsigned first_entry,
 271                                       enum i915_cache_level cache_level)
 272 {
 273         struct i915_hw_ppgtt *ppgtt =
 274                 container_of(vm, struct i915_hw_ppgtt, base);
 275         gen6_gtt_pte_t *pt_vaddr;
 276         unsigned act_pt = first_entry / I915_PPGTT_PT_ENTRIES;
 277         unsigned act_pte = first_entry % I915_PPGTT_PT_ENTRIES;
 278         struct sg_page_iter sg_iter;
 279
 280         pt_vaddr = kmap_atomic(ppgtt->pt_pages[act_pt]);
 281         for_each_sg_page(pages->sgl, &sg_iter, pages->nents, 0) {
 282                 dma_addr_t page_addr;
 283
 284                 page_addr = sg_page_iter_dma_address(&sg_iter);
 285                 pt_vaddr[act_pte] = vm->pte_encode(page_addr, cache_level);
 286                 if (++act_pte == I915_PPGTT_PT_ENTRIES) {
 287                         kunmap_atomic(pt_vaddr);
 288                         act_pt++;
 289                         pt_vaddr = kmap_atomic(ppgtt->pt_pages[act_pt]);
 290                         act_pte = 0;
 291
 292                 }
 293         }
 294         kunmap_atomic(pt_vaddr);
 295 }
 296
 297 static void gen6_ppgtt_cleanup(struct i915_address_space *vm)
 298 {
 299         struct i915_hw_ppgtt *ppgtt =
 300                 container_of(vm, struct i915_hw_ppgtt, base);
 301         int i;
 302
 303         drm_mm_takedown(&ppgtt->base.mm);
 304
 305         if (ppgtt->pt_dma_addr) {
 306                 for (i = 0; i < ppgtt->num_pd_entries; i++)
 307                         pci_unmap_page(ppgtt->base.dev->pdev,
 308                                        ppgtt->pt_dma_addr[i],
 309                                        4096, PCI_DMA_BIDIRECTIONAL);
 310         }
 311
 312         kfree(ppgtt->pt_dma_addr);
 313         for (i = 0; i < ppgtt->num_pd_entries; i++)
 314                 __free_page(ppgtt->pt_pages[i]);
 315         kfree(ppgtt->pt_pages);
 316         kfree(ppgtt);
 317 }
 318
 319 static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
 320 {
 321         struct drm_device *dev = ppgtt->base.dev;
 322         struct drm_i915_private *dev_priv = dev->dev_private;
 323         unsigned first_pd_entry_in_global_pt;
 324         int i;
 325         int ret = -ENOMEM;
 326
 327         /* ppgtt PDEs reside in the global gtt pagetable, which has 512*1024
 328          * entries. For aliasing ppgtt support we just steal them at the end for
 329          * now. */
 330         first_pd_entry_in_global_pt = gtt_total_entries(dev_priv->gtt);
 331
 332         ppgtt->base.pte_encode = dev_priv->gtt.base.pte_encode;
 333         ppgtt->num_pd_entries = GEN6_PPGTT_PD_ENTRIES;
 334         ppgtt->enable = gen6_ppgtt_enable;
 335         ppgtt->base.clear_range = gen6_ppgtt_clear_range;
 336         ppgtt->base.insert_entries = gen6_ppgtt_insert_entries;
 337         ppgtt->base.cleanup = gen6_ppgtt_cleanup;
 338         ppgtt->base.scratch = dev_priv->gtt.base.scratch;
 339         ppgtt->pt_pages = kzalloc(sizeof(struct page *)*ppgtt->num_pd_entries,
 340                                   GFP_KERNEL);
 341         if (!ppgtt->pt_pages)
 342                 return -ENOMEM;
 343
 344         for (i = 0; i < ppgtt->num_pd_entries; i++) {
 345                 ppgtt->pt_pages[i] = alloc_page(GFP_KERNEL);
 346                 if (!ppgtt->pt_pages[i])
 347                         goto err_pt_alloc;
 348         }
 349
 350         ppgtt->pt_dma_addr = kzalloc(sizeof(dma_addr_t) *ppgtt->num_pd_entries,
 351                                      GFP_KERNEL);
 352         if (!ppgtt->pt_dma_addr)
 353                 goto err_pt_alloc;
 354
 355         for (i = 0; i < ppgtt->num_pd_entries; i++) {
 356                 dma_addr_t pt_addr;
 357
 358                 pt_addr = pci_map_page(dev->pdev, ppgtt->pt_pages[i], 0, 4096,
 359                                        PCI_DMA_BIDIRECTIONAL);
 360
 361                 if (pci_dma_mapping_error(dev->pdev, pt_addr)) {
 362                         ret = -EIO;
 363                         goto err_pd_pin;
 364
 365                 }
 366                 ppgtt->pt_dma_addr[i] = pt_addr;
 367         }
 368
 369         ppgtt->base.clear_range(&ppgtt->base, 0,
 370                                 ppgtt->num_pd_entries * I915_PPGTT_PT_ENTRIES);
 371
 372         ppgtt->pd_offset = first_pd_entry_in_global_pt * sizeof(gen6_gtt_pte_t);
 373
 374         return 0;
 375
 376 err_pd_pin:
 377         if (ppgtt->pt_dma_addr) {
 378                 for (i--; i >= 0; i--)
 379                         pci_unmap_page(dev->pdev, ppgtt->pt_dma_addr[i],
 380                                        4096, PCI_DMA_BIDIRECTIONAL);
 381         }
 382 err_pt_alloc:
 383         kfree(ppgtt->pt_dma_addr);
 384         for (i = 0; i < ppgtt->num_pd_entries; i++) {
 385                 if (ppgtt->pt_pages[i])
 386                         __free_page(ppgtt->pt_pages[i]);
 387         }
 388         kfree(ppgtt->pt_pages);
 389
 390         return ret;
 391 }
 392
 393 static int i915_gem_init_aliasing_ppgtt(struct drm_device *dev)
 394 {
 395         struct drm_i915_private *dev_priv = dev->dev_private;
 396         struct i915_hw_ppgtt *ppgtt;
 397         int ret;
 398
 399         ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
 400         if (!ppgtt)
 401                 return -ENOMEM;
 402
 403         ppgtt->base.dev = dev;
 404
 405         if (INTEL_INFO(dev)->gen < 8)
 406                 ret = gen6_ppgtt_init(ppgtt);
 407         else
 408                 BUG();
 409
 410         if (ret)
 411                 kfree(ppgtt);
 412         else {
 413                 dev_priv->mm.aliasing_ppgtt = ppgtt;
 414                 drm_mm_init(&ppgtt->base.mm, ppgtt->base.start,
 415                             ppgtt->base.total);
 416         }
 417
 418         return ret;
 419 }
 420
 421 void i915_gem_cleanup_aliasing_ppgtt(struct drm_device *dev)
 422 {
 423         struct drm_i915_private *dev_priv = dev->dev_private;
 424         struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
 425
 426         if (!ppgtt)
 427                 return;
 428
 429         ppgtt->base.cleanup(&ppgtt->base);
 430         dev_priv->mm.aliasing_ppgtt = NULL;
 431 }
 432
 433 void i915_ppgtt_bind_object(struct i915_hw_ppgtt *ppgtt,
 434                             struct drm_i915_gem_object *obj,
 435                             enum i915_cache_level cache_level)
 436 {
 437         ppgtt->base.insert_entries(&ppgtt->base, obj->pages,
 438                                    i915_gem_obj_ggtt_offset(obj) >> PAGE_SHIFT,
 439                                    cache_level);
 440 }
 441
 442 void i915_ppgtt_unbind_object(struct i915_hw_ppgtt *ppgtt,
 443                               struct drm_i915_gem_object *obj)
 444 {
 445         ppgtt->base.clear_range(&ppgtt->base,
 446                                 i915_gem_obj_ggtt_offset(obj) >> PAGE_SHIFT,
 447                                 obj->base.size >> PAGE_SHIFT);
 448 }
 449
 450 extern int intel_iommu_gfx_mapped;
 451 /* Certain Gen5 chipsets require require idling the GPU before
 452  * unmapping anything from the GTT when VT-d is enabled.
 453  */
 454 static inline bool needs_idle_maps(struct drm_device *dev)
 455 {
 456 #ifdef CONFIG_INTEL_IOMMU
 457         /* Query intel_iommu to see if we need the workaround. Presumably that
 458          * was loaded first.
 459          */
 460         if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
 461                 return true;
 462 #endif
 463         return false;
 464 }
 465
 466 static bool do_idling(struct drm_i915_private *dev_priv)
 467 {
 468         bool ret = dev_priv->mm.interruptible;
 469
 470         if (unlikely(dev_priv->gtt.do_idle_maps)) {
 471                 dev_priv->mm.interruptible = false;
 472                 if (i915_gpu_idle(dev_priv->dev)) {
 473                         DRM_ERROR("Couldn't idle GPU\n");
 474                         /* Wait a bit, in hopes it avoids the hang */
 475                         udelay(10);
 476                 }
 477         }
 478
 479         return ret;
 480 }
 481
 482 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
 483 {
 484         if (unlikely(dev_priv->gtt.do_idle_maps))
 485                 dev_priv->mm.interruptible = interruptible;
 486 }
 487
 488 void i915_gem_restore_gtt_mappings(struct drm_device *dev)
 489 {
 490         struct drm_i915_private *dev_priv = dev->dev_private;
 491         struct drm_i915_gem_object *obj;
 492
 493         /* First fill our portion of the GTT with scratch pages */
 494         dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
 495                                        dev_priv->gtt.base.start / PAGE_SIZE,
 496                                        dev_priv->gtt.base.total / PAGE_SIZE);
 497
 498         list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
 499                 i915_gem_clflush_object(obj, obj->pin_display);
 500                 i915_gem_gtt_bind_object(obj, obj->cache_level);
 501         }
 502
 503         i915_gem_chipset_flush(dev);
 504 }
 505
 506 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
 507 {
 508         if (obj->has_dma_mapping)
 509                 return 0;
 510
 511         if (!dma_map_sg(&obj->base.dev->pdev->dev,
 512                         obj->pages->sgl, obj->pages->nents,
 513                         PCI_DMA_BIDIRECTIONAL))
 514                 return -ENOSPC;
 515
 516         return 0;
 517 }
 518
 519 /*
 520  * Binds an object into the global gtt with the specified cache level. The object
 521  * will be accessible to the GPU via commands whose operands reference offsets
 522  * within the global GTT as well as accessible by the GPU through the GMADR
 523  * mapped BAR (dev_priv->mm.gtt->gtt).
 524  */
 525 static void gen6_ggtt_insert_entries(struct i915_address_space *vm,
 526                                      struct sg_table *st,
 527                                      unsigned int first_entry,
 528                                      enum i915_cache_level level)
 529 {
 530         struct drm_i915_private *dev_priv = vm->dev->dev_private;
 531         gen6_gtt_pte_t __iomem *gtt_entries =
 532                 (gen6_gtt_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
 533         int i = 0;
 534         struct sg_page_iter sg_iter;
 535         dma_addr_t addr;
 536
 537         for_each_sg_page(st->sgl, &sg_iter, st->nents, 0) {
 538                 addr = sg_page_iter_dma_address(&sg_iter);
 539                 iowrite32(vm->pte_encode(addr, level), &gtt_entries[i]);
 540                 i++;
 541         }
 542
 543         /* XXX: This serves as a posting read to make sure that the PTE has
 544          * actually been updated. There is some concern that even though
 545          * registers and PTEs are within the same BAR that they are potentially
 546          * of NUMA access patterns. Therefore, even with the way we assume
 547          * hardware should work, we must keep this posting read for paranoia.
 548          */
 549         if (i != 0)
 550                 WARN_ON(readl(&gtt_entries[i-1]) !=
 551                         vm->pte_encode(addr, level));
 552
 553         /* This next bit makes the above posting read even more important. We
 554          * want to flush the TLBs only after we're certain all the PTE updates
 555          * have finished.
 556          */
 557         I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
 558         POSTING_READ(GFX_FLSH_CNTL_GEN6);
 559 }
 560
 561 static void gen6_ggtt_clear_range(struct i915_address_space *vm,
 562                                   unsigned int first_entry,
 563                                   unsigned int num_entries)
 564 {
 565         struct drm_i915_private *dev_priv = vm->dev->dev_private;
 566         gen6_gtt_pte_t scratch_pte, __iomem *gtt_base =
 567                 (gen6_gtt_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
 568         const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
 569         int i;
 570
 571         if (WARN(num_entries > max_entries,
 572                  "First entry = %d; Num entries = %d (max=%d)\n",
 573                  first_entry, num_entries, max_entries))
 574                 num_entries = max_entries;
 575
 576         scratch_pte = vm->pte_encode(vm->scratch.addr, I915_CACHE_LLC);
 577         for (i = 0; i < num_entries; i++)
 578                 iowrite32(scratch_pte, &gtt_base[i]);
 579         readl(gtt_base);
 580 }
 581
 582
 583 static void i915_ggtt_insert_entries(struct i915_address_space *vm,
 584                                      struct sg_table *st,
 585                                      unsigned int pg_start,
 586                                      enum i915_cache_level cache_level)
 587 {
 588         unsigned int flags = (cache_level == I915_CACHE_NONE) ?
 589                 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
 590
 591         intel_gtt_insert_sg_entries(st, pg_start, flags);
 592
 593 }
 594
 595 static void i915_ggtt_clear_range(struct i915_address_space *vm,
 596                                   unsigned int first_entry,
 597                                   unsigned int num_entries)
 598 {
 599         intel_gtt_clear_range(first_entry, num_entries);
 600 }
 601
 602
 603 void i915_gem_gtt_bind_object(struct drm_i915_gem_object *obj,
 604                               enum i915_cache_level cache_level)
 605 {
 606         struct drm_device *dev = obj->base.dev;
 607         struct drm_i915_private *dev_priv = dev->dev_private;
 608         const unsigned long entry = i915_gem_obj_ggtt_offset(obj) >> PAGE_SHIFT;
 609
 610         dev_priv->gtt.base.insert_entries(&dev_priv->gtt.base, obj->pages,
 611                                           entry,
 612                                           cache_level);
 613
 614         obj->has_global_gtt_mapping = 1;
 615 }
 616
 617 void i915_gem_gtt_unbind_object(struct drm_i915_gem_object *obj)
 618 {
 619         struct drm_device *dev = obj->base.dev;
 620         struct drm_i915_private *dev_priv = dev->dev_private;
 621         const unsigned long entry = i915_gem_obj_ggtt_offset(obj) >> PAGE_SHIFT;
 622
 623         dev_priv->gtt.base.clear_range(&dev_priv->gtt.base,
 624                                        entry,
 625                                        obj->base.size >> PAGE_SHIFT);
 626
 627         obj->has_global_gtt_mapping = 0;
 628 }
 629
 630 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
 631 {
 632         struct drm_device *dev = obj->base.dev;
 633         struct drm_i915_private *dev_priv = dev->dev_private;
 634         bool interruptible;
 635
 636         interruptible = do_idling(dev_priv);
 637
 638         if (!obj->has_dma_mapping)
 639                 dma_unmap_sg(&dev->pdev->dev,
 640                              obj->pages->sgl, obj->pages->nents,
 641                              PCI_DMA_BIDIRECTIONAL);
 642
 643         undo_idling(dev_priv, interruptible);
 644 }
 645
 646 static void i915_gtt_color_adjust(struct drm_mm_node *node,
 647                                   unsigned long color,
 648                                   unsigned long *start,
 649                                   unsigned long *end)
 650 {
 651         if (node->color != color)
 652                 *start += 4096;
 653
 654         if (!list_empty(&node->node_list)) {
 655                 node = list_entry(node->node_list.next,
 656                                   struct drm_mm_node,
 657                                   node_list);
 658                 if (node->allocated && node->color != color)
 659                         *end -= 4096;
 660         }
 661 }
 662 void i915_gem_setup_global_gtt(struct drm_device *dev,
 663                                unsigned long start,
 664                                unsigned long mappable_end,
 665                                unsigned long end)
 666 {
 667         /* Let GEM Manage all of the aperture.
 668          *
 669          * However, leave one page at the end still bound to the scratch page.
 670          * There are a number of places where the hardware apparently prefetches
 671          * past the end of the object, and we've seen multiple hangs with the
 672          * GPU head pointer stuck in a batchbuffer bound at the last page of the
 673          * aperture.  One page should be enough to keep any prefetching inside
 674          * of the aperture.
 675          */
 676         struct drm_i915_private *dev_priv = dev->dev_private;
 677         struct i915_address_space *ggtt_vm = &dev_priv->gtt.base;
 678         struct drm_mm_node *entry;
 679         struct drm_i915_gem_object *obj;
 680         unsigned long hole_start, hole_end;
 681
 682         BUG_ON(mappable_end > end);
 683
 684         /* Subtract the guard page ... */
 685         drm_mm_init(&ggtt_vm->mm, start, end - start - PAGE_SIZE);
 686         if (!HAS_LLC(dev))
 687                 dev_priv->gtt.base.mm.color_adjust = i915_gtt_color_adjust;
 688
 689         /* Mark any preallocated objects as occupied */
 690         list_for_each_entry(obj, &dev_priv->mm.bound_list, global_list) {
 691                 struct i915_vma *vma = i915_gem_obj_to_vma(obj, ggtt_vm);
 692                 int ret;
 693                 DRM_DEBUG_KMS("reserving preallocated space: %lx + %zx\n",
 694                               i915_gem_obj_ggtt_offset(obj), obj->base.size);
 695
 696                 WARN_ON(i915_gem_obj_ggtt_bound(obj));
 697                 ret = drm_mm_reserve_node(&ggtt_vm->mm, &vma->node);
 698                 if (ret)
 699                         DRM_DEBUG_KMS("Reservation failed\n");
 700                 obj->has_global_gtt_mapping = 1;
 701                 list_add(&vma->vma_link, &obj->vma_list);
 702         }
 703
 704         dev_priv->gtt.base.start = start;
 705         dev_priv->gtt.base.total = end - start;
 706
 707         /* Clear any non-preallocated blocks */
 708         drm_mm_for_each_hole(entry, &ggtt_vm->mm, hole_start, hole_end) {
 709                 const unsigned long count = (hole_end - hole_start) / PAGE_SIZE;
 710                 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
 711                               hole_start, hole_end);
 712                 ggtt_vm->clear_range(ggtt_vm, hole_start / PAGE_SIZE, count);
 713         }
 714
 715         /* And finally clear the reserved guard page */
 716         ggtt_vm->clear_range(ggtt_vm, end / PAGE_SIZE - 1, 1);
 717 }
 718
 719 static bool
 720 intel_enable_ppgtt(struct drm_device *dev)
 721 {
 722         if (i915_enable_ppgtt >= 0)
 723                 return i915_enable_ppgtt;
 724
 725 #ifdef CONFIG_INTEL_IOMMU
 726         /* Disable ppgtt on SNB if VT-d is on. */
 727         if (INTEL_INFO(dev)->gen == 6 && intel_iommu_gfx_mapped)
 728                 return false;
 729 #endif
 730
 731         return true;
 732 }
 733
 734 void i915_gem_init_global_gtt(struct drm_device *dev)
 735 {
 736         struct drm_i915_private *dev_priv = dev->dev_private;
 737         unsigned long gtt_size, mappable_size;
 738
 739         gtt_size = dev_priv->gtt.base.total;
 740         mappable_size = dev_priv->gtt.mappable_end;
 741
 742         if (intel_enable_ppgtt(dev) && HAS_ALIASING_PPGTT(dev)) {
 743                 int ret;
 744
 745                 if (INTEL_INFO(dev)->gen <= 7) {
 746                         /* PPGTT pdes are stolen from global gtt ptes, so shrink the
 747                          * aperture accordingly when using aliasing ppgtt. */
 748                         gtt_size -= GEN6_PPGTT_PD_ENTRIES * PAGE_SIZE;
 749                 }
 750
 751                 i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
 752
 753                 ret = i915_gem_init_aliasing_ppgtt(dev);
 754                 if (!ret)
 755                         return;
 756
 757                 DRM_ERROR("Aliased PPGTT setup failed %d\n", ret);
 758                 drm_mm_takedown(&dev_priv->gtt.base.mm);
 759                 gtt_size += GEN6_PPGTT_PD_ENTRIES * PAGE_SIZE;
 760         }
 761         i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
 762 }
 763
 764 static int setup_scratch_page(struct drm_device *dev)
 765 {
 766         struct drm_i915_private *dev_priv = dev->dev_private;
 767         struct page *page;
 768         dma_addr_t dma_addr;
 769
 770         page = alloc_page(GFP_KERNEL | GFP_DMA32 | __GFP_ZERO);
 771         if (page == NULL)
 772                 return -ENOMEM;
 773         get_page(page);
 774         set_pages_uc(page, 1);
 775
 776 #ifdef CONFIG_INTEL_IOMMU
 777         dma_addr = pci_map_page(dev->pdev, page, 0, PAGE_SIZE,
 778                                 PCI_DMA_BIDIRECTIONAL);
 779         if (pci_dma_mapping_error(dev->pdev, dma_addr))
 780                 return -EINVAL;
 781 #else
 782         dma_addr = page_to_phys(page);
 783 #endif
 784         dev_priv->gtt.base.scratch.page = page;
 785         dev_priv->gtt.base.scratch.addr = dma_addr;
 786
 787         return 0;
 788 }
 789
 790 static void teardown_scratch_page(struct drm_device *dev)
 791 {
 792         struct drm_i915_private *dev_priv = dev->dev_private;
 793         struct page *page = dev_priv->gtt.base.scratch.page;
 794
 795         set_pages_wb(page, 1);
 796         pci_unmap_page(dev->pdev, dev_priv->gtt.base.scratch.addr,
 797                        PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 798         put_page(page);
 799         __free_page(page);
 800 }
 801
 802 static inline unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
 803 {
 804         snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
 805         snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
 806         return snb_gmch_ctl << 20;
 807 }
 808
 809 static inline size_t gen6_get_stolen_size(u16 snb_gmch_ctl)
 810 {
 811         snb_gmch_ctl >>= SNB_GMCH_GMS_SHIFT;
 812         snb_gmch_ctl &= SNB_GMCH_GMS_MASK;
 813         return snb_gmch_ctl << 25; /* 32 MB units */
 814 }
 815
 816 static int gen6_gmch_probe(struct drm_device *dev,
 817                            size_t *gtt_total,
 818                            size_t *stolen,
 819                            phys_addr_t *mappable_base,
 820                            unsigned long *mappable_end)
 821 {
 822         struct drm_i915_private *dev_priv = dev->dev_private;
 823         phys_addr_t gtt_bus_addr;
 824         unsigned int gtt_size;
 825         u16 snb_gmch_ctl;
 826         int ret;
 827
 828         *mappable_base = pci_resource_start(dev->pdev, 2);
 829         *mappable_end = pci_resource_len(dev->pdev, 2);
 830
 831         /* 64/512MB is the current min/max we actually know of, but this is just
 832          * a coarse sanity check.
 833          */
 834         if ((*mappable_end < (64<<20) || (*mappable_end > (512<<20)))) {
 835                 DRM_ERROR("Unknown GMADR size (%lx)\n",
 836                           dev_priv->gtt.mappable_end);
 837                 return -ENXIO;
 838         }
 839
 840         if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(40)))
 841                 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(40));
 842         pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
 843         gtt_size = gen6_get_total_gtt_size(snb_gmch_ctl);
 844
 845         *stolen = gen6_get_stolen_size(snb_gmch_ctl);
 846         *gtt_total = (gtt_size / sizeof(gen6_gtt_pte_t)) << PAGE_SHIFT;
 847
 848         /* For Modern GENs the PTEs and register space are split in the BAR */
 849         gtt_bus_addr = pci_resource_start(dev->pdev, 0) +
 850                 (pci_resource_len(dev->pdev, 0) / 2);
 851
 852         dev_priv->gtt.gsm = ioremap_wc(gtt_bus_addr, gtt_size);
 853         if (!dev_priv->gtt.gsm) {
 854                 DRM_ERROR("Failed to map the gtt page table\n");
 855                 return -ENOMEM;
 856         }
 857
 858         ret = setup_scratch_page(dev);
 859         if (ret)
 860                 DRM_ERROR("Scratch setup failed\n");
 861
 862         dev_priv->gtt.base.clear_range = gen6_ggtt_clear_range;
 863         dev_priv->gtt.base.insert_entries = gen6_ggtt_insert_entries;
 864
 865         return ret;
 866 }
 867
 868 static void gen6_gmch_remove(struct i915_address_space *vm)
 869 {
 870
 871         struct i915_gtt *gtt = container_of(vm, struct i915_gtt, base);
 872         iounmap(gtt->gsm);
 873         teardown_scratch_page(vm->dev);
 874 }
 875
 876 static int i915_gmch_probe(struct drm_device *dev,
 877                            size_t *gtt_total,
 878                            size_t *stolen,
 879                            phys_addr_t *mappable_base,
 880                            unsigned long *mappable_end)
 881 {
 882         struct drm_i915_private *dev_priv = dev->dev_private;
 883         int ret;
 884
 885         ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->dev->pdev, NULL);
 886         if (!ret) {
 887                 DRM_ERROR("failed to set up gmch\n");
 888                 return -EIO;
 889         }
 890
 891         intel_gtt_get(gtt_total, stolen, mappable_base, mappable_end);
 892
 893         dev_priv->gtt.do_idle_maps = needs_idle_maps(dev_priv->dev);
 894         dev_priv->gtt.base.clear_range = i915_ggtt_clear_range;
 895         dev_priv->gtt.base.insert_entries = i915_ggtt_insert_entries;
 896
 897         return 0;
 898 }
 899
 900 static void i915_gmch_remove(struct i915_address_space *vm)
 901 {
 902         intel_gmch_remove();
 903 }
 904
 905 int i915_gem_gtt_init(struct drm_device *dev)
 906 {
 907         struct drm_i915_private *dev_priv = dev->dev_private;
 908         struct i915_gtt *gtt = &dev_priv->gtt;
 909         int ret;
 910
 911         if (INTEL_INFO(dev)->gen <= 5) {
 912                 gtt->gtt_probe = i915_gmch_probe;
 913                 gtt->base.cleanup = i915_gmch_remove;
 914         } else {
 915                 gtt->gtt_probe = gen6_gmch_probe;
 916                 gtt->base.cleanup = gen6_gmch_remove;
 917                 if (IS_HASWELL(dev) && dev_priv->ellc_size)
 918                         gtt->base.pte_encode = iris_pte_encode;
 919                 else if (IS_HASWELL(dev))
 920                         gtt->base.pte_encode = hsw_pte_encode;
 921                 else if (IS_VALLEYVIEW(dev))
 922                         gtt->base.pte_encode = byt_pte_encode;
 923                 else if (INTEL_INFO(dev)->gen >= 7)
 924                         gtt->base.pte_encode = ivb_pte_encode;
 925                 else
 926                         gtt->base.pte_encode = snb_pte_encode;
 927         }
 928
 929         ret = gtt->gtt_probe(dev, &gtt->base.total, &gtt->stolen_size,
 930                              &gtt->mappable_base, &gtt->mappable_end);
 931         if (ret)
 932                 return ret;
 933
 934         gtt->base.dev = dev;
 935
 936         /* GMADR is the PCI mmio aperture into the global GTT. */
 937         DRM_INFO("Memory usable by graphics device = %zdM\n",
 938                  gtt->base.total >> 20);
 939         DRM_DEBUG_DRIVER("GMADR size = %ldM\n", gtt->mappable_end >> 20);
 940         DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n", gtt->stolen_size >> 20);
 941
 942         return 0;
 943 }