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 = dev_priv->mm.gtt->gtt_total_entries - I915_PPGTT_PD_ENTRIES;
 186
 187         ppgtt->num_pd_entries = I915_PPGTT_PD_ENTRIES;
 188         ppgtt->clear_range = gen6_ppgtt_clear_range;
 189         ppgtt->insert_entries = gen6_ppgtt_insert_entries;
 190         ppgtt->cleanup = gen6_ppgtt_cleanup;
 191         ppgtt->pt_pages = kzalloc(sizeof(struct page *)*ppgtt->num_pd_entries,
 192                                   GFP_KERNEL);
 193         if (!ppgtt->pt_pages)
 194                 return -ENOMEM;
 195
 196         for (i = 0; i < ppgtt->num_pd_entries; i++) {
 197                 ppgtt->pt_pages[i] = alloc_page(GFP_KERNEL);
 198                 if (!ppgtt->pt_pages[i])
 199                         goto err_pt_alloc;
 200         }
 201
 202         ppgtt->pt_dma_addr = kzalloc(sizeof(dma_addr_t) *ppgtt->num_pd_entries,
 203                                      GFP_KERNEL);
 204         if (!ppgtt->pt_dma_addr)
 205                 goto err_pt_alloc;
 206
 207         for (i = 0; i < ppgtt->num_pd_entries; i++) {
 208                 dma_addr_t pt_addr;
 209
 210                 pt_addr = pci_map_page(dev->pdev, ppgtt->pt_pages[i], 0, 4096,
 211                                        PCI_DMA_BIDIRECTIONAL);
 212
 213                 if (pci_dma_mapping_error(dev->pdev, pt_addr)) {
 214                         ret = -EIO;
 215                         goto err_pd_pin;
 216
 217                 }
 218                 ppgtt->pt_dma_addr[i] = pt_addr;
 219         }
 220
 221         ppgtt->scratch_page_dma_addr = dev_priv->gtt.scratch_page_dma;
 222
 223         ppgtt->clear_range(ppgtt, 0,
 224                            ppgtt->num_pd_entries*I915_PPGTT_PT_ENTRIES);
 225
 226         ppgtt->pd_offset = (first_pd_entry_in_global_pt)*sizeof(gtt_pte_t);
 227
 228         return 0;
 229
 230 err_pd_pin:
 231         if (ppgtt->pt_dma_addr) {
 232                 for (i--; i >= 0; i--)
 233                         pci_unmap_page(dev->pdev, ppgtt->pt_dma_addr[i],
 234                                        4096, PCI_DMA_BIDIRECTIONAL);
 235         }
 236 err_pt_alloc:
 237         kfree(ppgtt->pt_dma_addr);
 238         for (i = 0; i < ppgtt->num_pd_entries; i++) {
 239                 if (ppgtt->pt_pages[i])
 240                         __free_page(ppgtt->pt_pages[i]);
 241         }
 242         kfree(ppgtt->pt_pages);
 243
 244         return ret;
 245 }
 246
 247 static int i915_gem_init_aliasing_ppgtt(struct drm_device *dev)
 248 {
 249         struct drm_i915_private *dev_priv = dev->dev_private;
 250         struct i915_hw_ppgtt *ppgtt;
 251         int ret;
 252
 253         ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
 254         if (!ppgtt)
 255                 return -ENOMEM;
 256
 257         ppgtt->dev = dev;
 258
 259         ret = gen6_ppgtt_init(ppgtt);
 260         if (ret)
 261                 kfree(ppgtt);
 262         else
 263                 dev_priv->mm.aliasing_ppgtt = ppgtt;
 264
 265         return ret;
 266 }
 267
 268 void i915_gem_cleanup_aliasing_ppgtt(struct drm_device *dev)
 269 {
 270         struct drm_i915_private *dev_priv = dev->dev_private;
 271         struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
 272
 273         if (!ppgtt)
 274                 return;
 275
 276         ppgtt->cleanup(ppgtt);
 277 }
 278
 279 void i915_ppgtt_bind_object(struct i915_hw_ppgtt *ppgtt,
 280                             struct drm_i915_gem_object *obj,
 281                             enum i915_cache_level cache_level)
 282 {
 283         ppgtt->insert_entries(ppgtt, obj->pages,
 284                               obj->gtt_space->start >> PAGE_SHIFT,
 285                               cache_level);
 286 }
 287
 288 void i915_ppgtt_unbind_object(struct i915_hw_ppgtt *ppgtt,
 289                               struct drm_i915_gem_object *obj)
 290 {
 291         ppgtt->clear_range(ppgtt,
 292                            obj->gtt_space->start >> PAGE_SHIFT,
 293                            obj->base.size >> PAGE_SHIFT);
 294 }
 295
 296 void i915_gem_init_ppgtt(struct drm_device *dev)
 297 {
 298         drm_i915_private_t *dev_priv = dev->dev_private;
 299         uint32_t pd_offset;
 300         struct intel_ring_buffer *ring;
 301         struct i915_hw_ppgtt *ppgtt = dev_priv->mm.aliasing_ppgtt;
 302         gtt_pte_t __iomem *pd_addr;
 303         uint32_t pd_entry;
 304         int i;
 305
 306         if (!dev_priv->mm.aliasing_ppgtt)
 307                 return;
 308
 309
 310         pd_addr = (gtt_pte_t __iomem*)dev_priv->gtt.gsm + ppgtt->pd_offset/sizeof(gtt_pte_t);
 311         for (i = 0; i < ppgtt->num_pd_entries; i++) {
 312                 dma_addr_t pt_addr;
 313
 314                 pt_addr = ppgtt->pt_dma_addr[i];
 315                 pd_entry = GEN6_PDE_ADDR_ENCODE(pt_addr);
 316                 pd_entry |= GEN6_PDE_VALID;
 317
 318                 writel(pd_entry, pd_addr + i);
 319         }
 320         readl(pd_addr);
 321
 322         pd_offset = ppgtt->pd_offset;
 323         pd_offset /= 64; /* in cachelines, */
 324         pd_offset <<= 16;
 325
 326         if (INTEL_INFO(dev)->gen == 6) {
 327                 uint32_t ecochk, gab_ctl, ecobits;
 328
 329                 ecobits = I915_READ(GAC_ECO_BITS);
 330                 I915_WRITE(GAC_ECO_BITS, ecobits | ECOBITS_PPGTT_CACHE64B);
 331
 332                 gab_ctl = I915_READ(GAB_CTL);
 333                 I915_WRITE(GAB_CTL, gab_ctl | GAB_CTL_CONT_AFTER_PAGEFAULT);
 334
 335                 ecochk = I915_READ(GAM_ECOCHK);
 336                 I915_WRITE(GAM_ECOCHK, ecochk | ECOCHK_SNB_BIT |
 337                                        ECOCHK_PPGTT_CACHE64B);
 338                 I915_WRITE(GFX_MODE, _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
 339         } else if (INTEL_INFO(dev)->gen >= 7) {
 340                 I915_WRITE(GAM_ECOCHK, ECOCHK_PPGTT_CACHE64B);
 341                 /* GFX_MODE is per-ring on gen7+ */
 342         }
 343
 344         for_each_ring(ring, dev_priv, i) {
 345                 if (INTEL_INFO(dev)->gen >= 7)
 346                         I915_WRITE(RING_MODE_GEN7(ring),
 347                                    _MASKED_BIT_ENABLE(GFX_PPGTT_ENABLE));
 348
 349                 I915_WRITE(RING_PP_DIR_DCLV(ring), PP_DIR_DCLV_2G);
 350                 I915_WRITE(RING_PP_DIR_BASE(ring), pd_offset);
 351         }
 352 }
 353
 354 extern int intel_iommu_gfx_mapped;
 355 /* Certain Gen5 chipsets require require idling the GPU before
 356  * unmapping anything from the GTT when VT-d is enabled.
 357  */
 358 static inline bool needs_idle_maps(struct drm_device *dev)
 359 {
 360 #ifdef CONFIG_INTEL_IOMMU
 361         /* Query intel_iommu to see if we need the workaround. Presumably that
 362          * was loaded first.
 363          */
 364         if (IS_GEN5(dev) && IS_MOBILE(dev) && intel_iommu_gfx_mapped)
 365                 return true;
 366 #endif
 367         return false;
 368 }
 369
 370 static bool do_idling(struct drm_i915_private *dev_priv)
 371 {
 372         bool ret = dev_priv->mm.interruptible;
 373
 374         if (unlikely(dev_priv->gtt.do_idle_maps)) {
 375                 dev_priv->mm.interruptible = false;
 376                 if (i915_gpu_idle(dev_priv->dev)) {
 377                         DRM_ERROR("Couldn't idle GPU\n");
 378                         /* Wait a bit, in hopes it avoids the hang */
 379                         udelay(10);
 380                 }
 381         }
 382
 383         return ret;
 384 }
 385
 386 static void undo_idling(struct drm_i915_private *dev_priv, bool interruptible)
 387 {
 388         if (unlikely(dev_priv->gtt.do_idle_maps))
 389                 dev_priv->mm.interruptible = interruptible;
 390 }
 391
 392 void i915_gem_restore_gtt_mappings(struct drm_device *dev)
 393 {
 394         struct drm_i915_private *dev_priv = dev->dev_private;
 395         struct drm_i915_gem_object *obj;
 396
 397         /* First fill our portion of the GTT with scratch pages */
 398         dev_priv->gtt.gtt_clear_range(dev, dev_priv->gtt.start / PAGE_SIZE,
 399                                       dev_priv->gtt.total / PAGE_SIZE);
 400
 401         list_for_each_entry(obj, &dev_priv->mm.bound_list, gtt_list) {
 402                 i915_gem_clflush_object(obj);
 403                 i915_gem_gtt_bind_object(obj, obj->cache_level);
 404         }
 405
 406         i915_gem_chipset_flush(dev);
 407 }
 408
 409 int i915_gem_gtt_prepare_object(struct drm_i915_gem_object *obj)
 410 {
 411         if (obj->has_dma_mapping)
 412                 return 0;
 413
 414         if (!dma_map_sg(&obj->base.dev->pdev->dev,
 415                         obj->pages->sgl, obj->pages->nents,
 416                         PCI_DMA_BIDIRECTIONAL))
 417                 return -ENOSPC;
 418
 419         return 0;
 420 }
 421
 422 /*
 423  * Binds an object into the global gtt with the specified cache level. The object
 424  * will be accessible to the GPU via commands whose operands reference offsets
 425  * within the global GTT as well as accessible by the GPU through the GMADR
 426  * mapped BAR (dev_priv->mm.gtt->gtt).
 427  */
 428 static void gen6_ggtt_insert_entries(struct drm_device *dev,
 429                                      struct sg_table *st,
 430                                      unsigned int first_entry,
 431                                      enum i915_cache_level level)
 432 {
 433         struct drm_i915_private *dev_priv = dev->dev_private;
 434         struct scatterlist *sg = st->sgl;
 435         gtt_pte_t __iomem *gtt_entries =
 436                 (gtt_pte_t __iomem *)dev_priv->gtt.gsm + first_entry;
 437         int unused, i = 0;
 438         unsigned int len, m = 0;
 439         dma_addr_t addr;
 440
 441         for_each_sg(st->sgl, sg, st->nents, unused) {
 442                 len = sg_dma_len(sg) >> PAGE_SHIFT;
 443                 for (m = 0; m < len; m++) {
 444                         addr = sg_dma_address(sg) + (m << PAGE_SHIFT);
 445                         iowrite32(gen6_pte_encode(dev, addr, level),
 446                                   &gtt_entries[i]);
 447                         i++;
 448                 }
 449         }
 450
 451         /* XXX: This serves as a posting read to make sure that the PTE has
 452          * actually been updated. There is some concern that even though
 453          * registers and PTEs are within the same BAR that they are potentially
 454          * of NUMA access patterns. Therefore, even with the way we assume
 455          * hardware should work, we must keep this posting read for paranoia.
 456          */
 457         if (i != 0)
 458                 WARN_ON(readl(&gtt_entries[i-1])
 459                         != gen6_pte_encode(dev, addr, level));
 460
 461         /* This next bit makes the above posting read even more important. We
 462          * want to flush the TLBs only after we're certain all the PTE updates
 463          * have finished.
 464          */
 465         I915_WRITE(GFX_FLSH_CNTL_GEN6, GFX_FLSH_CNTL_EN);
 466         POSTING_READ(GFX_FLSH_CNTL_GEN6);
 467 }
 468
 469 static void gen6_ggtt_clear_range(struct drm_device *dev,
 470                                   unsigned int first_entry,
 471                                   unsigned int num_entries)
 472 {
 473         struct drm_i915_private *dev_priv = dev->dev_private;
 474         gtt_pte_t scratch_pte;
 475         gtt_pte_t __iomem *gtt_base = (gtt_pte_t __iomem *) dev_priv->gtt.gsm + first_entry;
 476         const int max_entries = dev_priv->mm.gtt->gtt_total_entries - first_entry;
 477         int i;
 478
 479         if (WARN(num_entries > max_entries,
 480                  "First entry = %d; Num entries = %d (max=%d)\n",
 481                  first_entry, num_entries, max_entries))
 482                 num_entries = max_entries;
 483
 484         scratch_pte = gen6_pte_encode(dev, dev_priv->gtt.scratch_page_dma,
 485                                       I915_CACHE_LLC);
 486         for (i = 0; i < num_entries; i++)
 487                 iowrite32(scratch_pte, &gtt_base[i]);
 488         readl(gtt_base);
 489 }
 490
 491
 492 static void i915_ggtt_insert_entries(struct drm_device *dev,
 493                                      struct sg_table *st,
 494                                      unsigned int pg_start,
 495                                      enum i915_cache_level cache_level)
 496 {
 497         unsigned int flags = (cache_level == I915_CACHE_NONE) ?
 498                 AGP_USER_MEMORY : AGP_USER_CACHED_MEMORY;
 499
 500         intel_gtt_insert_sg_entries(st, pg_start, flags);
 501
 502 }
 503
 504 static void i915_ggtt_clear_range(struct drm_device *dev,
 505                                   unsigned int first_entry,
 506                                   unsigned int num_entries)
 507 {
 508         intel_gtt_clear_range(first_entry, num_entries);
 509 }
 510
 511
 512 void i915_gem_gtt_bind_object(struct drm_i915_gem_object *obj,
 513                               enum i915_cache_level cache_level)
 514 {
 515         struct drm_device *dev = obj->base.dev;
 516         struct drm_i915_private *dev_priv = dev->dev_private;
 517
 518         dev_priv->gtt.gtt_insert_entries(dev, obj->pages,
 519                                          obj->gtt_space->start >> PAGE_SHIFT,
 520                                          cache_level);
 521
 522         obj->has_global_gtt_mapping = 1;
 523 }
 524
 525 void i915_gem_gtt_unbind_object(struct drm_i915_gem_object *obj)
 526 {
 527         struct drm_device *dev = obj->base.dev;
 528         struct drm_i915_private *dev_priv = dev->dev_private;
 529
 530         dev_priv->gtt.gtt_clear_range(obj->base.dev,
 531                                       obj->gtt_space->start >> PAGE_SHIFT,
 532                                       obj->base.size >> PAGE_SHIFT);
 533
 534         obj->has_global_gtt_mapping = 0;
 535 }
 536
 537 void i915_gem_gtt_finish_object(struct drm_i915_gem_object *obj)
 538 {
 539         struct drm_device *dev = obj->base.dev;
 540         struct drm_i915_private *dev_priv = dev->dev_private;
 541         bool interruptible;
 542
 543         interruptible = do_idling(dev_priv);
 544
 545         if (!obj->has_dma_mapping)
 546                 dma_unmap_sg(&dev->pdev->dev,
 547                              obj->pages->sgl, obj->pages->nents,
 548                              PCI_DMA_BIDIRECTIONAL);
 549
 550         undo_idling(dev_priv, interruptible);
 551 }
 552
 553 static void i915_gtt_color_adjust(struct drm_mm_node *node,
 554                                   unsigned long color,
 555                                   unsigned long *start,
 556                                   unsigned long *end)
 557 {
 558         if (node->color != color)
 559                 *start += 4096;
 560
 561         if (!list_empty(&node->node_list)) {
 562                 node = list_entry(node->node_list.next,
 563                                   struct drm_mm_node,
 564                                   node_list);
 565                 if (node->allocated && node->color != color)
 566                         *end -= 4096;
 567         }
 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         drm_i915_private_t *dev_priv = dev->dev_private;
 576         struct drm_mm_node *entry;
 577         struct drm_i915_gem_object *obj;
 578         unsigned long hole_start, hole_end;
 579
 580         BUG_ON(mappable_end > end);
 581
 582         /* Subtract the guard page ... */
 583         drm_mm_init(&dev_priv->mm.gtt_space, start, end - start - PAGE_SIZE);
 584         if (!HAS_LLC(dev))
 585                 dev_priv->mm.gtt_space.color_adjust = i915_gtt_color_adjust;
 586
 587         /* Mark any preallocated objects as occupied */
 588         list_for_each_entry(obj, &dev_priv->mm.bound_list, gtt_list) {
 589                 DRM_DEBUG_KMS("reserving preallocated space: %x + %zx\n",
 590                               obj->gtt_offset, obj->base.size);
 591
 592                 BUG_ON(obj->gtt_space != I915_GTT_RESERVED);
 593                 obj->gtt_space = drm_mm_create_block(&dev_priv->mm.gtt_space,
 594                                                      obj->gtt_offset,
 595                                                      obj->base.size,
 596                                                      false);
 597                 obj->has_global_gtt_mapping = 1;
 598         }
 599
 600         dev_priv->gtt.start = start;
 601         dev_priv->gtt.total = end - start;
 602
 603         /* Clear any non-preallocated blocks */
 604         drm_mm_for_each_hole(entry, &dev_priv->mm.gtt_space,
 605                              hole_start, hole_end) {
 606                 DRM_DEBUG_KMS("clearing unused GTT space: [%lx, %lx]\n",
 607                               hole_start, hole_end);
 608                 dev_priv->gtt.gtt_clear_range(dev, hole_start / PAGE_SIZE,
 609                                               (hole_end-hole_start) / PAGE_SIZE);
 610         }
 611
 612         /* And finally clear the reserved guard page */
 613         dev_priv->gtt.gtt_clear_range(dev, end / PAGE_SIZE - 1, 1);
 614 }
 615
 616 static bool
 617 intel_enable_ppgtt(struct drm_device *dev)
 618 {
 619         if (i915_enable_ppgtt >= 0)
 620                 return i915_enable_ppgtt;
 621
 622 #ifdef CONFIG_INTEL_IOMMU
 623         /* Disable ppgtt on SNB if VT-d is on. */
 624         if (INTEL_INFO(dev)->gen == 6 && intel_iommu_gfx_mapped)
 625                 return false;
 626 #endif
 627
 628         return true;
 629 }
 630
 631 void i915_gem_init_global_gtt(struct drm_device *dev)
 632 {
 633         struct drm_i915_private *dev_priv = dev->dev_private;
 634         unsigned long gtt_size, mappable_size;
 635         int ret;
 636
 637         gtt_size = dev_priv->mm.gtt->gtt_total_entries << PAGE_SHIFT;
 638         mappable_size = dev_priv->gtt.mappable_end;
 639
 640         if (intel_enable_ppgtt(dev) && HAS_ALIASING_PPGTT(dev)) {
 641                 /* PPGTT pdes are stolen from global gtt ptes, so shrink the
 642                  * aperture accordingly when using aliasing ppgtt. */
 643                 gtt_size -= I915_PPGTT_PD_ENTRIES*PAGE_SIZE;
 644
 645                 i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
 646
 647                 ret = i915_gem_init_aliasing_ppgtt(dev);
 648                 if (ret) {
 649                         mutex_unlock(&dev->struct_mutex);
 650                         return;
 651                 }
 652         } else {
 653                 /* Let GEM Manage all of the aperture.
 654                  *
 655                  * However, leave one page at the end still bound to the scratch
 656                  * page.  There are a number of places where the hardware
 657                  * apparently prefetches past the end of the object, and we've
 658                  * seen multiple hangs with the GPU head pointer stuck in a
 659                  * batchbuffer bound at the last page of the aperture.  One page
 660                  * should be enough to keep any prefetching inside of the
 661                  * aperture.
 662                  */
 663                 i915_gem_setup_global_gtt(dev, 0, mappable_size, gtt_size);
 664         }
 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 {
 730         struct drm_i915_private *dev_priv = dev->dev_private;
 731         phys_addr_t gtt_bus_addr;
 732         unsigned int gtt_size;
 733         u16 snb_gmch_ctl;
 734         int ret;
 735
 736         /* 64/512MB is the current min/max we actually know of, but this is just
 737          * a coarse sanity check.
 738          */
 739         if ((dev_priv->gtt.mappable_end < (64<<20) ||
 740              (dev_priv->gtt.mappable_end > (512<<20)))) {
 741                 DRM_ERROR("Unknown GMADR size (%lx)\n",
 742                           dev_priv->gtt.mappable_end);
 743                 return -ENXIO;
 744         }
 745
 746         if (!pci_set_dma_mask(dev->pdev, DMA_BIT_MASK(40)))
 747                 pci_set_consistent_dma_mask(dev->pdev, DMA_BIT_MASK(40));
 748         pci_read_config_word(dev->pdev, SNB_GMCH_CTRL, &snb_gmch_ctl);
 749         gtt_size = gen6_get_total_gtt_size(snb_gmch_ctl);
 750
 751         if (IS_GEN7(dev))
 752                 *stolen = gen7_get_stolen_size(snb_gmch_ctl);
 753         else
 754                 *stolen = gen6_get_stolen_size(snb_gmch_ctl);
 755
 756         *gtt_total = (gtt_size / sizeof(gtt_pte_t)) << PAGE_SHIFT;
 757
 758         /* For GEN6+ the PTEs for the ggtt live at 2MB + BAR0 */
 759         gtt_bus_addr = pci_resource_start(dev->pdev, 0) + (2<<20);
 760         dev_priv->gtt.gsm = ioremap_wc(gtt_bus_addr, gtt_size);
 761         if (!dev_priv->gtt.gsm) {
 762                 DRM_ERROR("Failed to map the gtt page table\n");
 763                 return -ENOMEM;
 764         }
 765
 766         ret = setup_scratch_page(dev);
 767         if (ret)
 768                 DRM_ERROR("Scratch setup failed\n");
 769
 770         dev_priv->gtt.gtt_clear_range = gen6_ggtt_clear_range;
 771         dev_priv->gtt.gtt_insert_entries = gen6_ggtt_insert_entries;
 772
 773         return ret;
 774 }
 775
 776 void gen6_gmch_remove(struct drm_device *dev)
 777 {
 778         struct drm_i915_private *dev_priv = dev->dev_private;
 779         iounmap(dev_priv->gtt.gsm);
 780         teardown_scratch_page(dev_priv->dev);
 781         kfree(dev_priv->mm.gtt);
 782 }
 783
 784 static int i915_gmch_probe(struct drm_device *dev,
 785                            size_t *gtt_total,
 786                            size_t *stolen)
 787 {
 788         struct drm_i915_private *dev_priv = dev->dev_private;
 789         int ret;
 790
 791         /* This is a temporary hack to make the code cleaner in
 792          * i915_gem_gtt_init. I promise it will go away very shortly. */
 793         kfree(dev_priv->mm.gtt);
 794
 795         ret = intel_gmch_probe(dev_priv->bridge_dev, dev_priv->dev->pdev, NULL);
 796         if (!ret) {
 797                 DRM_ERROR("failed to set up gmch\n");
 798                 return -EIO;
 799         }
 800
 801         dev_priv->mm.gtt = intel_gtt_get();
 802         if (!dev_priv->mm.gtt) {
 803                 DRM_ERROR("Failed to initialize GTT\n");
 804                 intel_gmch_remove();
 805                 return -ENODEV;
 806         }
 807
 808         dev_priv->gtt.do_idle_maps = needs_idle_maps(dev_priv->dev);
 809         dev_priv->gtt.gtt_clear_range = i915_ggtt_clear_range;
 810         dev_priv->gtt.gtt_insert_entries = i915_ggtt_insert_entries;
 811
 812         return 0;
 813 }
 814
 815 static void i915_gmch_remove(struct drm_device *dev)
 816 {
 817         intel_gmch_remove();
 818 }
 819
 820 int i915_gem_gtt_init(struct drm_device *dev)
 821 {
 822         struct drm_i915_private *dev_priv = dev->dev_private;
 823         struct i915_gtt *gtt = &dev_priv->gtt;
 824         unsigned long gtt_size;
 825         int ret;
 826
 827         dev_priv->mm.gtt = kzalloc(sizeof(*dev_priv->mm.gtt), GFP_KERNEL);
 828         if (!dev_priv->mm.gtt)
 829                 return -ENOMEM;
 830
 831         gtt->mappable_base = pci_resource_start(dev->pdev, 2);
 832         gtt->mappable_end = pci_resource_len(dev->pdev, 2);
 833
 834         if (INTEL_INFO(dev)->gen <= 5) {
 835                 dev_priv->gtt.gtt_probe = i915_gmch_probe;
 836                 dev_priv->gtt.gtt_remove = i915_gmch_remove;
 837         } else {
 838                 dev_priv->gtt.gtt_probe = gen6_gmch_probe;
 839                 dev_priv->gtt.gtt_remove = gen6_gmch_remove;
 840         }
 841
 842         ret = dev_priv->gtt.gtt_probe(dev, &dev_priv->gtt.total,
 843                                      &dev_priv->gtt.stolen_size);
 844         if (ret) {
 845                 kfree(dev_priv->mm.gtt);
 846                 return ret;
 847         }
 848
 849         dev_priv->mm.gtt->gtt_total_entries = dev_priv->gtt.total >> PAGE_SHIFT;
 850         dev_priv->mm.gtt->stolen_size = dev_priv->gtt.stolen_size;
 851
 852         gtt_size = (dev_priv->gtt.total >> PAGE_SHIFT) * sizeof(gtt_pte_t);
 853
 854         /* GMADR is the PCI mmio aperture into the global GTT. */
 855         DRM_INFO("Memory usable by graphics device = %zdM\n",
 856                  dev_priv->gtt.total >> 20);
 857         DRM_DEBUG_DRIVER("GMADR size = %ldM\n",
 858                          dev_priv->gtt.mappable_end >> 20);
 859         DRM_DEBUG_DRIVER("GTT stolen size = %zdM\n",
 860                          dev_priv->gtt.stolen_size >> 20);
 861
 862         return 0;
 863 }