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
  36 #define  __EXEC_OBJECT_HAS_PIN (1<<31)
  37 #define  __EXEC_OBJECT_HAS_FENCE (1<<30)
  38 #define  __EXEC_OBJECT_NEEDS_MAP (1<<29)
  39 #define  __EXEC_OBJECT_NEEDS_BIAS (1<<28)
  40 #define  __EXEC_OBJECT_PURGEABLE (1<<27)
  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 hlist_node *node;
 197
 198                 head = &eb->buckets[handle & eb->and];
 199                 hlist_for_each(node, head) {
 200                         struct i915_vma *vma;
 201
 202                         vma = hlist_entry(node, struct i915_vma, exec_node);
 203                         if (vma->exec_handle == handle)
 204                                 return vma;
 205                 }
 206                 return NULL;
 207         }
 208 }
 209
 210 static void
 211 i915_gem_execbuffer_unreserve_vma(struct i915_vma *vma)
 212 {
 213         struct drm_i915_gem_exec_object2 *entry;
 214         struct drm_i915_gem_object *obj = vma->obj;
 215
 216         if (!drm_mm_node_allocated(&vma->node))
 217                 return;
 218
 219         entry = vma->exec_entry;
 220
 221         if (entry->flags & __EXEC_OBJECT_HAS_FENCE)
 222                 i915_gem_object_unpin_fence(obj);
 223
 224         if (entry->flags & __EXEC_OBJECT_HAS_PIN)
 225                 vma->pin_count--;
 226
 227         if (entry->flags & __EXEC_OBJECT_PURGEABLE)
 228                 obj->madv = I915_MADV_DONTNEED;
 229
 230         entry->flags &= ~(__EXEC_OBJECT_HAS_FENCE |
 231                           __EXEC_OBJECT_HAS_PIN |
 232                           __EXEC_OBJECT_PURGEABLE);
 233 }
 234
 235 static void eb_destroy(struct eb_vmas *eb)
 236 {
 237         while (!list_empty(&eb->vmas)) {
 238                 struct i915_vma *vma;
 239
 240                 vma = list_first_entry(&eb->vmas,
 241                                        struct i915_vma,
 242                                        exec_list);
 243                 list_del_init(&vma->exec_list);
 244                 i915_gem_execbuffer_unreserve_vma(vma);
 245                 drm_gem_object_unreference(&vma->obj->base);
 246         }
 247         kfree(eb);
 248 }
 249
 250 static inline int use_cpu_reloc(struct drm_i915_gem_object *obj)
 251 {
 252         return (HAS_LLC(obj->base.dev) ||
 253                 obj->base.write_domain == I915_GEM_DOMAIN_CPU ||
 254                 !obj->map_and_fenceable ||
 255                 obj->cache_level != I915_CACHE_NONE);
 256 }
 257
 258 static int
 259 relocate_entry_cpu(struct drm_i915_gem_object *obj,
 260                    struct drm_i915_gem_relocation_entry *reloc,
 261                    uint64_t target_offset)
 262 {
 263         struct drm_device *dev = obj->base.dev;
 264         uint32_t page_offset = offset_in_page(reloc->offset);
 265         uint64_t delta = reloc->delta + target_offset;
 266         char *vaddr;
 267         int ret;
 268
 269         ret = i915_gem_object_set_to_cpu_domain(obj, true);
 270         if (ret)
 271                 return ret;
 272
 273         vaddr = kmap_atomic(i915_gem_object_get_page(obj,
 274                                 reloc->offset >> PAGE_SHIFT));
 275         *(uint32_t *)(vaddr + page_offset) = lower_32_bits(delta);
 276
 277         if (INTEL_INFO(dev)->gen >= 8) {
 278                 page_offset = offset_in_page(page_offset + sizeof(uint32_t));
 279
 280                 if (page_offset == 0) {
 281                         kunmap_atomic(vaddr);
 282                         vaddr = kmap_atomic(i915_gem_object_get_page(obj,
 283                             (reloc->offset + sizeof(uint32_t)) >> PAGE_SHIFT));
 284                 }
 285
 286                 *(uint32_t *)(vaddr + page_offset) = upper_32_bits(delta);
 287         }
 288
 289         kunmap_atomic(vaddr);
 290
 291         return 0;
 292 }
 293
 294 static int
 295 relocate_entry_gtt(struct drm_i915_gem_object *obj,
 296                    struct drm_i915_gem_relocation_entry *reloc,
 297                    uint64_t target_offset)
 298 {
 299         struct drm_device *dev = obj->base.dev;
 300         struct drm_i915_private *dev_priv = dev->dev_private;
 301         uint64_t delta = reloc->delta + target_offset;
 302         uint64_t offset;
 303         void __iomem *reloc_page;
 304         int ret;
 305
 306         ret = i915_gem_object_set_to_gtt_domain(obj, true);
 307         if (ret)
 308                 return ret;
 309
 310         ret = i915_gem_object_put_fence(obj);
 311         if (ret)
 312                 return ret;
 313
 314         /* Map the page containing the relocation we're going to perform.  */
 315         offset = i915_gem_obj_ggtt_offset(obj);
 316         offset += reloc->offset;
 317         reloc_page = io_mapping_map_atomic_wc(dev_priv->gtt.mappable,
 318                                               offset & PAGE_MASK);
 319         iowrite32(lower_32_bits(delta), reloc_page + offset_in_page(offset));
 320
 321         if (INTEL_INFO(dev)->gen >= 8) {
 322                 offset += sizeof(uint32_t);
 323
 324                 if (offset_in_page(offset) == 0) {
 325                         io_mapping_unmap_atomic(reloc_page);
 326                         reloc_page =
 327                                 io_mapping_map_atomic_wc(dev_priv->gtt.mappable,
 328                                                          offset);
 329                 }
 330
 331                 iowrite32(upper_32_bits(delta),
 332                           reloc_page + offset_in_page(offset));
 333         }
 334
 335         io_mapping_unmap_atomic(reloc_page);
 336
 337         return 0;
 338 }
 339
 340 static int
 341 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
 342                                    struct eb_vmas *eb,
 343                                    struct drm_i915_gem_relocation_entry *reloc)
 344 {
 345         struct drm_device *dev = obj->base.dev;
 346         struct drm_gem_object *target_obj;
 347         struct drm_i915_gem_object *target_i915_obj;
 348         struct i915_vma *target_vma;
 349         uint64_t target_offset;
 350         int ret;
 351
 352         /* we've already hold a reference to all valid objects */
 353         target_vma = eb_get_vma(eb, reloc->target_handle);
 354         if (unlikely(target_vma == NULL))
 355                 return -ENOENT;
 356         target_i915_obj = target_vma->obj;
 357         target_obj = &target_vma->obj->base;
 358
 359         target_offset = target_vma->node.start;
 360
 361         /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
 362          * pipe_control writes because the gpu doesn't properly redirect them
 363          * through the ppgtt for non_secure batchbuffers. */
 364         if (unlikely(IS_GEN6(dev) &&
 365             reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION &&
 366             !(target_vma->bound & GLOBAL_BIND))) {
 367                 ret = i915_vma_bind(target_vma, target_i915_obj->cache_level,
 368                                     GLOBAL_BIND);
 369                 if (WARN_ONCE(ret, "Unexpected failure to bind target VMA!"))
 370                         return ret;
 371         }
 372
 373         /* Validate that the target is in a valid r/w GPU domain */
 374         if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
 375                 DRM_DEBUG("reloc with multiple write domains: "
 376                           "obj %p target %d offset %d "
 377                           "read %08x write %08x",
 378                           obj, reloc->target_handle,
 379                           (int) reloc->offset,
 380                           reloc->read_domains,
 381                           reloc->write_domain);
 382                 return -EINVAL;
 383         }
 384         if (unlikely((reloc->write_domain | reloc->read_domains)
 385                      & ~I915_GEM_GPU_DOMAINS)) {
 386                 DRM_DEBUG("reloc with read/write non-GPU domains: "
 387                           "obj %p target %d offset %d "
 388                           "read %08x write %08x",
 389                           obj, reloc->target_handle,
 390                           (int) reloc->offset,
 391                           reloc->read_domains,
 392                           reloc->write_domain);
 393                 return -EINVAL;
 394         }
 395
 396         target_obj->pending_read_domains |= reloc->read_domains;
 397         target_obj->pending_write_domain |= reloc->write_domain;
 398
 399         /* If the relocation already has the right value in it, no
 400          * more work needs to be done.
 401          */
 402         if (target_offset == reloc->presumed_offset)
 403                 return 0;
 404
 405         /* Check that the relocation address is valid... */
 406         if (unlikely(reloc->offset >
 407                 obj->base.size - (INTEL_INFO(dev)->gen >= 8 ? 8 : 4))) {
 408                 DRM_DEBUG("Relocation beyond object bounds: "
 409                           "obj %p target %d offset %d size %d.\n",
 410                           obj, reloc->target_handle,
 411                           (int) reloc->offset,
 412                           (int) obj->base.size);
 413                 return -EINVAL;
 414         }
 415         if (unlikely(reloc->offset & 3)) {
 416                 DRM_DEBUG("Relocation not 4-byte aligned: "
 417                           "obj %p target %d offset %d.\n",
 418                           obj, reloc->target_handle,
 419                           (int) reloc->offset);
 420                 return -EINVAL;
 421         }
 422
 423         /* We can't wait for rendering with pagefaults disabled */
 424         if (obj->active && in_atomic())
 425                 return -EFAULT;
 426
 427         if (use_cpu_reloc(obj))
 428                 ret = relocate_entry_cpu(obj, reloc, target_offset);
 429         else
 430                 ret = relocate_entry_gtt(obj, reloc, target_offset);
 431
 432         if (ret)
 433                 return ret;
 434
 435         /* and update the user's relocation entry */
 436         reloc->presumed_offset = target_offset;
 437
 438         return 0;
 439 }
 440
 441 static int
 442 i915_gem_execbuffer_relocate_vma(struct i915_vma *vma,
 443                                  struct eb_vmas *eb)
 444 {
 445 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
 446         struct drm_i915_gem_relocation_entry stack_reloc[N_RELOC(512)];
 447         struct drm_i915_gem_relocation_entry __user *user_relocs;
 448         struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 449         int remain, ret;
 450
 451         user_relocs = to_user_ptr(entry->relocs_ptr);
 452
 453         remain = entry->relocation_count;
 454         while (remain) {
 455                 struct drm_i915_gem_relocation_entry *r = stack_reloc;
 456                 int count = remain;
 457                 if (count > ARRAY_SIZE(stack_reloc))
 458                         count = ARRAY_SIZE(stack_reloc);
 459                 remain -= count;
 460
 461                 if (__copy_from_user_inatomic(r, user_relocs, count*sizeof(r[0])))
 462                         return -EFAULT;
 463
 464                 do {
 465                         u64 offset = r->presumed_offset;
 466
 467                         ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, r);
 468                         if (ret)
 469                                 return ret;
 470
 471                         if (r->presumed_offset != offset &&
 472                             __copy_to_user_inatomic(&user_relocs->presumed_offset,
 473                                                     &r->presumed_offset,
 474                                                     sizeof(r->presumed_offset))) {
 475                                 return -EFAULT;
 476                         }
 477
 478                         user_relocs++;
 479                         r++;
 480                 } while (--count);
 481         }
 482
 483         return 0;
 484 #undef N_RELOC
 485 }
 486
 487 static int
 488 i915_gem_execbuffer_relocate_vma_slow(struct i915_vma *vma,
 489                                       struct eb_vmas *eb,
 490                                       struct drm_i915_gem_relocation_entry *relocs)
 491 {
 492         const struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 493         int i, ret;
 494
 495         for (i = 0; i < entry->relocation_count; i++) {
 496                 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, &relocs[i]);
 497                 if (ret)
 498                         return ret;
 499         }
 500
 501         return 0;
 502 }
 503
 504 static int
 505 i915_gem_execbuffer_relocate(struct eb_vmas *eb)
 506 {
 507         struct i915_vma *vma;
 508         int ret = 0;
 509
 510         /* This is the fast path and we cannot handle a pagefault whilst
 511          * holding the struct mutex lest the user pass in the relocations
 512          * contained within a mmaped bo. For in such a case we, the page
 513          * fault handler would call i915_gem_fault() and we would try to
 514          * acquire the struct mutex again. Obviously this is bad and so
 515          * lockdep complains vehemently.
 516          */
 517         pagefault_disable();
 518         list_for_each_entry(vma, &eb->vmas, exec_list) {
 519                 ret = i915_gem_execbuffer_relocate_vma(vma, eb);
 520                 if (ret)
 521                         break;
 522         }
 523         pagefault_enable();
 524
 525         return ret;
 526 }
 527
 528 static int
 529 i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,
 530                                 struct intel_engine_cs *ring,
 531                                 bool *need_reloc)
 532 {
 533         struct drm_i915_gem_object *obj = vma->obj;
 534         struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 535         uint64_t flags;
 536         int ret;
 537
 538         flags = 0;
 539         if (entry->flags & __EXEC_OBJECT_NEEDS_MAP)
 540                 flags |= PIN_GLOBAL | PIN_MAPPABLE;
 541         if (entry->flags & EXEC_OBJECT_NEEDS_GTT)
 542                 flags |= PIN_GLOBAL;
 543         if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS)
 544                 flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS;
 545
 546         ret = i915_gem_object_pin(obj, vma->vm, entry->alignment, flags);
 547         if (ret)
 548                 return ret;
 549
 550         entry->flags |= __EXEC_OBJECT_HAS_PIN;
 551
 552         if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
 553                 ret = i915_gem_object_get_fence(obj);
 554                 if (ret)
 555                         return ret;
 556
 557                 if (i915_gem_object_pin_fence(obj))
 558                         entry->flags |= __EXEC_OBJECT_HAS_FENCE;
 559         }
 560
 561         if (entry->offset != vma->node.start) {
 562                 entry->offset = vma->node.start;
 563                 *need_reloc = true;
 564         }
 565
 566         if (entry->flags & EXEC_OBJECT_WRITE) {
 567                 obj->base.pending_read_domains = I915_GEM_DOMAIN_RENDER;
 568                 obj->base.pending_write_domain = I915_GEM_DOMAIN_RENDER;
 569         }
 570
 571         return 0;
 572 }
 573
 574 static bool
 575 need_reloc_mappable(struct i915_vma *vma)
 576 {
 577         struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 578
 579         if (entry->relocation_count == 0)
 580                 return false;
 581
 582         if (!i915_is_ggtt(vma->vm))
 583                 return false;
 584
 585         /* See also use_cpu_reloc() */
 586         if (HAS_LLC(vma->obj->base.dev))
 587                 return false;
 588
 589         if (vma->obj->base.write_domain == I915_GEM_DOMAIN_CPU)
 590                 return false;
 591
 592         return true;
 593 }
 594
 595 static bool
 596 eb_vma_misplaced(struct i915_vma *vma)
 597 {
 598         struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 599         struct drm_i915_gem_object *obj = vma->obj;
 600
 601         WARN_ON(entry->flags & __EXEC_OBJECT_NEEDS_MAP &&
 602                !i915_is_ggtt(vma->vm));
 603
 604         if (entry->alignment &&
 605             vma->node.start & (entry->alignment - 1))
 606                 return true;
 607
 608         if (entry->flags & __EXEC_OBJECT_NEEDS_MAP && !obj->map_and_fenceable)
 609                 return true;
 610
 611         if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS &&
 612             vma->node.start < BATCH_OFFSET_BIAS)
 613                 return true;
 614
 615         return false;
 616 }
 617
 618 static int
 619 i915_gem_execbuffer_reserve(struct intel_engine_cs *ring,
 620                             struct list_head *vmas,
 621                             bool *need_relocs)
 622 {
 623         struct drm_i915_gem_object *obj;
 624         struct i915_vma *vma;
 625         struct i915_address_space *vm;
 626         struct list_head ordered_vmas;
 627         bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
 628         int retry;
 629
 630         i915_gem_retire_requests_ring(ring);
 631
 632         vm = list_first_entry(vmas, struct i915_vma, exec_list)->vm;
 633
 634         INIT_LIST_HEAD(&ordered_vmas);
 635         while (!list_empty(vmas)) {
 636                 struct drm_i915_gem_exec_object2 *entry;
 637                 bool need_fence, need_mappable;
 638
 639                 vma = list_first_entry(vmas, struct i915_vma, exec_list);
 640                 obj = vma->obj;
 641                 entry = vma->exec_entry;
 642
 643                 if (!has_fenced_gpu_access)
 644                         entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE;
 645                 need_fence =
 646                         entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
 647                         obj->tiling_mode != I915_TILING_NONE;
 648                 need_mappable = need_fence || need_reloc_mappable(vma);
 649
 650                 if (need_mappable) {
 651                         entry->flags |= __EXEC_OBJECT_NEEDS_MAP;
 652                         list_move(&vma->exec_list, &ordered_vmas);
 653                 } else
 654                         list_move_tail(&vma->exec_list, &ordered_vmas);
 655
 656                 obj->base.pending_read_domains = I915_GEM_GPU_DOMAINS & ~I915_GEM_DOMAIN_COMMAND;
 657                 obj->base.pending_write_domain = 0;
 658         }
 659         list_splice(&ordered_vmas, vmas);
 660
 661         /* Attempt to pin all of the buffers into the GTT.
 662          * This is done in 3 phases:
 663          *
 664          * 1a. Unbind all objects that do not match the GTT constraints for
 665          *     the execbuffer (fenceable, mappable, alignment etc).
 666          * 1b. Increment pin count for already bound objects.
 667          * 2.  Bind new objects.
 668          * 3.  Decrement pin count.
 669          *
 670          * This avoid unnecessary unbinding of later objects in order to make
 671          * room for the earlier objects *unless* we need to defragment.
 672          */
 673         retry = 0;
 674         do {
 675                 int ret = 0;
 676
 677                 /* Unbind any ill-fitting objects or pin. */
 678                 list_for_each_entry(vma, vmas, exec_list) {
 679                         if (!drm_mm_node_allocated(&vma->node))
 680                                 continue;
 681
 682                         if (eb_vma_misplaced(vma))
 683                                 ret = i915_vma_unbind(vma);
 684                         else
 685                                 ret = i915_gem_execbuffer_reserve_vma(vma, ring, need_relocs);
 686                         if (ret)
 687                                 goto err;
 688                 }
 689
 690                 /* Bind fresh objects */
 691                 list_for_each_entry(vma, vmas, exec_list) {
 692                         if (drm_mm_node_allocated(&vma->node))
 693                                 continue;
 694
 695                         ret = i915_gem_execbuffer_reserve_vma(vma, ring, need_relocs);
 696                         if (ret)
 697                                 goto err;
 698                 }
 699
 700 err:
 701                 if (ret != -ENOSPC || retry++)
 702                         return ret;
 703
 704                 /* Decrement pin count for bound objects */
 705                 list_for_each_entry(vma, vmas, exec_list)
 706                         i915_gem_execbuffer_unreserve_vma(vma);
 707
 708                 ret = i915_gem_evict_vm(vm, true);
 709                 if (ret)
 710                         return ret;
 711         } while (1);
 712 }
 713
 714 static int
 715 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
 716                                   struct drm_i915_gem_execbuffer2 *args,
 717                                   struct drm_file *file,
 718                                   struct intel_engine_cs *ring,
 719                                   struct eb_vmas *eb,
 720                                   struct drm_i915_gem_exec_object2 *exec)
 721 {
 722         struct drm_i915_gem_relocation_entry *reloc;
 723         struct i915_address_space *vm;
 724         struct i915_vma *vma;
 725         bool need_relocs;
 726         int *reloc_offset;
 727         int i, total, ret;
 728         unsigned count = args->buffer_count;
 729
 730         vm = list_first_entry(&eb->vmas, struct i915_vma, exec_list)->vm;
 731
 732         /* We may process another execbuffer during the unlock... */
 733         while (!list_empty(&eb->vmas)) {
 734                 vma = list_first_entry(&eb->vmas, struct i915_vma, exec_list);
 735                 list_del_init(&vma->exec_list);
 736                 i915_gem_execbuffer_unreserve_vma(vma);
 737                 drm_gem_object_unreference(&vma->obj->base);
 738         }
 739
 740         mutex_unlock(&dev->struct_mutex);
 741
 742         total = 0;
 743         for (i = 0; i < count; i++)
 744                 total += exec[i].relocation_count;
 745
 746         reloc_offset = drm_malloc_ab(count, sizeof(*reloc_offset));
 747         reloc = drm_malloc_ab(total, sizeof(*reloc));
 748         if (reloc == NULL || reloc_offset == NULL) {
 749                 drm_free_large(reloc);
 750                 drm_free_large(reloc_offset);
 751                 mutex_lock(&dev->struct_mutex);
 752                 return -ENOMEM;
 753         }
 754
 755         total = 0;
 756         for (i = 0; i < count; i++) {
 757                 struct drm_i915_gem_relocation_entry __user *user_relocs;
 758                 u64 invalid_offset = (u64)-1;
 759                 int j;
 760
 761                 user_relocs = to_user_ptr(exec[i].relocs_ptr);
 762
 763                 if (copy_from_user(reloc+total, user_relocs,
 764                                    exec[i].relocation_count * sizeof(*reloc))) {
 765                         ret = -EFAULT;
 766                         mutex_lock(&dev->struct_mutex);
 767                         goto err;
 768                 }
 769
 770                 /* As we do not update the known relocation offsets after
 771                  * relocating (due to the complexities in lock handling),
 772                  * we need to mark them as invalid now so that we force the
 773                  * relocation processing next time. Just in case the target
 774                  * object is evicted and then rebound into its old
 775                  * presumed_offset before the next execbuffer - if that
 776                  * happened we would make the mistake of assuming that the
 777                  * relocations were valid.
 778                  */
 779                 for (j = 0; j < exec[i].relocation_count; j++) {
 780                         if (__copy_to_user(&user_relocs[j].presumed_offset,
 781                                            &invalid_offset,
 782                                            sizeof(invalid_offset))) {
 783                                 ret = -EFAULT;
 784                                 mutex_lock(&dev->struct_mutex);
 785                                 goto err;
 786                         }
 787                 }
 788
 789                 reloc_offset[i] = total;
 790                 total += exec[i].relocation_count;
 791         }
 792
 793         ret = i915_mutex_lock_interruptible(dev);
 794         if (ret) {
 795                 mutex_lock(&dev->struct_mutex);
 796                 goto err;
 797         }
 798
 799         /* reacquire the objects */
 800         eb_reset(eb);
 801         ret = eb_lookup_vmas(eb, exec, args, vm, file);
 802         if (ret)
 803                 goto err;
 804
 805         need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
 806         ret = i915_gem_execbuffer_reserve(ring, &eb->vmas, &need_relocs);
 807         if (ret)
 808                 goto err;
 809
 810         list_for_each_entry(vma, &eb->vmas, exec_list) {
 811                 int offset = vma->exec_entry - exec;
 812                 ret = i915_gem_execbuffer_relocate_vma_slow(vma, eb,
 813                                                             reloc + reloc_offset[offset]);
 814                 if (ret)
 815                         goto err;
 816         }
 817
 818         /* Leave the user relocations as are, this is the painfully slow path,
 819          * and we want to avoid the complication of dropping the lock whilst
 820          * having buffers reserved in the aperture and so causing spurious
 821          * ENOSPC for random operations.
 822          */
 823
 824 err:
 825         drm_free_large(reloc);
 826         drm_free_large(reloc_offset);
 827         return ret;
 828 }
 829
 830 static int
 831 i915_gem_execbuffer_move_to_gpu(struct intel_engine_cs *ring,
 832                                 struct list_head *vmas)
 833 {
 834         struct i915_vma *vma;
 835         uint32_t flush_domains = 0;
 836         bool flush_chipset = false;
 837         int ret;
 838
 839         list_for_each_entry(vma, vmas, exec_list) {
 840                 struct drm_i915_gem_object *obj = vma->obj;
 841                 ret = i915_gem_object_sync(obj, ring);
 842                 if (ret)
 843                         return ret;
 844
 845                 if (obj->base.write_domain & I915_GEM_DOMAIN_CPU)
 846                         flush_chipset |= i915_gem_clflush_object(obj, false);
 847
 848                 flush_domains |= obj->base.write_domain;
 849         }
 850
 851         if (flush_chipset)
 852                 i915_gem_chipset_flush(ring->dev);
 853
 854         if (flush_domains & I915_GEM_DOMAIN_GTT)
 855                 wmb();
 856
 857         /* Unconditionally invalidate gpu caches and ensure that we do flush
 858          * any residual writes from the previous batch.
 859          */
 860         return intel_ring_invalidate_all_caches(ring);
 861 }
 862
 863 static bool
 864 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
 865 {
 866         if (exec->flags & __I915_EXEC_UNKNOWN_FLAGS)
 867                 return false;
 868
 869         return ((exec->batch_start_offset | exec->batch_len) & 0x7) == 0;
 870 }
 871
 872 static int
 873 validate_exec_list(struct drm_device *dev,
 874                    struct drm_i915_gem_exec_object2 *exec,
 875                    int count)
 876 {
 877         unsigned relocs_total = 0;
 878         unsigned relocs_max = UINT_MAX / sizeof(struct drm_i915_gem_relocation_entry);
 879         unsigned invalid_flags;
 880         int i;
 881
 882         invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS;
 883         if (USES_FULL_PPGTT(dev))
 884                 invalid_flags |= EXEC_OBJECT_NEEDS_GTT;
 885
 886         for (i = 0; i < count; i++) {
 887                 char __user *ptr = to_user_ptr(exec[i].relocs_ptr);
 888                 int length; /* limited by fault_in_pages_readable() */
 889
 890                 if (exec[i].flags & invalid_flags)
 891                         return -EINVAL;
 892
 893                 /* First check for malicious input causing overflow in
 894                  * the worst case where we need to allocate the entire
 895                  * relocation tree as a single array.
 896                  */
 897                 if (exec[i].relocation_count > relocs_max - relocs_total)
 898                         return -EINVAL;
 899                 relocs_total += exec[i].relocation_count;
 900
 901                 length = exec[i].relocation_count *
 902                         sizeof(struct drm_i915_gem_relocation_entry);
 903                 /*
 904                  * We must check that the entire relocation array is safe
 905                  * to read, but since we may need to update the presumed
 906                  * offsets during execution, check for full write access.
 907                  */
 908                 if (!access_ok(VERIFY_WRITE, ptr, length))
 909                         return -EFAULT;
 910
 911                 if (likely(!i915.prefault_disable)) {
 912                         if (fault_in_multipages_readable(ptr, length))
 913                                 return -EFAULT;
 914                 }
 915         }
 916
 917         return 0;
 918 }
 919
 920 static struct intel_context *
 921 i915_gem_validate_context(struct drm_device *dev, struct drm_file *file,
 922                           struct intel_engine_cs *ring, const u32 ctx_id)
 923 {
 924         struct intel_context *ctx = NULL;
 925         struct i915_ctx_hang_stats *hs;
 926
 927         if (ring->id != RCS && ctx_id != DEFAULT_CONTEXT_HANDLE)
 928                 return ERR_PTR(-EINVAL);
 929
 930         ctx = i915_gem_context_get(file->driver_priv, ctx_id);
 931         if (IS_ERR(ctx))
 932                 return ctx;
 933
 934         hs = &ctx->hang_stats;
 935         if (hs->banned) {
 936                 DRM_DEBUG("Context %u tried to submit while banned\n", ctx_id);
 937                 return ERR_PTR(-EIO);
 938         }
 939
 940         if (i915.enable_execlists && !ctx->engine[ring->id].state) {
 941                 int ret = intel_lr_context_deferred_create(ctx, ring);
 942                 if (ret) {
 943                         DRM_DEBUG("Could not create LRC %u: %d\n", ctx_id, ret);
 944                         return ERR_PTR(ret);
 945                 }
 946         }
 947
 948         return ctx;
 949 }
 950
 951 void
 952 i915_gem_execbuffer_move_to_active(struct list_head *vmas,
 953                                    struct intel_engine_cs *ring)
 954 {
 955         struct drm_i915_gem_request *req = intel_ring_get_request(ring);
 956         struct i915_vma *vma;
 957
 958         list_for_each_entry(vma, vmas, exec_list) {
 959                 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 960                 struct drm_i915_gem_object *obj = vma->obj;
 961                 u32 old_read = obj->base.read_domains;
 962                 u32 old_write = obj->base.write_domain;
 963
 964                 obj->base.write_domain = obj->base.pending_write_domain;
 965                 if (obj->base.write_domain == 0)
 966                         obj->base.pending_read_domains |= obj->base.read_domains;
 967                 obj->base.read_domains = obj->base.pending_read_domains;
 968
 969                 i915_vma_move_to_active(vma, ring);
 970                 if (obj->base.write_domain) {
 971                         obj->dirty = 1;
 972                         i915_gem_request_assign(&obj->last_write_req, req);
 973
 974                         intel_fb_obj_invalidate(obj, ring);
 975
 976                         /* update for the implicit flush after a batch */
 977                         obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
 978                 }
 979                 if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
 980                         i915_gem_request_assign(&obj->last_fenced_req, req);
 981                         if (entry->flags & __EXEC_OBJECT_HAS_FENCE) {
 982                                 struct drm_i915_private *dev_priv = to_i915(ring->dev);
 983                                 list_move_tail(&dev_priv->fence_regs[obj->fence_reg].lru_list,
 984                                                &dev_priv->mm.fence_list);
 985                         }
 986                 }
 987
 988                 trace_i915_gem_object_change_domain(obj, old_read, old_write);
 989         }
 990 }
 991
 992 void
 993 i915_gem_execbuffer_retire_commands(struct drm_device *dev,
 994                                     struct drm_file *file,
 995                                     struct intel_engine_cs *ring,
 996                                     struct drm_i915_gem_object *obj)
 997 {
 998         /* Unconditionally force add_request to emit a full flush. */
 999         ring->gpu_caches_dirty = true;
1000
1001         /* Add a breadcrumb for the completion of the batch buffer */
1002         (void)__i915_add_request(ring, file, obj);
1003 }
1004
1005 static int
1006 i915_reset_gen7_sol_offsets(struct drm_device *dev,
1007                             struct intel_engine_cs *ring)
1008 {
1009         struct drm_i915_private *dev_priv = dev->dev_private;
1010         int ret, i;
1011
1012         if (!IS_GEN7(dev) || ring != &dev_priv->ring[RCS]) {
1013                 DRM_DEBUG("sol reset is gen7/rcs only\n");
1014                 return -EINVAL;
1015         }
1016
1017         ret = intel_ring_begin(ring, 4 * 3);
1018         if (ret)
1019                 return ret;
1020
1021         for (i = 0; i < 4; i++) {
1022                 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1023                 intel_ring_emit(ring, GEN7_SO_WRITE_OFFSET(i));
1024                 intel_ring_emit(ring, 0);
1025         }
1026
1027         intel_ring_advance(ring);
1028
1029         return 0;
1030 }
1031
1032 static int
1033 i915_emit_box(struct intel_engine_cs *ring,
1034               struct drm_clip_rect *box,
1035               int DR1, int DR4)
1036 {
1037         int ret;
1038
1039         if (box->y2 <= box->y1 || box->x2 <= box->x1 ||
1040             box->y2 <= 0 || box->x2 <= 0) {
1041                 DRM_ERROR("Bad box %d,%d..%d,%d\n",
1042                           box->x1, box->y1, box->x2, box->y2);
1043                 return -EINVAL;
1044         }
1045
1046         if (INTEL_INFO(ring->dev)->gen >= 4) {
1047                 ret = intel_ring_begin(ring, 4);
1048                 if (ret)
1049                         return ret;
1050
1051                 intel_ring_emit(ring, GFX_OP_DRAWRECT_INFO_I965);
1052                 intel_ring_emit(ring, (box->x1 & 0xffff) | box->y1 << 16);
1053                 intel_ring_emit(ring, ((box->x2 - 1) & 0xffff) | (box->y2 - 1) << 16);
1054                 intel_ring_emit(ring, DR4);
1055         } else {
1056                 ret = intel_ring_begin(ring, 6);
1057                 if (ret)
1058                         return ret;
1059
1060                 intel_ring_emit(ring, GFX_OP_DRAWRECT_INFO);
1061                 intel_ring_emit(ring, DR1);
1062                 intel_ring_emit(ring, (box->x1 & 0xffff) | box->y1 << 16);
1063                 intel_ring_emit(ring, ((box->x2 - 1) & 0xffff) | (box->y2 - 1) << 16);
1064                 intel_ring_emit(ring, DR4);
1065                 intel_ring_emit(ring, 0);
1066         }
1067         intel_ring_advance(ring);
1068
1069         return 0;
1070 }
1071
1072 static struct drm_i915_gem_object*
1073 i915_gem_execbuffer_parse(struct intel_engine_cs *ring,
1074                           struct drm_i915_gem_exec_object2 *shadow_exec_entry,
1075                           struct eb_vmas *eb,
1076                           struct drm_i915_gem_object *batch_obj,
1077                           u32 batch_start_offset,
1078                           u32 batch_len,
1079                           bool is_master,
1080                           u32 *flags)
1081 {
1082         struct drm_i915_private *dev_priv = to_i915(batch_obj->base.dev);
1083         struct drm_i915_gem_object *shadow_batch_obj;
1084         int ret;
1085
1086         shadow_batch_obj = i915_gem_batch_pool_get(&dev_priv->mm.batch_pool,
1087                                                    batch_obj->base.size);
1088         if (IS_ERR(shadow_batch_obj))
1089                 return shadow_batch_obj;
1090
1091         ret = i915_parse_cmds(ring,
1092                               batch_obj,
1093                               shadow_batch_obj,
1094                               batch_start_offset,
1095                               batch_len,
1096                               is_master);
1097         if (ret) {
1098                 if (ret == -EACCES)
1099                         return batch_obj;
1100         } else {
1101                 struct i915_vma *vma;
1102
1103                 memset(shadow_exec_entry, 0, sizeof(*shadow_exec_entry));
1104
1105                 vma = i915_gem_obj_to_ggtt(shadow_batch_obj);
1106                 vma->exec_entry = shadow_exec_entry;
1107                 vma->exec_entry->flags = __EXEC_OBJECT_PURGEABLE;
1108                 drm_gem_object_reference(&shadow_batch_obj->base);
1109                 list_add_tail(&vma->exec_list, &eb->vmas);
1110
1111                 shadow_batch_obj->base.pending_read_domains =
1112                         batch_obj->base.pending_read_domains;
1113
1114                 /*
1115                  * Set the DISPATCH_SECURE bit to remove the NON_SECURE
1116                  * bit from MI_BATCH_BUFFER_START commands issued in the
1117                  * dispatch_execbuffer implementations. We specifically
1118                  * don't want that set when the command parser is
1119                  * enabled.
1120                  *
1121                  * FIXME: with aliasing ppgtt, buffers that should only
1122                  * be in ggtt still end up in the aliasing ppgtt. remove
1123                  * this check when that is fixed.
1124                  */
1125                 if (USES_FULL_PPGTT(dev))
1126                         *flags |= I915_DISPATCH_SECURE;
1127         }
1128
1129         return ret ? ERR_PTR(ret) : shadow_batch_obj;
1130 }
1131
1132 int
1133 i915_gem_ringbuffer_submission(struct drm_device *dev, struct drm_file *file,
1134                                struct intel_engine_cs *ring,
1135                                struct intel_context *ctx,
1136                                struct drm_i915_gem_execbuffer2 *args,
1137                                struct list_head *vmas,
1138                                struct drm_i915_gem_object *batch_obj,
1139                                u64 exec_start, u32 flags)
1140 {
1141         struct drm_clip_rect *cliprects = NULL;
1142         struct drm_i915_private *dev_priv = dev->dev_private;
1143         u64 exec_len;
1144         int instp_mode;
1145         u32 instp_mask;
1146         int i, ret = 0;
1147
1148         if (args->num_cliprects != 0) {
1149                 if (ring != &dev_priv->ring[RCS]) {
1150                         DRM_DEBUG("clip rectangles are only valid with the render ring\n");
1151                         return -EINVAL;
1152                 }
1153
1154                 if (INTEL_INFO(dev)->gen >= 5) {
1155                         DRM_DEBUG("clip rectangles are only valid on pre-gen5\n");
1156                         return -EINVAL;
1157                 }
1158
1159                 if (args->num_cliprects > UINT_MAX / sizeof(*cliprects)) {
1160                         DRM_DEBUG("execbuf with %u cliprects\n",
1161                                   args->num_cliprects);
1162                         return -EINVAL;
1163                 }
1164
1165                 cliprects = kcalloc(args->num_cliprects,
1166                                     sizeof(*cliprects),
1167                                     GFP_KERNEL);
1168                 if (cliprects == NULL) {
1169                         ret = -ENOMEM;
1170                         goto error;
1171                 }
1172
1173                 if (copy_from_user(cliprects,
1174                                    to_user_ptr(args->cliprects_ptr),
1175                                    sizeof(*cliprects)*args->num_cliprects)) {
1176                         ret = -EFAULT;
1177                         goto error;
1178                 }
1179         } else {
1180                 if (args->DR4 == 0xffffffff) {
1181                         DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
1182                         args->DR4 = 0;
1183                 }
1184
1185                 if (args->DR1 || args->DR4 || args->cliprects_ptr) {
1186                         DRM_DEBUG("0 cliprects but dirt in cliprects fields\n");
1187                         return -EINVAL;
1188                 }
1189         }
1190
1191         ret = i915_gem_execbuffer_move_to_gpu(ring, vmas);
1192         if (ret)
1193                 goto error;
1194
1195         ret = i915_switch_context(ring, ctx);
1196         if (ret)
1197                 goto error;
1198
1199         instp_mode = args->flags & I915_EXEC_CONSTANTS_MASK;
1200         instp_mask = I915_EXEC_CONSTANTS_MASK;
1201         switch (instp_mode) {
1202         case I915_EXEC_CONSTANTS_REL_GENERAL:
1203         case I915_EXEC_CONSTANTS_ABSOLUTE:
1204         case I915_EXEC_CONSTANTS_REL_SURFACE:
1205                 if (instp_mode != 0 && ring != &dev_priv->ring[RCS]) {
1206                         DRM_DEBUG("non-0 rel constants mode on non-RCS\n");
1207                         ret = -EINVAL;
1208                         goto error;
1209                 }
1210
1211                 if (instp_mode != dev_priv->relative_constants_mode) {
1212                         if (INTEL_INFO(dev)->gen < 4) {
1213                                 DRM_DEBUG("no rel constants on pre-gen4\n");
1214                                 ret = -EINVAL;
1215                                 goto error;
1216                         }
1217
1218                         if (INTEL_INFO(dev)->gen > 5 &&
1219                             instp_mode == I915_EXEC_CONSTANTS_REL_SURFACE) {
1220                                 DRM_DEBUG("rel surface constants mode invalid on gen5+\n");
1221                                 ret = -EINVAL;
1222                                 goto error;
1223                         }
1224
1225                         /* The HW changed the meaning on this bit on gen6 */
1226                         if (INTEL_INFO(dev)->gen >= 6)
1227                                 instp_mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
1228                 }
1229                 break;
1230         default:
1231                 DRM_DEBUG("execbuf with unknown constants: %d\n", instp_mode);
1232                 ret = -EINVAL;
1233                 goto error;
1234         }
1235
1236         if (ring == &dev_priv->ring[RCS] &&
1237                         instp_mode != dev_priv->relative_constants_mode) {
1238                 ret = intel_ring_begin(ring, 4);
1239                 if (ret)
1240                         goto error;
1241
1242                 intel_ring_emit(ring, MI_NOOP);
1243                 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1244                 intel_ring_emit(ring, INSTPM);
1245                 intel_ring_emit(ring, instp_mask << 16 | instp_mode);
1246                 intel_ring_advance(ring);
1247
1248                 dev_priv->relative_constants_mode = instp_mode;
1249         }
1250
1251         if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
1252                 ret = i915_reset_gen7_sol_offsets(dev, ring);
1253                 if (ret)
1254                         goto error;
1255         }
1256
1257         exec_len = args->batch_len;
1258         if (cliprects) {
1259                 for (i = 0; i < args->num_cliprects; i++) {
1260                         ret = i915_emit_box(ring, &cliprects[i],
1261                                             args->DR1, args->DR4);
1262                         if (ret)
1263                                 goto error;
1264
1265                         ret = ring->dispatch_execbuffer(ring,
1266                                                         exec_start, exec_len,
1267                                                         flags);
1268                         if (ret)
1269                                 goto error;
1270                 }
1271         } else {
1272                 ret = ring->dispatch_execbuffer(ring,
1273                                                 exec_start, exec_len,
1274                                                 flags);
1275                 if (ret)
1276                         return ret;
1277         }
1278
1279         trace_i915_gem_ring_dispatch(intel_ring_get_request(ring), flags);
1280
1281         i915_gem_execbuffer_move_to_active(vmas, ring);
1282         i915_gem_execbuffer_retire_commands(dev, file, ring, batch_obj);
1283
1284 error:
1285         kfree(cliprects);
1286         return ret;
1287 }
1288
1289 /**
1290  * Find one BSD ring to dispatch the corresponding BSD command.
1291  * The Ring ID is returned.
1292  */
1293 static int gen8_dispatch_bsd_ring(struct drm_device *dev,
1294                                   struct drm_file *file)
1295 {
1296         struct drm_i915_private *dev_priv = dev->dev_private;
1297         struct drm_i915_file_private *file_priv = file->driver_priv;
1298
1299         /* Check whether the file_priv is using one ring */
1300         if (file_priv->bsd_ring)
1301                 return file_priv->bsd_ring->id;
1302         else {
1303                 /* If no, use the ping-pong mechanism to select one ring */
1304                 int ring_id;
1305
1306                 mutex_lock(&dev->struct_mutex);
1307                 if (dev_priv->mm.bsd_ring_dispatch_index == 0) {
1308                         ring_id = VCS;
1309                         dev_priv->mm.bsd_ring_dispatch_index = 1;
1310                 } else {
1311                         ring_id = VCS2;
1312                         dev_priv->mm.bsd_ring_dispatch_index = 0;
1313                 }
1314                 file_priv->bsd_ring = &dev_priv->ring[ring_id];
1315                 mutex_unlock(&dev->struct_mutex);
1316                 return ring_id;
1317         }
1318 }
1319
1320 static struct drm_i915_gem_object *
1321 eb_get_batch(struct eb_vmas *eb)
1322 {
1323         struct i915_vma *vma = list_entry(eb->vmas.prev, typeof(*vma), exec_list);
1324
1325         /*
1326          * SNA is doing fancy tricks with compressing batch buffers, which leads
1327          * to negative relocation deltas. Usually that works out ok since the
1328          * relocate address is still positive, except when the batch is placed
1329          * very low in the GTT. Ensure this doesn't happen.
1330          *
1331          * Note that actual hangs have only been observed on gen7, but for
1332          * paranoia do it everywhere.
1333          */
1334         vma->exec_entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
1335
1336         return vma->obj;
1337 }
1338
1339 static int
1340 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1341                        struct drm_file *file,
1342                        struct drm_i915_gem_execbuffer2 *args,
1343                        struct drm_i915_gem_exec_object2 *exec)
1344 {
1345         struct drm_i915_private *dev_priv = dev->dev_private;
1346         struct eb_vmas *eb;
1347         struct drm_i915_gem_object *batch_obj;
1348         struct drm_i915_gem_exec_object2 shadow_exec_entry;
1349         struct intel_engine_cs *ring;
1350         struct intel_context *ctx;
1351         struct i915_address_space *vm;
1352         const u32 ctx_id = i915_execbuffer2_get_context_id(*args);
1353         u64 exec_start = args->batch_start_offset;
1354         u32 flags;
1355         int ret;
1356         bool need_relocs;
1357
1358         if (!i915_gem_check_execbuffer(args))
1359                 return -EINVAL;
1360
1361         ret = validate_exec_list(dev, exec, args->buffer_count);
1362         if (ret)
1363                 return ret;
1364
1365         flags = 0;
1366         if (args->flags & I915_EXEC_SECURE) {
1367                 if (!file->is_master || !capable(CAP_SYS_ADMIN))
1368                     return -EPERM;
1369
1370                 flags |= I915_DISPATCH_SECURE;
1371         }
1372         if (args->flags & I915_EXEC_IS_PINNED)
1373                 flags |= I915_DISPATCH_PINNED;
1374
1375         if ((args->flags & I915_EXEC_RING_MASK) > LAST_USER_RING) {
1376                 DRM_DEBUG("execbuf with unknown ring: %d\n",
1377                           (int)(args->flags & I915_EXEC_RING_MASK));
1378                 return -EINVAL;
1379         }
1380
1381         if ((args->flags & I915_EXEC_RING_MASK) == I915_EXEC_DEFAULT)
1382                 ring = &dev_priv->ring[RCS];
1383         else if ((args->flags & I915_EXEC_RING_MASK) == I915_EXEC_BSD) {
1384                 if (HAS_BSD2(dev)) {
1385                         int ring_id;
1386                         ring_id = gen8_dispatch_bsd_ring(dev, file);
1387                         ring = &dev_priv->ring[ring_id];
1388                 } else
1389                         ring = &dev_priv->ring[VCS];
1390         } else
1391                 ring = &dev_priv->ring[(args->flags & I915_EXEC_RING_MASK) - 1];
1392
1393         if (!intel_ring_initialized(ring)) {
1394                 DRM_DEBUG("execbuf with invalid ring: %d\n",
1395                           (int)(args->flags & I915_EXEC_RING_MASK));
1396                 return -EINVAL;
1397         }
1398
1399         if (args->buffer_count < 1) {
1400                 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1401                 return -EINVAL;
1402         }
1403
1404         intel_runtime_pm_get(dev_priv);
1405
1406         ret = i915_mutex_lock_interruptible(dev);
1407         if (ret)
1408                 goto pre_mutex_err;
1409
1410         ctx = i915_gem_validate_context(dev, file, ring, ctx_id);
1411         if (IS_ERR(ctx)) {
1412                 mutex_unlock(&dev->struct_mutex);
1413                 ret = PTR_ERR(ctx);
1414                 goto pre_mutex_err;
1415         }
1416
1417         i915_gem_context_reference(ctx);
1418
1419         if (ctx->ppgtt)
1420                 vm = &ctx->ppgtt->base;
1421         else
1422                 vm = &dev_priv->gtt.base;
1423
1424         eb = eb_create(args);
1425         if (eb == NULL) {
1426                 i915_gem_context_unreference(ctx);
1427                 mutex_unlock(&dev->struct_mutex);
1428                 ret = -ENOMEM;
1429                 goto pre_mutex_err;
1430         }
1431
1432         /* Look up object handles */
1433         ret = eb_lookup_vmas(eb, exec, args, vm, file);
1434         if (ret)
1435                 goto err;
1436
1437         /* take note of the batch buffer before we might reorder the lists */
1438         batch_obj = eb_get_batch(eb);
1439
1440         /* Move the objects en-masse into the GTT, evicting if necessary. */
1441         need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
1442         ret = i915_gem_execbuffer_reserve(ring, &eb->vmas, &need_relocs);
1443         if (ret)
1444                 goto err;
1445
1446         /* The objects are in their final locations, apply the relocations. */
1447         if (need_relocs)
1448                 ret = i915_gem_execbuffer_relocate(eb);
1449         if (ret) {
1450                 if (ret == -EFAULT) {
1451                         ret = i915_gem_execbuffer_relocate_slow(dev, args, file, ring,
1452                                                                 eb, exec);
1453                         BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1454                 }
1455                 if (ret)
1456                         goto err;
1457         }
1458
1459         /* Set the pending read domains for the batch buffer to COMMAND */
1460         if (batch_obj->base.pending_write_domain) {
1461                 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1462                 ret = -EINVAL;
1463                 goto err;
1464         }
1465
1466         if (i915_needs_cmd_parser(ring)) {
1467                 batch_obj = i915_gem_execbuffer_parse(ring,
1468                                                       &shadow_exec_entry,
1469                                                       eb,
1470                                                       batch_obj,
1471                                                       args->batch_start_offset,
1472                                                       args->batch_len,
1473                                                       file->is_master,
1474                                                       &flags);
1475                 if (IS_ERR(batch_obj)) {
1476                         ret = PTR_ERR(batch_obj);
1477                         goto err;
1478                 }
1479         }
1480
1481         batch_obj->base.pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
1482
1483         /* snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
1484          * batch" bit. Hence we need to pin secure batches into the global gtt.
1485          * hsw should have this fixed, but bdw mucks it up again. */
1486         if (flags & I915_DISPATCH_SECURE) {
1487                 /*
1488                  * So on first glance it looks freaky that we pin the batch here
1489                  * outside of the reservation loop. But:
1490                  * - The batch is already pinned into the relevant ppgtt, so we
1491                  *   already have the backing storage fully allocated.
1492                  * - No other BO uses the global gtt (well contexts, but meh),
1493                  *   so we don't really have issues with mutliple objects not
1494                  *   fitting due to fragmentation.
1495                  * So this is actually safe.
1496                  */
1497                 ret = i915_gem_obj_ggtt_pin(batch_obj, 0, 0);
1498                 if (ret)
1499                         goto err;
1500
1501                 exec_start += i915_gem_obj_ggtt_offset(batch_obj);
1502         } else
1503                 exec_start += i915_gem_obj_offset(batch_obj, vm);
1504
1505         ret = dev_priv->gt.do_execbuf(dev, file, ring, ctx, args,
1506                                       &eb->vmas, batch_obj, exec_start, flags);
1507
1508         /*
1509          * FIXME: We crucially rely upon the active tracking for the (ppgtt)
1510          * batch vma for correctness. For less ugly and less fragility this
1511          * needs to be adjusted to also track the ggtt batch vma properly as
1512          * active.
1513          */
1514         if (flags & I915_DISPATCH_SECURE)
1515                 i915_gem_object_ggtt_unpin(batch_obj);
1516 err:
1517         /* the request owns the ref now */
1518         i915_gem_context_unreference(ctx);
1519         eb_destroy(eb);
1520
1521         mutex_unlock(&dev->struct_mutex);
1522
1523 pre_mutex_err:
1524         /* intel_gpu_busy should also get a ref, so it will free when the device
1525          * is really idle. */
1526         intel_runtime_pm_put(dev_priv);
1527         return ret;
1528 }
1529
1530 /*
1531  * Legacy execbuffer just creates an exec2 list from the original exec object
1532  * list array and passes it to the real function.
1533  */
1534 int
1535 i915_gem_execbuffer(struct drm_device *dev, void *data,
1536                     struct drm_file *file)
1537 {
1538         struct drm_i915_gem_execbuffer *args = data;
1539         struct drm_i915_gem_execbuffer2 exec2;
1540         struct drm_i915_gem_exec_object *exec_list = NULL;
1541         struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1542         int ret, i;
1543
1544         if (args->buffer_count < 1) {
1545                 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1546                 return -EINVAL;
1547         }
1548
1549         /* Copy in the exec list from userland */
1550         exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
1551         exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
1552         if (exec_list == NULL || exec2_list == NULL) {
1553                 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1554                           args->buffer_count);
1555                 drm_free_large(exec_list);
1556                 drm_free_large(exec2_list);
1557                 return -ENOMEM;
1558         }
1559         ret = copy_from_user(exec_list,
1560                              to_user_ptr(args->buffers_ptr),
1561                              sizeof(*exec_list) * args->buffer_count);
1562         if (ret != 0) {
1563                 DRM_DEBUG("copy %d exec entries failed %d\n",
1564                           args->buffer_count, ret);
1565                 drm_free_large(exec_list);
1566                 drm_free_large(exec2_list);
1567                 return -EFAULT;
1568         }
1569
1570         for (i = 0; i < args->buffer_count; i++) {
1571                 exec2_list[i].handle = exec_list[i].handle;
1572                 exec2_list[i].relocation_count = exec_list[i].relocation_count;
1573                 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
1574                 exec2_list[i].alignment = exec_list[i].alignment;
1575                 exec2_list[i].offset = exec_list[i].offset;
1576                 if (INTEL_INFO(dev)->gen < 4)
1577                         exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
1578                 else
1579                         exec2_list[i].flags = 0;
1580         }
1581
1582         exec2.buffers_ptr = args->buffers_ptr;
1583         exec2.buffer_count = args->buffer_count;
1584         exec2.batch_start_offset = args->batch_start_offset;
1585         exec2.batch_len = args->batch_len;
1586         exec2.DR1 = args->DR1;
1587         exec2.DR4 = args->DR4;
1588         exec2.num_cliprects = args->num_cliprects;
1589         exec2.cliprects_ptr = args->cliprects_ptr;
1590         exec2.flags = I915_EXEC_RENDER;
1591         i915_execbuffer2_set_context_id(exec2, 0);
1592
1593         ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list);
1594         if (!ret) {
1595                 struct drm_i915_gem_exec_object __user *user_exec_list =
1596                         to_user_ptr(args->buffers_ptr);
1597
1598                 /* Copy the new buffer offsets back to the user's exec list. */
1599                 for (i = 0; i < args->buffer_count; i++) {
1600                         ret = __copy_to_user(&user_exec_list[i].offset,
1601                                              &exec2_list[i].offset,
1602                                              sizeof(user_exec_list[i].offset));
1603                         if (ret) {
1604                                 ret = -EFAULT;
1605                                 DRM_DEBUG("failed to copy %d exec entries "
1606                                           "back to user (%d)\n",
1607                                           args->buffer_count, ret);
1608                                 break;
1609                         }
1610                 }
1611         }
1612
1613         drm_free_large(exec_list);
1614         drm_free_large(exec2_list);
1615         return ret;
1616 }
1617
1618 int
1619 i915_gem_execbuffer2(struct drm_device *dev, void *data,
1620                      struct drm_file *file)
1621 {
1622         struct drm_i915_gem_execbuffer2 *args = data;
1623         struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1624         int ret;
1625
1626         if (args->buffer_count < 1 ||
1627             args->buffer_count > UINT_MAX / sizeof(*exec2_list)) {
1628                 DRM_DEBUG("execbuf2 with %d buffers\n", args->buffer_count);
1629                 return -EINVAL;
1630         }
1631
1632         if (args->rsvd2 != 0) {
1633                 DRM_DEBUG("dirty rvsd2 field\n");
1634                 return -EINVAL;
1635         }
1636
1637         exec2_list = kmalloc(sizeof(*exec2_list)*args->buffer_count,
1638                              GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
1639         if (exec2_list == NULL)
1640                 exec2_list = drm_malloc_ab(sizeof(*exec2_list),
1641                                            args->buffer_count);
1642         if (exec2_list == NULL) {
1643                 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1644                           args->buffer_count);
1645                 return -ENOMEM;
1646         }
1647         ret = copy_from_user(exec2_list,
1648                              to_user_ptr(args->buffers_ptr),
1649                              sizeof(*exec2_list) * args->buffer_count);
1650         if (ret != 0) {
1651                 DRM_DEBUG("copy %d exec entries failed %d\n",
1652                           args->buffer_count, ret);
1653                 drm_free_large(exec2_list);
1654                 return -EFAULT;
1655         }
1656
1657         ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list);
1658         if (!ret) {
1659                 /* Copy the new buffer offsets back to the user's exec list. */
1660                 struct drm_i915_gem_exec_object2 __user *user_exec_list =
1661                                    to_user_ptr(args->buffers_ptr);
1662                 int i;
1663
1664                 for (i = 0; i < args->buffer_count; i++) {
1665                         ret = __copy_to_user(&user_exec_list[i].offset,
1666                                              &exec2_list[i].offset,
1667                                              sizeof(user_exec_list[i].offset));
1668                         if (ret) {
1669                                 ret = -EFAULT;
1670                                 DRM_DEBUG("failed to copy %d exec entries "
1671                                           "back to user\n",
1672                                           args->buffer_count);
1673                                 break;
1674                         }
1675                 }
1676         }
1677
1678         drm_free_large(exec2_list);
1679         return ret;
1680 }