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 = to_i915(dev);
 317         struct i915_ggtt *ggtt = &dev_priv->ggtt;
 318         uint64_t delta = relocation_target(reloc, target_offset);
 319         uint64_t offset;
 320         void __iomem *reloc_page;
 321         int ret;
 322
 323         ret = i915_gem_object_set_to_gtt_domain(obj, true);
 324         if (ret)
 325                 return ret;
 326
 327         ret = i915_gem_object_put_fence(obj);
 328         if (ret)
 329                 return ret;
 330
 331         /* Map the page containing the relocation we're going to perform.  */
 332         offset = i915_gem_obj_ggtt_offset(obj);
 333         offset += reloc->offset;
 334         reloc_page = io_mapping_map_atomic_wc(ggtt->mappable,
 335                                               offset & PAGE_MASK);
 336         iowrite32(lower_32_bits(delta), reloc_page + offset_in_page(offset));
 337
 338         if (INTEL_INFO(dev)->gen >= 8) {
 339                 offset += sizeof(uint32_t);
 340
 341                 if (offset_in_page(offset) == 0) {
 342                         io_mapping_unmap_atomic(reloc_page);
 343                         reloc_page =
 344                                 io_mapping_map_atomic_wc(ggtt->mappable,
 345                                                          offset);
 346                 }
 347
 348                 iowrite32(upper_32_bits(delta),
 349                           reloc_page + offset_in_page(offset));
 350         }
 351
 352         io_mapping_unmap_atomic(reloc_page);
 353
 354         return 0;
 355 }
 356
 357 static void
 358 clflush_write32(void *addr, uint32_t value)
 359 {
 360         /* This is not a fast path, so KISS. */
 361         drm_clflush_virt_range(addr, sizeof(uint32_t));
 362         *(uint32_t *)addr = value;
 363         drm_clflush_virt_range(addr, sizeof(uint32_t));
 364 }
 365
 366 static int
 367 relocate_entry_clflush(struct drm_i915_gem_object *obj,
 368                        struct drm_i915_gem_relocation_entry *reloc,
 369                        uint64_t target_offset)
 370 {
 371         struct drm_device *dev = obj->base.dev;
 372         uint32_t page_offset = offset_in_page(reloc->offset);
 373         uint64_t delta = relocation_target(reloc, target_offset);
 374         char *vaddr;
 375         int ret;
 376
 377         ret = i915_gem_object_set_to_gtt_domain(obj, true);
 378         if (ret)
 379                 return ret;
 380
 381         vaddr = kmap_atomic(i915_gem_object_get_dirty_page(obj,
 382                                 reloc->offset >> PAGE_SHIFT));
 383         clflush_write32(vaddr + page_offset, lower_32_bits(delta));
 384
 385         if (INTEL_INFO(dev)->gen >= 8) {
 386                 page_offset = offset_in_page(page_offset + sizeof(uint32_t));
 387
 388                 if (page_offset == 0) {
 389                         kunmap_atomic(vaddr);
 390                         vaddr = kmap_atomic(i915_gem_object_get_dirty_page(obj,
 391                             (reloc->offset + sizeof(uint32_t)) >> PAGE_SHIFT));
 392                 }
 393
 394                 clflush_write32(vaddr + page_offset, upper_32_bits(delta));
 395         }
 396
 397         kunmap_atomic(vaddr);
 398
 399         return 0;
 400 }
 401
 402 static int
 403 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
 404                                    struct eb_vmas *eb,
 405                                    struct drm_i915_gem_relocation_entry *reloc)
 406 {
 407         struct drm_device *dev = obj->base.dev;
 408         struct drm_gem_object *target_obj;
 409         struct drm_i915_gem_object *target_i915_obj;
 410         struct i915_vma *target_vma;
 411         uint64_t target_offset;
 412         int ret;
 413
 414         /* we've already hold a reference to all valid objects */
 415         target_vma = eb_get_vma(eb, reloc->target_handle);
 416         if (unlikely(target_vma == NULL))
 417                 return -ENOENT;
 418         target_i915_obj = target_vma->obj;
 419         target_obj = &target_vma->obj->base;
 420
 421         target_offset = gen8_canonical_addr(target_vma->node.start);
 422
 423         /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
 424          * pipe_control writes because the gpu doesn't properly redirect them
 425          * through the ppgtt for non_secure batchbuffers. */
 426         if (unlikely(IS_GEN6(dev) &&
 427             reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION)) {
 428                 ret = i915_vma_bind(target_vma, target_i915_obj->cache_level,
 429                                     PIN_GLOBAL);
 430                 if (WARN_ONCE(ret, "Unexpected failure to bind target VMA!"))
 431                         return ret;
 432         }
 433
 434         /* Validate that the target is in a valid r/w GPU domain */
 435         if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
 436                 DRM_DEBUG("reloc with multiple write domains: "
 437                           "obj %p target %d offset %d "
 438                           "read %08x write %08x",
 439                           obj, reloc->target_handle,
 440                           (int) reloc->offset,
 441                           reloc->read_domains,
 442                           reloc->write_domain);
 443                 return -EINVAL;
 444         }
 445         if (unlikely((reloc->write_domain | reloc->read_domains)
 446                      & ~I915_GEM_GPU_DOMAINS)) {
 447                 DRM_DEBUG("reloc with read/write non-GPU domains: "
 448                           "obj %p target %d offset %d "
 449                           "read %08x write %08x",
 450                           obj, reloc->target_handle,
 451                           (int) reloc->offset,
 452                           reloc->read_domains,
 453                           reloc->write_domain);
 454                 return -EINVAL;
 455         }
 456
 457         target_obj->pending_read_domains |= reloc->read_domains;
 458         target_obj->pending_write_domain |= reloc->write_domain;
 459
 460         /* If the relocation already has the right value in it, no
 461          * more work needs to be done.
 462          */
 463         if (target_offset == reloc->presumed_offset)
 464                 return 0;
 465
 466         /* Check that the relocation address is valid... */
 467         if (unlikely(reloc->offset >
 468                 obj->base.size - (INTEL_INFO(dev)->gen >= 8 ? 8 : 4))) {
 469                 DRM_DEBUG("Relocation beyond object bounds: "
 470                           "obj %p target %d offset %d size %d.\n",
 471                           obj, reloc->target_handle,
 472                           (int) reloc->offset,
 473                           (int) obj->base.size);
 474                 return -EINVAL;
 475         }
 476         if (unlikely(reloc->offset & 3)) {
 477                 DRM_DEBUG("Relocation not 4-byte aligned: "
 478                           "obj %p target %d offset %d.\n",
 479                           obj, reloc->target_handle,
 480                           (int) reloc->offset);
 481                 return -EINVAL;
 482         }
 483
 484         /* We can't wait for rendering with pagefaults disabled */
 485         if (obj->active && pagefault_disabled())
 486                 return -EFAULT;
 487
 488         if (use_cpu_reloc(obj))
 489                 ret = relocate_entry_cpu(obj, reloc, target_offset);
 490         else if (obj->map_and_fenceable)
 491                 ret = relocate_entry_gtt(obj, reloc, target_offset);
 492         else if (static_cpu_has(X86_FEATURE_CLFLUSH))
 493                 ret = relocate_entry_clflush(obj, reloc, target_offset);
 494         else {
 495                 WARN_ONCE(1, "Impossible case in relocation handling\n");
 496                 ret = -ENODEV;
 497         }
 498
 499         if (ret)
 500                 return ret;
 501
 502         /* and update the user's relocation entry */
 503         reloc->presumed_offset = target_offset;
 504
 505         return 0;
 506 }
 507
 508 static int
 509 i915_gem_execbuffer_relocate_vma(struct i915_vma *vma,
 510                                  struct eb_vmas *eb)
 511 {
 512 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
 513         struct drm_i915_gem_relocation_entry stack_reloc[N_RELOC(512)];
 514         struct drm_i915_gem_relocation_entry __user *user_relocs;
 515         struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 516         int remain, ret;
 517
 518         user_relocs = u64_to_user_ptr(entry->relocs_ptr);
 519
 520         remain = entry->relocation_count;
 521         while (remain) {
 522                 struct drm_i915_gem_relocation_entry *r = stack_reloc;
 523                 int count = remain;
 524                 if (count > ARRAY_SIZE(stack_reloc))
 525                         count = ARRAY_SIZE(stack_reloc);
 526                 remain -= count;
 527
 528                 if (__copy_from_user_inatomic(r, user_relocs, count*sizeof(r[0])))
 529                         return -EFAULT;
 530
 531                 do {
 532                         u64 offset = r->presumed_offset;
 533
 534                         ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, r);
 535                         if (ret)
 536                                 return ret;
 537
 538                         if (r->presumed_offset != offset &&
 539                             __put_user(r->presumed_offset, &user_relocs->presumed_offset)) {
 540                                 return -EFAULT;
 541                         }
 542
 543                         user_relocs++;
 544                         r++;
 545                 } while (--count);
 546         }
 547
 548         return 0;
 549 #undef N_RELOC
 550 }
 551
 552 static int
 553 i915_gem_execbuffer_relocate_vma_slow(struct i915_vma *vma,
 554                                       struct eb_vmas *eb,
 555                                       struct drm_i915_gem_relocation_entry *relocs)
 556 {
 557         const struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 558         int i, ret;
 559
 560         for (i = 0; i < entry->relocation_count; i++) {
 561                 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, &relocs[i]);
 562                 if (ret)
 563                         return ret;
 564         }
 565
 566         return 0;
 567 }
 568
 569 static int
 570 i915_gem_execbuffer_relocate(struct eb_vmas *eb)
 571 {
 572         struct i915_vma *vma;
 573         int ret = 0;
 574
 575         /* This is the fast path and we cannot handle a pagefault whilst
 576          * holding the struct mutex lest the user pass in the relocations
 577          * contained within a mmaped bo. For in such a case we, the page
 578          * fault handler would call i915_gem_fault() and we would try to
 579          * acquire the struct mutex again. Obviously this is bad and so
 580          * lockdep complains vehemently.
 581          */
 582         pagefault_disable();
 583         list_for_each_entry(vma, &eb->vmas, exec_list) {
 584                 ret = i915_gem_execbuffer_relocate_vma(vma, eb);
 585                 if (ret)
 586                         break;
 587         }
 588         pagefault_enable();
 589
 590         return ret;
 591 }
 592
 593 static bool only_mappable_for_reloc(unsigned int flags)
 594 {
 595         return (flags & (EXEC_OBJECT_NEEDS_FENCE | __EXEC_OBJECT_NEEDS_MAP)) ==
 596                 __EXEC_OBJECT_NEEDS_MAP;
 597 }
 598
 599 static int
 600 i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,
 601                                 struct intel_engine_cs *engine,
 602                                 bool *need_reloc)
 603 {
 604         struct drm_i915_gem_object *obj = vma->obj;
 605         struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 606         uint64_t flags;
 607         int ret;
 608
 609         flags = PIN_USER;
 610         if (entry->flags & EXEC_OBJECT_NEEDS_GTT)
 611                 flags |= PIN_GLOBAL;
 612
 613         if (!drm_mm_node_allocated(&vma->node)) {
 614                 /* Wa32bitGeneralStateOffset & Wa32bitInstructionBaseOffset,
 615                  * limit address to the first 4GBs for unflagged objects.
 616                  */
 617                 if ((entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) == 0)
 618                         flags |= PIN_ZONE_4G;
 619                 if (entry->flags & __EXEC_OBJECT_NEEDS_MAP)
 620                         flags |= PIN_GLOBAL | PIN_MAPPABLE;
 621                 if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS)
 622                         flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS;
 623                 if (entry->flags & EXEC_OBJECT_PINNED)
 624                         flags |= entry->offset | PIN_OFFSET_FIXED;
 625                 if ((flags & PIN_MAPPABLE) == 0)
 626                         flags |= PIN_HIGH;
 627         }
 628
 629         ret = i915_gem_object_pin(obj, vma->vm, entry->alignment, flags);
 630         if ((ret == -ENOSPC  || ret == -E2BIG) &&
 631             only_mappable_for_reloc(entry->flags))
 632                 ret = i915_gem_object_pin(obj, vma->vm,
 633                                           entry->alignment,
 634                                           flags & ~PIN_MAPPABLE);
 635         if (ret)
 636                 return ret;
 637
 638         entry->flags |= __EXEC_OBJECT_HAS_PIN;
 639
 640         if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
 641                 ret = i915_gem_object_get_fence(obj);
 642                 if (ret)
 643                         return ret;
 644
 645                 if (i915_gem_object_pin_fence(obj))
 646                         entry->flags |= __EXEC_OBJECT_HAS_FENCE;
 647         }
 648
 649         if (entry->offset != vma->node.start) {
 650                 entry->offset = vma->node.start;
 651                 *need_reloc = true;
 652         }
 653
 654         if (entry->flags & EXEC_OBJECT_WRITE) {
 655                 obj->base.pending_read_domains = I915_GEM_DOMAIN_RENDER;
 656                 obj->base.pending_write_domain = I915_GEM_DOMAIN_RENDER;
 657         }
 658
 659         return 0;
 660 }
 661
 662 static bool
 663 need_reloc_mappable(struct i915_vma *vma)
 664 {
 665         struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 666
 667         if (entry->relocation_count == 0)
 668                 return false;
 669
 670         if (!vma->is_ggtt)
 671                 return false;
 672
 673         /* See also use_cpu_reloc() */
 674         if (HAS_LLC(vma->obj->base.dev))
 675                 return false;
 676
 677         if (vma->obj->base.write_domain == I915_GEM_DOMAIN_CPU)
 678                 return false;
 679
 680         return true;
 681 }
 682
 683 static bool
 684 eb_vma_misplaced(struct i915_vma *vma)
 685 {
 686         struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 687         struct drm_i915_gem_object *obj = vma->obj;
 688
 689         WARN_ON(entry->flags & __EXEC_OBJECT_NEEDS_MAP && !vma->is_ggtt);
 690
 691         if (entry->alignment &&
 692             vma->node.start & (entry->alignment - 1))
 693                 return true;
 694
 695         if (entry->flags & EXEC_OBJECT_PINNED &&
 696             vma->node.start != entry->offset)
 697                 return true;
 698
 699         if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS &&
 700             vma->node.start < BATCH_OFFSET_BIAS)
 701                 return true;
 702
 703         /* avoid costly ping-pong once a batch bo ended up non-mappable */
 704         if (entry->flags & __EXEC_OBJECT_NEEDS_MAP && !obj->map_and_fenceable)
 705                 return !only_mappable_for_reloc(entry->flags);
 706
 707         if ((entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) == 0 &&
 708             (vma->node.start + vma->node.size - 1) >> 32)
 709                 return true;
 710
 711         return false;
 712 }
 713
 714 static int
 715 i915_gem_execbuffer_reserve(struct intel_engine_cs *engine,
 716                             struct list_head *vmas,
 717                             struct i915_gem_context *ctx,
 718                             bool *need_relocs)
 719 {
 720         struct drm_i915_gem_object *obj;
 721         struct i915_vma *vma;
 722         struct i915_address_space *vm;
 723         struct list_head ordered_vmas;
 724         struct list_head pinned_vmas;
 725         bool has_fenced_gpu_access = INTEL_GEN(engine->i915) < 4;
 726         int retry;
 727
 728         i915_gem_retire_requests_ring(engine);
 729
 730         vm = list_first_entry(vmas, struct i915_vma, exec_list)->vm;
 731
 732         INIT_LIST_HEAD(&ordered_vmas);
 733         INIT_LIST_HEAD(&pinned_vmas);
 734         while (!list_empty(vmas)) {
 735                 struct drm_i915_gem_exec_object2 *entry;
 736                 bool need_fence, need_mappable;
 737
 738                 vma = list_first_entry(vmas, struct i915_vma, exec_list);
 739                 obj = vma->obj;
 740                 entry = vma->exec_entry;
 741
 742                 if (ctx->flags & CONTEXT_NO_ZEROMAP)
 743                         entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
 744
 745                 if (!has_fenced_gpu_access)
 746                         entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE;
 747                 need_fence =
 748                         entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
 749                         obj->tiling_mode != I915_TILING_NONE;
 750                 need_mappable = need_fence || need_reloc_mappable(vma);
 751
 752                 if (entry->flags & EXEC_OBJECT_PINNED)
 753                         list_move_tail(&vma->exec_list, &pinned_vmas);
 754                 else if (need_mappable) {
 755                         entry->flags |= __EXEC_OBJECT_NEEDS_MAP;
 756                         list_move(&vma->exec_list, &ordered_vmas);
 757                 } else
 758                         list_move_tail(&vma->exec_list, &ordered_vmas);
 759
 760                 obj->base.pending_read_domains = I915_GEM_GPU_DOMAINS & ~I915_GEM_DOMAIN_COMMAND;
 761                 obj->base.pending_write_domain = 0;
 762         }
 763         list_splice(&ordered_vmas, vmas);
 764         list_splice(&pinned_vmas, vmas);
 765
 766         /* Attempt to pin all of the buffers into the GTT.
 767          * This is done in 3 phases:
 768          *
 769          * 1a. Unbind all objects that do not match the GTT constraints for
 770          *     the execbuffer (fenceable, mappable, alignment etc).
 771          * 1b. Increment pin count for already bound objects.
 772          * 2.  Bind new objects.
 773          * 3.  Decrement pin count.
 774          *
 775          * This avoid unnecessary unbinding of later objects in order to make
 776          * room for the earlier objects *unless* we need to defragment.
 777          */
 778         retry = 0;
 779         do {
 780                 int ret = 0;
 781
 782                 /* Unbind any ill-fitting objects or pin. */
 783                 list_for_each_entry(vma, vmas, exec_list) {
 784                         if (!drm_mm_node_allocated(&vma->node))
 785                                 continue;
 786
 787                         if (eb_vma_misplaced(vma))
 788                                 ret = i915_vma_unbind(vma);
 789                         else
 790                                 ret = i915_gem_execbuffer_reserve_vma(vma,
 791                                                                       engine,
 792                                                                       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, engine,
 803                                                               need_relocs);
 804                         if (ret)
 805                                 goto err;
 806                 }
 807
 808 err:
 809                 if (ret != -ENOSPC || retry++)
 810                         return ret;
 811
 812                 /* Decrement pin count for bound objects */
 813                 list_for_each_entry(vma, vmas, exec_list)
 814                         i915_gem_execbuffer_unreserve_vma(vma);
 815
 816                 ret = i915_gem_evict_vm(vm, true);
 817                 if (ret)
 818                         return ret;
 819         } while (1);
 820 }
 821
 822 static int
 823 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
 824                                   struct drm_i915_gem_execbuffer2 *args,
 825                                   struct drm_file *file,
 826                                   struct intel_engine_cs *engine,
 827                                   struct eb_vmas *eb,
 828                                   struct drm_i915_gem_exec_object2 *exec,
 829                                   struct i915_gem_context *ctx)
 830 {
 831         struct drm_i915_gem_relocation_entry *reloc;
 832         struct i915_address_space *vm;
 833         struct i915_vma *vma;
 834         bool need_relocs;
 835         int *reloc_offset;
 836         int i, total, ret;
 837         unsigned count = args->buffer_count;
 838
 839         vm = list_first_entry(&eb->vmas, struct i915_vma, exec_list)->vm;
 840
 841         /* We may process another execbuffer during the unlock... */
 842         while (!list_empty(&eb->vmas)) {
 843                 vma = list_first_entry(&eb->vmas, struct i915_vma, exec_list);
 844                 list_del_init(&vma->exec_list);
 845                 i915_gem_execbuffer_unreserve_vma(vma);
 846                 drm_gem_object_unreference(&vma->obj->base);
 847         }
 848
 849         mutex_unlock(&dev->struct_mutex);
 850
 851         total = 0;
 852         for (i = 0; i < count; i++)
 853                 total += exec[i].relocation_count;
 854
 855         reloc_offset = drm_malloc_ab(count, sizeof(*reloc_offset));
 856         reloc = drm_malloc_ab(total, sizeof(*reloc));
 857         if (reloc == NULL || reloc_offset == NULL) {
 858                 drm_free_large(reloc);
 859                 drm_free_large(reloc_offset);
 860                 mutex_lock(&dev->struct_mutex);
 861                 return -ENOMEM;
 862         }
 863
 864         total = 0;
 865         for (i = 0; i < count; i++) {
 866                 struct drm_i915_gem_relocation_entry __user *user_relocs;
 867                 u64 invalid_offset = (u64)-1;
 868                 int j;
 869
 870                 user_relocs = u64_to_user_ptr(exec[i].relocs_ptr);
 871
 872                 if (copy_from_user(reloc+total, user_relocs,
 873                                    exec[i].relocation_count * sizeof(*reloc))) {
 874                         ret = -EFAULT;
 875                         mutex_lock(&dev->struct_mutex);
 876                         goto err;
 877                 }
 878
 879                 /* As we do not update the known relocation offsets after
 880                  * relocating (due to the complexities in lock handling),
 881                  * we need to mark them as invalid now so that we force the
 882                  * relocation processing next time. Just in case the target
 883                  * object is evicted and then rebound into its old
 884                  * presumed_offset before the next execbuffer - if that
 885                  * happened we would make the mistake of assuming that the
 886                  * relocations were valid.
 887                  */
 888                 for (j = 0; j < exec[i].relocation_count; j++) {
 889                         if (__copy_to_user(&user_relocs[j].presumed_offset,
 890                                            &invalid_offset,
 891                                            sizeof(invalid_offset))) {
 892                                 ret = -EFAULT;
 893                                 mutex_lock(&dev->struct_mutex);
 894                                 goto err;
 895                         }
 896                 }
 897
 898                 reloc_offset[i] = total;
 899                 total += exec[i].relocation_count;
 900         }
 901
 902         ret = i915_mutex_lock_interruptible(dev);
 903         if (ret) {
 904                 mutex_lock(&dev->struct_mutex);
 905                 goto err;
 906         }
 907
 908         /* reacquire the objects */
 909         eb_reset(eb);
 910         ret = eb_lookup_vmas(eb, exec, args, vm, file);
 911         if (ret)
 912                 goto err;
 913
 914         need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
 915         ret = i915_gem_execbuffer_reserve(engine, &eb->vmas, ctx,
 916                                           &need_relocs);
 917         if (ret)
 918                 goto err;
 919
 920         list_for_each_entry(vma, &eb->vmas, exec_list) {
 921                 int offset = vma->exec_entry - exec;
 922                 ret = i915_gem_execbuffer_relocate_vma_slow(vma, eb,
 923                                                             reloc + reloc_offset[offset]);
 924                 if (ret)
 925                         goto err;
 926         }
 927
 928         /* Leave the user relocations as are, this is the painfully slow path,
 929          * and we want to avoid the complication of dropping the lock whilst
 930          * having buffers reserved in the aperture and so causing spurious
 931          * ENOSPC for random operations.
 932          */
 933
 934 err:
 935         drm_free_large(reloc);
 936         drm_free_large(reloc_offset);
 937         return ret;
 938 }
 939
 940 static int
 941 i915_gem_execbuffer_move_to_gpu(struct drm_i915_gem_request *req,
 942                                 struct list_head *vmas)
 943 {
 944         const unsigned other_rings = ~intel_engine_flag(req->engine);
 945         struct i915_vma *vma;
 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->engine, &req);
 953                         if (ret)
 954                                 return ret;
 955                 }
 956
 957                 if (obj->base.write_domain & I915_GEM_DOMAIN_CPU)
 958                         i915_gem_clflush_object(obj, false);
 959         }
 960
 961         /* Unconditionally flush any chipset caches (for streaming writes). */
 962         i915_gem_chipset_flush(req->engine->i915);
 963
 964         /* Unconditionally invalidate gpu caches and ensure that we do flush
 965          * any residual writes from the previous batch.
 966          */
 967         return intel_ring_invalidate_all_caches(req);
 968 }
 969
 970 static bool
 971 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
 972 {
 973         if (exec->flags & __I915_EXEC_UNKNOWN_FLAGS)
 974                 return false;
 975
 976         /* Kernel clipping was a DRI1 misfeature */
 977         if (exec->num_cliprects || exec->cliprects_ptr)
 978                 return false;
 979
 980         if (exec->DR4 == 0xffffffff) {
 981                 DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
 982                 exec->DR4 = 0;
 983         }
 984         if (exec->DR1 || exec->DR4)
 985                 return false;
 986
 987         if ((exec->batch_start_offset | exec->batch_len) & 0x7)
 988                 return false;
 989
 990         return true;
 991 }
 992
 993 static int
 994 validate_exec_list(struct drm_device *dev,
 995                    struct drm_i915_gem_exec_object2 *exec,
 996                    int count)
 997 {
 998         unsigned relocs_total = 0;
 999         unsigned relocs_max = UINT_MAX / sizeof(struct drm_i915_gem_relocation_entry);
1000         unsigned invalid_flags;
1001         int i;
1002
1003         invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS;
1004         if (USES_FULL_PPGTT(dev))
1005                 invalid_flags |= EXEC_OBJECT_NEEDS_GTT;
1006
1007         for (i = 0; i < count; i++) {
1008                 char __user *ptr = u64_to_user_ptr(exec[i].relocs_ptr);
1009                 int length; /* limited by fault_in_pages_readable() */
1010
1011                 if (exec[i].flags & invalid_flags)
1012                         return -EINVAL;
1013
1014                 /* Offset can be used as input (EXEC_OBJECT_PINNED), reject
1015                  * any non-page-aligned or non-canonical addresses.
1016                  */
1017                 if (exec[i].flags & EXEC_OBJECT_PINNED) {
1018                         if (exec[i].offset !=
1019                             gen8_canonical_addr(exec[i].offset & PAGE_MASK))
1020                                 return -EINVAL;
1021
1022                         /* From drm_mm perspective address space is continuous,
1023                          * so from this point we're always using non-canonical
1024                          * form internally.
1025                          */
1026                         exec[i].offset = gen8_noncanonical_addr(exec[i].offset);
1027                 }
1028
1029                 if (exec[i].alignment && !is_power_of_2(exec[i].alignment))
1030                         return -EINVAL;
1031
1032                 /* First check for malicious input causing overflow in
1033                  * the worst case where we need to allocate the entire
1034                  * relocation tree as a single array.
1035                  */
1036                 if (exec[i].relocation_count > relocs_max - relocs_total)
1037                         return -EINVAL;
1038                 relocs_total += exec[i].relocation_count;
1039
1040                 length = exec[i].relocation_count *
1041                         sizeof(struct drm_i915_gem_relocation_entry);
1042                 /*
1043                  * We must check that the entire relocation array is safe
1044                  * to read, but since we may need to update the presumed
1045                  * offsets during execution, check for full write access.
1046                  */
1047                 if (!access_ok(VERIFY_WRITE, ptr, length))
1048                         return -EFAULT;
1049
1050                 if (likely(!i915.prefault_disable)) {
1051                         if (fault_in_multipages_readable(ptr, length))
1052                                 return -EFAULT;
1053                 }
1054         }
1055
1056         return 0;
1057 }
1058
1059 static struct i915_gem_context *
1060 i915_gem_validate_context(struct drm_device *dev, struct drm_file *file,
1061                           struct intel_engine_cs *engine, const u32 ctx_id)
1062 {
1063         struct i915_gem_context *ctx = NULL;
1064         struct i915_ctx_hang_stats *hs;
1065
1066         if (engine->id != RCS && ctx_id != DEFAULT_CONTEXT_HANDLE)
1067                 return ERR_PTR(-EINVAL);
1068
1069         ctx = i915_gem_context_lookup(file->driver_priv, ctx_id);
1070         if (IS_ERR(ctx))
1071                 return ctx;
1072
1073         hs = &ctx->hang_stats;
1074         if (hs->banned) {
1075                 DRM_DEBUG("Context %u tried to submit while banned\n", ctx_id);
1076                 return ERR_PTR(-EIO);
1077         }
1078
1079         return ctx;
1080 }
1081
1082 void
1083 i915_gem_execbuffer_move_to_active(struct list_head *vmas,
1084                                    struct drm_i915_gem_request *req)
1085 {
1086         struct intel_engine_cs *engine = i915_gem_request_get_engine(req);
1087         struct i915_vma *vma;
1088
1089         list_for_each_entry(vma, vmas, exec_list) {
1090                 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
1091                 struct drm_i915_gem_object *obj = vma->obj;
1092                 u32 old_read = obj->base.read_domains;
1093                 u32 old_write = obj->base.write_domain;
1094
1095                 obj->dirty = 1; /* be paranoid  */
1096                 obj->base.write_domain = obj->base.pending_write_domain;
1097                 if (obj->base.write_domain == 0)
1098                         obj->base.pending_read_domains |= obj->base.read_domains;
1099                 obj->base.read_domains = obj->base.pending_read_domains;
1100
1101                 i915_vma_move_to_active(vma, req);
1102                 if (obj->base.write_domain) {
1103                         i915_gem_request_assign(&obj->last_write_req, req);
1104
1105                         intel_fb_obj_invalidate(obj, ORIGIN_CS);
1106
1107                         /* update for the implicit flush after a batch */
1108                         obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
1109                 }
1110                 if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
1111                         i915_gem_request_assign(&obj->last_fenced_req, req);
1112                         if (entry->flags & __EXEC_OBJECT_HAS_FENCE) {
1113                                 struct drm_i915_private *dev_priv = engine->i915;
1114                                 list_move_tail(&dev_priv->fence_regs[obj->fence_reg].lru_list,
1115                                                &dev_priv->mm.fence_list);
1116                         }
1117                 }
1118
1119                 trace_i915_gem_object_change_domain(obj, old_read, old_write);
1120         }
1121 }
1122
1123 static void
1124 i915_gem_execbuffer_retire_commands(struct i915_execbuffer_params *params)
1125 {
1126         /* Unconditionally force add_request to emit a full flush. */
1127         params->engine->gpu_caches_dirty = true;
1128
1129         /* Add a breadcrumb for the completion of the batch buffer */
1130         __i915_add_request(params->request, params->batch_obj, true);
1131 }
1132
1133 static int
1134 i915_reset_gen7_sol_offsets(struct drm_device *dev,
1135                             struct drm_i915_gem_request *req)
1136 {
1137         struct intel_engine_cs *engine = req->engine;
1138         struct drm_i915_private *dev_priv = to_i915(dev);
1139         int ret, i;
1140
1141         if (!IS_GEN7(dev) || engine != &dev_priv->engine[RCS]) {
1142                 DRM_DEBUG("sol reset is gen7/rcs only\n");
1143                 return -EINVAL;
1144         }
1145
1146         ret = intel_ring_begin(req, 4 * 3);
1147         if (ret)
1148                 return ret;
1149
1150         for (i = 0; i < 4; i++) {
1151                 intel_ring_emit(engine, MI_LOAD_REGISTER_IMM(1));
1152                 intel_ring_emit_reg(engine, GEN7_SO_WRITE_OFFSET(i));
1153                 intel_ring_emit(engine, 0);
1154         }
1155
1156         intel_ring_advance(engine);
1157
1158         return 0;
1159 }
1160
1161 static struct drm_i915_gem_object*
1162 i915_gem_execbuffer_parse(struct intel_engine_cs *engine,
1163                           struct drm_i915_gem_exec_object2 *shadow_exec_entry,
1164                           struct eb_vmas *eb,
1165                           struct drm_i915_gem_object *batch_obj,
1166                           u32 batch_start_offset,
1167                           u32 batch_len,
1168                           bool is_master)
1169 {
1170         struct drm_i915_gem_object *shadow_batch_obj;
1171         struct i915_vma *vma;
1172         int ret;
1173
1174         shadow_batch_obj = i915_gem_batch_pool_get(&engine->batch_pool,
1175                                                    PAGE_ALIGN(batch_len));
1176         if (IS_ERR(shadow_batch_obj))
1177                 return shadow_batch_obj;
1178
1179         ret = i915_parse_cmds(engine,
1180                               batch_obj,
1181                               shadow_batch_obj,
1182                               batch_start_offset,
1183                               batch_len,
1184                               is_master);
1185         if (ret)
1186                 goto err;
1187
1188         ret = i915_gem_obj_ggtt_pin(shadow_batch_obj, 0, 0);
1189         if (ret)
1190                 goto err;
1191
1192         i915_gem_object_unpin_pages(shadow_batch_obj);
1193
1194         memset(shadow_exec_entry, 0, sizeof(*shadow_exec_entry));
1195
1196         vma = i915_gem_obj_to_ggtt(shadow_batch_obj);
1197         vma->exec_entry = shadow_exec_entry;
1198         vma->exec_entry->flags = __EXEC_OBJECT_HAS_PIN;
1199         drm_gem_object_reference(&shadow_batch_obj->base);
1200         list_add_tail(&vma->exec_list, &eb->vmas);
1201
1202         shadow_batch_obj->base.pending_read_domains = I915_GEM_DOMAIN_COMMAND;
1203
1204         return shadow_batch_obj;
1205
1206 err:
1207         i915_gem_object_unpin_pages(shadow_batch_obj);
1208         if (ret == -EACCES) /* unhandled chained batch */
1209                 return batch_obj;
1210         else
1211                 return ERR_PTR(ret);
1212 }
1213
1214 int
1215 i915_gem_ringbuffer_submission(struct i915_execbuffer_params *params,
1216                                struct drm_i915_gem_execbuffer2 *args,
1217                                struct list_head *vmas)
1218 {
1219         struct drm_device *dev = params->dev;
1220         struct intel_engine_cs *engine = params->engine;
1221         struct drm_i915_private *dev_priv = to_i915(dev);
1222         u64 exec_start, exec_len;
1223         int instp_mode;
1224         u32 instp_mask;
1225         int ret;
1226
1227         ret = i915_gem_execbuffer_move_to_gpu(params->request, vmas);
1228         if (ret)
1229                 return ret;
1230
1231         ret = i915_switch_context(params->request);
1232         if (ret)
1233                 return ret;
1234
1235         WARN(params->ctx->ppgtt && params->ctx->ppgtt->pd_dirty_rings & (1<<engine->id),
1236              "%s didn't clear reload\n", engine->name);
1237
1238         instp_mode = args->flags & I915_EXEC_CONSTANTS_MASK;
1239         instp_mask = I915_EXEC_CONSTANTS_MASK;
1240         switch (instp_mode) {
1241         case I915_EXEC_CONSTANTS_REL_GENERAL:
1242         case I915_EXEC_CONSTANTS_ABSOLUTE:
1243         case I915_EXEC_CONSTANTS_REL_SURFACE:
1244                 if (instp_mode != 0 && engine != &dev_priv->engine[RCS]) {
1245                         DRM_DEBUG("non-0 rel constants mode on non-RCS\n");
1246                         return -EINVAL;
1247                 }
1248
1249                 if (instp_mode != dev_priv->relative_constants_mode) {
1250                         if (INTEL_INFO(dev)->gen < 4) {
1251                                 DRM_DEBUG("no rel constants on pre-gen4\n");
1252                                 return -EINVAL;
1253                         }
1254
1255                         if (INTEL_INFO(dev)->gen > 5 &&
1256                             instp_mode == I915_EXEC_CONSTANTS_REL_SURFACE) {
1257                                 DRM_DEBUG("rel surface constants mode invalid on gen5+\n");
1258                                 return -EINVAL;
1259                         }
1260
1261                         /* The HW changed the meaning on this bit on gen6 */
1262                         if (INTEL_INFO(dev)->gen >= 6)
1263                                 instp_mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
1264                 }
1265                 break;
1266         default:
1267                 DRM_DEBUG("execbuf with unknown constants: %d\n", instp_mode);
1268                 return -EINVAL;
1269         }
1270
1271         if (engine == &dev_priv->engine[RCS] &&
1272             instp_mode != dev_priv->relative_constants_mode) {
1273                 ret = intel_ring_begin(params->request, 4);
1274                 if (ret)
1275                         return ret;
1276
1277                 intel_ring_emit(engine, MI_NOOP);
1278                 intel_ring_emit(engine, MI_LOAD_REGISTER_IMM(1));
1279                 intel_ring_emit_reg(engine, INSTPM);
1280                 intel_ring_emit(engine, instp_mask << 16 | instp_mode);
1281                 intel_ring_advance(engine);
1282
1283                 dev_priv->relative_constants_mode = instp_mode;
1284         }
1285
1286         if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
1287                 ret = i915_reset_gen7_sol_offsets(dev, params->request);
1288                 if (ret)
1289                         return ret;
1290         }
1291
1292         exec_len   = args->batch_len;
1293         exec_start = params->batch_obj_vm_offset +
1294                      params->args_batch_start_offset;
1295
1296         if (exec_len == 0)
1297                 exec_len = params->batch_obj->base.size;
1298
1299         ret = engine->dispatch_execbuffer(params->request,
1300                                         exec_start, exec_len,
1301                                         params->dispatch_flags);
1302         if (ret)
1303                 return ret;
1304
1305         trace_i915_gem_ring_dispatch(params->request, params->dispatch_flags);
1306
1307         i915_gem_execbuffer_move_to_active(vmas, params->request);
1308
1309         return 0;
1310 }
1311
1312 /**
1313  * Find one BSD ring to dispatch the corresponding BSD command.
1314  * The ring index is returned.
1315  */
1316 static unsigned int
1317 gen8_dispatch_bsd_ring(struct drm_i915_private *dev_priv, struct drm_file *file)
1318 {
1319         struct drm_i915_file_private *file_priv = file->driver_priv;
1320
1321         /* Check whether the file_priv has already selected one ring. */
1322         if ((int)file_priv->bsd_ring < 0) {
1323                 /* If not, use the ping-pong mechanism to select one. */
1324                 mutex_lock(&dev_priv->drm.struct_mutex);
1325                 file_priv->bsd_ring = dev_priv->mm.bsd_ring_dispatch_index;
1326                 dev_priv->mm.bsd_ring_dispatch_index ^= 1;
1327                 mutex_unlock(&dev_priv->drm.struct_mutex);
1328         }
1329
1330         return file_priv->bsd_ring;
1331 }
1332
1333 static struct drm_i915_gem_object *
1334 eb_get_batch(struct eb_vmas *eb)
1335 {
1336         struct i915_vma *vma = list_entry(eb->vmas.prev, typeof(*vma), exec_list);
1337
1338         /*
1339          * SNA is doing fancy tricks with compressing batch buffers, which leads
1340          * to negative relocation deltas. Usually that works out ok since the
1341          * relocate address is still positive, except when the batch is placed
1342          * very low in the GTT. Ensure this doesn't happen.
1343          *
1344          * Note that actual hangs have only been observed on gen7, but for
1345          * paranoia do it everywhere.
1346          */
1347         if ((vma->exec_entry->flags & EXEC_OBJECT_PINNED) == 0)
1348                 vma->exec_entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
1349
1350         return vma->obj;
1351 }
1352
1353 #define I915_USER_RINGS (4)
1354
1355 static const enum intel_engine_id user_ring_map[I915_USER_RINGS + 1] = {
1356         [I915_EXEC_DEFAULT]     = RCS,
1357         [I915_EXEC_RENDER]      = RCS,
1358         [I915_EXEC_BLT]         = BCS,
1359         [I915_EXEC_BSD]         = VCS,
1360         [I915_EXEC_VEBOX]       = VECS
1361 };
1362
1363 static int
1364 eb_select_ring(struct drm_i915_private *dev_priv,
1365                struct drm_file *file,
1366                struct drm_i915_gem_execbuffer2 *args,
1367                struct intel_engine_cs **ring)
1368 {
1369         unsigned int user_ring_id = args->flags & I915_EXEC_RING_MASK;
1370
1371         if (user_ring_id > I915_USER_RINGS) {
1372                 DRM_DEBUG("execbuf with unknown ring: %u\n", user_ring_id);
1373                 return -EINVAL;
1374         }
1375
1376         if ((user_ring_id != I915_EXEC_BSD) &&
1377             ((args->flags & I915_EXEC_BSD_MASK) != 0)) {
1378                 DRM_DEBUG("execbuf with non bsd ring but with invalid "
1379                           "bsd dispatch flags: %d\n", (int)(args->flags));
1380                 return -EINVAL;
1381         }
1382
1383         if (user_ring_id == I915_EXEC_BSD && HAS_BSD2(dev_priv)) {
1384                 unsigned int bsd_idx = args->flags & I915_EXEC_BSD_MASK;
1385
1386                 if (bsd_idx == I915_EXEC_BSD_DEFAULT) {
1387                         bsd_idx = gen8_dispatch_bsd_ring(dev_priv, file);
1388                 } else if (bsd_idx >= I915_EXEC_BSD_RING1 &&
1389                            bsd_idx <= I915_EXEC_BSD_RING2) {
1390                         bsd_idx >>= I915_EXEC_BSD_SHIFT;
1391                         bsd_idx--;
1392                 } else {
1393                         DRM_DEBUG("execbuf with unknown bsd ring: %u\n",
1394                                   bsd_idx);
1395                         return -EINVAL;
1396                 }
1397
1398                 *ring = &dev_priv->engine[_VCS(bsd_idx)];
1399         } else {
1400                 *ring = &dev_priv->engine[user_ring_map[user_ring_id]];
1401         }
1402
1403         if (!intel_engine_initialized(*ring)) {
1404                 DRM_DEBUG("execbuf with invalid ring: %u\n", user_ring_id);
1405                 return -EINVAL;
1406         }
1407
1408         return 0;
1409 }
1410
1411 static int
1412 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1413                        struct drm_file *file,
1414                        struct drm_i915_gem_execbuffer2 *args,
1415                        struct drm_i915_gem_exec_object2 *exec)
1416 {
1417         struct drm_i915_private *dev_priv = to_i915(dev);
1418         struct i915_ggtt *ggtt = &dev_priv->ggtt;
1419         struct drm_i915_gem_request *req = NULL;
1420         struct eb_vmas *eb;
1421         struct drm_i915_gem_object *batch_obj;
1422         struct drm_i915_gem_exec_object2 shadow_exec_entry;
1423         struct intel_engine_cs *engine;
1424         struct i915_gem_context *ctx;
1425         struct i915_address_space *vm;
1426         struct i915_execbuffer_params params_master; /* XXX: will be removed later */
1427         struct i915_execbuffer_params *params = &params_master;
1428         const u32 ctx_id = i915_execbuffer2_get_context_id(*args);
1429         u32 dispatch_flags;
1430         int ret;
1431         bool need_relocs;
1432
1433         if (!i915_gem_check_execbuffer(args))
1434                 return -EINVAL;
1435
1436         ret = validate_exec_list(dev, exec, args->buffer_count);
1437         if (ret)
1438                 return ret;
1439
1440         dispatch_flags = 0;
1441         if (args->flags & I915_EXEC_SECURE) {
1442                 if (!drm_is_current_master(file) || !capable(CAP_SYS_ADMIN))
1443                     return -EPERM;
1444
1445                 dispatch_flags |= I915_DISPATCH_SECURE;
1446         }
1447         if (args->flags & I915_EXEC_IS_PINNED)
1448                 dispatch_flags |= I915_DISPATCH_PINNED;
1449
1450         ret = eb_select_ring(dev_priv, file, args, &engine);
1451         if (ret)
1452                 return ret;
1453
1454         if (args->buffer_count < 1) {
1455                 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1456                 return -EINVAL;
1457         }
1458
1459         if (args->flags & I915_EXEC_RESOURCE_STREAMER) {
1460                 if (!HAS_RESOURCE_STREAMER(dev)) {
1461                         DRM_DEBUG("RS is only allowed for Haswell, Gen8 and above\n");
1462                         return -EINVAL;
1463                 }
1464                 if (engine->id != RCS) {
1465                         DRM_DEBUG("RS is not available on %s\n",
1466                                  engine->name);
1467                         return -EINVAL;
1468                 }
1469
1470                 dispatch_flags |= I915_DISPATCH_RS;
1471         }
1472
1473         /* Take a local wakeref for preparing to dispatch the execbuf as
1474          * we expect to access the hardware fairly frequently in the
1475          * process. Upon first dispatch, we acquire another prolonged
1476          * wakeref that we hold until the GPU has been idle for at least
1477          * 100ms.
1478          */
1479         intel_runtime_pm_get(dev_priv);
1480
1481         ret = i915_mutex_lock_interruptible(dev);
1482         if (ret)
1483                 goto pre_mutex_err;
1484
1485         ctx = i915_gem_validate_context(dev, file, engine, ctx_id);
1486         if (IS_ERR(ctx)) {
1487                 mutex_unlock(&dev->struct_mutex);
1488                 ret = PTR_ERR(ctx);
1489                 goto pre_mutex_err;
1490         }
1491
1492         i915_gem_context_reference(ctx);
1493
1494         if (ctx->ppgtt)
1495                 vm = &ctx->ppgtt->base;
1496         else
1497                 vm = &ggtt->base;
1498
1499         memset(&params_master, 0x00, sizeof(params_master));
1500
1501         eb = eb_create(args);
1502         if (eb == NULL) {
1503                 i915_gem_context_unreference(ctx);
1504                 mutex_unlock(&dev->struct_mutex);
1505                 ret = -ENOMEM;
1506                 goto pre_mutex_err;
1507         }
1508
1509         /* Look up object handles */
1510         ret = eb_lookup_vmas(eb, exec, args, vm, file);
1511         if (ret)
1512                 goto err;
1513
1514         /* take note of the batch buffer before we might reorder the lists */
1515         batch_obj = eb_get_batch(eb);
1516
1517         /* Move the objects en-masse into the GTT, evicting if necessary. */
1518         need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
1519         ret = i915_gem_execbuffer_reserve(engine, &eb->vmas, ctx,
1520                                           &need_relocs);
1521         if (ret)
1522                 goto err;
1523
1524         /* The objects are in their final locations, apply the relocations. */
1525         if (need_relocs)
1526                 ret = i915_gem_execbuffer_relocate(eb);
1527         if (ret) {
1528                 if (ret == -EFAULT) {
1529                         ret = i915_gem_execbuffer_relocate_slow(dev, args, file,
1530                                                                 engine,
1531                                                                 eb, exec, ctx);
1532                         BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1533                 }
1534                 if (ret)
1535                         goto err;
1536         }
1537
1538         /* Set the pending read domains for the batch buffer to COMMAND */
1539         if (batch_obj->base.pending_write_domain) {
1540                 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1541                 ret = -EINVAL;
1542                 goto err;
1543         }
1544
1545         params->args_batch_start_offset = args->batch_start_offset;
1546         if (i915_needs_cmd_parser(engine) && args->batch_len) {
1547                 struct drm_i915_gem_object *parsed_batch_obj;
1548
1549                 parsed_batch_obj = i915_gem_execbuffer_parse(engine,
1550                                                              &shadow_exec_entry,
1551                                                              eb,
1552                                                              batch_obj,
1553                                                              args->batch_start_offset,
1554                                                              args->batch_len,
1555                                                              drm_is_current_master(file));
1556                 if (IS_ERR(parsed_batch_obj)) {
1557                         ret = PTR_ERR(parsed_batch_obj);
1558                         goto err;
1559                 }
1560
1561                 /*
1562                  * parsed_batch_obj == batch_obj means batch not fully parsed:
1563                  * Accept, but don't promote to secure.
1564                  */
1565
1566                 if (parsed_batch_obj != batch_obj) {
1567                         /*
1568                          * Batch parsed and accepted:
1569                          *
1570                          * Set the DISPATCH_SECURE bit to remove the NON_SECURE
1571                          * bit from MI_BATCH_BUFFER_START commands issued in
1572                          * the dispatch_execbuffer implementations. We
1573                          * specifically don't want that set on batches the
1574                          * command parser has accepted.
1575                          */
1576                         dispatch_flags |= I915_DISPATCH_SECURE;
1577                         params->args_batch_start_offset = 0;
1578                         batch_obj = parsed_batch_obj;
1579                 }
1580         }
1581
1582         batch_obj->base.pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
1583
1584         /* snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
1585          * batch" bit. Hence we need to pin secure batches into the global gtt.
1586          * hsw should have this fixed, but bdw mucks it up again. */
1587         if (dispatch_flags & I915_DISPATCH_SECURE) {
1588                 /*
1589                  * So on first glance it looks freaky that we pin the batch here
1590                  * outside of the reservation loop. But:
1591                  * - The batch is already pinned into the relevant ppgtt, so we
1592                  *   already have the backing storage fully allocated.
1593                  * - No other BO uses the global gtt (well contexts, but meh),
1594                  *   so we don't really have issues with multiple objects not
1595                  *   fitting due to fragmentation.
1596                  * So this is actually safe.
1597                  */
1598                 ret = i915_gem_obj_ggtt_pin(batch_obj, 0, 0);
1599                 if (ret)
1600                         goto err;
1601
1602                 params->batch_obj_vm_offset = i915_gem_obj_ggtt_offset(batch_obj);
1603         } else
1604                 params->batch_obj_vm_offset = i915_gem_obj_offset(batch_obj, vm);
1605
1606         /* Allocate a request for this batch buffer nice and early. */
1607         req = i915_gem_request_alloc(engine, ctx);
1608         if (IS_ERR(req)) {
1609                 ret = PTR_ERR(req);
1610                 goto err_batch_unpin;
1611         }
1612
1613         ret = i915_gem_request_add_to_client(req, file);
1614         if (ret)
1615                 goto err_request;
1616
1617         /*
1618          * Save assorted stuff away to pass through to *_submission().
1619          * NB: This data should be 'persistent' and not local as it will
1620          * kept around beyond the duration of the IOCTL once the GPU
1621          * scheduler arrives.
1622          */
1623         params->dev                     = dev;
1624         params->file                    = file;
1625         params->engine                    = engine;
1626         params->dispatch_flags          = dispatch_flags;
1627         params->batch_obj               = batch_obj;
1628         params->ctx                     = ctx;
1629         params->request                 = req;
1630
1631         ret = dev_priv->gt.execbuf_submit(params, args, &eb->vmas);
1632 err_request:
1633         i915_gem_execbuffer_retire_commands(params);
1634
1635 err_batch_unpin:
1636         /*
1637          * FIXME: We crucially rely upon the active tracking for the (ppgtt)
1638          * batch vma for correctness. For less ugly and less fragility this
1639          * needs to be adjusted to also track the ggtt batch vma properly as
1640          * active.
1641          */
1642         if (dispatch_flags & I915_DISPATCH_SECURE)
1643                 i915_gem_object_ggtt_unpin(batch_obj);
1644
1645 err:
1646         /* the request owns the ref now */
1647         i915_gem_context_unreference(ctx);
1648         eb_destroy(eb);
1649
1650         mutex_unlock(&dev->struct_mutex);
1651
1652 pre_mutex_err:
1653         /* intel_gpu_busy should also get a ref, so it will free when the device
1654          * is really idle. */
1655         intel_runtime_pm_put(dev_priv);
1656         return ret;
1657 }
1658
1659 /*
1660  * Legacy execbuffer just creates an exec2 list from the original exec object
1661  * list array and passes it to the real function.
1662  */
1663 int
1664 i915_gem_execbuffer(struct drm_device *dev, void *data,
1665                     struct drm_file *file)
1666 {
1667         struct drm_i915_gem_execbuffer *args = data;
1668         struct drm_i915_gem_execbuffer2 exec2;
1669         struct drm_i915_gem_exec_object *exec_list = NULL;
1670         struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1671         int ret, i;
1672
1673         if (args->buffer_count < 1) {
1674                 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1675                 return -EINVAL;
1676         }
1677
1678         /* Copy in the exec list from userland */
1679         exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
1680         exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
1681         if (exec_list == NULL || exec2_list == NULL) {
1682                 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1683                           args->buffer_count);
1684                 drm_free_large(exec_list);
1685                 drm_free_large(exec2_list);
1686                 return -ENOMEM;
1687         }
1688         ret = copy_from_user(exec_list,
1689                              u64_to_user_ptr(args->buffers_ptr),
1690                              sizeof(*exec_list) * args->buffer_count);
1691         if (ret != 0) {
1692                 DRM_DEBUG("copy %d exec entries failed %d\n",
1693                           args->buffer_count, ret);
1694                 drm_free_large(exec_list);
1695                 drm_free_large(exec2_list);
1696                 return -EFAULT;
1697         }
1698
1699         for (i = 0; i < args->buffer_count; i++) {
1700                 exec2_list[i].handle = exec_list[i].handle;
1701                 exec2_list[i].relocation_count = exec_list[i].relocation_count;
1702                 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
1703                 exec2_list[i].alignment = exec_list[i].alignment;
1704                 exec2_list[i].offset = exec_list[i].offset;
1705                 if (INTEL_INFO(dev)->gen < 4)
1706                         exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
1707                 else
1708                         exec2_list[i].flags = 0;
1709         }
1710
1711         exec2.buffers_ptr = args->buffers_ptr;
1712         exec2.buffer_count = args->buffer_count;
1713         exec2.batch_start_offset = args->batch_start_offset;
1714         exec2.batch_len = args->batch_len;
1715         exec2.DR1 = args->DR1;
1716         exec2.DR4 = args->DR4;
1717         exec2.num_cliprects = args->num_cliprects;
1718         exec2.cliprects_ptr = args->cliprects_ptr;
1719         exec2.flags = I915_EXEC_RENDER;
1720         i915_execbuffer2_set_context_id(exec2, 0);
1721
1722         ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list);
1723         if (!ret) {
1724                 struct drm_i915_gem_exec_object __user *user_exec_list =
1725                         u64_to_user_ptr(args->buffers_ptr);
1726
1727                 /* Copy the new buffer offsets back to the user's exec list. */
1728                 for (i = 0; i < args->buffer_count; i++) {
1729                         exec2_list[i].offset =
1730                                 gen8_canonical_addr(exec2_list[i].offset);
1731                         ret = __copy_to_user(&user_exec_list[i].offset,
1732                                              &exec2_list[i].offset,
1733                                              sizeof(user_exec_list[i].offset));
1734                         if (ret) {
1735                                 ret = -EFAULT;
1736                                 DRM_DEBUG("failed to copy %d exec entries "
1737                                           "back to user (%d)\n",
1738                                           args->buffer_count, ret);
1739                                 break;
1740                         }
1741                 }
1742         }
1743
1744         drm_free_large(exec_list);
1745         drm_free_large(exec2_list);
1746         return ret;
1747 }
1748
1749 int
1750 i915_gem_execbuffer2(struct drm_device *dev, void *data,
1751                      struct drm_file *file)
1752 {
1753         struct drm_i915_gem_execbuffer2 *args = data;
1754         struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1755         int ret;
1756
1757         if (args->buffer_count < 1 ||
1758             args->buffer_count > UINT_MAX / sizeof(*exec2_list)) {
1759                 DRM_DEBUG("execbuf2 with %d buffers\n", args->buffer_count);
1760                 return -EINVAL;
1761         }
1762
1763         if (args->rsvd2 != 0) {
1764                 DRM_DEBUG("dirty rvsd2 field\n");
1765                 return -EINVAL;
1766         }
1767
1768         exec2_list = drm_malloc_gfp(args->buffer_count,
1769                                     sizeof(*exec2_list),
1770                                     GFP_TEMPORARY);
1771         if (exec2_list == NULL) {
1772                 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1773                           args->buffer_count);
1774                 return -ENOMEM;
1775         }
1776         ret = copy_from_user(exec2_list,
1777                              u64_to_user_ptr(args->buffers_ptr),
1778                              sizeof(*exec2_list) * args->buffer_count);
1779         if (ret != 0) {
1780                 DRM_DEBUG("copy %d exec entries failed %d\n",
1781                           args->buffer_count, ret);
1782                 drm_free_large(exec2_list);
1783                 return -EFAULT;
1784         }
1785
1786         ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list);
1787         if (!ret) {
1788                 /* Copy the new buffer offsets back to the user's exec list. */
1789                 struct drm_i915_gem_exec_object2 __user *user_exec_list =
1790                                    u64_to_user_ptr(args->buffers_ptr);
1791                 int i;
1792
1793                 for (i = 0; i < args->buffer_count; i++) {
1794                         exec2_list[i].offset =
1795                                 gen8_canonical_addr(exec2_list[i].offset);
1796                         ret = __copy_to_user(&user_exec_list[i].offset,
1797                                              &exec2_list[i].offset,
1798                                              sizeof(user_exec_list[i].offset));
1799                         if (ret) {
1800                                 ret = -EFAULT;
1801                                 DRM_DEBUG("failed to copy %d exec entries "
1802                                           "back to user\n",
1803                                           args->buffer_count);
1804                                 break;
1805                         }
1806                 }
1807         }
1808
1809         drm_free_large(exec2_list);
1810         return ret;
1811 }