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 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         entry->flags &= ~(__EXEC_OBJECT_HAS_FENCE | __EXEC_OBJECT_HAS_PIN);
 228 }
 229
 230 static void eb_destroy(struct eb_vmas *eb)
 231 {
 232         while (!list_empty(&eb->vmas)) {
 233                 struct i915_vma *vma;
 234
 235                 vma = list_first_entry(&eb->vmas,
 236                                        struct i915_vma,
 237                                        exec_list);
 238                 list_del_init(&vma->exec_list);
 239                 i915_gem_execbuffer_unreserve_vma(vma);
 240                 drm_gem_object_unreference(&vma->obj->base);
 241         }
 242         kfree(eb);
 243 }
 244
 245 static inline int use_cpu_reloc(struct drm_i915_gem_object *obj)
 246 {
 247         return (HAS_LLC(obj->base.dev) ||
 248                 obj->base.write_domain == I915_GEM_DOMAIN_CPU ||
 249                 obj->cache_level != I915_CACHE_NONE);
 250 }
 251
 252 static int
 253 relocate_entry_cpu(struct drm_i915_gem_object *obj,
 254                    struct drm_i915_gem_relocation_entry *reloc,
 255                    uint64_t target_offset)
 256 {
 257         struct drm_device *dev = obj->base.dev;
 258         uint32_t page_offset = offset_in_page(reloc->offset);
 259         uint64_t delta = reloc->delta + target_offset;
 260         char *vaddr;
 261         int ret;
 262
 263         ret = i915_gem_object_set_to_cpu_domain(obj, true);
 264         if (ret)
 265                 return ret;
 266
 267         vaddr = kmap_atomic(i915_gem_object_get_page(obj,
 268                                 reloc->offset >> PAGE_SHIFT));
 269         *(uint32_t *)(vaddr + page_offset) = lower_32_bits(delta);
 270
 271         if (INTEL_INFO(dev)->gen >= 8) {
 272                 page_offset = offset_in_page(page_offset + sizeof(uint32_t));
 273
 274                 if (page_offset == 0) {
 275                         kunmap_atomic(vaddr);
 276                         vaddr = kmap_atomic(i915_gem_object_get_page(obj,
 277                             (reloc->offset + sizeof(uint32_t)) >> PAGE_SHIFT));
 278                 }
 279
 280                 *(uint32_t *)(vaddr + page_offset) = upper_32_bits(delta);
 281         }
 282
 283         kunmap_atomic(vaddr);
 284
 285         return 0;
 286 }
 287
 288 static int
 289 relocate_entry_gtt(struct drm_i915_gem_object *obj,
 290                    struct drm_i915_gem_relocation_entry *reloc,
 291                    uint64_t target_offset)
 292 {
 293         struct drm_device *dev = obj->base.dev;
 294         struct drm_i915_private *dev_priv = dev->dev_private;
 295         uint64_t delta = reloc->delta + target_offset;
 296         uint64_t offset;
 297         void __iomem *reloc_page;
 298         int ret;
 299
 300         ret = i915_gem_object_set_to_gtt_domain(obj, true);
 301         if (ret)
 302                 return ret;
 303
 304         ret = i915_gem_object_put_fence(obj);
 305         if (ret)
 306                 return ret;
 307
 308         /* Map the page containing the relocation we're going to perform.  */
 309         offset = i915_gem_obj_ggtt_offset(obj);
 310         offset += reloc->offset;
 311         reloc_page = io_mapping_map_atomic_wc(dev_priv->gtt.mappable,
 312                                               offset & PAGE_MASK);
 313         iowrite32(lower_32_bits(delta), reloc_page + offset_in_page(offset));
 314
 315         if (INTEL_INFO(dev)->gen >= 8) {
 316                 offset += sizeof(uint32_t);
 317
 318                 if (offset_in_page(offset) == 0) {
 319                         io_mapping_unmap_atomic(reloc_page);
 320                         reloc_page =
 321                                 io_mapping_map_atomic_wc(dev_priv->gtt.mappable,
 322                                                          offset);
 323                 }
 324
 325                 iowrite32(upper_32_bits(delta),
 326                           reloc_page + offset_in_page(offset));
 327         }
 328
 329         io_mapping_unmap_atomic(reloc_page);
 330
 331         return 0;
 332 }
 333
 334 static void
 335 clflush_write32(void *addr, uint32_t value)
 336 {
 337         /* This is not a fast path, so KISS. */
 338         drm_clflush_virt_range(addr, sizeof(uint32_t));
 339         *(uint32_t *)addr = value;
 340         drm_clflush_virt_range(addr, sizeof(uint32_t));
 341 }
 342
 343 static int
 344 relocate_entry_clflush(struct drm_i915_gem_object *obj,
 345                        struct drm_i915_gem_relocation_entry *reloc,
 346                        uint64_t target_offset)
 347 {
 348         struct drm_device *dev = obj->base.dev;
 349         uint32_t page_offset = offset_in_page(reloc->offset);
 350         uint64_t delta = (int)reloc->delta + target_offset;
 351         char *vaddr;
 352         int ret;
 353
 354         ret = i915_gem_object_set_to_gtt_domain(obj, true);
 355         if (ret)
 356                 return ret;
 357
 358         vaddr = kmap_atomic(i915_gem_object_get_page(obj,
 359                                 reloc->offset >> PAGE_SHIFT));
 360         clflush_write32(vaddr + page_offset, lower_32_bits(delta));
 361
 362         if (INTEL_INFO(dev)->gen >= 8) {
 363                 page_offset = offset_in_page(page_offset + sizeof(uint32_t));
 364
 365                 if (page_offset == 0) {
 366                         kunmap_atomic(vaddr);
 367                         vaddr = kmap_atomic(i915_gem_object_get_page(obj,
 368                             (reloc->offset + sizeof(uint32_t)) >> PAGE_SHIFT));
 369                 }
 370
 371                 clflush_write32(vaddr + page_offset, upper_32_bits(delta));
 372         }
 373
 374         kunmap_atomic(vaddr);
 375
 376         return 0;
 377 }
 378
 379 static int
 380 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
 381                                    struct eb_vmas *eb,
 382                                    struct drm_i915_gem_relocation_entry *reloc)
 383 {
 384         struct drm_device *dev = obj->base.dev;
 385         struct drm_gem_object *target_obj;
 386         struct drm_i915_gem_object *target_i915_obj;
 387         struct i915_vma *target_vma;
 388         uint64_t target_offset;
 389         int ret;
 390
 391         /* we've already hold a reference to all valid objects */
 392         target_vma = eb_get_vma(eb, reloc->target_handle);
 393         if (unlikely(target_vma == NULL))
 394                 return -ENOENT;
 395         target_i915_obj = target_vma->obj;
 396         target_obj = &target_vma->obj->base;
 397
 398         target_offset = target_vma->node.start;
 399
 400         /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
 401          * pipe_control writes because the gpu doesn't properly redirect them
 402          * through the ppgtt for non_secure batchbuffers. */
 403         if (unlikely(IS_GEN6(dev) &&
 404             reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION)) {
 405                 ret = i915_vma_bind(target_vma, target_i915_obj->cache_level,
 406                                     PIN_GLOBAL);
 407                 if (WARN_ONCE(ret, "Unexpected failure to bind target VMA!"))
 408                         return ret;
 409         }
 410
 411         /* Validate that the target is in a valid r/w GPU domain */
 412         if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
 413                 DRM_DEBUG("reloc with multiple write domains: "
 414                           "obj %p target %d offset %d "
 415                           "read %08x write %08x",
 416                           obj, reloc->target_handle,
 417                           (int) reloc->offset,
 418                           reloc->read_domains,
 419                           reloc->write_domain);
 420                 return -EINVAL;
 421         }
 422         if (unlikely((reloc->write_domain | reloc->read_domains)
 423                      & ~I915_GEM_GPU_DOMAINS)) {
 424                 DRM_DEBUG("reloc with read/write non-GPU domains: "
 425                           "obj %p target %d offset %d "
 426                           "read %08x write %08x",
 427                           obj, reloc->target_handle,
 428                           (int) reloc->offset,
 429                           reloc->read_domains,
 430                           reloc->write_domain);
 431                 return -EINVAL;
 432         }
 433
 434         target_obj->pending_read_domains |= reloc->read_domains;
 435         target_obj->pending_write_domain |= reloc->write_domain;
 436
 437         /* If the relocation already has the right value in it, no
 438          * more work needs to be done.
 439          */
 440         if (target_offset == reloc->presumed_offset)
 441                 return 0;
 442
 443         /* Check that the relocation address is valid... */
 444         if (unlikely(reloc->offset >
 445                 obj->base.size - (INTEL_INFO(dev)->gen >= 8 ? 8 : 4))) {
 446                 DRM_DEBUG("Relocation beyond object bounds: "
 447                           "obj %p target %d offset %d size %d.\n",
 448                           obj, reloc->target_handle,
 449                           (int) reloc->offset,
 450                           (int) obj->base.size);
 451                 return -EINVAL;
 452         }
 453         if (unlikely(reloc->offset & 3)) {
 454                 DRM_DEBUG("Relocation not 4-byte aligned: "
 455                           "obj %p target %d offset %d.\n",
 456                           obj, reloc->target_handle,
 457                           (int) reloc->offset);
 458                 return -EINVAL;
 459         }
 460
 461         /* We can't wait for rendering with pagefaults disabled */
 462         if (obj->active && pagefault_disabled())
 463                 return -EFAULT;
 464
 465         if (use_cpu_reloc(obj))
 466                 ret = relocate_entry_cpu(obj, reloc, target_offset);
 467         else if (obj->map_and_fenceable)
 468                 ret = relocate_entry_gtt(obj, reloc, target_offset);
 469         else if (cpu_has_clflush)
 470                 ret = relocate_entry_clflush(obj, reloc, target_offset);
 471         else {
 472                 WARN_ONCE(1, "Impossible case in relocation handling\n");
 473                 ret = -ENODEV;
 474         }
 475
 476         if (ret)
 477                 return ret;
 478
 479         /* and update the user's relocation entry */
 480         reloc->presumed_offset = target_offset;
 481
 482         return 0;
 483 }
 484
 485 static int
 486 i915_gem_execbuffer_relocate_vma(struct i915_vma *vma,
 487                                  struct eb_vmas *eb)
 488 {
 489 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
 490         struct drm_i915_gem_relocation_entry stack_reloc[N_RELOC(512)];
 491         struct drm_i915_gem_relocation_entry __user *user_relocs;
 492         struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 493         int remain, ret;
 494
 495         user_relocs = to_user_ptr(entry->relocs_ptr);
 496
 497         remain = entry->relocation_count;
 498         while (remain) {
 499                 struct drm_i915_gem_relocation_entry *r = stack_reloc;
 500                 int count = remain;
 501                 if (count > ARRAY_SIZE(stack_reloc))
 502                         count = ARRAY_SIZE(stack_reloc);
 503                 remain -= count;
 504
 505                 if (__copy_from_user_inatomic(r, user_relocs, count*sizeof(r[0])))
 506                         return -EFAULT;
 507
 508                 do {
 509                         u64 offset = r->presumed_offset;
 510
 511                         ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, r);
 512                         if (ret)
 513                                 return ret;
 514
 515                         if (r->presumed_offset != offset &&
 516                             __copy_to_user_inatomic(&user_relocs->presumed_offset,
 517                                                     &r->presumed_offset,
 518                                                     sizeof(r->presumed_offset))) {
 519                                 return -EFAULT;
 520                         }
 521
 522                         user_relocs++;
 523                         r++;
 524                 } while (--count);
 525         }
 526
 527         return 0;
 528 #undef N_RELOC
 529 }
 530
 531 static int
 532 i915_gem_execbuffer_relocate_vma_slow(struct i915_vma *vma,
 533                                       struct eb_vmas *eb,
 534                                       struct drm_i915_gem_relocation_entry *relocs)
 535 {
 536         const struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 537         int i, ret;
 538
 539         for (i = 0; i < entry->relocation_count; i++) {
 540                 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, &relocs[i]);
 541                 if (ret)
 542                         return ret;
 543         }
 544
 545         return 0;
 546 }
 547
 548 static int
 549 i915_gem_execbuffer_relocate(struct eb_vmas *eb)
 550 {
 551         struct i915_vma *vma;
 552         int ret = 0;
 553
 554         /* This is the fast path and we cannot handle a pagefault whilst
 555          * holding the struct mutex lest the user pass in the relocations
 556          * contained within a mmaped bo. For in such a case we, the page
 557          * fault handler would call i915_gem_fault() and we would try to
 558          * acquire the struct mutex again. Obviously this is bad and so
 559          * lockdep complains vehemently.
 560          */
 561         pagefault_disable();
 562         list_for_each_entry(vma, &eb->vmas, exec_list) {
 563                 ret = i915_gem_execbuffer_relocate_vma(vma, eb);
 564                 if (ret)
 565                         break;
 566         }
 567         pagefault_enable();
 568
 569         return ret;
 570 }
 571
 572 static bool only_mappable_for_reloc(unsigned int flags)
 573 {
 574         return (flags & (EXEC_OBJECT_NEEDS_FENCE | __EXEC_OBJECT_NEEDS_MAP)) ==
 575                 __EXEC_OBJECT_NEEDS_MAP;
 576 }
 577
 578 static int
 579 i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,
 580                                 struct intel_engine_cs *ring,
 581                                 bool *need_reloc)
 582 {
 583         struct drm_i915_gem_object *obj = vma->obj;
 584         struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 585         uint64_t flags;
 586         int ret;
 587
 588         flags = PIN_USER;
 589         if (entry->flags & EXEC_OBJECT_NEEDS_GTT)
 590                 flags |= PIN_GLOBAL;
 591
 592         if (!drm_mm_node_allocated(&vma->node)) {
 593                 if (entry->flags & __EXEC_OBJECT_NEEDS_MAP)
 594                         flags |= PIN_GLOBAL | PIN_MAPPABLE;
 595                 if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS)
 596                         flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS;
 597         }
 598
 599         ret = i915_gem_object_pin(obj, vma->vm, entry->alignment, flags);
 600         if ((ret == -ENOSPC  || ret == -E2BIG) &&
 601             only_mappable_for_reloc(entry->flags))
 602                 ret = i915_gem_object_pin(obj, vma->vm,
 603                                           entry->alignment,
 604                                           flags & ~PIN_MAPPABLE);
 605         if (ret)
 606                 return ret;
 607
 608         entry->flags |= __EXEC_OBJECT_HAS_PIN;
 609
 610         if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
 611                 ret = i915_gem_object_get_fence(obj);
 612                 if (ret)
 613                         return ret;
 614
 615                 if (i915_gem_object_pin_fence(obj))
 616                         entry->flags |= __EXEC_OBJECT_HAS_FENCE;
 617         }
 618
 619         if (entry->offset != vma->node.start) {
 620                 entry->offset = vma->node.start;
 621                 *need_reloc = true;
 622         }
 623
 624         if (entry->flags & EXEC_OBJECT_WRITE) {
 625                 obj->base.pending_read_domains = I915_GEM_DOMAIN_RENDER;
 626                 obj->base.pending_write_domain = I915_GEM_DOMAIN_RENDER;
 627         }
 628
 629         return 0;
 630 }
 631
 632 static bool
 633 need_reloc_mappable(struct i915_vma *vma)
 634 {
 635         struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 636
 637         if (entry->relocation_count == 0)
 638                 return false;
 639
 640         if (!i915_is_ggtt(vma->vm))
 641                 return false;
 642
 643         /* See also use_cpu_reloc() */
 644         if (HAS_LLC(vma->obj->base.dev))
 645                 return false;
 646
 647         if (vma->obj->base.write_domain == I915_GEM_DOMAIN_CPU)
 648                 return false;
 649
 650         return true;
 651 }
 652
 653 static bool
 654 eb_vma_misplaced(struct i915_vma *vma)
 655 {
 656         struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 657         struct drm_i915_gem_object *obj = vma->obj;
 658
 659         WARN_ON(entry->flags & __EXEC_OBJECT_NEEDS_MAP &&
 660                !i915_is_ggtt(vma->vm));
 661
 662         if (entry->alignment &&
 663             vma->node.start & (entry->alignment - 1))
 664                 return true;
 665
 666         if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS &&
 667             vma->node.start < BATCH_OFFSET_BIAS)
 668                 return true;
 669
 670         /* avoid costly ping-pong once a batch bo ended up non-mappable */
 671         if (entry->flags & __EXEC_OBJECT_NEEDS_MAP && !obj->map_and_fenceable)
 672                 return !only_mappable_for_reloc(entry->flags);
 673
 674         return false;
 675 }
 676
 677 static int
 678 i915_gem_execbuffer_reserve(struct intel_engine_cs *ring,
 679                             struct list_head *vmas,
 680                             struct intel_context *ctx,
 681                             bool *need_relocs)
 682 {
 683         struct drm_i915_gem_object *obj;
 684         struct i915_vma *vma;
 685         struct i915_address_space *vm;
 686         struct list_head ordered_vmas;
 687         bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
 688         int retry;
 689
 690         i915_gem_retire_requests_ring(ring);
 691
 692         vm = list_first_entry(vmas, struct i915_vma, exec_list)->vm;
 693
 694         INIT_LIST_HEAD(&ordered_vmas);
 695         while (!list_empty(vmas)) {
 696                 struct drm_i915_gem_exec_object2 *entry;
 697                 bool need_fence, need_mappable;
 698
 699                 vma = list_first_entry(vmas, struct i915_vma, exec_list);
 700                 obj = vma->obj;
 701                 entry = vma->exec_entry;
 702
 703                 if (ctx->flags & CONTEXT_NO_ZEROMAP)
 704                         entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
 705
 706                 if (!has_fenced_gpu_access)
 707                         entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE;
 708                 need_fence =
 709                         entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
 710                         obj->tiling_mode != I915_TILING_NONE;
 711                 need_mappable = need_fence || need_reloc_mappable(vma);
 712
 713                 if (need_mappable) {
 714                         entry->flags |= __EXEC_OBJECT_NEEDS_MAP;
 715                         list_move(&vma->exec_list, &ordered_vmas);
 716                 } else
 717                         list_move_tail(&vma->exec_list, &ordered_vmas);
 718
 719                 obj->base.pending_read_domains = I915_GEM_GPU_DOMAINS & ~I915_GEM_DOMAIN_COMMAND;
 720                 obj->base.pending_write_domain = 0;
 721         }
 722         list_splice(&ordered_vmas, vmas);
 723
 724         /* Attempt to pin all of the buffers into the GTT.
 725          * This is done in 3 phases:
 726          *
 727          * 1a. Unbind all objects that do not match the GTT constraints for
 728          *     the execbuffer (fenceable, mappable, alignment etc).
 729          * 1b. Increment pin count for already bound objects.
 730          * 2.  Bind new objects.
 731          * 3.  Decrement pin count.
 732          *
 733          * This avoid unnecessary unbinding of later objects in order to make
 734          * room for the earlier objects *unless* we need to defragment.
 735          */
 736         retry = 0;
 737         do {
 738                 int ret = 0;
 739
 740                 /* Unbind any ill-fitting objects or pin. */
 741                 list_for_each_entry(vma, vmas, exec_list) {
 742                         if (!drm_mm_node_allocated(&vma->node))
 743                                 continue;
 744
 745                         if (eb_vma_misplaced(vma))
 746                                 ret = i915_vma_unbind(vma);
 747                         else
 748                                 ret = i915_gem_execbuffer_reserve_vma(vma, ring, need_relocs);
 749                         if (ret)
 750                                 goto err;
 751                 }
 752
 753                 /* Bind fresh objects */
 754                 list_for_each_entry(vma, vmas, exec_list) {
 755                         if (drm_mm_node_allocated(&vma->node))
 756                                 continue;
 757
 758                         ret = i915_gem_execbuffer_reserve_vma(vma, ring, need_relocs);
 759                         if (ret)
 760                                 goto err;
 761                 }
 762
 763 err:
 764                 if (ret != -ENOSPC || retry++)
 765                         return ret;
 766
 767                 /* Decrement pin count for bound objects */
 768                 list_for_each_entry(vma, vmas, exec_list)
 769                         i915_gem_execbuffer_unreserve_vma(vma);
 770
 771                 ret = i915_gem_evict_vm(vm, true);
 772                 if (ret)
 773                         return ret;
 774         } while (1);
 775 }
 776
 777 static int
 778 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
 779                                   struct drm_i915_gem_execbuffer2 *args,
 780                                   struct drm_file *file,
 781                                   struct intel_engine_cs *ring,
 782                                   struct eb_vmas *eb,
 783                                   struct drm_i915_gem_exec_object2 *exec,
 784                                   struct intel_context *ctx)
 785 {
 786         struct drm_i915_gem_relocation_entry *reloc;
 787         struct i915_address_space *vm;
 788         struct i915_vma *vma;
 789         bool need_relocs;
 790         int *reloc_offset;
 791         int i, total, ret;
 792         unsigned count = args->buffer_count;
 793
 794         vm = list_first_entry(&eb->vmas, struct i915_vma, exec_list)->vm;
 795
 796         /* We may process another execbuffer during the unlock... */
 797         while (!list_empty(&eb->vmas)) {
 798                 vma = list_first_entry(&eb->vmas, struct i915_vma, exec_list);
 799                 list_del_init(&vma->exec_list);
 800                 i915_gem_execbuffer_unreserve_vma(vma);
 801                 drm_gem_object_unreference(&vma->obj->base);
 802         }
 803
 804         mutex_unlock(&dev->struct_mutex);
 805
 806         total = 0;
 807         for (i = 0; i < count; i++)
 808                 total += exec[i].relocation_count;
 809
 810         reloc_offset = drm_malloc_ab(count, sizeof(*reloc_offset));
 811         reloc = drm_malloc_ab(total, sizeof(*reloc));
 812         if (reloc == NULL || reloc_offset == NULL) {
 813                 drm_free_large(reloc);
 814                 drm_free_large(reloc_offset);
 815                 mutex_lock(&dev->struct_mutex);
 816                 return -ENOMEM;
 817         }
 818
 819         total = 0;
 820         for (i = 0; i < count; i++) {
 821                 struct drm_i915_gem_relocation_entry __user *user_relocs;
 822                 u64 invalid_offset = (u64)-1;
 823                 int j;
 824
 825                 user_relocs = to_user_ptr(exec[i].relocs_ptr);
 826
 827                 if (copy_from_user(reloc+total, user_relocs,
 828                                    exec[i].relocation_count * sizeof(*reloc))) {
 829                         ret = -EFAULT;
 830                         mutex_lock(&dev->struct_mutex);
 831                         goto err;
 832                 }
 833
 834                 /* As we do not update the known relocation offsets after
 835                  * relocating (due to the complexities in lock handling),
 836                  * we need to mark them as invalid now so that we force the
 837                  * relocation processing next time. Just in case the target
 838                  * object is evicted and then rebound into its old
 839                  * presumed_offset before the next execbuffer - if that
 840                  * happened we would make the mistake of assuming that the
 841                  * relocations were valid.
 842                  */
 843                 for (j = 0; j < exec[i].relocation_count; j++) {
 844                         if (__copy_to_user(&user_relocs[j].presumed_offset,
 845                                            &invalid_offset,
 846                                            sizeof(invalid_offset))) {
 847                                 ret = -EFAULT;
 848                                 mutex_lock(&dev->struct_mutex);
 849                                 goto err;
 850                         }
 851                 }
 852
 853                 reloc_offset[i] = total;
 854                 total += exec[i].relocation_count;
 855         }
 856
 857         ret = i915_mutex_lock_interruptible(dev);
 858         if (ret) {
 859                 mutex_lock(&dev->struct_mutex);
 860                 goto err;
 861         }
 862
 863         /* reacquire the objects */
 864         eb_reset(eb);
 865         ret = eb_lookup_vmas(eb, exec, args, vm, file);
 866         if (ret)
 867                 goto err;
 868
 869         need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
 870         ret = i915_gem_execbuffer_reserve(ring, &eb->vmas, ctx, &need_relocs);
 871         if (ret)
 872                 goto err;
 873
 874         list_for_each_entry(vma, &eb->vmas, exec_list) {
 875                 int offset = vma->exec_entry - exec;
 876                 ret = i915_gem_execbuffer_relocate_vma_slow(vma, eb,
 877                                                             reloc + reloc_offset[offset]);
 878                 if (ret)
 879                         goto err;
 880         }
 881
 882         /* Leave the user relocations as are, this is the painfully slow path,
 883          * and we want to avoid the complication of dropping the lock whilst
 884          * having buffers reserved in the aperture and so causing spurious
 885          * ENOSPC for random operations.
 886          */
 887
 888 err:
 889         drm_free_large(reloc);
 890         drm_free_large(reloc_offset);
 891         return ret;
 892 }
 893
 894 static int
 895 i915_gem_execbuffer_move_to_gpu(struct drm_i915_gem_request *req,
 896                                 struct list_head *vmas)
 897 {
 898         const unsigned other_rings = ~intel_ring_flag(req->ring);
 899         struct i915_vma *vma;
 900         uint32_t flush_domains = 0;
 901         bool flush_chipset = false;
 902         int ret;
 903
 904         list_for_each_entry(vma, vmas, exec_list) {
 905                 struct drm_i915_gem_object *obj = vma->obj;
 906
 907                 if (obj->active & other_rings) {
 908                         ret = i915_gem_object_sync(obj, req->ring, &req);
 909                         if (ret)
 910                                 return ret;
 911                 }
 912
 913                 if (obj->base.write_domain & I915_GEM_DOMAIN_CPU)
 914                         flush_chipset |= i915_gem_clflush_object(obj, false);
 915
 916                 flush_domains |= obj->base.write_domain;
 917         }
 918
 919         if (flush_chipset)
 920                 i915_gem_chipset_flush(req->ring->dev);
 921
 922         if (flush_domains & I915_GEM_DOMAIN_GTT)
 923                 wmb();
 924
 925         /* Unconditionally invalidate gpu caches and ensure that we do flush
 926          * any residual writes from the previous batch.
 927          */
 928         return intel_ring_invalidate_all_caches(req);
 929 }
 930
 931 static bool
 932 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
 933 {
 934         if (exec->flags & __I915_EXEC_UNKNOWN_FLAGS)
 935                 return false;
 936
 937         return ((exec->batch_start_offset | exec->batch_len) & 0x7) == 0;
 938 }
 939
 940 static int
 941 validate_exec_list(struct drm_device *dev,
 942                    struct drm_i915_gem_exec_object2 *exec,
 943                    int count)
 944 {
 945         unsigned relocs_total = 0;
 946         unsigned relocs_max = UINT_MAX / sizeof(struct drm_i915_gem_relocation_entry);
 947         unsigned invalid_flags;
 948         int i;
 949
 950         invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS;
 951         if (USES_FULL_PPGTT(dev))
 952                 invalid_flags |= EXEC_OBJECT_NEEDS_GTT;
 953
 954         for (i = 0; i < count; i++) {
 955                 char __user *ptr = to_user_ptr(exec[i].relocs_ptr);
 956                 int length; /* limited by fault_in_pages_readable() */
 957
 958                 if (exec[i].flags & invalid_flags)
 959                         return -EINVAL;
 960
 961                 if (exec[i].alignment && !is_power_of_2(exec[i].alignment))
 962                         return -EINVAL;
 963
 964                 /* First check for malicious input causing overflow in
 965                  * the worst case where we need to allocate the entire
 966                  * relocation tree as a single array.
 967                  */
 968                 if (exec[i].relocation_count > relocs_max - relocs_total)
 969                         return -EINVAL;
 970                 relocs_total += exec[i].relocation_count;
 971
 972                 length = exec[i].relocation_count *
 973                         sizeof(struct drm_i915_gem_relocation_entry);
 974                 /*
 975                  * We must check that the entire relocation array is safe
 976                  * to read, but since we may need to update the presumed
 977                  * offsets during execution, check for full write access.
 978                  */
 979                 if (!access_ok(VERIFY_WRITE, ptr, length))
 980                         return -EFAULT;
 981
 982                 if (likely(!i915.prefault_disable)) {
 983                         if (fault_in_multipages_readable(ptr, length))
 984                                 return -EFAULT;
 985                 }
 986         }
 987
 988         return 0;
 989 }
 990
 991 static struct intel_context *
 992 i915_gem_validate_context(struct drm_device *dev, struct drm_file *file,
 993                           struct intel_engine_cs *ring, const u32 ctx_id)
 994 {
 995         struct intel_context *ctx = NULL;
 996         struct i915_ctx_hang_stats *hs;
 997
 998         if (ring->id != RCS && ctx_id != DEFAULT_CONTEXT_HANDLE)
 999                 return ERR_PTR(-EINVAL);
1000
1001         ctx = i915_gem_context_get(file->driver_priv, ctx_id);
1002         if (IS_ERR(ctx))
1003                 return ctx;
1004
1005         hs = &ctx->hang_stats;
1006         if (hs->banned) {
1007                 DRM_DEBUG("Context %u tried to submit while banned\n", ctx_id);
1008                 return ERR_PTR(-EIO);
1009         }
1010
1011         if (i915.enable_execlists && !ctx->engine[ring->id].state) {
1012                 int ret = intel_lr_context_deferred_create(ctx, ring);
1013                 if (ret) {
1014                         DRM_DEBUG("Could not create LRC %u: %d\n", ctx_id, ret);
1015                         return ERR_PTR(ret);
1016                 }
1017         }
1018
1019         return ctx;
1020 }
1021
1022 void
1023 i915_gem_execbuffer_move_to_active(struct list_head *vmas,
1024                                    struct drm_i915_gem_request *req)
1025 {
1026         struct intel_engine_cs *ring = i915_gem_request_get_ring(req);
1027         struct i915_vma *vma;
1028
1029         list_for_each_entry(vma, vmas, exec_list) {
1030                 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
1031                 struct drm_i915_gem_object *obj = vma->obj;
1032                 u32 old_read = obj->base.read_domains;
1033                 u32 old_write = obj->base.write_domain;
1034
1035                 obj->base.write_domain = obj->base.pending_write_domain;
1036                 if (obj->base.write_domain == 0)
1037                         obj->base.pending_read_domains |= obj->base.read_domains;
1038                 obj->base.read_domains = obj->base.pending_read_domains;
1039
1040                 i915_vma_move_to_active(vma, req);
1041                 if (obj->base.write_domain) {
1042                         obj->dirty = 1;
1043                         i915_gem_request_assign(&obj->last_write_req, req);
1044
1045                         intel_fb_obj_invalidate(obj, ORIGIN_CS);
1046
1047                         /* update for the implicit flush after a batch */
1048                         obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
1049                 }
1050                 if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
1051                         i915_gem_request_assign(&obj->last_fenced_req, req);
1052                         if (entry->flags & __EXEC_OBJECT_HAS_FENCE) {
1053                                 struct drm_i915_private *dev_priv = to_i915(ring->dev);
1054                                 list_move_tail(&dev_priv->fence_regs[obj->fence_reg].lru_list,
1055                                                &dev_priv->mm.fence_list);
1056                         }
1057                 }
1058
1059                 trace_i915_gem_object_change_domain(obj, old_read, old_write);
1060         }
1061 }
1062
1063 void
1064 i915_gem_execbuffer_retire_commands(struct i915_execbuffer_params *params)
1065 {
1066         /* Unconditionally force add_request to emit a full flush. */
1067         params->ring->gpu_caches_dirty = true;
1068
1069         /* Add a breadcrumb for the completion of the batch buffer */
1070         __i915_add_request(params->request, params->batch_obj, true);
1071 }
1072
1073 static int
1074 i915_reset_gen7_sol_offsets(struct drm_device *dev,
1075                             struct drm_i915_gem_request *req)
1076 {
1077         struct intel_engine_cs *ring = req->ring;
1078         struct drm_i915_private *dev_priv = dev->dev_private;
1079         int ret, i;
1080
1081         if (!IS_GEN7(dev) || ring != &dev_priv->ring[RCS]) {
1082                 DRM_DEBUG("sol reset is gen7/rcs only\n");
1083                 return -EINVAL;
1084         }
1085
1086         ret = intel_ring_begin(req, 4 * 3);
1087         if (ret)
1088                 return ret;
1089
1090         for (i = 0; i < 4; i++) {
1091                 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1092                 intel_ring_emit(ring, GEN7_SO_WRITE_OFFSET(i));
1093                 intel_ring_emit(ring, 0);
1094         }
1095
1096         intel_ring_advance(ring);
1097
1098         return 0;
1099 }
1100
1101 static int
1102 i915_emit_box(struct drm_i915_gem_request *req,
1103               struct drm_clip_rect *box,
1104               int DR1, int DR4)
1105 {
1106         struct intel_engine_cs *ring = req->ring;
1107         int ret;
1108
1109         if (box->y2 <= box->y1 || box->x2 <= box->x1 ||
1110             box->y2 <= 0 || box->x2 <= 0) {
1111                 DRM_ERROR("Bad box %d,%d..%d,%d\n",
1112                           box->x1, box->y1, box->x2, box->y2);
1113                 return -EINVAL;
1114         }
1115
1116         if (INTEL_INFO(ring->dev)->gen >= 4) {
1117                 ret = intel_ring_begin(req, 4);
1118                 if (ret)
1119                         return ret;
1120
1121                 intel_ring_emit(ring, GFX_OP_DRAWRECT_INFO_I965);
1122                 intel_ring_emit(ring, (box->x1 & 0xffff) | box->y1 << 16);
1123                 intel_ring_emit(ring, ((box->x2 - 1) & 0xffff) | (box->y2 - 1) << 16);
1124                 intel_ring_emit(ring, DR4);
1125         } else {
1126                 ret = intel_ring_begin(req, 6);
1127                 if (ret)
1128                         return ret;
1129
1130                 intel_ring_emit(ring, GFX_OP_DRAWRECT_INFO);
1131                 intel_ring_emit(ring, DR1);
1132                 intel_ring_emit(ring, (box->x1 & 0xffff) | box->y1 << 16);
1133                 intel_ring_emit(ring, ((box->x2 - 1) & 0xffff) | (box->y2 - 1) << 16);
1134                 intel_ring_emit(ring, DR4);
1135                 intel_ring_emit(ring, 0);
1136         }
1137         intel_ring_advance(ring);
1138
1139         return 0;
1140 }
1141
1142 static struct drm_i915_gem_object*
1143 i915_gem_execbuffer_parse(struct intel_engine_cs *ring,
1144                           struct drm_i915_gem_exec_object2 *shadow_exec_entry,
1145                           struct eb_vmas *eb,
1146                           struct drm_i915_gem_object *batch_obj,
1147                           u32 batch_start_offset,
1148                           u32 batch_len,
1149                           bool is_master)
1150 {
1151         struct drm_i915_gem_object *shadow_batch_obj;
1152         struct i915_vma *vma;
1153         int ret;
1154
1155         shadow_batch_obj = i915_gem_batch_pool_get(&ring->batch_pool,
1156                                                    PAGE_ALIGN(batch_len));
1157         if (IS_ERR(shadow_batch_obj))
1158                 return shadow_batch_obj;
1159
1160         ret = i915_parse_cmds(ring,
1161                               batch_obj,
1162                               shadow_batch_obj,
1163                               batch_start_offset,
1164                               batch_len,
1165                               is_master);
1166         if (ret)
1167                 goto err;
1168
1169         ret = i915_gem_obj_ggtt_pin(shadow_batch_obj, 0, 0);
1170         if (ret)
1171                 goto err;
1172
1173         i915_gem_object_unpin_pages(shadow_batch_obj);
1174
1175         memset(shadow_exec_entry, 0, sizeof(*shadow_exec_entry));
1176
1177         vma = i915_gem_obj_to_ggtt(shadow_batch_obj);
1178         vma->exec_entry = shadow_exec_entry;
1179         vma->exec_entry->flags = __EXEC_OBJECT_HAS_PIN;
1180         drm_gem_object_reference(&shadow_batch_obj->base);
1181         list_add_tail(&vma->exec_list, &eb->vmas);
1182
1183         shadow_batch_obj->base.pending_read_domains = I915_GEM_DOMAIN_COMMAND;
1184
1185         return shadow_batch_obj;
1186
1187 err:
1188         i915_gem_object_unpin_pages(shadow_batch_obj);
1189         if (ret == -EACCES) /* unhandled chained batch */
1190                 return batch_obj;
1191         else
1192                 return ERR_PTR(ret);
1193 }
1194
1195 int
1196 i915_gem_ringbuffer_submission(struct i915_execbuffer_params *params,
1197                                struct drm_i915_gem_execbuffer2 *args,
1198                                struct list_head *vmas)
1199 {
1200         struct drm_clip_rect *cliprects = NULL;
1201         struct drm_device *dev = params->dev;
1202         struct intel_engine_cs *ring = params->ring;
1203         struct drm_i915_private *dev_priv = dev->dev_private;
1204         u64 exec_start, exec_len;
1205         int instp_mode;
1206         u32 instp_mask;
1207         int i, ret = 0;
1208
1209         if (args->num_cliprects != 0) {
1210                 if (ring != &dev_priv->ring[RCS]) {
1211                         DRM_DEBUG("clip rectangles are only valid with the render ring\n");
1212                         return -EINVAL;
1213                 }
1214
1215                 if (INTEL_INFO(dev)->gen >= 5) {
1216                         DRM_DEBUG("clip rectangles are only valid on pre-gen5\n");
1217                         return -EINVAL;
1218                 }
1219
1220                 if (args->num_cliprects > UINT_MAX / sizeof(*cliprects)) {
1221                         DRM_DEBUG("execbuf with %u cliprects\n",
1222                                   args->num_cliprects);
1223                         return -EINVAL;
1224                 }
1225
1226                 cliprects = kcalloc(args->num_cliprects,
1227                                     sizeof(*cliprects),
1228                                     GFP_KERNEL);
1229                 if (cliprects == NULL) {
1230                         ret = -ENOMEM;
1231                         goto error;
1232                 }
1233
1234                 if (copy_from_user(cliprects,
1235                                    to_user_ptr(args->cliprects_ptr),
1236                                    sizeof(*cliprects)*args->num_cliprects)) {
1237                         ret = -EFAULT;
1238                         goto error;
1239                 }
1240         } else {
1241                 if (args->DR4 == 0xffffffff) {
1242                         DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
1243                         args->DR4 = 0;
1244                 }
1245
1246                 if (args->DR1 || args->DR4 || args->cliprects_ptr) {
1247                         DRM_DEBUG("0 cliprects but dirt in cliprects fields\n");
1248                         return -EINVAL;
1249                 }
1250         }
1251
1252         ret = i915_gem_execbuffer_move_to_gpu(params->request, vmas);
1253         if (ret)
1254                 goto error;
1255
1256         ret = i915_switch_context(params->request);
1257         if (ret)
1258                 goto error;
1259
1260         WARN(params->ctx->ppgtt && params->ctx->ppgtt->pd_dirty_rings & (1<<ring->id),
1261              "%s didn't clear reload\n", ring->name);
1262
1263         instp_mode = args->flags & I915_EXEC_CONSTANTS_MASK;
1264         instp_mask = I915_EXEC_CONSTANTS_MASK;
1265         switch (instp_mode) {
1266         case I915_EXEC_CONSTANTS_REL_GENERAL:
1267         case I915_EXEC_CONSTANTS_ABSOLUTE:
1268         case I915_EXEC_CONSTANTS_REL_SURFACE:
1269                 if (instp_mode != 0 && ring != &dev_priv->ring[RCS]) {
1270                         DRM_DEBUG("non-0 rel constants mode on non-RCS\n");
1271                         ret = -EINVAL;
1272                         goto error;
1273                 }
1274
1275                 if (instp_mode != dev_priv->relative_constants_mode) {
1276                         if (INTEL_INFO(dev)->gen < 4) {
1277                                 DRM_DEBUG("no rel constants on pre-gen4\n");
1278                                 ret = -EINVAL;
1279                                 goto error;
1280                         }
1281
1282                         if (INTEL_INFO(dev)->gen > 5 &&
1283                             instp_mode == I915_EXEC_CONSTANTS_REL_SURFACE) {
1284                                 DRM_DEBUG("rel surface constants mode invalid on gen5+\n");
1285                                 ret = -EINVAL;
1286                                 goto error;
1287                         }
1288
1289                         /* The HW changed the meaning on this bit on gen6 */
1290                         if (INTEL_INFO(dev)->gen >= 6)
1291                                 instp_mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
1292                 }
1293                 break;
1294         default:
1295                 DRM_DEBUG("execbuf with unknown constants: %d\n", instp_mode);
1296                 ret = -EINVAL;
1297                 goto error;
1298         }
1299
1300         if (ring == &dev_priv->ring[RCS] &&
1301                         instp_mode != dev_priv->relative_constants_mode) {
1302                 ret = intel_ring_begin(params->request, 4);
1303                 if (ret)
1304                         goto error;
1305
1306                 intel_ring_emit(ring, MI_NOOP);
1307                 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1308                 intel_ring_emit(ring, INSTPM);
1309                 intel_ring_emit(ring, instp_mask << 16 | instp_mode);
1310                 intel_ring_advance(ring);
1311
1312                 dev_priv->relative_constants_mode = instp_mode;
1313         }
1314
1315         if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
1316                 ret = i915_reset_gen7_sol_offsets(dev, params->request);
1317                 if (ret)
1318                         goto error;
1319         }
1320
1321         exec_len   = args->batch_len;
1322         exec_start = params->batch_obj_vm_offset +
1323                      params->args_batch_start_offset;
1324
1325         if (cliprects) {
1326                 for (i = 0; i < args->num_cliprects; i++) {
1327                         ret = i915_emit_box(params->request, &cliprects[i],
1328                                             args->DR1, args->DR4);
1329                         if (ret)
1330                                 goto error;
1331
1332                         ret = ring->dispatch_execbuffer(params->request,
1333                                                         exec_start, exec_len,
1334                                                         params->dispatch_flags);
1335                         if (ret)
1336                                 goto error;
1337                 }
1338         } else {
1339                 ret = ring->dispatch_execbuffer(params->request,
1340                                                 exec_start, exec_len,
1341                                                 params->dispatch_flags);
1342                 if (ret)
1343                         return ret;
1344         }
1345
1346         trace_i915_gem_ring_dispatch(params->request, params->dispatch_flags);
1347
1348         i915_gem_execbuffer_move_to_active(vmas, params->request);
1349         i915_gem_execbuffer_retire_commands(params);
1350
1351 error:
1352         kfree(cliprects);
1353         return ret;
1354 }
1355
1356 /**
1357  * Find one BSD ring to dispatch the corresponding BSD command.
1358  * The Ring ID is returned.
1359  */
1360 static int gen8_dispatch_bsd_ring(struct drm_device *dev,
1361                                   struct drm_file *file)
1362 {
1363         struct drm_i915_private *dev_priv = dev->dev_private;
1364         struct drm_i915_file_private *file_priv = file->driver_priv;
1365
1366         /* Check whether the file_priv is using one ring */
1367         if (file_priv->bsd_ring)
1368                 return file_priv->bsd_ring->id;
1369         else {
1370                 /* If no, use the ping-pong mechanism to select one ring */
1371                 int ring_id;
1372
1373                 mutex_lock(&dev->struct_mutex);
1374                 if (dev_priv->mm.bsd_ring_dispatch_index == 0) {
1375                         ring_id = VCS;
1376                         dev_priv->mm.bsd_ring_dispatch_index = 1;
1377                 } else {
1378                         ring_id = VCS2;
1379                         dev_priv->mm.bsd_ring_dispatch_index = 0;
1380                 }
1381                 file_priv->bsd_ring = &dev_priv->ring[ring_id];
1382                 mutex_unlock(&dev->struct_mutex);
1383                 return ring_id;
1384         }
1385 }
1386
1387 static struct drm_i915_gem_object *
1388 eb_get_batch(struct eb_vmas *eb)
1389 {
1390         struct i915_vma *vma = list_entry(eb->vmas.prev, typeof(*vma), exec_list);
1391
1392         /*
1393          * SNA is doing fancy tricks with compressing batch buffers, which leads
1394          * to negative relocation deltas. Usually that works out ok since the
1395          * relocate address is still positive, except when the batch is placed
1396          * very low in the GTT. Ensure this doesn't happen.
1397          *
1398          * Note that actual hangs have only been observed on gen7, but for
1399          * paranoia do it everywhere.
1400          */
1401         vma->exec_entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
1402
1403         return vma->obj;
1404 }
1405
1406 static int
1407 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1408                        struct drm_file *file,
1409                        struct drm_i915_gem_execbuffer2 *args,
1410                        struct drm_i915_gem_exec_object2 *exec)
1411 {
1412         struct drm_i915_private *dev_priv = dev->dev_private;
1413         struct eb_vmas *eb;
1414         struct drm_i915_gem_object *batch_obj;
1415         struct drm_i915_gem_exec_object2 shadow_exec_entry;
1416         struct intel_engine_cs *ring;
1417         struct intel_context *ctx;
1418         struct i915_address_space *vm;
1419         struct i915_execbuffer_params params_master; /* XXX: will be removed later */
1420         struct i915_execbuffer_params *params = &params_master;
1421         const u32 ctx_id = i915_execbuffer2_get_context_id(*args);
1422         u32 dispatch_flags;
1423         int ret;
1424         bool need_relocs;
1425
1426         if (!i915_gem_check_execbuffer(args))
1427                 return -EINVAL;
1428
1429         ret = validate_exec_list(dev, exec, args->buffer_count);
1430         if (ret)
1431                 return ret;
1432
1433         dispatch_flags = 0;
1434         if (args->flags & I915_EXEC_SECURE) {
1435                 if (!file->is_master || !capable(CAP_SYS_ADMIN))
1436                     return -EPERM;
1437
1438                 dispatch_flags |= I915_DISPATCH_SECURE;
1439         }
1440         if (args->flags & I915_EXEC_IS_PINNED)
1441                 dispatch_flags |= I915_DISPATCH_PINNED;
1442
1443         if ((args->flags & I915_EXEC_RING_MASK) > LAST_USER_RING) {
1444                 DRM_DEBUG("execbuf with unknown ring: %d\n",
1445                           (int)(args->flags & I915_EXEC_RING_MASK));
1446                 return -EINVAL;
1447         }
1448
1449         if (((args->flags & I915_EXEC_RING_MASK) != I915_EXEC_BSD) &&
1450             ((args->flags & I915_EXEC_BSD_MASK) != 0)) {
1451                 DRM_DEBUG("execbuf with non bsd ring but with invalid "
1452                         "bsd dispatch flags: %d\n", (int)(args->flags));
1453                 return -EINVAL;
1454         }
1455
1456         if ((args->flags & I915_EXEC_RING_MASK) == I915_EXEC_DEFAULT)
1457                 ring = &dev_priv->ring[RCS];
1458         else if ((args->flags & I915_EXEC_RING_MASK) == I915_EXEC_BSD) {
1459                 if (HAS_BSD2(dev)) {
1460                         int ring_id;
1461
1462                         switch (args->flags & I915_EXEC_BSD_MASK) {
1463                         case I915_EXEC_BSD_DEFAULT:
1464                                 ring_id = gen8_dispatch_bsd_ring(dev, file);
1465                                 ring = &dev_priv->ring[ring_id];
1466                                 break;
1467                         case I915_EXEC_BSD_RING1:
1468                                 ring = &dev_priv->ring[VCS];
1469                                 break;
1470                         case I915_EXEC_BSD_RING2:
1471                                 ring = &dev_priv->ring[VCS2];
1472                                 break;
1473                         default:
1474                                 DRM_DEBUG("execbuf with unknown bsd ring: %d\n",
1475                                           (int)(args->flags & I915_EXEC_BSD_MASK));
1476                                 return -EINVAL;
1477                         }
1478                 } else
1479                         ring = &dev_priv->ring[VCS];
1480         } else
1481                 ring = &dev_priv->ring[(args->flags & I915_EXEC_RING_MASK) - 1];
1482
1483         if (!intel_ring_initialized(ring)) {
1484                 DRM_DEBUG("execbuf with invalid ring: %d\n",
1485                           (int)(args->flags & I915_EXEC_RING_MASK));
1486                 return -EINVAL;
1487         }
1488
1489         if (args->buffer_count < 1) {
1490                 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1491                 return -EINVAL;
1492         }
1493
1494         if (args->flags & I915_EXEC_RESOURCE_STREAMER) {
1495                 if (!HAS_RESOURCE_STREAMER(dev)) {
1496                         DRM_DEBUG("RS is only allowed for Haswell, Gen8 and above\n");
1497                         return -EINVAL;
1498                 }
1499                 if (ring->id != RCS) {
1500                         DRM_DEBUG("RS is not available on %s\n",
1501                                  ring->name);
1502                         return -EINVAL;
1503                 }
1504
1505                 dispatch_flags |= I915_DISPATCH_RS;
1506         }
1507
1508         intel_runtime_pm_get(dev_priv);
1509
1510         ret = i915_mutex_lock_interruptible(dev);
1511         if (ret)
1512                 goto pre_mutex_err;
1513
1514         ctx = i915_gem_validate_context(dev, file, ring, ctx_id);
1515         if (IS_ERR(ctx)) {
1516                 mutex_unlock(&dev->struct_mutex);
1517                 ret = PTR_ERR(ctx);
1518                 goto pre_mutex_err;
1519         }
1520
1521         i915_gem_context_reference(ctx);
1522
1523         if (ctx->ppgtt)
1524                 vm = &ctx->ppgtt->base;
1525         else
1526                 vm = &dev_priv->gtt.base;
1527
1528         memset(&params_master, 0x00, sizeof(params_master));
1529
1530         eb = eb_create(args);
1531         if (eb == NULL) {
1532                 i915_gem_context_unreference(ctx);
1533                 mutex_unlock(&dev->struct_mutex);
1534                 ret = -ENOMEM;
1535                 goto pre_mutex_err;
1536         }
1537
1538         /* Look up object handles */
1539         ret = eb_lookup_vmas(eb, exec, args, vm, file);
1540         if (ret)
1541                 goto err;
1542
1543         /* take note of the batch buffer before we might reorder the lists */
1544         batch_obj = eb_get_batch(eb);
1545
1546         /* Move the objects en-masse into the GTT, evicting if necessary. */
1547         need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
1548         ret = i915_gem_execbuffer_reserve(ring, &eb->vmas, ctx, &need_relocs);
1549         if (ret)
1550                 goto err;
1551
1552         /* The objects are in their final locations, apply the relocations. */
1553         if (need_relocs)
1554                 ret = i915_gem_execbuffer_relocate(eb);
1555         if (ret) {
1556                 if (ret == -EFAULT) {
1557                         ret = i915_gem_execbuffer_relocate_slow(dev, args, file, ring,
1558                                                                 eb, exec, ctx);
1559                         BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1560                 }
1561                 if (ret)
1562                         goto err;
1563         }
1564
1565         /* Set the pending read domains for the batch buffer to COMMAND */
1566         if (batch_obj->base.pending_write_domain) {
1567                 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1568                 ret = -EINVAL;
1569                 goto err;
1570         }
1571
1572         params->args_batch_start_offset = args->batch_start_offset;
1573         if (i915_needs_cmd_parser(ring) && args->batch_len) {
1574                 struct drm_i915_gem_object *parsed_batch_obj;
1575
1576                 parsed_batch_obj = i915_gem_execbuffer_parse(ring,
1577                                                       &shadow_exec_entry,
1578                                                       eb,
1579                                                       batch_obj,
1580                                                       args->batch_start_offset,
1581                                                       args->batch_len,
1582                                                       file->is_master);
1583                 if (IS_ERR(parsed_batch_obj)) {
1584                         ret = PTR_ERR(parsed_batch_obj);
1585                         goto err;
1586                 }
1587
1588                 /*
1589                  * parsed_batch_obj == batch_obj means batch not fully parsed:
1590                  * Accept, but don't promote to secure.
1591                  */
1592
1593                 if (parsed_batch_obj != batch_obj) {
1594                         /*
1595                          * Batch parsed and accepted:
1596                          *
1597                          * Set the DISPATCH_SECURE bit to remove the NON_SECURE
1598                          * bit from MI_BATCH_BUFFER_START commands issued in
1599                          * the dispatch_execbuffer implementations. We
1600                          * specifically don't want that set on batches the
1601                          * command parser has accepted.
1602                          */
1603                         dispatch_flags |= I915_DISPATCH_SECURE;
1604                         params->args_batch_start_offset = 0;
1605                         batch_obj = parsed_batch_obj;
1606                 }
1607         }
1608
1609         batch_obj->base.pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
1610
1611         /* snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
1612          * batch" bit. Hence we need to pin secure batches into the global gtt.
1613          * hsw should have this fixed, but bdw mucks it up again. */
1614         if (dispatch_flags & I915_DISPATCH_SECURE) {
1615                 /*
1616                  * So on first glance it looks freaky that we pin the batch here
1617                  * outside of the reservation loop. But:
1618                  * - The batch is already pinned into the relevant ppgtt, so we
1619                  *   already have the backing storage fully allocated.
1620                  * - No other BO uses the global gtt (well contexts, but meh),
1621                  *   so we don't really have issues with multiple objects not
1622                  *   fitting due to fragmentation.
1623                  * So this is actually safe.
1624                  */
1625                 ret = i915_gem_obj_ggtt_pin(batch_obj, 0, 0);
1626                 if (ret)
1627                         goto err;
1628
1629                 params->batch_obj_vm_offset = i915_gem_obj_ggtt_offset(batch_obj);
1630         } else
1631                 params->batch_obj_vm_offset = i915_gem_obj_offset(batch_obj, vm);
1632
1633         /* Allocate a request for this batch buffer nice and early. */
1634         ret = i915_gem_request_alloc(ring, ctx, &params->request);
1635         if (ret)
1636                 goto err_batch_unpin;
1637
1638         ret = i915_gem_request_add_to_client(params->request, file);
1639         if (ret)
1640                 goto err_batch_unpin;
1641
1642         /*
1643          * Save assorted stuff away to pass through to *_submission().
1644          * NB: This data should be 'persistent' and not local as it will
1645          * kept around beyond the duration of the IOCTL once the GPU
1646          * scheduler arrives.
1647          */
1648         params->dev                     = dev;
1649         params->file                    = file;
1650         params->ring                    = ring;
1651         params->dispatch_flags          = dispatch_flags;
1652         params->batch_obj               = batch_obj;
1653         params->ctx                     = ctx;
1654
1655         ret = dev_priv->gt.execbuf_submit(params, args, &eb->vmas);
1656
1657 err_batch_unpin:
1658         /*
1659          * FIXME: We crucially rely upon the active tracking for the (ppgtt)
1660          * batch vma for correctness. For less ugly and less fragility this
1661          * needs to be adjusted to also track the ggtt batch vma properly as
1662          * active.
1663          */
1664         if (dispatch_flags & I915_DISPATCH_SECURE)
1665                 i915_gem_object_ggtt_unpin(batch_obj);
1666
1667 err:
1668         /* the request owns the ref now */
1669         i915_gem_context_unreference(ctx);
1670         eb_destroy(eb);
1671
1672         /*
1673          * If the request was created but not successfully submitted then it
1674          * must be freed again. If it was submitted then it is being tracked
1675          * on the active request list and no clean up is required here.
1676          */
1677         if (ret && params->request)
1678                 i915_gem_request_cancel(params->request);
1679
1680         mutex_unlock(&dev->struct_mutex);
1681
1682 pre_mutex_err:
1683         /* intel_gpu_busy should also get a ref, so it will free when the device
1684          * is really idle. */
1685         intel_runtime_pm_put(dev_priv);
1686         return ret;
1687 }
1688
1689 /*
1690  * Legacy execbuffer just creates an exec2 list from the original exec object
1691  * list array and passes it to the real function.
1692  */
1693 int
1694 i915_gem_execbuffer(struct drm_device *dev, void *data,
1695                     struct drm_file *file)
1696 {
1697         struct drm_i915_gem_execbuffer *args = data;
1698         struct drm_i915_gem_execbuffer2 exec2;
1699         struct drm_i915_gem_exec_object *exec_list = NULL;
1700         struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1701         int ret, i;
1702
1703         if (args->buffer_count < 1) {
1704                 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1705                 return -EINVAL;
1706         }
1707
1708         /* Copy in the exec list from userland */
1709         exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
1710         exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
1711         if (exec_list == NULL || exec2_list == NULL) {
1712                 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1713                           args->buffer_count);
1714                 drm_free_large(exec_list);
1715                 drm_free_large(exec2_list);
1716                 return -ENOMEM;
1717         }
1718         ret = copy_from_user(exec_list,
1719                              to_user_ptr(args->buffers_ptr),
1720                              sizeof(*exec_list) * args->buffer_count);
1721         if (ret != 0) {
1722                 DRM_DEBUG("copy %d exec entries failed %d\n",
1723                           args->buffer_count, ret);
1724                 drm_free_large(exec_list);
1725                 drm_free_large(exec2_list);
1726                 return -EFAULT;
1727         }
1728
1729         for (i = 0; i < args->buffer_count; i++) {
1730                 exec2_list[i].handle = exec_list[i].handle;
1731                 exec2_list[i].relocation_count = exec_list[i].relocation_count;
1732                 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
1733                 exec2_list[i].alignment = exec_list[i].alignment;
1734                 exec2_list[i].offset = exec_list[i].offset;
1735                 if (INTEL_INFO(dev)->gen < 4)
1736                         exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
1737                 else
1738                         exec2_list[i].flags = 0;
1739         }
1740
1741         exec2.buffers_ptr = args->buffers_ptr;
1742         exec2.buffer_count = args->buffer_count;
1743         exec2.batch_start_offset = args->batch_start_offset;
1744         exec2.batch_len = args->batch_len;
1745         exec2.DR1 = args->DR1;
1746         exec2.DR4 = args->DR4;
1747         exec2.num_cliprects = args->num_cliprects;
1748         exec2.cliprects_ptr = args->cliprects_ptr;
1749         exec2.flags = I915_EXEC_RENDER;
1750         i915_execbuffer2_set_context_id(exec2, 0);
1751
1752         ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list);
1753         if (!ret) {
1754                 struct drm_i915_gem_exec_object __user *user_exec_list =
1755                         to_user_ptr(args->buffers_ptr);
1756
1757                 /* Copy the new buffer offsets back to the user's exec list. */
1758                 for (i = 0; i < args->buffer_count; i++) {
1759                         ret = __copy_to_user(&user_exec_list[i].offset,
1760                                              &exec2_list[i].offset,
1761                                              sizeof(user_exec_list[i].offset));
1762                         if (ret) {
1763                                 ret = -EFAULT;
1764                                 DRM_DEBUG("failed to copy %d exec entries "
1765                                           "back to user (%d)\n",
1766                                           args->buffer_count, ret);
1767                                 break;
1768                         }
1769                 }
1770         }
1771
1772         drm_free_large(exec_list);
1773         drm_free_large(exec2_list);
1774         return ret;
1775 }
1776
1777 int
1778 i915_gem_execbuffer2(struct drm_device *dev, void *data,
1779                      struct drm_file *file)
1780 {
1781         struct drm_i915_gem_execbuffer2 *args = data;
1782         struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1783         int ret;
1784
1785         if (args->buffer_count < 1 ||
1786             args->buffer_count > UINT_MAX / sizeof(*exec2_list)) {
1787                 DRM_DEBUG("execbuf2 with %d buffers\n", args->buffer_count);
1788                 return -EINVAL;
1789         }
1790
1791         if (args->rsvd2 != 0) {
1792                 DRM_DEBUG("dirty rvsd2 field\n");
1793                 return -EINVAL;
1794         }
1795
1796         exec2_list = kmalloc(sizeof(*exec2_list)*args->buffer_count,
1797                              GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
1798         if (exec2_list == NULL)
1799                 exec2_list = drm_malloc_ab(sizeof(*exec2_list),
1800                                            args->buffer_count);
1801         if (exec2_list == NULL) {
1802                 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1803                           args->buffer_count);
1804                 return -ENOMEM;
1805         }
1806         ret = copy_from_user(exec2_list,
1807                              to_user_ptr(args->buffers_ptr),
1808                              sizeof(*exec2_list) * args->buffer_count);
1809         if (ret != 0) {
1810                 DRM_DEBUG("copy %d exec entries failed %d\n",
1811                           args->buffer_count, ret);
1812                 drm_free_large(exec2_list);
1813                 return -EFAULT;
1814         }
1815
1816         ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list);
1817         if (!ret) {
1818                 /* Copy the new buffer offsets back to the user's exec list. */
1819                 struct drm_i915_gem_exec_object2 __user *user_exec_list =
1820                                    to_user_ptr(args->buffers_ptr);
1821                 int i;
1822
1823                 for (i = 0; i < args->buffer_count; i++) {
1824                         ret = __copy_to_user(&user_exec_list[i].offset,
1825                                              &exec2_list[i].offset,
1826                                              sizeof(user_exec_list[i].offset));
1827                         if (ret) {
1828                                 ret = -EFAULT;
1829                                 DRM_DEBUG("failed to copy %d exec entries "
1830                                           "back to user\n",
1831                                           args->buffer_count);
1832                                 break;
1833                         }
1834                 }
1835         }
1836
1837         drm_free_large(exec2_list);
1838         return ret;
1839 }