drivers/gpu/drm/i915/i915_gem_execbuffer.c

   1 /*
   2  * Copyright © 2008,2010 Intel Corporation
   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  * Authors:
  24  *    Eric Anholt <eric@anholt.net>
  25  *    Chris Wilson <chris@chris-wilson.co.uk>
  26  *
  27  */
  28
  29 #include <drm/drmP.h>
  30 #include <drm/i915_drm.h>
  31 #include "i915_drv.h"
  32 #include "i915_trace.h"
  33 #include "intel_drv.h"
  34 #include <linux/dma_remapping.h>
  35 #include <linux/uaccess.h>
  36
  37 #define  __EXEC_OBJECT_HAS_PIN (1<<31)
  38 #define  __EXEC_OBJECT_HAS_FENCE (1<<30)
  39 #define  __EXEC_OBJECT_NEEDS_MAP (1<<29)
  40 #define  __EXEC_OBJECT_NEEDS_BIAS (1<<28)
  41
  42 #define BATCH_OFFSET_BIAS (256*1024)
  43
  44 struct eb_vmas {
  45         struct list_head vmas;
  46         int and;
  47         union {
  48                 struct i915_vma *lut[0];
  49                 struct hlist_head buckets[0];
  50         };
  51 };
  52
  53 static struct eb_vmas *
  54 eb_create(struct drm_i915_gem_execbuffer2 *args)
  55 {
  56         struct eb_vmas *eb = NULL;
  57
  58         if (args->flags & I915_EXEC_HANDLE_LUT) {
  59                 unsigned size = args->buffer_count;
  60                 size *= sizeof(struct i915_vma *);
  61                 size += sizeof(struct eb_vmas);
  62                 eb = kmalloc(size, GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
  63         }
  64
  65         if (eb == NULL) {
  66                 unsigned size = args->buffer_count;
  67                 unsigned count = PAGE_SIZE / sizeof(struct hlist_head) / 2;
  68                 BUILD_BUG_ON_NOT_POWER_OF_2(PAGE_SIZE / sizeof(struct hlist_head));
  69                 while (count > 2*size)
  70                         count >>= 1;
  71                 eb = kzalloc(count*sizeof(struct hlist_head) +
  72                              sizeof(struct eb_vmas),
  73                              GFP_TEMPORARY);
  74                 if (eb == NULL)
  75                         return eb;
  76
  77                 eb->and = count - 1;
  78         } else
  79                 eb->and = -args->buffer_count;
  80
  81         INIT_LIST_HEAD(&eb->vmas);
  82         return eb;
  83 }
  84
  85 static void
  86 eb_reset(struct eb_vmas *eb)
  87 {
  88         if (eb->and >= 0)
  89                 memset(eb->buckets, 0, (eb->and+1)*sizeof(struct hlist_head));
  90 }
  91
  92 static int
  93 eb_lookup_vmas(struct eb_vmas *eb,
  94                struct drm_i915_gem_exec_object2 *exec,
  95                const struct drm_i915_gem_execbuffer2 *args,
  96                struct i915_address_space *vm,
  97                struct drm_file *file)
  98 {
  99         struct drm_i915_gem_object *obj;
 100         struct list_head objects;
 101         int i, ret;
 102
 103         INIT_LIST_HEAD(&objects);
 104         spin_lock(&file->table_lock);
 105         /* Grab a reference to the object and release the lock so we can lookup
 106          * or create the VMA without using GFP_ATOMIC */
 107         for (i = 0; i < args->buffer_count; i++) {
 108                 obj = to_intel_bo(idr_find(&file->object_idr, exec[i].handle));
 109                 if (obj == NULL) {
 110                         spin_unlock(&file->table_lock);
 111                         DRM_DEBUG("Invalid object handle %d at index %d\n",
 112                                    exec[i].handle, i);
 113                         ret = -ENOENT;
 114                         goto err;
 115                 }
 116
 117                 if (!list_empty(&obj->obj_exec_link)) {
 118                         spin_unlock(&file->table_lock);
 119                         DRM_DEBUG("Object %p [handle %d, index %d] appears more than once in object list\n",
 120                                    obj, exec[i].handle, i);
 121                         ret = -EINVAL;
 122                         goto err;
 123                 }
 124
 125                 drm_gem_object_reference(&obj->base);
 126                 list_add_tail(&obj->obj_exec_link, &objects);
 127         }
 128         spin_unlock(&file->table_lock);
 129
 130         i = 0;
 131         while (!list_empty(&objects)) {
 132                 struct i915_vma *vma;
 133
 134                 obj = list_first_entry(&objects,
 135                                        struct drm_i915_gem_object,
 136                                        obj_exec_link);
 137
 138                 /*
 139                  * NOTE: We can leak any vmas created here when something fails
 140                  * later on. But that's no issue since vma_unbind can deal with
 141                  * vmas which are not actually bound. And since only
 142                  * lookup_or_create exists as an interface to get at the vma
 143                  * from the (obj, vm) we don't run the risk of creating
 144                  * duplicated vmas for the same vm.
 145                  */
 146                 vma = i915_gem_obj_lookup_or_create_vma(obj, vm);
 147                 if (IS_ERR(vma)) {
 148                         DRM_DEBUG("Failed to lookup VMA\n");
 149                         ret = PTR_ERR(vma);
 150                         goto err;
 151                 }
 152
 153                 /* Transfer ownership from the objects list to the vmas list. */
 154                 list_add_tail(&vma->exec_list, &eb->vmas);
 155                 list_del_init(&obj->obj_exec_link);
 156
 157                 vma->exec_entry = &exec[i];
 158                 if (eb->and < 0) {
 159                         eb->lut[i] = vma;
 160                 } else {
 161                         uint32_t handle = args->flags & I915_EXEC_HANDLE_LUT ? i : exec[i].handle;
 162                         vma->exec_handle = handle;
 163                         hlist_add_head(&vma->exec_node,
 164                                        &eb->buckets[handle & eb->and]);
 165                 }
 166                 ++i;
 167         }
 168
 169         return 0;
 170
 171
 172 err:
 173         while (!list_empty(&objects)) {
 174                 obj = list_first_entry(&objects,
 175                                        struct drm_i915_gem_object,
 176                                        obj_exec_link);
 177                 list_del_init(&obj->obj_exec_link);
 178                 drm_gem_object_unreference(&obj->base);
 179         }
 180         /*
 181          * Objects already transfered to the vmas list will be unreferenced by
 182          * eb_destroy.
 183          */
 184
 185         return ret;
 186 }
 187
 188 static struct i915_vma *eb_get_vma(struct eb_vmas *eb, unsigned long handle)
 189 {
 190         if (eb->and < 0) {
 191                 if (handle >= -eb->and)
 192                         return NULL;
 193                 return eb->lut[handle];
 194         } else {
 195                 struct hlist_head *head;
 196                 struct i915_vma *vma;
 197
 198                 head = &eb->buckets[handle & eb->and];
 199                 hlist_for_each_entry(vma, head, exec_node) {
 200                         if (vma->exec_handle == handle)
 201                                 return vma;
 202                 }
 203                 return NULL;
 204         }
 205 }
 206
 207 static void
 208 i915_gem_execbuffer_unreserve_vma(struct i915_vma *vma)
 209 {
 210         struct drm_i915_gem_exec_object2 *entry;
 211         struct drm_i915_gem_object *obj = vma->obj;
 212
 213         if (!drm_mm_node_allocated(&vma->node))
 214                 return;
 215
 216         entry = vma->exec_entry;
 217
 218         if (entry->flags & __EXEC_OBJECT_HAS_FENCE)
 219                 i915_gem_object_unpin_fence(obj);
 220
 221         if (entry->flags & __EXEC_OBJECT_HAS_PIN)
 222                 vma->pin_count--;
 223
 224         entry->flags &= ~(__EXEC_OBJECT_HAS_FENCE | __EXEC_OBJECT_HAS_PIN);
 225 }
 226
 227 static void eb_destroy(struct eb_vmas *eb)
 228 {
 229         while (!list_empty(&eb->vmas)) {
 230                 struct i915_vma *vma;
 231
 232                 vma = list_first_entry(&eb->vmas,
 233                                        struct i915_vma,
 234                                        exec_list);
 235                 list_del_init(&vma->exec_list);
 236                 i915_gem_execbuffer_unreserve_vma(vma);
 237                 drm_gem_object_unreference(&vma->obj->base);
 238         }
 239         kfree(eb);
 240 }
 241
 242 static inline int use_cpu_reloc(struct drm_i915_gem_object *obj)
 243 {
 244         return (HAS_LLC(obj->base.dev) ||
 245                 obj->base.write_domain == I915_GEM_DOMAIN_CPU ||
 246                 obj->cache_level != I915_CACHE_NONE);
 247 }
 248
 249 /* Used to convert any address to canonical form.
 250  * Starting from gen8, some commands (e.g. STATE_BASE_ADDRESS,
 251  * MI_LOAD_REGISTER_MEM and others, see Broadwell PRM Vol2a) require the
 252  * addresses to be in a canonical form:
 253  * "GraphicsAddress[63:48] are ignored by the HW and assumed to be in correct
 254  * canonical form [63:48] == [47]."
 255  */
 256 #define GEN8_HIGH_ADDRESS_BIT 47
 257 static inline uint64_t gen8_canonical_addr(uint64_t address)
 258 {
 259         return sign_extend64(address, GEN8_HIGH_ADDRESS_BIT);
 260 }
 261
 262 static inline uint64_t gen8_noncanonical_addr(uint64_t address)
 263 {
 264         return address & ((1ULL << (GEN8_HIGH_ADDRESS_BIT + 1)) - 1);
 265 }
 266
 267 static inline uint64_t
 268 relocation_target(struct drm_i915_gem_relocation_entry *reloc,
 269                   uint64_t target_offset)
 270 {
 271         return gen8_canonical_addr((int)reloc->delta + target_offset);
 272 }
 273
 274 static int
 275 relocate_entry_cpu(struct drm_i915_gem_object *obj,
 276                    struct drm_i915_gem_relocation_entry *reloc,
 277                    uint64_t target_offset)
 278 {
 279         struct drm_device *dev = obj->base.dev;
 280         uint32_t page_offset = offset_in_page(reloc->offset);
 281         uint64_t delta = relocation_target(reloc, target_offset);
 282         char *vaddr;
 283         int ret;
 284
 285         ret = i915_gem_object_set_to_cpu_domain(obj, true);
 286         if (ret)
 287                 return ret;
 288
 289         vaddr = kmap_atomic(i915_gem_object_get_dirty_page(obj,
 290                                 reloc->offset >> PAGE_SHIFT));
 291         *(uint32_t *)(vaddr + page_offset) = lower_32_bits(delta);
 292
 293         if (INTEL_INFO(dev)->gen >= 8) {
 294                 page_offset = offset_in_page(page_offset + sizeof(uint32_t));
 295
 296                 if (page_offset == 0) {
 297                         kunmap_atomic(vaddr);
 298                         vaddr = kmap_atomic(i915_gem_object_get_dirty_page(obj,
 299                             (reloc->offset + sizeof(uint32_t)) >> PAGE_SHIFT));
 300                 }
 301
 302                 *(uint32_t *)(vaddr + page_offset) = upper_32_bits(delta);
 303         }
 304
 305         kunmap_atomic(vaddr);
 306
 307         return 0;
 308 }
 309
 310 static int
 311 relocate_entry_gtt(struct drm_i915_gem_object *obj,
 312                    struct drm_i915_gem_relocation_entry *reloc,
 313                    uint64_t target_offset)
 314 {
 315         struct drm_device *dev = obj->base.dev;
 316         struct drm_i915_private *dev_priv = dev->dev_private;
 317         uint64_t delta = relocation_target(reloc, target_offset);
 318         uint64_t offset;
 319         void __iomem *reloc_page;
 320         int ret;
 321
 322         ret = i915_gem_object_set_to_gtt_domain(obj, true);
 323         if (ret)
 324                 return ret;
 325
 326         ret = i915_gem_object_put_fence(obj);
 327         if (ret)
 328                 return ret;
 329
 330         /* Map the page containing the relocation we're going to perform.  */
 331         offset = i915_gem_obj_ggtt_offset(obj);
 332         offset += reloc->offset;
 333         reloc_page = io_mapping_map_atomic_wc(dev_priv->gtt.mappable,
 334                                               offset & PAGE_MASK);
 335         iowrite32(lower_32_bits(delta), reloc_page + offset_in_page(offset));
 336
 337         if (INTEL_INFO(dev)->gen >= 8) {
 338                 offset += sizeof(uint32_t);
 339
 340                 if (offset_in_page(offset) == 0) {
 341                         io_mapping_unmap_atomic(reloc_page);
 342                         reloc_page =
 343                                 io_mapping_map_atomic_wc(dev_priv->gtt.mappable,
 344                                                          offset);
 345                 }
 346
 347                 iowrite32(upper_32_bits(delta),
 348                           reloc_page + offset_in_page(offset));
 349         }
 350
 351         io_mapping_unmap_atomic(reloc_page);
 352
 353         return 0;
 354 }
 355
 356 static void
 357 clflush_write32(void *addr, uint32_t value)
 358 {
 359         /* This is not a fast path, so KISS. */
 360         drm_clflush_virt_range(addr, sizeof(uint32_t));
 361         *(uint32_t *)addr = value;
 362         drm_clflush_virt_range(addr, sizeof(uint32_t));
 363 }
 364
 365 static int
 366 relocate_entry_clflush(struct drm_i915_gem_object *obj,
 367                        struct drm_i915_gem_relocation_entry *reloc,
 368                        uint64_t target_offset)
 369 {
 370         struct drm_device *dev = obj->base.dev;
 371         uint32_t page_offset = offset_in_page(reloc->offset);
 372         uint64_t delta = relocation_target(reloc, target_offset);
 373         char *vaddr;
 374         int ret;
 375
 376         ret = i915_gem_object_set_to_gtt_domain(obj, true);
 377         if (ret)
 378                 return ret;
 379
 380         vaddr = kmap_atomic(i915_gem_object_get_dirty_page(obj,
 381                                 reloc->offset >> PAGE_SHIFT));
 382         clflush_write32(vaddr + page_offset, lower_32_bits(delta));
 383
 384         if (INTEL_INFO(dev)->gen >= 8) {
 385                 page_offset = offset_in_page(page_offset + sizeof(uint32_t));
 386
 387                 if (page_offset == 0) {
 388                         kunmap_atomic(vaddr);
 389                         vaddr = kmap_atomic(i915_gem_object_get_dirty_page(obj,
 390                             (reloc->offset + sizeof(uint32_t)) >> PAGE_SHIFT));
 391                 }
 392
 393                 clflush_write32(vaddr + page_offset, upper_32_bits(delta));
 394         }
 395
 396         kunmap_atomic(vaddr);
 397
 398         return 0;
 399 }
 400
 401 static int
 402 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
 403                                    struct eb_vmas *eb,
 404                                    struct drm_i915_gem_relocation_entry *reloc)
 405 {
 406         struct drm_device *dev = obj->base.dev;
 407         struct drm_gem_object *target_obj;
 408         struct drm_i915_gem_object *target_i915_obj;
 409         struct i915_vma *target_vma;
 410         uint64_t target_offset;
 411         int ret;
 412
 413         /* we've already hold a reference to all valid objects */
 414         target_vma = eb_get_vma(eb, reloc->target_handle);
 415         if (unlikely(target_vma == NULL))
 416                 return -ENOENT;
 417         target_i915_obj = target_vma->obj;
 418         target_obj = &target_vma->obj->base;
 419
 420         target_offset = gen8_canonical_addr(target_vma->node.start);
 421
 422         /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
 423          * pipe_control writes because the gpu doesn't properly redirect them
 424          * through the ppgtt for non_secure batchbuffers. */
 425         if (unlikely(IS_GEN6(dev) &&
 426             reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION)) {
 427                 ret = i915_vma_bind(target_vma, target_i915_obj->cache_level,
 428                                     PIN_GLOBAL);
 429                 if (WARN_ONCE(ret, "Unexpected failure to bind target VMA!"))
 430                         return ret;
 431         }
 432
 433         /* Validate that the target is in a valid r/w GPU domain */
 434         if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
 435                 DRM_DEBUG("reloc with multiple write domains: "
 436                           "obj %p target %d offset %d "
 437                           "read %08x write %08x",
 438                           obj, reloc->target_handle,
 439                           (int) reloc->offset,
 440                           reloc->read_domains,
 441                           reloc->write_domain);
 442                 return -EINVAL;
 443         }
 444         if (unlikely((reloc->write_domain | reloc->read_domains)
 445                      & ~I915_GEM_GPU_DOMAINS)) {
 446                 DRM_DEBUG("reloc with read/write non-GPU domains: "
 447                           "obj %p target %d offset %d "
 448                           "read %08x write %08x",
 449                           obj, reloc->target_handle,
 450                           (int) reloc->offset,
 451                           reloc->read_domains,
 452                           reloc->write_domain);
 453                 return -EINVAL;
 454         }
 455
 456         target_obj->pending_read_domains |= reloc->read_domains;
 457         target_obj->pending_write_domain |= reloc->write_domain;
 458
 459         /* If the relocation already has the right value in it, no
 460          * more work needs to be done.
 461          */
 462         if (target_offset == reloc->presumed_offset)
 463                 return 0;
 464
 465         /* Check that the relocation address is valid... */
 466         if (unlikely(reloc->offset >
 467                 obj->base.size - (INTEL_INFO(dev)->gen >= 8 ? 8 : 4))) {
 468                 DRM_DEBUG("Relocation beyond object bounds: "
 469                           "obj %p target %d offset %d size %d.\n",
 470                           obj, reloc->target_handle,
 471                           (int) reloc->offset,
 472                           (int) obj->base.size);
 473                 return -EINVAL;
 474         }
 475         if (unlikely(reloc->offset & 3)) {
 476                 DRM_DEBUG("Relocation not 4-byte aligned: "
 477                           "obj %p target %d offset %d.\n",
 478                           obj, reloc->target_handle,
 479                           (int) reloc->offset);
 480                 return -EINVAL;
 481         }
 482
 483         /* We can't wait for rendering with pagefaults disabled */
 484         if (obj->active && pagefault_disabled())
 485                 return -EFAULT;
 486
 487         if (use_cpu_reloc(obj))
 488                 ret = relocate_entry_cpu(obj, reloc, target_offset);
 489         else if (obj->map_and_fenceable)
 490                 ret = relocate_entry_gtt(obj, reloc, target_offset);
 491         else if (cpu_has_clflush)
 492                 ret = relocate_entry_clflush(obj, reloc, target_offset);
 493         else {
 494                 WARN_ONCE(1, "Impossible case in relocation handling\n");
 495                 ret = -ENODEV;
 496         }
 497
 498         if (ret)
 499                 return ret;
 500
 501         /* and update the user's relocation entry */
 502         reloc->presumed_offset = target_offset;
 503
 504         return 0;
 505 }
 506
 507 static int
 508 i915_gem_execbuffer_relocate_vma(struct i915_vma *vma,
 509                                  struct eb_vmas *eb)
 510 {
 511 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
 512         struct drm_i915_gem_relocation_entry stack_reloc[N_RELOC(512)];
 513         struct drm_i915_gem_relocation_entry __user *user_relocs;
 514         struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 515         int remain, ret;
 516
 517         user_relocs = to_user_ptr(entry->relocs_ptr);
 518
 519         remain = entry->relocation_count;
 520         while (remain) {
 521                 struct drm_i915_gem_relocation_entry *r = stack_reloc;
 522                 int count = remain;
 523                 if (count > ARRAY_SIZE(stack_reloc))
 524                         count = ARRAY_SIZE(stack_reloc);
 525                 remain -= count;
 526
 527                 if (__copy_from_user_inatomic(r, user_relocs, count*sizeof(r[0])))
 528                         return -EFAULT;
 529
 530                 do {
 531                         u64 offset = r->presumed_offset;
 532
 533                         ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, r);
 534                         if (ret)
 535                                 return ret;
 536
 537                         if (r->presumed_offset != offset &&
 538                             __copy_to_user_inatomic(&user_relocs->presumed_offset,
 539                                                     &r->presumed_offset,
 540                                                     sizeof(r->presumed_offset))) {
 541                                 return -EFAULT;
 542                         }
 543
 544                         user_relocs++;
 545                         r++;
 546                 } while (--count);
 547         }
 548
 549         return 0;
 550 #undef N_RELOC
 551 }
 552
 553 static int
 554 i915_gem_execbuffer_relocate_vma_slow(struct i915_vma *vma,
 555                                       struct eb_vmas *eb,
 556                                       struct drm_i915_gem_relocation_entry *relocs)
 557 {
 558         const struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 559         int i, ret;
 560
 561         for (i = 0; i < entry->relocation_count; i++) {
 562                 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, &relocs[i]);
 563                 if (ret)
 564                         return ret;
 565         }
 566
 567         return 0;
 568 }
 569
 570 static int
 571 i915_gem_execbuffer_relocate(struct eb_vmas *eb)
 572 {
 573         struct i915_vma *vma;
 574         int ret = 0;
 575
 576         /* This is the fast path and we cannot handle a pagefault whilst
 577          * holding the struct mutex lest the user pass in the relocations
 578          * contained within a mmaped bo. For in such a case we, the page
 579          * fault handler would call i915_gem_fault() and we would try to
 580          * acquire the struct mutex again. Obviously this is bad and so
 581          * lockdep complains vehemently.
 582          */
 583         pagefault_disable();
 584         list_for_each_entry(vma, &eb->vmas, exec_list) {
 585                 ret = i915_gem_execbuffer_relocate_vma(vma, eb);
 586                 if (ret)
 587                         break;
 588         }
 589         pagefault_enable();
 590
 591         return ret;
 592 }
 593
 594 static bool only_mappable_for_reloc(unsigned int flags)
 595 {
 596         return (flags & (EXEC_OBJECT_NEEDS_FENCE | __EXEC_OBJECT_NEEDS_MAP)) ==
 597                 __EXEC_OBJECT_NEEDS_MAP;
 598 }
 599
 600 static int
 601 i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,
 602                                 struct intel_engine_cs *ring,
 603                                 bool *need_reloc)
 604 {
 605         struct drm_i915_gem_object *obj = vma->obj;
 606         struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 607         uint64_t flags;
 608         int ret;
 609
 610         flags = PIN_USER;
 611         if (entry->flags & EXEC_OBJECT_NEEDS_GTT)
 612                 flags |= PIN_GLOBAL;
 613
 614         if (!drm_mm_node_allocated(&vma->node)) {
 615                 /* Wa32bitGeneralStateOffset & Wa32bitInstructionBaseOffset,
 616                  * limit address to the first 4GBs for unflagged objects.
 617                  */
 618                 if ((entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) == 0)
 619                         flags |= PIN_ZONE_4G;
 620                 if (entry->flags & __EXEC_OBJECT_NEEDS_MAP)
 621                         flags |= PIN_GLOBAL | PIN_MAPPABLE;
 622                 if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS)
 623                         flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS;
 624                 if (entry->flags & EXEC_OBJECT_PINNED)
 625                         flags |= entry->offset | PIN_OFFSET_FIXED;
 626                 if ((flags & PIN_MAPPABLE) == 0)
 627                         flags |= PIN_HIGH;
 628         }
 629
 630         ret = i915_gem_object_pin(obj, vma->vm, entry->alignment, flags);
 631         if ((ret == -ENOSPC  || ret == -E2BIG) &&
 632             only_mappable_for_reloc(entry->flags))
 633                 ret = i915_gem_object_pin(obj, vma->vm,
 634                                           entry->alignment,
 635                                           flags & ~PIN_MAPPABLE);
 636         if (ret)
 637                 return ret;
 638
 639         entry->flags |= __EXEC_OBJECT_HAS_PIN;
 640
 641         if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
 642                 ret = i915_gem_object_get_fence(obj);
 643                 if (ret)
 644                         return ret;
 645
 646                 if (i915_gem_object_pin_fence(obj))
 647                         entry->flags |= __EXEC_OBJECT_HAS_FENCE;
 648         }
 649
 650         if (entry->offset != vma->node.start) {
 651                 entry->offset = vma->node.start;
 652                 *need_reloc = true;
 653         }
 654
 655         if (entry->flags & EXEC_OBJECT_WRITE) {
 656                 obj->base.pending_read_domains = I915_GEM_DOMAIN_RENDER;
 657                 obj->base.pending_write_domain = I915_GEM_DOMAIN_RENDER;
 658         }
 659
 660         return 0;
 661 }
 662
 663 static bool
 664 need_reloc_mappable(struct i915_vma *vma)
 665 {
 666         struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 667
 668         if (entry->relocation_count == 0)
 669                 return false;
 670
 671         if (!i915_is_ggtt(vma->vm))
 672                 return false;
 673
 674         /* See also use_cpu_reloc() */
 675         if (HAS_LLC(vma->obj->base.dev))
 676                 return false;
 677
 678         if (vma->obj->base.write_domain == I915_GEM_DOMAIN_CPU)
 679                 return false;
 680
 681         return true;
 682 }
 683
 684 static bool
 685 eb_vma_misplaced(struct i915_vma *vma)
 686 {
 687         struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 688         struct drm_i915_gem_object *obj = vma->obj;
 689
 690         WARN_ON(entry->flags & __EXEC_OBJECT_NEEDS_MAP &&
 691                !i915_is_ggtt(vma->vm));
 692
 693         if (entry->alignment &&
 694             vma->node.start & (entry->alignment - 1))
 695                 return true;
 696
 697         if (entry->flags & EXEC_OBJECT_PINNED &&
 698             vma->node.start != entry->offset)
 699                 return true;
 700
 701         if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS &&
 702             vma->node.start < BATCH_OFFSET_BIAS)
 703                 return true;
 704
 705         /* avoid costly ping-pong once a batch bo ended up non-mappable */
 706         if (entry->flags & __EXEC_OBJECT_NEEDS_MAP && !obj->map_and_fenceable)
 707                 return !only_mappable_for_reloc(entry->flags);
 708
 709         if ((entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) == 0 &&
 710             (vma->node.start + vma->node.size - 1) >> 32)
 711                 return true;
 712
 713         return false;
 714 }
 715
 716 static int
 717 i915_gem_execbuffer_reserve(struct intel_engine_cs *ring,
 718                             struct list_head *vmas,
 719                             struct intel_context *ctx,
 720                             bool *need_relocs)
 721 {
 722         struct drm_i915_gem_object *obj;
 723         struct i915_vma *vma;
 724         struct i915_address_space *vm;
 725         struct list_head ordered_vmas;
 726         struct list_head pinned_vmas;
 727         bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
 728         int retry;
 729
 730         i915_gem_retire_requests_ring(ring);
 731
 732         vm = list_first_entry(vmas, struct i915_vma, exec_list)->vm;
 733
 734         INIT_LIST_HEAD(&ordered_vmas);
 735         INIT_LIST_HEAD(&pinned_vmas);
 736         while (!list_empty(vmas)) {
 737                 struct drm_i915_gem_exec_object2 *entry;
 738                 bool need_fence, need_mappable;
 739
 740                 vma = list_first_entry(vmas, struct i915_vma, exec_list);
 741                 obj = vma->obj;
 742                 entry = vma->exec_entry;
 743
 744                 if (ctx->flags & CONTEXT_NO_ZEROMAP)
 745                         entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
 746
 747                 if (!has_fenced_gpu_access)
 748                         entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE;
 749                 need_fence =
 750                         entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
 751                         obj->tiling_mode != I915_TILING_NONE;
 752                 need_mappable = need_fence || need_reloc_mappable(vma);
 753
 754                 if (entry->flags & EXEC_OBJECT_PINNED)
 755                         list_move_tail(&vma->exec_list, &pinned_vmas);
 756                 else if (need_mappable) {
 757                         entry->flags |= __EXEC_OBJECT_NEEDS_MAP;
 758                         list_move(&vma->exec_list, &ordered_vmas);
 759                 } else
 760                         list_move_tail(&vma->exec_list, &ordered_vmas);
 761
 762                 obj->base.pending_read_domains = I915_GEM_GPU_DOMAINS & ~I915_GEM_DOMAIN_COMMAND;
 763                 obj->base.pending_write_domain = 0;
 764         }
 765         list_splice(&ordered_vmas, vmas);
 766         list_splice(&pinned_vmas, vmas);
 767
 768         /* Attempt to pin all of the buffers into the GTT.
 769          * This is done in 3 phases:
 770          *
 771          * 1a. Unbind all objects that do not match the GTT constraints for
 772          *     the execbuffer (fenceable, mappable, alignment etc).
 773          * 1b. Increment pin count for already bound objects.
 774          * 2.  Bind new objects.
 775          * 3.  Decrement pin count.
 776          *
 777          * This avoid unnecessary unbinding of later objects in order to make
 778          * room for the earlier objects *unless* we need to defragment.
 779          */
 780         retry = 0;
 781         do {
 782                 int ret = 0;
 783
 784                 /* Unbind any ill-fitting objects or pin. */
 785                 list_for_each_entry(vma, vmas, exec_list) {
 786                         if (!drm_mm_node_allocated(&vma->node))
 787                                 continue;
 788
 789                         if (eb_vma_misplaced(vma))
 790                                 ret = i915_vma_unbind(vma);
 791                         else
 792                                 ret = i915_gem_execbuffer_reserve_vma(vma, ring, need_relocs);
 793                         if (ret)
 794                                 goto err;
 795                 }
 796
 797                 /* Bind fresh objects */
 798                 list_for_each_entry(vma, vmas, exec_list) {
 799                         if (drm_mm_node_allocated(&vma->node))
 800                                 continue;
 801
 802                         ret = i915_gem_execbuffer_reserve_vma(vma, ring, need_relocs);
 803                         if (ret)
 804                                 goto err;
 805                 }
 806
 807 err:
 808                 if (ret != -ENOSPC || retry++)
 809                         return ret;
 810
 811                 /* Decrement pin count for bound objects */
 812                 list_for_each_entry(vma, vmas, exec_list)
 813                         i915_gem_execbuffer_unreserve_vma(vma);
 814
 815                 ret = i915_gem_evict_vm(vm, true);
 816                 if (ret)
 817                         return ret;
 818         } while (1);
 819 }
 820
 821 static int
 822 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
 823                                   struct drm_i915_gem_execbuffer2 *args,
 824                                   struct drm_file *file,
 825                                   struct intel_engine_cs *ring,
 826                                   struct eb_vmas *eb,
 827                                   struct drm_i915_gem_exec_object2 *exec,
 828                                   struct intel_context *ctx)
 829 {
 830         struct drm_i915_gem_relocation_entry *reloc;
 831         struct i915_address_space *vm;
 832         struct i915_vma *vma;
 833         bool need_relocs;
 834         int *reloc_offset;
 835         int i, total, ret;
 836         unsigned count = args->buffer_count;
 837
 838         vm = list_first_entry(&eb->vmas, struct i915_vma, exec_list)->vm;
 839
 840         /* We may process another execbuffer during the unlock... */
 841         while (!list_empty(&eb->vmas)) {
 842                 vma = list_first_entry(&eb->vmas, struct i915_vma, exec_list);
 843                 list_del_init(&vma->exec_list);
 844                 i915_gem_execbuffer_unreserve_vma(vma);
 845                 drm_gem_object_unreference(&vma->obj->base);
 846         }
 847
 848         mutex_unlock(&dev->struct_mutex);
 849
 850         total = 0;
 851         for (i = 0; i < count; i++)
 852                 total += exec[i].relocation_count;
 853
 854         reloc_offset = drm_malloc_ab(count, sizeof(*reloc_offset));
 855         reloc = drm_malloc_ab(total, sizeof(*reloc));
 856         if (reloc == NULL || reloc_offset == NULL) {
 857                 drm_free_large(reloc);
 858                 drm_free_large(reloc_offset);
 859                 mutex_lock(&dev->struct_mutex);
 860                 return -ENOMEM;
 861         }
 862
 863         total = 0;
 864         for (i = 0; i < count; i++) {
 865                 struct drm_i915_gem_relocation_entry __user *user_relocs;
 866                 u64 invalid_offset = (u64)-1;
 867                 int j;
 868
 869                 user_relocs = to_user_ptr(exec[i].relocs_ptr);
 870
 871                 if (copy_from_user(reloc+total, user_relocs,
 872                                    exec[i].relocation_count * sizeof(*reloc))) {
 873                         ret = -EFAULT;
 874                         mutex_lock(&dev->struct_mutex);
 875                         goto err;
 876                 }
 877
 878                 /* As we do not update the known relocation offsets after
 879                  * relocating (due to the complexities in lock handling),
 880                  * we need to mark them as invalid now so that we force the
 881                  * relocation processing next time. Just in case the target
 882                  * object is evicted and then rebound into its old
 883                  * presumed_offset before the next execbuffer - if that
 884                  * happened we would make the mistake of assuming that the
 885                  * relocations were valid.
 886                  */
 887                 for (j = 0; j < exec[i].relocation_count; j++) {
 888                         if (__copy_to_user(&user_relocs[j].presumed_offset,
 889                                            &invalid_offset,
 890                                            sizeof(invalid_offset))) {
 891                                 ret = -EFAULT;
 892                                 mutex_lock(&dev->struct_mutex);
 893                                 goto err;
 894                         }
 895                 }
 896
 897                 reloc_offset[i] = total;
 898                 total += exec[i].relocation_count;
 899         }
 900
 901         ret = i915_mutex_lock_interruptible(dev);
 902         if (ret) {
 903                 mutex_lock(&dev->struct_mutex);
 904                 goto err;
 905         }
 906
 907         /* reacquire the objects */
 908         eb_reset(eb);
 909         ret = eb_lookup_vmas(eb, exec, args, vm, file);
 910         if (ret)
 911                 goto err;
 912
 913         need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
 914         ret = i915_gem_execbuffer_reserve(ring, &eb->vmas, ctx, &need_relocs);
 915         if (ret)
 916                 goto err;
 917
 918         list_for_each_entry(vma, &eb->vmas, exec_list) {
 919                 int offset = vma->exec_entry - exec;
 920                 ret = i915_gem_execbuffer_relocate_vma_slow(vma, eb,
 921                                                             reloc + reloc_offset[offset]);
 922                 if (ret)
 923                         goto err;
 924         }
 925
 926         /* Leave the user relocations as are, this is the painfully slow path,
 927          * and we want to avoid the complication of dropping the lock whilst
 928          * having buffers reserved in the aperture and so causing spurious
 929          * ENOSPC for random operations.
 930          */
 931
 932 err:
 933         drm_free_large(reloc);
 934         drm_free_large(reloc_offset);
 935         return ret;
 936 }
 937
 938 static int
 939 i915_gem_execbuffer_move_to_gpu(struct drm_i915_gem_request *req,
 940                                 struct list_head *vmas)
 941 {
 942         const unsigned other_rings = ~intel_ring_flag(req->ring);
 943         struct i915_vma *vma;
 944         uint32_t flush_domains = 0;
 945         bool flush_chipset = false;
 946         int ret;
 947
 948         list_for_each_entry(vma, vmas, exec_list) {
 949                 struct drm_i915_gem_object *obj = vma->obj;
 950
 951                 if (obj->active & other_rings) {
 952                         ret = i915_gem_object_sync(obj, req->ring, &req);
 953                         if (ret)
 954                                 return ret;
 955                 }
 956
 957                 if (obj->base.write_domain & I915_GEM_DOMAIN_CPU)
 958                         flush_chipset |= i915_gem_clflush_object(obj, false);
 959
 960                 flush_domains |= obj->base.write_domain;
 961         }
 962
 963         if (flush_chipset)
 964                 i915_gem_chipset_flush(req->ring->dev);
 965
 966         if (flush_domains & I915_GEM_DOMAIN_GTT)
 967                 wmb();
 968
 969         /* Unconditionally invalidate gpu caches and ensure that we do flush
 970          * any residual writes from the previous batch.
 971          */
 972         return intel_ring_invalidate_all_caches(req);
 973 }
 974
 975 static bool
 976 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
 977 {
 978         if (exec->flags & __I915_EXEC_UNKNOWN_FLAGS)
 979                 return false;
 980
 981         /* Kernel clipping was a DRI1 misfeature */
 982         if (exec->num_cliprects || exec->cliprects_ptr)
 983                 return false;
 984
 985         if (exec->DR4 == 0xffffffff) {
 986                 DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
 987                 exec->DR4 = 0;
 988         }
 989         if (exec->DR1 || exec->DR4)
 990                 return false;
 991
 992         if ((exec->batch_start_offset | exec->batch_len) & 0x7)
 993                 return false;
 994
 995         return true;
 996 }
 997
 998 static int
 999 validate_exec_list(struct drm_device *dev,
1000                    struct drm_i915_gem_exec_object2 *exec,
1001                    int count)
1002 {
1003         unsigned relocs_total = 0;
1004         unsigned relocs_max = UINT_MAX / sizeof(struct drm_i915_gem_relocation_entry);
1005         unsigned invalid_flags;
1006         int i;
1007
1008         invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS;
1009         if (USES_FULL_PPGTT(dev))
1010                 invalid_flags |= EXEC_OBJECT_NEEDS_GTT;
1011
1012         for (i = 0; i < count; i++) {
1013                 char __user *ptr = to_user_ptr(exec[i].relocs_ptr);
1014                 int length; /* limited by fault_in_pages_readable() */
1015
1016                 if (exec[i].flags & invalid_flags)
1017                         return -EINVAL;
1018
1019                 /* Offset can be used as input (EXEC_OBJECT_PINNED), reject
1020                  * any non-page-aligned or non-canonical addresses.
1021                  */
1022                 if (exec[i].flags & EXEC_OBJECT_PINNED) {
1023                         if (exec[i].offset !=
1024                             gen8_canonical_addr(exec[i].offset & PAGE_MASK))
1025                                 return -EINVAL;
1026
1027                         /* From drm_mm perspective address space is continuous,
1028                          * so from this point we're always using non-canonical
1029                          * form internally.
1030                          */
1031                         exec[i].offset = gen8_noncanonical_addr(exec[i].offset);
1032                 }
1033
1034                 if (exec[i].alignment && !is_power_of_2(exec[i].alignment))
1035                         return -EINVAL;
1036
1037                 /* First check for malicious input causing overflow in
1038                  * the worst case where we need to allocate the entire
1039                  * relocation tree as a single array.
1040                  */
1041                 if (exec[i].relocation_count > relocs_max - relocs_total)
1042                         return -EINVAL;
1043                 relocs_total += exec[i].relocation_count;
1044
1045                 length = exec[i].relocation_count *
1046                         sizeof(struct drm_i915_gem_relocation_entry);
1047                 /*
1048                  * We must check that the entire relocation array is safe
1049                  * to read, but since we may need to update the presumed
1050                  * offsets during execution, check for full write access.
1051                  */
1052                 if (!access_ok(VERIFY_WRITE, ptr, length))
1053                         return -EFAULT;
1054
1055                 if (likely(!i915.prefault_disable)) {
1056                         if (fault_in_multipages_readable(ptr, length))
1057                                 return -EFAULT;
1058                 }
1059         }
1060
1061         return 0;
1062 }
1063
1064 static struct intel_context *
1065 i915_gem_validate_context(struct drm_device *dev, struct drm_file *file,
1066                           struct intel_engine_cs *ring, const u32 ctx_id)
1067 {
1068         struct intel_context *ctx = NULL;
1069         struct i915_ctx_hang_stats *hs;
1070
1071         if (ring->id != RCS && ctx_id != DEFAULT_CONTEXT_HANDLE)
1072                 return ERR_PTR(-EINVAL);
1073
1074         ctx = i915_gem_context_get(file->driver_priv, ctx_id);
1075         if (IS_ERR(ctx))
1076                 return ctx;
1077
1078         hs = &ctx->hang_stats;
1079         if (hs->banned) {
1080                 DRM_DEBUG("Context %u tried to submit while banned\n", ctx_id);
1081                 return ERR_PTR(-EIO);
1082         }
1083
1084         if (i915.enable_execlists && !ctx->engine[ring->id].state) {
1085                 int ret = intel_lr_context_deferred_alloc(ctx, ring);
1086                 if (ret) {
1087                         DRM_DEBUG("Could not create LRC %u: %d\n", ctx_id, ret);
1088                         return ERR_PTR(ret);
1089                 }
1090         }
1091
1092         return ctx;
1093 }
1094
1095 void
1096 i915_gem_execbuffer_move_to_active(struct list_head *vmas,
1097                                    struct drm_i915_gem_request *req)
1098 {
1099         struct intel_engine_cs *ring = i915_gem_request_get_ring(req);
1100         struct i915_vma *vma;
1101
1102         list_for_each_entry(vma, vmas, exec_list) {
1103                 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
1104                 struct drm_i915_gem_object *obj = vma->obj;
1105                 u32 old_read = obj->base.read_domains;
1106                 u32 old_write = obj->base.write_domain;
1107
1108                 obj->dirty = 1; /* be paranoid  */
1109                 obj->base.write_domain = obj->base.pending_write_domain;
1110                 if (obj->base.write_domain == 0)
1111                         obj->base.pending_read_domains |= obj->base.read_domains;
1112                 obj->base.read_domains = obj->base.pending_read_domains;
1113
1114                 i915_vma_move_to_active(vma, req);
1115                 if (obj->base.write_domain) {
1116                         i915_gem_request_assign(&obj->last_write_req, req);
1117
1118                         intel_fb_obj_invalidate(obj, ORIGIN_CS);
1119
1120                         /* update for the implicit flush after a batch */
1121                         obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
1122                 }
1123                 if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
1124                         i915_gem_request_assign(&obj->last_fenced_req, req);
1125                         if (entry->flags & __EXEC_OBJECT_HAS_FENCE) {
1126                                 struct drm_i915_private *dev_priv = to_i915(ring->dev);
1127                                 list_move_tail(&dev_priv->fence_regs[obj->fence_reg].lru_list,
1128                                                &dev_priv->mm.fence_list);
1129                         }
1130                 }
1131
1132                 trace_i915_gem_object_change_domain(obj, old_read, old_write);
1133         }
1134 }
1135
1136 void
1137 i915_gem_execbuffer_retire_commands(struct i915_execbuffer_params *params)
1138 {
1139         /* Unconditionally force add_request to emit a full flush. */
1140         params->ring->gpu_caches_dirty = true;
1141
1142         /* Add a breadcrumb for the completion of the batch buffer */
1143         __i915_add_request(params->request, params->batch_obj, true);
1144 }
1145
1146 static int
1147 i915_reset_gen7_sol_offsets(struct drm_device *dev,
1148                             struct drm_i915_gem_request *req)
1149 {
1150         struct intel_engine_cs *ring = req->ring;
1151         struct drm_i915_private *dev_priv = dev->dev_private;
1152         int ret, i;
1153
1154         if (!IS_GEN7(dev) || ring != &dev_priv->ring[RCS]) {
1155                 DRM_DEBUG("sol reset is gen7/rcs only\n");
1156                 return -EINVAL;
1157         }
1158
1159         ret = intel_ring_begin(req, 4 * 3);
1160         if (ret)
1161                 return ret;
1162
1163         for (i = 0; i < 4; i++) {
1164                 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1165                 intel_ring_emit_reg(ring, GEN7_SO_WRITE_OFFSET(i));
1166                 intel_ring_emit(ring, 0);
1167         }
1168
1169         intel_ring_advance(ring);
1170
1171         return 0;
1172 }
1173
1174 static struct drm_i915_gem_object*
1175 i915_gem_execbuffer_parse(struct intel_engine_cs *ring,
1176                           struct drm_i915_gem_exec_object2 *shadow_exec_entry,
1177                           struct eb_vmas *eb,
1178                           struct drm_i915_gem_object *batch_obj,
1179                           u32 batch_start_offset,
1180                           u32 batch_len,
1181                           bool is_master)
1182 {
1183         struct drm_i915_gem_object *shadow_batch_obj;
1184         struct i915_vma *vma;
1185         int ret;
1186
1187         shadow_batch_obj = i915_gem_batch_pool_get(&ring->batch_pool,
1188                                                    PAGE_ALIGN(batch_len));
1189         if (IS_ERR(shadow_batch_obj))
1190                 return shadow_batch_obj;
1191
1192         ret = i915_parse_cmds(ring,
1193                               batch_obj,
1194                               shadow_batch_obj,
1195                               batch_start_offset,
1196                               batch_len,
1197                               is_master);
1198         if (ret)
1199                 goto err;
1200
1201         ret = i915_gem_obj_ggtt_pin(shadow_batch_obj, 0, 0);
1202         if (ret)
1203                 goto err;
1204
1205         i915_gem_object_unpin_pages(shadow_batch_obj);
1206
1207         memset(shadow_exec_entry, 0, sizeof(*shadow_exec_entry));
1208
1209         vma = i915_gem_obj_to_ggtt(shadow_batch_obj);
1210         vma->exec_entry = shadow_exec_entry;
1211         vma->exec_entry->flags = __EXEC_OBJECT_HAS_PIN;
1212         drm_gem_object_reference(&shadow_batch_obj->base);
1213         list_add_tail(&vma->exec_list, &eb->vmas);
1214
1215         shadow_batch_obj->base.pending_read_domains = I915_GEM_DOMAIN_COMMAND;
1216
1217         return shadow_batch_obj;
1218
1219 err:
1220         i915_gem_object_unpin_pages(shadow_batch_obj);
1221         if (ret == -EACCES) /* unhandled chained batch */
1222                 return batch_obj;
1223         else
1224                 return ERR_PTR(ret);
1225 }
1226
1227 int
1228 i915_gem_ringbuffer_submission(struct i915_execbuffer_params *params,
1229                                struct drm_i915_gem_execbuffer2 *args,
1230                                struct list_head *vmas)
1231 {
1232         struct drm_device *dev = params->dev;
1233         struct intel_engine_cs *ring = params->ring;
1234         struct drm_i915_private *dev_priv = dev->dev_private;
1235         u64 exec_start, exec_len;
1236         int instp_mode;
1237         u32 instp_mask;
1238         int ret;
1239
1240         ret = i915_gem_execbuffer_move_to_gpu(params->request, vmas);
1241         if (ret)
1242                 return ret;
1243
1244         ret = i915_switch_context(params->request);
1245         if (ret)
1246                 return ret;
1247
1248         WARN(params->ctx->ppgtt && params->ctx->ppgtt->pd_dirty_rings & (1<<ring->id),
1249              "%s didn't clear reload\n", ring->name);
1250
1251         instp_mode = args->flags & I915_EXEC_CONSTANTS_MASK;
1252         instp_mask = I915_EXEC_CONSTANTS_MASK;
1253         switch (instp_mode) {
1254         case I915_EXEC_CONSTANTS_REL_GENERAL:
1255         case I915_EXEC_CONSTANTS_ABSOLUTE:
1256         case I915_EXEC_CONSTANTS_REL_SURFACE:
1257                 if (instp_mode != 0 && ring != &dev_priv->ring[RCS]) {
1258                         DRM_DEBUG("non-0 rel constants mode on non-RCS\n");
1259                         return -EINVAL;
1260                 }
1261
1262                 if (instp_mode != dev_priv->relative_constants_mode) {
1263                         if (INTEL_INFO(dev)->gen < 4) {
1264                                 DRM_DEBUG("no rel constants on pre-gen4\n");
1265                                 return -EINVAL;
1266                         }
1267
1268                         if (INTEL_INFO(dev)->gen > 5 &&
1269                             instp_mode == I915_EXEC_CONSTANTS_REL_SURFACE) {
1270                                 DRM_DEBUG("rel surface constants mode invalid on gen5+\n");
1271                                 return -EINVAL;
1272                         }
1273
1274                         /* The HW changed the meaning on this bit on gen6 */
1275                         if (INTEL_INFO(dev)->gen >= 6)
1276                                 instp_mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
1277                 }
1278                 break;
1279         default:
1280                 DRM_DEBUG("execbuf with unknown constants: %d\n", instp_mode);
1281                 return -EINVAL;
1282         }
1283
1284         if (ring == &dev_priv->ring[RCS] &&
1285             instp_mode != dev_priv->relative_constants_mode) {
1286                 ret = intel_ring_begin(params->request, 4);
1287                 if (ret)
1288                         return ret;
1289
1290                 intel_ring_emit(ring, MI_NOOP);
1291                 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1292                 intel_ring_emit_reg(ring, INSTPM);
1293                 intel_ring_emit(ring, instp_mask << 16 | instp_mode);
1294                 intel_ring_advance(ring);
1295
1296                 dev_priv->relative_constants_mode = instp_mode;
1297         }
1298
1299         if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
1300                 ret = i915_reset_gen7_sol_offsets(dev, params->request);
1301                 if (ret)
1302                         return ret;
1303         }
1304
1305         exec_len   = args->batch_len;
1306         exec_start = params->batch_obj_vm_offset +
1307                      params->args_batch_start_offset;
1308
1309         if (exec_len == 0)
1310                 exec_len = params->batch_obj->base.size;
1311
1312         ret = ring->dispatch_execbuffer(params->request,
1313                                         exec_start, exec_len,
1314                                         params->dispatch_flags);
1315         if (ret)
1316                 return ret;
1317
1318         trace_i915_gem_ring_dispatch(params->request, params->dispatch_flags);
1319
1320         i915_gem_execbuffer_move_to_active(vmas, params->request);
1321         i915_gem_execbuffer_retire_commands(params);
1322
1323         return 0;
1324 }
1325
1326 /**
1327  * Find one BSD ring to dispatch the corresponding BSD command.
1328  * The ring index is returned.
1329  */
1330 static unsigned int
1331 gen8_dispatch_bsd_ring(struct drm_i915_private *dev_priv, struct drm_file *file)
1332 {
1333         struct drm_i915_file_private *file_priv = file->driver_priv;
1334
1335         /* Check whether the file_priv has already selected one ring. */
1336         if ((int)file_priv->bsd_ring < 0) {
1337                 /* If not, use the ping-pong mechanism to select one. */
1338                 mutex_lock(&dev_priv->dev->struct_mutex);
1339                 file_priv->bsd_ring = dev_priv->mm.bsd_ring_dispatch_index;
1340                 dev_priv->mm.bsd_ring_dispatch_index ^= 1;
1341                 mutex_unlock(&dev_priv->dev->struct_mutex);
1342         }
1343
1344         return file_priv->bsd_ring;
1345 }
1346
1347 static struct drm_i915_gem_object *
1348 eb_get_batch(struct eb_vmas *eb)
1349 {
1350         struct i915_vma *vma = list_entry(eb->vmas.prev, typeof(*vma), exec_list);
1351
1352         /*
1353          * SNA is doing fancy tricks with compressing batch buffers, which leads
1354          * to negative relocation deltas. Usually that works out ok since the
1355          * relocate address is still positive, except when the batch is placed
1356          * very low in the GTT. Ensure this doesn't happen.
1357          *
1358          * Note that actual hangs have only been observed on gen7, but for
1359          * paranoia do it everywhere.
1360          */
1361         if ((vma->exec_entry->flags & EXEC_OBJECT_PINNED) == 0)
1362                 vma->exec_entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
1363
1364         return vma->obj;
1365 }
1366
1367 #define I915_USER_RINGS (4)
1368
1369 static const enum intel_ring_id user_ring_map[I915_USER_RINGS + 1] = {
1370         [I915_EXEC_DEFAULT]     = RCS,
1371         [I915_EXEC_RENDER]      = RCS,
1372         [I915_EXEC_BLT]         = BCS,
1373         [I915_EXEC_BSD]         = VCS,
1374         [I915_EXEC_VEBOX]       = VECS
1375 };
1376
1377 static int
1378 eb_select_ring(struct drm_i915_private *dev_priv,
1379                struct drm_file *file,
1380                struct drm_i915_gem_execbuffer2 *args,
1381                struct intel_engine_cs **ring)
1382 {
1383         unsigned int user_ring_id = args->flags & I915_EXEC_RING_MASK;
1384
1385         if (user_ring_id > I915_USER_RINGS) {
1386                 DRM_DEBUG("execbuf with unknown ring: %u\n", user_ring_id);
1387                 return -EINVAL;
1388         }
1389
1390         if ((user_ring_id != I915_EXEC_BSD) &&
1391             ((args->flags & I915_EXEC_BSD_MASK) != 0)) {
1392                 DRM_DEBUG("execbuf with non bsd ring but with invalid "
1393                           "bsd dispatch flags: %d\n", (int)(args->flags));
1394                 return -EINVAL;
1395         }
1396
1397         if (user_ring_id == I915_EXEC_BSD && HAS_BSD2(dev_priv)) {
1398                 unsigned int bsd_idx = args->flags & I915_EXEC_BSD_MASK;
1399
1400                 if (bsd_idx == I915_EXEC_BSD_DEFAULT) {
1401                         bsd_idx = gen8_dispatch_bsd_ring(dev_priv, file);
1402                 } else if (bsd_idx >= I915_EXEC_BSD_RING1 &&
1403                            bsd_idx <= I915_EXEC_BSD_RING2) {
1404                         bsd_idx >>= I915_EXEC_BSD_SHIFT;
1405                         bsd_idx--;
1406                 } else {
1407                         DRM_DEBUG("execbuf with unknown bsd ring: %u\n",
1408                                   bsd_idx);
1409                         return -EINVAL;
1410                 }
1411
1412                 *ring = &dev_priv->ring[_VCS(bsd_idx)];
1413         } else {
1414                 *ring = &dev_priv->ring[user_ring_map[user_ring_id]];
1415         }
1416
1417         if (!intel_ring_initialized(*ring)) {
1418                 DRM_DEBUG("execbuf with invalid ring: %u\n", user_ring_id);
1419                 return -EINVAL;
1420         }
1421
1422         return 0;
1423 }
1424
1425 static int
1426 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1427                        struct drm_file *file,
1428                        struct drm_i915_gem_execbuffer2 *args,
1429                        struct drm_i915_gem_exec_object2 *exec)
1430 {
1431         struct drm_i915_private *dev_priv = dev->dev_private;
1432         struct drm_i915_gem_request *req = NULL;
1433         struct eb_vmas *eb;
1434         struct drm_i915_gem_object *batch_obj;
1435         struct drm_i915_gem_exec_object2 shadow_exec_entry;
1436         struct intel_engine_cs *ring;
1437         struct intel_context *ctx;
1438         struct i915_address_space *vm;
1439         struct i915_execbuffer_params params_master; /* XXX: will be removed later */
1440         struct i915_execbuffer_params *params = &params_master;
1441         const u32 ctx_id = i915_execbuffer2_get_context_id(*args);
1442         u32 dispatch_flags;
1443         int ret;
1444         bool need_relocs;
1445
1446         if (!i915_gem_check_execbuffer(args))
1447                 return -EINVAL;
1448
1449         ret = validate_exec_list(dev, exec, args->buffer_count);
1450         if (ret)
1451                 return ret;
1452
1453         dispatch_flags = 0;
1454         if (args->flags & I915_EXEC_SECURE) {
1455                 if (!file->is_master || !capable(CAP_SYS_ADMIN))
1456                     return -EPERM;
1457
1458                 dispatch_flags |= I915_DISPATCH_SECURE;
1459         }
1460         if (args->flags & I915_EXEC_IS_PINNED)
1461                 dispatch_flags |= I915_DISPATCH_PINNED;
1462
1463         ret = eb_select_ring(dev_priv, file, args, &ring);
1464         if (ret)
1465                 return ret;
1466
1467         if (args->buffer_count < 1) {
1468                 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1469                 return -EINVAL;
1470         }
1471
1472         if (args->flags & I915_EXEC_RESOURCE_STREAMER) {
1473                 if (!HAS_RESOURCE_STREAMER(dev)) {
1474                         DRM_DEBUG("RS is only allowed for Haswell, Gen8 and above\n");
1475                         return -EINVAL;
1476                 }
1477                 if (ring->id != RCS) {
1478                         DRM_DEBUG("RS is not available on %s\n",
1479                                  ring->name);
1480                         return -EINVAL;
1481                 }
1482
1483                 dispatch_flags |= I915_DISPATCH_RS;
1484         }
1485
1486         intel_runtime_pm_get(dev_priv);
1487
1488         ret = i915_mutex_lock_interruptible(dev);
1489         if (ret)
1490                 goto pre_mutex_err;
1491
1492         ctx = i915_gem_validate_context(dev, file, ring, ctx_id);
1493         if (IS_ERR(ctx)) {
1494                 mutex_unlock(&dev->struct_mutex);
1495                 ret = PTR_ERR(ctx);
1496                 goto pre_mutex_err;
1497         }
1498
1499         i915_gem_context_reference(ctx);
1500
1501         if (ctx->ppgtt)
1502                 vm = &ctx->ppgtt->base;
1503         else
1504                 vm = &dev_priv->gtt.base;
1505
1506         memset(&params_master, 0x00, sizeof(params_master));
1507
1508         eb = eb_create(args);
1509         if (eb == NULL) {
1510                 i915_gem_context_unreference(ctx);
1511                 mutex_unlock(&dev->struct_mutex);
1512                 ret = -ENOMEM;
1513                 goto pre_mutex_err;
1514         }
1515
1516         /* Look up object handles */
1517         ret = eb_lookup_vmas(eb, exec, args, vm, file);
1518         if (ret)
1519                 goto err;
1520
1521         /* take note of the batch buffer before we might reorder the lists */
1522         batch_obj = eb_get_batch(eb);
1523
1524         /* Move the objects en-masse into the GTT, evicting if necessary. */
1525         need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
1526         ret = i915_gem_execbuffer_reserve(ring, &eb->vmas, ctx, &need_relocs);
1527         if (ret)
1528                 goto err;
1529
1530         /* The objects are in their final locations, apply the relocations. */
1531         if (need_relocs)
1532                 ret = i915_gem_execbuffer_relocate(eb);
1533         if (ret) {
1534                 if (ret == -EFAULT) {
1535                         ret = i915_gem_execbuffer_relocate_slow(dev, args, file, ring,
1536                                                                 eb, exec, ctx);
1537                         BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1538                 }
1539                 if (ret)
1540                         goto err;
1541         }
1542
1543         /* Set the pending read domains for the batch buffer to COMMAND */
1544         if (batch_obj->base.pending_write_domain) {
1545                 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1546                 ret = -EINVAL;
1547                 goto err;
1548         }
1549
1550         params->args_batch_start_offset = args->batch_start_offset;
1551         if (i915_needs_cmd_parser(ring) && args->batch_len) {
1552                 struct drm_i915_gem_object *parsed_batch_obj;
1553
1554                 parsed_batch_obj = i915_gem_execbuffer_parse(ring,
1555                                                       &shadow_exec_entry,
1556                                                       eb,
1557                                                       batch_obj,
1558                                                       args->batch_start_offset,
1559                                                       args->batch_len,
1560                                                       file->is_master);
1561                 if (IS_ERR(parsed_batch_obj)) {
1562                         ret = PTR_ERR(parsed_batch_obj);
1563                         goto err;
1564                 }
1565
1566                 /*
1567                  * parsed_batch_obj == batch_obj means batch not fully parsed:
1568                  * Accept, but don't promote to secure.
1569                  */
1570
1571                 if (parsed_batch_obj != batch_obj) {
1572                         /*
1573                          * Batch parsed and accepted:
1574                          *
1575                          * Set the DISPATCH_SECURE bit to remove the NON_SECURE
1576                          * bit from MI_BATCH_BUFFER_START commands issued in
1577                          * the dispatch_execbuffer implementations. We
1578                          * specifically don't want that set on batches the
1579                          * command parser has accepted.
1580                          */
1581                         dispatch_flags |= I915_DISPATCH_SECURE;
1582                         params->args_batch_start_offset = 0;
1583                         batch_obj = parsed_batch_obj;
1584                 }
1585         }
1586
1587         batch_obj->base.pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
1588
1589         /* snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
1590          * batch" bit. Hence we need to pin secure batches into the global gtt.
1591          * hsw should have this fixed, but bdw mucks it up again. */
1592         if (dispatch_flags & I915_DISPATCH_SECURE) {
1593                 /*
1594                  * So on first glance it looks freaky that we pin the batch here
1595                  * outside of the reservation loop. But:
1596                  * - The batch is already pinned into the relevant ppgtt, so we
1597                  *   already have the backing storage fully allocated.
1598                  * - No other BO uses the global gtt (well contexts, but meh),
1599                  *   so we don't really have issues with multiple objects not
1600                  *   fitting due to fragmentation.
1601                  * So this is actually safe.
1602                  */
1603                 ret = i915_gem_obj_ggtt_pin(batch_obj, 0, 0);
1604                 if (ret)
1605                         goto err;
1606
1607                 params->batch_obj_vm_offset = i915_gem_obj_ggtt_offset(batch_obj);
1608         } else
1609                 params->batch_obj_vm_offset = i915_gem_obj_offset(batch_obj, vm);
1610
1611         /* Allocate a request for this batch buffer nice and early. */
1612         req = i915_gem_request_alloc(ring, ctx);
1613         if (IS_ERR(req)) {
1614                 ret = PTR_ERR(req);
1615                 goto err_batch_unpin;
1616         }
1617
1618         ret = i915_gem_request_add_to_client(req, file);
1619         if (ret)
1620                 goto err_batch_unpin;
1621
1622         /*
1623          * Save assorted stuff away to pass through to *_submission().
1624          * NB: This data should be 'persistent' and not local as it will
1625          * kept around beyond the duration of the IOCTL once the GPU
1626          * scheduler arrives.
1627          */
1628         params->dev                     = dev;
1629         params->file                    = file;
1630         params->ring                    = ring;
1631         params->dispatch_flags          = dispatch_flags;
1632         params->batch_obj               = batch_obj;
1633         params->ctx                     = ctx;
1634         params->request                 = req;
1635
1636         ret = dev_priv->gt.execbuf_submit(params, args, &eb->vmas);
1637
1638 err_batch_unpin:
1639         /*
1640          * FIXME: We crucially rely upon the active tracking for the (ppgtt)
1641          * batch vma for correctness. For less ugly and less fragility this
1642          * needs to be adjusted to also track the ggtt batch vma properly as
1643          * active.
1644          */
1645         if (dispatch_flags & I915_DISPATCH_SECURE)
1646                 i915_gem_object_ggtt_unpin(batch_obj);
1647
1648 err:
1649         /* the request owns the ref now */
1650         i915_gem_context_unreference(ctx);
1651         eb_destroy(eb);
1652
1653         /*
1654          * If the request was created but not successfully submitted then it
1655          * must be freed again. If it was submitted then it is being tracked
1656          * on the active request list and no clean up is required here.
1657          */
1658         if (ret && !IS_ERR_OR_NULL(req))
1659                 i915_gem_request_cancel(req);
1660
1661         mutex_unlock(&dev->struct_mutex);
1662
1663 pre_mutex_err:
1664         /* intel_gpu_busy should also get a ref, so it will free when the device
1665          * is really idle. */
1666         intel_runtime_pm_put(dev_priv);
1667         return ret;
1668 }
1669
1670 /*
1671  * Legacy execbuffer just creates an exec2 list from the original exec object
1672  * list array and passes it to the real function.
1673  */
1674 int
1675 i915_gem_execbuffer(struct drm_device *dev, void *data,
1676                     struct drm_file *file)
1677 {
1678         struct drm_i915_gem_execbuffer *args = data;
1679         struct drm_i915_gem_execbuffer2 exec2;
1680         struct drm_i915_gem_exec_object *exec_list = NULL;
1681         struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1682         int ret, i;
1683
1684         if (args->buffer_count < 1) {
1685                 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1686                 return -EINVAL;
1687         }
1688
1689         /* Copy in the exec list from userland */
1690         exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
1691         exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
1692         if (exec_list == NULL || exec2_list == NULL) {
1693                 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1694                           args->buffer_count);
1695                 drm_free_large(exec_list);
1696                 drm_free_large(exec2_list);
1697                 return -ENOMEM;
1698         }
1699         ret = copy_from_user(exec_list,
1700                              to_user_ptr(args->buffers_ptr),
1701                              sizeof(*exec_list) * args->buffer_count);
1702         if (ret != 0) {
1703                 DRM_DEBUG("copy %d exec entries failed %d\n",
1704                           args->buffer_count, ret);
1705                 drm_free_large(exec_list);
1706                 drm_free_large(exec2_list);
1707                 return -EFAULT;
1708         }
1709
1710         for (i = 0; i < args->buffer_count; i++) {
1711                 exec2_list[i].handle = exec_list[i].handle;
1712                 exec2_list[i].relocation_count = exec_list[i].relocation_count;
1713                 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
1714                 exec2_list[i].alignment = exec_list[i].alignment;
1715                 exec2_list[i].offset = exec_list[i].offset;
1716                 if (INTEL_INFO(dev)->gen < 4)
1717                         exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
1718                 else
1719                         exec2_list[i].flags = 0;
1720         }
1721
1722         exec2.buffers_ptr = args->buffers_ptr;
1723         exec2.buffer_count = args->buffer_count;
1724         exec2.batch_start_offset = args->batch_start_offset;
1725         exec2.batch_len = args->batch_len;
1726         exec2.DR1 = args->DR1;
1727         exec2.DR4 = args->DR4;
1728         exec2.num_cliprects = args->num_cliprects;
1729         exec2.cliprects_ptr = args->cliprects_ptr;
1730         exec2.flags = I915_EXEC_RENDER;
1731         i915_execbuffer2_set_context_id(exec2, 0);
1732
1733         ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list);
1734         if (!ret) {
1735                 struct drm_i915_gem_exec_object __user *user_exec_list =
1736                         to_user_ptr(args->buffers_ptr);
1737
1738                 /* Copy the new buffer offsets back to the user's exec list. */
1739                 for (i = 0; i < args->buffer_count; i++) {
1740                         exec2_list[i].offset =
1741                                 gen8_canonical_addr(exec2_list[i].offset);
1742                         ret = __copy_to_user(&user_exec_list[i].offset,
1743                                              &exec2_list[i].offset,
1744                                              sizeof(user_exec_list[i].offset));
1745                         if (ret) {
1746                                 ret = -EFAULT;
1747                                 DRM_DEBUG("failed to copy %d exec entries "
1748                                           "back to user (%d)\n",
1749                                           args->buffer_count, ret);
1750                                 break;
1751                         }
1752                 }
1753         }
1754
1755         drm_free_large(exec_list);
1756         drm_free_large(exec2_list);
1757         return ret;
1758 }
1759
1760 int
1761 i915_gem_execbuffer2(struct drm_device *dev, void *data,
1762                      struct drm_file *file)
1763 {
1764         struct drm_i915_gem_execbuffer2 *args = data;
1765         struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1766         int ret;
1767
1768         if (args->buffer_count < 1 ||
1769             args->buffer_count > UINT_MAX / sizeof(*exec2_list)) {
1770                 DRM_DEBUG("execbuf2 with %d buffers\n", args->buffer_count);
1771                 return -EINVAL;
1772         }
1773
1774         if (args->rsvd2 != 0) {
1775                 DRM_DEBUG("dirty rvsd2 field\n");
1776                 return -EINVAL;
1777         }
1778
1779         exec2_list = kmalloc(sizeof(*exec2_list)*args->buffer_count,
1780                              GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
1781         if (exec2_list == NULL)
1782                 exec2_list = drm_malloc_ab(sizeof(*exec2_list),
1783                                            args->buffer_count);
1784         if (exec2_list == NULL) {
1785                 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1786                           args->buffer_count);
1787                 return -ENOMEM;
1788         }
1789         ret = copy_from_user(exec2_list,
1790                              to_user_ptr(args->buffers_ptr),
1791                              sizeof(*exec2_list) * args->buffer_count);
1792         if (ret != 0) {
1793                 DRM_DEBUG("copy %d exec entries failed %d\n",
1794                           args->buffer_count, ret);
1795                 drm_free_large(exec2_list);
1796                 return -EFAULT;
1797         }
1798
1799         ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list);
1800         if (!ret) {
1801                 /* Copy the new buffer offsets back to the user's exec list. */
1802                 struct drm_i915_gem_exec_object2 __user *user_exec_list =
1803                                    to_user_ptr(args->buffers_ptr);
1804                 int i;
1805
1806                 for (i = 0; i < args->buffer_count; i++) {
1807                         exec2_list[i].offset =
1808                                 gen8_canonical_addr(exec2_list[i].offset);
1809                         ret = __copy_to_user(&user_exec_list[i].offset,
1810                                              &exec2_list[i].offset,
1811                                              sizeof(user_exec_list[i].offset));
1812                         if (ret) {
1813                                 ret = -EFAULT;
1814                                 DRM_DEBUG("failed to copy %d exec entries "
1815                                           "back to user\n",
1816                                           args->buffer_count);
1817                                 break;
1818                         }
1819                 }
1820         }
1821
1822         drm_free_large(exec2_list);
1823         return ret;
1824 }