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 typedef uint32_t gtt_pte_t;
  32
  33 /* PPGTT stuff */
  34 #define GEN6_GTT_ADDR_ENCODE(addr)      ((addr) | (((addr) >> 28) & 0xff0))
  35
  36 #define GEN6_PDE_VALID                  (1 << 0)
  37 /* gen6+ has bit 11-4 for physical addr bit 39-32 */
  38 #define GEN6_PDE_ADDR_ENCODE(addr)      GEN6_GTT_ADDR_ENCODE(addr)
  39
  40 #define GEN6_PTE_VALID                  (1 << 0)
  41 #define GEN6_PTE_UNCACHED               (1 << 1)
  42 #define HSW_PTE_UNCACHED                (0)
  43 #define GEN6_PTE_CACHE_LLC              (2 << 1)
  44 #define GEN6_PTE_CACHE_LLC_MLC          (3 << 1)
  45 #define GEN6_PTE_ADDR_ENCODE(addr)      GEN6_GTT_ADDR_ENCODE(addr)
  46
  47 static inline gtt_pte_t gen6_pte_encode(struct drm_device *dev,
  48                                         dma_addr_t addr,
  49                                         enum i915_cache_level level)
  50 {
  51         gtt_pte_t pte = GEN6_PTE_VALID;
  52         pte |= GEN6_PTE_ADDR_ENCODE(addr);
  53
  54         switch (level) {
  55         case I915_CACHE_LLC_MLC:
  56                 /* Haswell doesn't set L3 this way */
  57                 if (IS_HASWELL(dev))
  58                         pte |= GEN6_PTE_CACHE_LLC;
  59                 else
  60                         pte |= GEN6_PTE_CACHE_LLC_MLC;
  61                 break;
  62         case I915_CACHE_LLC:
  63                 pte |= GEN6_PTE_CACHE_LLC;
  64                 break;
  65         case I915_CACHE_NONE:
  66                 if (IS_HASWELL(dev))
  67                         pte |= HSW_PTE_UNCACHED;
  68                 else
  69                         pte |= GEN6_PTE_UNCACHED;
  70                 break;
  71         default:
  72                 BUG();
  73         }
  74
  75
  76         return pte;
  77 }
  78
  79 /* PPGTT support for Sandybdrige/Gen6 and later */
  80 static void gen6_ppgtt_clear_range(struct i915_hw_ppgtt *ppgtt,
  81                                    unsigned first_entry,
  82                                    unsigned num_entries)
  83 {
  84         gtt_pte_t *pt_vaddr;
  85         gtt_pte_t scratch_pte;
  86         unsigned act_pd = first_entry / I915_PPGTT_PT_ENTRIES;
  87         unsigned first_pte = first_entry % I915_PPGTT_PT_ENTRIES;
  88         unsigned last_pte, i;
  89
  90         scratch_pte = gen6_pte_encode(ppgtt->dev,
  91                                       ppgtt->scratch_page_dma_addr,
  92                                       I915_CACHE_LLC);
  93
  94         while (num_entries) {
  95                 last_pte = first_pte + num_entries;
  96                 if (last_pte > I915_PPGTT_PT_ENTRIES)
  97                         last_pte = I915_PPGTT_PT_ENTRIES;
  98
  99                 pt_vaddr = kmap_atomic(ppgtt->pt_pages[act_pd]);
 100
 101                 for (i = first_pte; i < last_pte; i++)
 102                         pt_vaddr[i] = scratch_pte;
 103
 104                 kunmap_atomic(pt_vaddr);
 105
 106                 num_entries -= last_pte - first_pte;
 107                 first_pte = 0;
 108                 act_pd++;
 109         }
 110 }
 111
 112 static void gen6_ppgtt_insert_entries(struct i915_hw_ppgtt *ppgtt,
 113                                       struct sg_table *pages,
 114                                       unsigned first_entry,
 115                                       enum i915_cache_level cache_level)
 116 {
 117         gtt_pte_t *pt_vaddr;
 118         unsigned act_pd = first_entry / I915_PPGTT_PT_ENTRIES;
 119         unsigned first_pte = first_entry % I915_PPGTT_PT_ENTRIES;
 120         unsigned i, j, m, segment_len;
 121         dma_addr_t page_addr;
 122         struct scatterlist *sg;
 123
 124         /* init sg walking */
 125         sg = pages->sgl;
 126         i = 0;
 127         segment_len = sg_dma_len(sg) >> PAGE_SHIFT;
 128         m = 0;
 129
 130         while (i < pages->nents) {
 131                 pt_vaddr = kmap_atomic(ppgtt->pt_pages[act_pd]);
 132
 133                 for (j = first_pte; j < I915_PPGTT_PT_ENTRIES; j++) {
 134                         page_addr = sg_dma_address(sg) + (m << PAGE_SHIFT);
 135                         pt_vaddr[j] = gen6_pte_encode(ppgtt->dev, page_addr,
 136                                                       cache_level);
 137
 138                         /* grab the next page */
 139                         if (++m == segment_len) {
 140                                 if (++i == pages->nents)
 141                                         break;
 142
 143                                 sg = sg_next(sg);
 144                                 segment_len = sg_dma_len(sg) >> PAGE_SHIFT;
 145                                 m = 0;
 146                         }
 147                 }
 148
 149                 kunmap_atomic(pt_vaddr);
 150
 151                 first_pte = 0;
 152                 act_pd++;
 153         }
 154 }
 155
 156 static void gen6_ppgtt_cleanup(struct i915_hw_ppgtt *ppgtt)
 157 {
 158         int i;
 159
 160         if (ppgtt->pt_dma_addr) {
 161                 for (i = 0; i < ppgtt->num_pd_entries; i++)
 162                         pci_unmap_page(ppgtt->dev->pdev,
 163                                        ppgtt->pt_dma_addr[i],
 164                                        4096, PCI_DMA_BIDIRECTIONAL);
 165         }
 166
 167         kfree(ppgtt->pt_dma_addr);
 168         for (i = 0; i < ppgtt->num_pd_entries; i++)
 169                 __free_page(ppgtt->pt_pages[i]);
 170         kfree(ppgtt->pt_pages);
 171         kfree(ppgtt);
 172 }
 173
 174 static int gen6_ppgtt_init(struct i915_hw_ppgtt *ppgtt)
 175 {
 176         struct drm_device *dev = ppgtt->dev;
 177         struct drm_i915_private *dev_priv = dev->dev_private;
 178         unsigned first_pd_entry_in_global_pt;
 179         int i;
 180         int ret = -ENOMEM;
 181
 182         /* ppgtt PDEs reside in the global gtt pagetable, which has 512*1024
 183          * entries. For aliasing ppgtt support we just steal them at the end for
 184          * now. */
 185         first_pd_entry_in_global_pt =
 186                 gtt_total_entries(dev_priv->gtt) - I915_PPGTT_PD_ENTRIES;
 187
 188         ppgtt->num_pd_entries = I915_PPGTT_PD_ENTRIES;
 189         ppgtt->clear_range = gen6_ppgtt_clear_range;
 190         ppgtt->insert_entries = gen6_ppgtt_insert_entries;
 191         ppgtt->cleanup = gen6_ppgtt_cleanup;
 192         ppgtt->pt_pages = kzalloc(sizeof(struct page *)*ppgtt->num_pd_entries,
 193                                   GFP_KERNEL);
 194         if (!ppgtt->pt_pages)
 195                 return -ENOMEM;
 196
 197         for (i = 0; i < ppgtt->num_pd_entries; i++) {
 198                 ppgtt->pt_pages[i] = alloc_page(GFP_KERNEL);
 199                 if (!ppgtt->pt_pages[i])
 200                         goto err_pt_alloc;
 201         }
 202
 203         ppgtt->pt_dma_addr = kzalloc(sizeof(dma_addr_t) *ppgtt->num_pd_entries,
 204                                      GFP_KERNEL);
 205         if (!ppgtt->pt_dma_addr)
 206                 goto err_pt_alloc;
 207
 208         for (i = 0; i < ppgtt->num_pd_entries; i++) {
 209                 dma_addr_t pt_addr;
 210
 211                 pt_addr = pci_map_page(dev->pdev, ppgtt->pt_pages[i], 0, 4096,
 212                                        PCI_DMA_BIDIRECTIONAL);
 213
 214                 if (pci_dma_mapping_error(dev->pdev, pt_addr)) {
 215                         ret = -EIO;
 216                         goto err_pd_pin;
 217
 218                 }
 219                 ppgtt->pt_dma_addr[i] = pt_addr;
 220         }
 221
 222         ppgtt->scratch_page_dma_addr = dev_priv->gtt.scratch_page_dma;
 223
 224         ppgtt->clear_range(ppgtt, 0,
 225                            ppgtt->num_pd_entries*I915_PPGTT_PT_ENTRIES);
 226
 227         ppgtt->pd_offset = (first_pd_entry_in_global_pt)*sizeof(gtt_pte_t);
 228
 229         return 0;
 230
 231 err_pd_pin:
 232         if (ppgtt->pt_dma_addr) {
 233                 for (i--; i >= 0; i--)
 234                         pci_unmap_page(dev->pdev, ppgtt->pt_dma_addr[i],
 235                                        4096, PCI_DMA_BIDIRECTIONAL);
 236         }
 237 err_pt_alloc:
 238         kfree(ppgtt->pt_dma_addr);
 239         for (i = 0; i < ppgtt->num_pd_entries; i++) {
 240                 if (ppgtt->pt_pages[i])
 241                         __free_page(ppgtt->pt_pages[i]);
 242         }
 243         kfree(ppgtt->pt_pages);
 244
 245         return ret;
 246 }
 247
 248 static int i915_gem_init_aliasing_ppgtt(struct drm_device *dev)
 249 {
 250         struct drm_i915_private *dev_priv = dev->dev_private;
 251         struct i915_hw_ppgtt *ppgtt;
 252         int ret;
 253
 254         ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
 255         if (!ppgtt)
 256                 return -ENOMEM;
 257
 258         ppgtt->dev = dev;
 259
 260         ret = gen6_ppgtt_init(ppgtt);
 261         if (ret)
 262                 kfree(ppgtt);
 263         else
 264                 dev_priv->mm.aliasing_ppgtt = ppgtt;
 265
 266         return ret;
 267 }
 268
 269 void i915_gem_cleanup_aliasing_ppgtt(struct drm_device *dev)
 270 {
 271         struct drm_i915_private *dev_priv = dev->dev_private;
 272         struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
 273
 274         if (!ppgtt)
 275                 return;
 276
 277         ppgtt->cleanup(ppgtt);
 278 }
 279
 280 void i915_ppgtt_bind_object(struct i915_hw_ppgtt *ppgtt,
 281                             struct drm_i915_gem_object *obj,
 282                             enum i915_cache_level cache_level)
 283 {
 284         ppgtt->insert_entries(ppgtt, obj->pages,
 285                               obj->gtt_space->start >> PAGE_SHIFT,
 286                               cache_level);
 287 }
 288
 289 void i915_ppgtt_unbind_object(struct i915_hw_ppgtt *ppgtt,
 290                               struct drm_i915_gem_object *obj)
 291 {
 292         ppgtt->clear_range(ppgtt,
 293                            obj->gtt_space->start >> PAGE_SHIFT,
 294                            obj->base.size >> PAGE_SHIFT);
 295 }
 296
 297 void i915_gem_init_ppgtt(struct drm_device *dev)
 298 {
 299         drm_i915_private_t *dev_priv = dev->dev_private;
 300         uint32_t pd_offset;
 301         struct intel_ring_buffer *ring;
 302         struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
 303         gtt_pte_t __iomem *pd_addr;
 304         uint32_t pd_entry;
 305         int i;
 306
 307         if (!dev_priv->mm.aliasing_ppgtt)
 308                 return;
 309
 310
 311         pd_addr = (gtt_pte_t __iomem*)dev_priv->gtt.gsm + ppgtt->pd_offset/sizeof(gtt_pte_t);
 312         for (i = 0; i < ppgtt->num_pd_entries; i++) {
 313                 dma_addr_t pt_addr;
 314
 315                 pt_addr = ppgtt->pt_dma_addr[i];
 316                 pd_entry = GEN6_PDE_ADDR_ENCODE(pt_addr);
 317                 pd_entry |= GEN6_PDE_VALID;
 318
 319                 writel(pd_entry, pd_addr + i);
 320         }
 321         readl(pd_addr);
 322
 323         pd_offset = ppgtt->pd_offset;
 324         pd_offset /= 64; /* in cachelines, */
 325         pd_offset <<= 16;
 326
 327         if (INTEL_INFO(dev)->gen == 6) {
 328                 uint32_t ecochk, gab_ctl, ecobits;
 329
 330                 ecobits = I915_READ(GAC_ECO_BITS);
 331                 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
 332
 333                 gab_ctl = I915_READ(GAB_CTL);
 334                 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
 335
 336                 ecochk = I915_READ(GAM_ECOCHK);
 337                 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT |
 338                                        ECOCHK_PPGTT_CACHE64B);
 339                 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
 340         } else if (INTEL_INFO(dev)->gen >= 7) {
 341                 I915_WRITE(GAM_ECOCHK, ECOCHK_PPGTT_CACHE64B);
 342                 /* GFX_MODE is per-ring on gen7+ */
 343         }
 344
 345         for_each_ring(ring, dev_priv, i) {
 346                 if (INTEL_INFO(dev)->gen >= 7)
 347                         I915_WRITE(RING_MODE_GEN7(ring),
 348                                    _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
 349
 350                 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
 351                 I915_WRITE(RING_PP_DIR_BASE(ring), pd_offset);
 352         }
 353 }
 354
 355 extern int intel_iommu_gfx_mapped;
 356 /* Certain Gen5 chipsets require require idling the GPU before
 357  * unmapping anything from the GTT when VT-d is enabled.
 358  */
 359 static inline bool needs_idle_maps(struct drm_device *dev)
 360 {
 361 #ifdef CONFIG_INTEL_IOMMU
 362         /* Query intel_iommu to see if we need the workaround. Presumably that
 363          * was loaded first.
 364          */
 365         if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
 366                 return true;
 367 #endif
 368         return false;
 369 }
 370
 371 static bool do_idling(struct drm_i915_private *dev_priv)
 372 {
 373         bool ret = dev_priv->mm.interruptible;
 374
 375         if (unlikely(dev_priv->gtt.do_idle_maps)) {
 376                 dev_priv->mm.interruptible = false;
 377                 if (i915_gpu_idle(dev_priv->dev)) {
 378                         DRM_ERROR("Couldn't idle GPU\n");
 379                         /* Wait a bit, in hopes it avoids the hang */
 380                         udelay(10);
 381                 }
 382         }
 383
 384         return ret;
 385 }
 386
 387 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
 388 {
 389         if (unlikely(dev_priv->gtt.do_idle_maps))
 390                 dev_priv->mm.interruptible = interruptible;
 391 }
 392
 393 void i915_gem_restore_gtt_mappings(struct drm_device *dev)
 394 {
 395         struct drm_i915_private *dev_priv = dev->dev_private;
 396         struct drm_i915_gem_object *obj;
 397
 398         /* First fill our portion of the GTT with scratch pages */
 399         dev_priv->gtt.gtt_clear_range(dev, dev_priv->gtt.start / PAGE_SIZE,
 400                                       dev_priv->gtt.total / PAGE_SIZE);
 401
 402         list_for_each_entry(obj, &dev_priv->mm.bound_list, gtt_list) {
 403                 i915_gem_clflush_object(obj);
 404                 i915_gem_gtt_bind_object(obj, obj->cache_level);
 405         }
 406
 407         i915_gem_chipset_flush(dev);
 408 }
 409
 410 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
 411 {
 412         if (obj->has_dma_mapping)
 413                 return 0;
 414
 415         if (!dma_map_sg(&obj->base.dev->pdev->dev,
 416                         obj->pages->sgl, obj->pages->nents,
 417                         PCI_DMA_BIDIRECTIONAL))
 418                 return -ENOSPC;
 419
 420         return 0;
 421 }
 422
 423 /*
 424  * Binds an object into the global gtt with the specified cache level. The object
 425  * will be accessible to the GPU via commands whose operands reference offsets
 426  * within the global GTT as well as accessible by the GPU through the GMADR
 427  * mapped BAR (dev_priv->mm.gtt->gtt).
 428  */
 429 static void gen6_ggtt_insert_entries(struct drm_device *dev,
 430                                      struct sg_table *st,
 431                                      unsigned int first_entry,
 432                                      enum i915_cache_level level)
 433 {
 434         struct drm_i915_private *dev_priv = dev->dev_private;
 435         struct scatterlist *sg = st->sgl;
 436         gtt_pte_t __iomem *gtt_entries =
 437                 (gtt_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
 438         int unused, i = 0;
 439         unsigned int len, m = 0;
 440         dma_addr_t addr;
 441
 442         for_each_sg(st->sgl, sg, st->nents, unused) {
 443                 len = sg_dma_len(sg) >> PAGE_SHIFT;
 444                 for (m = 0; m < len; m++) {
 445                         addr = sg_dma_address(sg) + (m << PAGE_SHIFT);
 446                         iowrite32(gen6_pte_encode(dev, addr, level),
 447                                   &gtt_entries[i]);
 448                         i++;
 449                 }
 450         }
 451
 452         /* XXX: This serves as a posting read to make sure that the PTE has
 453          * actually been updated. There is some concern that even though
 454          * registers and PTEs are within the same BAR that they are potentially
 455          * of NUMA access patterns. Therefore, even with the way we assume
 456          * hardware should work, we must keep this posting read for paranoia.
 457          */
 458         if (i != 0)
 459                 WARN_ON(readl(&gtt_entries[i-1])
 460                         != gen6_pte_encode(dev, addr, level));
 461
 462         /* This next bit makes the above posting read even more important. We
 463          * want to flush the TLBs only after we're certain all the PTE updates
 464          * have finished.
 465          */
 466         I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
 467         POSTING_READ(GFX_FLSH_CNTL_GEN6);
 468 }
 469
 470 static void gen6_ggtt_clear_range(struct drm_device *dev,
 471                                   unsigned int first_entry,
 472                                   unsigned int num_entries)
 473 {
 474         struct drm_i915_private *dev_priv = dev->dev_private;
 475         gtt_pte_t scratch_pte;
 476         gtt_pte_t __iomem *gtt_base = (gtt_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
 477         const int max_entries = gtt_total_entries(dev_priv->gtt) - first_entry;
 478         int i;
 479
 480         if (WARN(num_entries > max_entries,
 481                  "First entry = %d; Num entries = %d (max=%d)\n",
 482                  first_entry, num_entries, max_entries))
 483                 num_entries = max_entries;
 484
 485         scratch_pte = gen6_pte_encode(dev, dev_priv->gtt.scratch_page_dma,
 486                                       I915_CACHE_LLC);
 487         for (i = 0; i < num_entries; i++)
 488                 iowrite32(scratch_pte, &gtt_base[i]);
 489         readl(gtt_base);
 490 }
 491
 492
 493 static void i915_ggtt_insert_entries(struct drm_device *dev,
 494                                      struct sg_table *st,
 495                                      unsigned int pg_start,
 496                                      enum i915_cache_level cache_level)
 497 {
 498         unsigned int flags = (cache_level == I915_CACHE_NONE) ?
 499                 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
 500
 501         intel_gtt_insert_sg_entries(st, pg_start, flags);
 502
 503 }
 504
 505 static void i915_ggtt_clear_range(struct drm_device *dev,
 506                                   unsigned int first_entry,
 507                                   unsigned int num_entries)
 508 {
 509         intel_gtt_clear_range(first_entry, num_entries);
 510 }
 511
 512
 513 void i915_gem_gtt_bind_object(struct drm_i915_gem_object *obj,
 514                               enum i915_cache_level cache_level)
 515 {
 516         struct drm_device *dev = obj->base.dev;
 517         struct drm_i915_private *dev_priv = dev->dev_private;
 518
 519         dev_priv->gtt.gtt_insert_entries(dev, obj->pages,
 520                                          obj->gtt_space->start >> PAGE_SHIFT,
 521                                          cache_level);
 522
 523         obj->has_global_gtt_mapping = 1;
 524 }
 525
 526 void i915_gem_gtt_unbind_object(struct drm_i915_gem_object *obj)
 527 {
 528         struct drm_device *dev = obj->base.dev;
 529         struct drm_i915_private *dev_priv = dev->dev_private;
 530
 531         dev_priv->gtt.gtt_clear_range(obj->base.dev,
 532                                       obj->gtt_space->start >> PAGE_SHIFT,
 533                                       obj->base.size >> PAGE_SHIFT);
 534
 535         obj->has_global_gtt_mapping = 0;
 536 }
 537
 538 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
 539 {
 540         struct drm_device *dev = obj->base.dev;
 541         struct drm_i915_private *dev_priv = dev->dev_private;
 542         bool interruptible;
 543
 544         interruptible = do_idling(dev_priv);
 545
 546         if (!obj->has_dma_mapping)
 547                 dma_unmap_sg(&dev->pdev->dev,
 548                              obj->pages->sgl, obj->pages->nents,
 549                              PCI_DMA_BIDIRECTIONAL);
 550
 551         undo_idling(dev_priv, interruptible);
 552 }
 553
 554 static void i915_gtt_color_adjust(struct drm_mm_node *node,
 555                                   unsigned long color,
 556                                   unsigned long *start,
 557                                   unsigned long *end)
 558 {
 559         if (node->color != color)
 560                 *start += 4096;
 561
 562         if (!list_empty(&node->node_list)) {
 563                 node = list_entry(node->node_list.next,
 564                                   struct drm_mm_node,
 565                                   node_list);
 566                 if (node->allocated && node->color != color)
 567                         *end -= 4096;
 568         }
 569 }
 570 void i915_gem_setup_global_gtt(struct drm_device *dev,
 571                                unsigned long start,
 572                                unsigned long mappable_end,
 573                                unsigned long end)
 574 {
 575         /* Let GEM Manage all of the aperture.
 576          *
 577          * However, leave one page at the end still bound to the scratch page.
 578          * There are a number of places where the hardware apparently prefetches
 579          * past the end of the object, and we've seen multiple hangs with the
 580          * GPU head pointer stuck in a batchbuffer bound at the last page of the
 581          * aperture.  One page should be enough to keep any prefetching inside
 582          * of the aperture.
 583          */
 584         drm_i915_private_t *dev_priv = dev->dev_private;
 585         struct drm_mm_node *entry;
 586         struct drm_i915_gem_object *obj;
 587         unsigned long hole_start, hole_end;
 588
 589         BUG_ON(mappable_end > end);
 590
 591         /* Subtract the guard page ... */
 592         drm_mm_init(&dev_priv->mm.gtt_space, start, end - start - PAGE_SIZE);
 593         if (!HAS_LLC(dev))
 594                 dev_priv->mm.gtt_space.color_adjust = i915_gtt_color_adjust;
 595
 596         /* Mark any preallocated objects as occupied */
 597         list_for_each_entry(obj, &dev_priv->mm.bound_list, gtt_list) {
 598                 DRM_DEBUG_KMS("reserving preallocated space: %x + %zx\n",
 599                               obj->gtt_offset, obj->base.size);
 600
 601                 BUG_ON(obj->gtt_space != I915_GTT_RESERVED);
 602                 obj->gtt_space = drm_mm_create_block(&dev_priv->mm.gtt_space,
 603                                                      obj->gtt_offset,
 604                                                      obj->base.size,
 605                                                      false);
 606                 obj->has_global_gtt_mapping = 1;
 607         }
 608
 609         dev_priv->gtt.start = start;
 610         dev_priv->gtt.total = end - start;
 611
 612         /* Clear any non-preallocated blocks */
 613         drm_mm_for_each_hole(entry, &dev_priv->mm.gtt_space,
 614                              hole_start, hole_end) {
 615                 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
 616                               hole_start, hole_end);
 617                 dev_priv->gtt.gtt_clear_range(dev, hole_start / PAGE_SIZE,
 618                                               (hole_end-hole_start) / PAGE_SIZE);
 619         }
 620
 621         /* And finally clear the reserved guard page */
 622         dev_priv->gtt.gtt_clear_range(dev, end / PAGE_SIZE - 1, 1);
 623 }
 624
 625 static bool
 626 intel_enable_ppgtt(struct drm_device *dev)
 627 {
 628         if (i915_enable_ppgtt >= 0)
 629                 return i915_enable_ppgtt;
 630
 631 #ifdef CONFIG_INTEL_IOMMU
 632         /* Disable ppgtt on SNB if VT-d is on. */
 633         if (INTEL_INFO(dev)->gen == 6 && intel_iommu_gfx_mapped)
 634                 return false;
 635 #endif
 636
 637         return true;
 638 }
 639
 640 void i915_gem_init_global_gtt(struct drm_device *dev)
 641 {
 642         struct drm_i915_private *dev_priv = dev->dev_private;
 643         unsigned long gtt_size, mappable_size;
 644
 645         gtt_size = dev_priv->gtt.total;
 646         mappable_size = dev_priv->gtt.mappable_end;
 647
 648         if (intel_enable_ppgtt(dev) && HAS_ALIASING_PPGTT(dev)) {
 649                 int ret;
 650                 /* PPGTT pdes are stolen from global gtt ptes, so shrink the
 651                  * aperture accordingly when using aliasing ppgtt. */
 652                 gtt_size -= I915_PPGTT_PD_ENTRIES*PAGE_SIZE;
 653
 654                 i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
 655
 656                 ret = i915_gem_init_aliasing_ppgtt(dev);
 657                 if (!ret)
 658                         return;
 659
 660                 DRM_ERROR("Aliased PPGTT setup failed %d\n", ret);
 661                 drm_mm_takedown(&dev_priv->mm.gtt_space);
 662                 gtt_size += I915_PPGTT_PD_ENTRIES*PAGE_SIZE;
 663         }
 664         i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
 665 }
 666
 667 static int setup_scratch_page(struct drm_device *dev)
 668 {
 669         struct drm_i915_private *dev_priv = dev->dev_private;
 670         struct page *page;
 671         dma_addr_t dma_addr;
 672
 673         page = alloc_page(GFP_KERNEL | GFP_DMA32 | __GFP_ZERO);
 674         if (page == NULL)
 675                 return -ENOMEM;
 676         get_page(page);
 677         set_pages_uc(page, 1);
 678
 679 #ifdef CONFIG_INTEL_IOMMU
 680         dma_addr = pci_map_page(dev->pdev, page, 0, PAGE_SIZE,
 681                                 PCI_DMA_BIDIRECTIONAL);
 682         if (pci_dma_mapping_error(dev->pdev, dma_addr))
 683                 return -EINVAL;
 684 #else
 685         dma_addr = page_to_phys(page);
 686 #endif
 687         dev_priv->gtt.scratch_page = page;
 688         dev_priv->gtt.scratch_page_dma = dma_addr;
 689
 690         return 0;
 691 }
 692
 693 static void teardown_scratch_page(struct drm_device *dev)
 694 {
 695         struct drm_i915_private *dev_priv = dev->dev_private;
 696         set_pages_wb(dev_priv->gtt.scratch_page, 1);
 697         pci_unmap_page(dev->pdev, dev_priv->gtt.scratch_page_dma,
 698                        PAGE_SIZE, PCI_DMA_BIDIRECTIONAL);
 699         put_page(dev_priv->gtt.scratch_page);
 700         __free_page(dev_priv->gtt.scratch_page);
 701 }
 702
 703 static inline unsigned int gen6_get_total_gtt_size(u16 snb_gmch_ctl)
 704 {
 705         snb_gmch_ctl >>= SNB_GMCH_GGMS_SHIFT;
 706         snb_gmch_ctl &= SNB_GMCH_GGMS_MASK;
 707         return snb_gmch_ctl << 20;
 708 }
 709
 710 static inline size_t gen6_get_stolen_size(u16 snb_gmch_ctl)
 711 {
 712         snb_gmch_ctl >>= SNB_GMCH_GMS_SHIFT;
 713         snb_gmch_ctl &= SNB_GMCH_GMS_MASK;
 714         return snb_gmch_ctl << 25; /* 32 MB units */
 715 }
 716
 717 static inline size_t gen7_get_stolen_size(u16 snb_gmch_ctl)
 718 {
 719         static const int stolen_decoder[] = {
 720                 0, 0, 0, 0, 0, 32, 48, 64, 128, 256, 96, 160, 224, 352};
 721         snb_gmch_ctl >>= IVB_GMCH_GMS_SHIFT;
 722         snb_gmch_ctl &= IVB_GMCH_GMS_MASK;
 723         return stolen_decoder[snb_gmch_ctl] << 20;
 724 }
 725
 726 static int gen6_gmch_probe(struct drm_device *dev,
 727                            size_t *gtt_total,
 728                            size_t *stolen,
 729                            phys_addr_t *mappable_base,
 730                            unsigned long *mappable_end)
 731 {
 732         struct drm_i915_private *dev_priv = dev->dev_private;
 733         phys_addr_t gtt_bus_addr;
 734         unsigned int gtt_size;
 735         u16 snb_gmch_ctl;
 736         int ret;
 737
 738         *mappable_base = pci_resource_start(dev->pdev, 2);
 739         *mappable_end = pci_resource_len(dev->pdev, 2);
 740
 741         /* 64/512MB is the current min/max we actually know of, but this is just
 742          * a coarse sanity check.
 743          */
 744         if ((*mappable_end < (64<<20) || (*mappable_end > (512<<20)))) {
 745                 DRM_ERROR("Unknown GMADR size (%lx)\n",
 746                           dev_priv->gtt.mappable_end);
 747                 return -ENXIO;
 748         }
 749
 750         if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(40)))
 751                 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(40));
 752         pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
 753         gtt_size = gen6_get_total_gtt_size(snb_gmch_ctl);
 754
 755         if (IS_GEN7(dev))
 756                 *stolen = gen7_get_stolen_size(snb_gmch_ctl);
 757         else
 758                 *stolen = gen6_get_stolen_size(snb_gmch_ctl);
 759
 760         *gtt_total = (gtt_size / sizeof(gtt_pte_t)) << PAGE_SHIFT;
 761
 762         /* For GEN6+ the PTEs for the ggtt live at 2MB + BAR0 */
 763         gtt_bus_addr = pci_resource_start(dev->pdev, 0) + (2<<20);
 764         dev_priv->gtt.gsm = ioremap_wc(gtt_bus_addr, gtt_size);
 765         if (!dev_priv->gtt.gsm) {
 766                 DRM_ERROR("Failed to map the gtt page table\n");
 767                 return -ENOMEM;
 768         }
 769
 770         ret = setup_scratch_page(dev);
 771         if (ret)
 772                 DRM_ERROR("Scratch setup failed\n");
 773
 774         dev_priv->gtt.gtt_clear_range = gen6_ggtt_clear_range;
 775         dev_priv->gtt.gtt_insert_entries = gen6_ggtt_insert_entries;
 776
 777         return ret;
 778 }
 779
 780 static void gen6_gmch_remove(struct drm_device *dev)
 781 {
 782         struct drm_i915_private *dev_priv = dev->dev_private;
 783         iounmap(dev_priv->gtt.gsm);
 784         teardown_scratch_page(dev_priv->dev);
 785 }
 786
 787 static int i915_gmch_probe(struct drm_device *dev,
 788                            size_t *gtt_total,
 789                            size_t *stolen,
 790                            phys_addr_t *mappable_base,
 791                            unsigned long *mappable_end)
 792 {
 793         struct drm_i915_private *dev_priv = dev->dev_private;
 794         int ret;
 795
 796         ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->dev->pdev, NULL);
 797         if (!ret) {
 798                 DRM_ERROR("failed to set up gmch\n");
 799                 return -EIO;
 800         }
 801
 802         intel_gtt_get(gtt_total, stolen, mappable_base, mappable_end);
 803
 804         dev_priv->gtt.do_idle_maps = needs_idle_maps(dev_priv->dev);
 805         dev_priv->gtt.gtt_clear_range = i915_ggtt_clear_range;
 806         dev_priv->gtt.gtt_insert_entries = i915_ggtt_insert_entries;
 807
 808         return 0;
 809 }
 810
 811 static void i915_gmch_remove(struct drm_device *dev)
 812 {
 813         intel_gmch_remove();
 814 }
 815
 816 int i915_gem_gtt_init(struct drm_device *dev)
 817 {
 818         struct drm_i915_private *dev_priv = dev->dev_private;
 819         struct i915_gtt *gtt = &dev_priv->gtt;
 820         unsigned long gtt_size;
 821         int ret;
 822
 823         if (INTEL_INFO(dev)->gen <= 5) {
 824                 dev_priv->gtt.gtt_probe = i915_gmch_probe;
 825                 dev_priv->gtt.gtt_remove = i915_gmch_remove;
 826         } else {
 827                 dev_priv->gtt.gtt_probe = gen6_gmch_probe;
 828                 dev_priv->gtt.gtt_remove = gen6_gmch_remove;
 829         }
 830
 831         ret = dev_priv->gtt.gtt_probe(dev, &dev_priv->gtt.total,
 832                                      &dev_priv->gtt.stolen_size,
 833                                      &gtt->mappable_base,
 834                                      &gtt->mappable_end);
 835         if (ret)
 836                 return ret;
 837
 838         gtt_size = (dev_priv->gtt.total >> PAGE_SHIFT) * sizeof(gtt_pte_t);
 839
 840         /* GMADR is the PCI mmio aperture into the global GTT. */
 841         DRM_INFO("Memory usable by graphics device = %zdM\n",
 842                  dev_priv->gtt.total >> 20);
 843         DRM_DEBUG_DRIVER("GMADR size = %ldM\n",
 844                          dev_priv->gtt.mappable_end >> 20);
 845         DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n",
 846                          dev_priv->gtt.stolen_size >> 20);
 847
 848         return 0;
 849 }