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                 /* Wa32bitGeneralStateOffset & Wa32bitInstructionBaseOffset,
 594                  * limit address to the first 4GBs for unflagged objects.
 595                  */
 596                 if ((entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) == 0)
 597                         flags |= PIN_ZONE_4G;
 598                 if (entry->flags & __EXEC_OBJECT_NEEDS_MAP)
 599                         flags |= PIN_GLOBAL | PIN_MAPPABLE;
 600                 if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS)
 601                         flags |= BATCH_OFFSET_BIAS | PIN_OFFSET_BIAS;
 602                 if ((flags & PIN_MAPPABLE) == 0)
 603                         flags |= PIN_HIGH;
 604         }
 605
 606         ret = i915_gem_object_pin(obj, vma->vm, entry->alignment, flags);
 607         if ((ret == -ENOSPC  || ret == -E2BIG) &&
 608             only_mappable_for_reloc(entry->flags))
 609                 ret = i915_gem_object_pin(obj, vma->vm,
 610                                           entry->alignment,
 611                                           flags & ~PIN_MAPPABLE);
 612         if (ret)
 613                 return ret;
 614
 615         entry->flags |= __EXEC_OBJECT_HAS_PIN;
 616
 617         if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
 618                 ret = i915_gem_object_get_fence(obj);
 619                 if (ret)
 620                         return ret;
 621
 622                 if (i915_gem_object_pin_fence(obj))
 623                         entry->flags |= __EXEC_OBJECT_HAS_FENCE;
 624         }
 625
 626         if (entry->offset != vma->node.start) {
 627                 entry->offset = vma->node.start;
 628                 *need_reloc = true;
 629         }
 630
 631         if (entry->flags & EXEC_OBJECT_WRITE) {
 632                 obj->base.pending_read_domains = I915_GEM_DOMAIN_RENDER;
 633                 obj->base.pending_write_domain = I915_GEM_DOMAIN_RENDER;
 634         }
 635
 636         return 0;
 637 }
 638
 639 static bool
 640 need_reloc_mappable(struct i915_vma *vma)
 641 {
 642         struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 643
 644         if (entry->relocation_count == 0)
 645                 return false;
 646
 647         if (!i915_is_ggtt(vma->vm))
 648                 return false;
 649
 650         /* See also use_cpu_reloc() */
 651         if (HAS_LLC(vma->obj->base.dev))
 652                 return false;
 653
 654         if (vma->obj->base.write_domain == I915_GEM_DOMAIN_CPU)
 655                 return false;
 656
 657         return true;
 658 }
 659
 660 static bool
 661 eb_vma_misplaced(struct i915_vma *vma)
 662 {
 663         struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
 664         struct drm_i915_gem_object *obj = vma->obj;
 665
 666         WARN_ON(entry->flags & __EXEC_OBJECT_NEEDS_MAP &&
 667                !i915_is_ggtt(vma->vm));
 668
 669         if (entry->alignment &&
 670             vma->node.start & (entry->alignment - 1))
 671                 return true;
 672
 673         if (entry->flags & __EXEC_OBJECT_NEEDS_BIAS &&
 674             vma->node.start < BATCH_OFFSET_BIAS)
 675                 return true;
 676
 677         /* avoid costly ping-pong once a batch bo ended up non-mappable */
 678         if (entry->flags & __EXEC_OBJECT_NEEDS_MAP && !obj->map_and_fenceable)
 679                 return !only_mappable_for_reloc(entry->flags);
 680
 681         if ((entry->flags & EXEC_OBJECT_SUPPORTS_48B_ADDRESS) == 0 &&
 682             (vma->node.start + vma->node.size - 1) >> 32)
 683                 return true;
 684
 685         return false;
 686 }
 687
 688 static int
 689 i915_gem_execbuffer_reserve(struct intel_engine_cs *ring,
 690                             struct list_head *vmas,
 691                             struct intel_context *ctx,
 692                             bool *need_relocs)
 693 {
 694         struct drm_i915_gem_object *obj;
 695         struct i915_vma *vma;
 696         struct i915_address_space *vm;
 697         struct list_head ordered_vmas;
 698         bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
 699         int retry;
 700
 701         i915_gem_retire_requests_ring(ring);
 702
 703         vm = list_first_entry(vmas, struct i915_vma, exec_list)->vm;
 704
 705         INIT_LIST_HEAD(&ordered_vmas);
 706         while (!list_empty(vmas)) {
 707                 struct drm_i915_gem_exec_object2 *entry;
 708                 bool need_fence, need_mappable;
 709
 710                 vma = list_first_entry(vmas, struct i915_vma, exec_list);
 711                 obj = vma->obj;
 712                 entry = vma->exec_entry;
 713
 714                 if (ctx->flags & CONTEXT_NO_ZEROMAP)
 715                         entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
 716
 717                 if (!has_fenced_gpu_access)
 718                         entry->flags &= ~EXEC_OBJECT_NEEDS_FENCE;
 719                 need_fence =
 720                         entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
 721                         obj->tiling_mode != I915_TILING_NONE;
 722                 need_mappable = need_fence || need_reloc_mappable(vma);
 723
 724                 if (need_mappable) {
 725                         entry->flags |= __EXEC_OBJECT_NEEDS_MAP;
 726                         list_move(&vma->exec_list, &ordered_vmas);
 727                 } else
 728                         list_move_tail(&vma->exec_list, &ordered_vmas);
 729
 730                 obj->base.pending_read_domains = I915_GEM_GPU_DOMAINS & ~I915_GEM_DOMAIN_COMMAND;
 731                 obj->base.pending_write_domain = 0;
 732         }
 733         list_splice(&ordered_vmas, vmas);
 734
 735         /* Attempt to pin all of the buffers into the GTT.
 736          * This is done in 3 phases:
 737          *
 738          * 1a. Unbind all objects that do not match the GTT constraints for
 739          *     the execbuffer (fenceable, mappable, alignment etc).
 740          * 1b. Increment pin count for already bound objects.
 741          * 2.  Bind new objects.
 742          * 3.  Decrement pin count.
 743          *
 744          * This avoid unnecessary unbinding of later objects in order to make
 745          * room for the earlier objects *unless* we need to defragment.
 746          */
 747         retry = 0;
 748         do {
 749                 int ret = 0;
 750
 751                 /* Unbind any ill-fitting objects or pin. */
 752                 list_for_each_entry(vma, vmas, exec_list) {
 753                         if (!drm_mm_node_allocated(&vma->node))
 754                                 continue;
 755
 756                         if (eb_vma_misplaced(vma))
 757                                 ret = i915_vma_unbind(vma);
 758                         else
 759                                 ret = i915_gem_execbuffer_reserve_vma(vma, ring, need_relocs);
 760                         if (ret)
 761                                 goto err;
 762                 }
 763
 764                 /* Bind fresh objects */
 765                 list_for_each_entry(vma, vmas, exec_list) {
 766                         if (drm_mm_node_allocated(&vma->node))
 767                                 continue;
 768
 769                         ret = i915_gem_execbuffer_reserve_vma(vma, ring, need_relocs);
 770                         if (ret)
 771                                 goto err;
 772                 }
 773
 774 err:
 775                 if (ret != -ENOSPC || retry++)
 776                         return ret;
 777
 778                 /* Decrement pin count for bound objects */
 779                 list_for_each_entry(vma, vmas, exec_list)
 780                         i915_gem_execbuffer_unreserve_vma(vma);
 781
 782                 ret = i915_gem_evict_vm(vm, true);
 783                 if (ret)
 784                         return ret;
 785         } while (1);
 786 }
 787
 788 static int
 789 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
 790                                   struct drm_i915_gem_execbuffer2 *args,
 791                                   struct drm_file *file,
 792                                   struct intel_engine_cs *ring,
 793                                   struct eb_vmas *eb,
 794                                   struct drm_i915_gem_exec_object2 *exec,
 795                                   struct intel_context *ctx)
 796 {
 797         struct drm_i915_gem_relocation_entry *reloc;
 798         struct i915_address_space *vm;
 799         struct i915_vma *vma;
 800         bool need_relocs;
 801         int *reloc_offset;
 802         int i, total, ret;
 803         unsigned count = args->buffer_count;
 804
 805         vm = list_first_entry(&eb->vmas, struct i915_vma, exec_list)->vm;
 806
 807         /* We may process another execbuffer during the unlock... */
 808         while (!list_empty(&eb->vmas)) {
 809                 vma = list_first_entry(&eb->vmas, struct i915_vma, exec_list);
 810                 list_del_init(&vma->exec_list);
 811                 i915_gem_execbuffer_unreserve_vma(vma);
 812                 drm_gem_object_unreference(&vma->obj->base);
 813         }
 814
 815         mutex_unlock(&dev->struct_mutex);
 816
 817         total = 0;
 818         for (i = 0; i < count; i++)
 819                 total += exec[i].relocation_count;
 820
 821         reloc_offset = drm_malloc_ab(count, sizeof(*reloc_offset));
 822         reloc = drm_malloc_ab(total, sizeof(*reloc));
 823         if (reloc == NULL || reloc_offset == NULL) {
 824                 drm_free_large(reloc);
 825                 drm_free_large(reloc_offset);
 826                 mutex_lock(&dev->struct_mutex);
 827                 return -ENOMEM;
 828         }
 829
 830         total = 0;
 831         for (i = 0; i < count; i++) {
 832                 struct drm_i915_gem_relocation_entry __user *user_relocs;
 833                 u64 invalid_offset = (u64)-1;
 834                 int j;
 835
 836                 user_relocs = to_user_ptr(exec[i].relocs_ptr);
 837
 838                 if (copy_from_user(reloc+total, user_relocs,
 839                                    exec[i].relocation_count * sizeof(*reloc))) {
 840                         ret = -EFAULT;
 841                         mutex_lock(&dev->struct_mutex);
 842                         goto err;
 843                 }
 844
 845                 /* As we do not update the known relocation offsets after
 846                  * relocating (due to the complexities in lock handling),
 847                  * we need to mark them as invalid now so that we force the
 848                  * relocation processing next time. Just in case the target
 849                  * object is evicted and then rebound into its old
 850                  * presumed_offset before the next execbuffer - if that
 851                  * happened we would make the mistake of assuming that the
 852                  * relocations were valid.
 853                  */
 854                 for (j = 0; j < exec[i].relocation_count; j++) {
 855                         if (__copy_to_user(&user_relocs[j].presumed_offset,
 856                                            &invalid_offset,
 857                                            sizeof(invalid_offset))) {
 858                                 ret = -EFAULT;
 859                                 mutex_lock(&dev->struct_mutex);
 860                                 goto err;
 861                         }
 862                 }
 863
 864                 reloc_offset[i] = total;
 865                 total += exec[i].relocation_count;
 866         }
 867
 868         ret = i915_mutex_lock_interruptible(dev);
 869         if (ret) {
 870                 mutex_lock(&dev->struct_mutex);
 871                 goto err;
 872         }
 873
 874         /* reacquire the objects */
 875         eb_reset(eb);
 876         ret = eb_lookup_vmas(eb, exec, args, vm, file);
 877         if (ret)
 878                 goto err;
 879
 880         need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
 881         ret = i915_gem_execbuffer_reserve(ring, &eb->vmas, ctx, &need_relocs);
 882         if (ret)
 883                 goto err;
 884
 885         list_for_each_entry(vma, &eb->vmas, exec_list) {
 886                 int offset = vma->exec_entry - exec;
 887                 ret = i915_gem_execbuffer_relocate_vma_slow(vma, eb,
 888                                                             reloc + reloc_offset[offset]);
 889                 if (ret)
 890                         goto err;
 891         }
 892
 893         /* Leave the user relocations as are, this is the painfully slow path,
 894          * and we want to avoid the complication of dropping the lock whilst
 895          * having buffers reserved in the aperture and so causing spurious
 896          * ENOSPC for random operations.
 897          */
 898
 899 err:
 900         drm_free_large(reloc);
 901         drm_free_large(reloc_offset);
 902         return ret;
 903 }
 904
 905 static int
 906 i915_gem_execbuffer_move_to_gpu(struct drm_i915_gem_request *req,
 907                                 struct list_head *vmas)
 908 {
 909         const unsigned other_rings = ~intel_ring_flag(req->ring);
 910         struct i915_vma *vma;
 911         uint32_t flush_domains = 0;
 912         bool flush_chipset = false;
 913         int ret;
 914
 915         list_for_each_entry(vma, vmas, exec_list) {
 916                 struct drm_i915_gem_object *obj = vma->obj;
 917
 918                 if (obj->active & other_rings) {
 919                         ret = i915_gem_object_sync(obj, req->ring, &req);
 920                         if (ret)
 921                                 return ret;
 922                 }
 923
 924                 if (obj->base.write_domain & I915_GEM_DOMAIN_CPU)
 925                         flush_chipset |= i915_gem_clflush_object(obj, false);
 926
 927                 flush_domains |= obj->base.write_domain;
 928         }
 929
 930         if (flush_chipset)
 931                 i915_gem_chipset_flush(req->ring->dev);
 932
 933         if (flush_domains & I915_GEM_DOMAIN_GTT)
 934                 wmb();
 935
 936         /* Unconditionally invalidate gpu caches and ensure that we do flush
 937          * any residual writes from the previous batch.
 938          */
 939         return intel_ring_invalidate_all_caches(req);
 940 }
 941
 942 static bool
 943 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
 944 {
 945         if (exec->flags & __I915_EXEC_UNKNOWN_FLAGS)
 946                 return false;
 947
 948         /* Kernel clipping was a DRI1 misfeature */
 949         if (exec->num_cliprects || exec->cliprects_ptr)
 950                 return false;
 951
 952         if (exec->DR4 == 0xffffffff) {
 953                 DRM_DEBUG("UXA submitting garbage DR4, fixing up\n");
 954                 exec->DR4 = 0;
 955         }
 956         if (exec->DR1 || exec->DR4)
 957                 return false;
 958
 959         if ((exec->batch_start_offset | exec->batch_len) & 0x7)
 960                 return false;
 961
 962         return true;
 963 }
 964
 965 static int
 966 validate_exec_list(struct drm_device *dev,
 967                    struct drm_i915_gem_exec_object2 *exec,
 968                    int count)
 969 {
 970         unsigned relocs_total = 0;
 971         unsigned relocs_max = UINT_MAX / sizeof(struct drm_i915_gem_relocation_entry);
 972         unsigned invalid_flags;
 973         int i;
 974
 975         invalid_flags = __EXEC_OBJECT_UNKNOWN_FLAGS;
 976         if (USES_FULL_PPGTT(dev))
 977                 invalid_flags |= EXEC_OBJECT_NEEDS_GTT;
 978
 979         for (i = 0; i < count; i++) {
 980                 char __user *ptr = to_user_ptr(exec[i].relocs_ptr);
 981                 int length; /* limited by fault_in_pages_readable() */
 982
 983                 if (exec[i].flags & invalid_flags)
 984                         return -EINVAL;
 985
 986                 if (exec[i].alignment && !is_power_of_2(exec[i].alignment))
 987                         return -EINVAL;
 988
 989                 /* First check for malicious input causing overflow in
 990                  * the worst case where we need to allocate the entire
 991                  * relocation tree as a single array.
 992                  */
 993                 if (exec[i].relocation_count > relocs_max - relocs_total)
 994                         return -EINVAL;
 995                 relocs_total += exec[i].relocation_count;
 996
 997                 length = exec[i].relocation_count *
 998                         sizeof(struct drm_i915_gem_relocation_entry);
 999                 /*
1000                  * We must check that the entire relocation array is safe
1001                  * to read, but since we may need to update the presumed
1002                  * offsets during execution, check for full write access.
1003                  */
1004                 if (!access_ok(VERIFY_WRITE, ptr, length))
1005                         return -EFAULT;
1006
1007                 if (likely(!i915.prefault_disable)) {
1008                         if (fault_in_multipages_readable(ptr, length))
1009                                 return -EFAULT;
1010                 }
1011         }
1012
1013         return 0;
1014 }
1015
1016 static struct intel_context *
1017 i915_gem_validate_context(struct drm_device *dev, struct drm_file *file,
1018                           struct intel_engine_cs *ring, const u32 ctx_id)
1019 {
1020         struct intel_context *ctx = NULL;
1021         struct i915_ctx_hang_stats *hs;
1022
1023         if (ring->id != RCS && ctx_id != DEFAULT_CONTEXT_HANDLE)
1024                 return ERR_PTR(-EINVAL);
1025
1026         ctx = i915_gem_context_get(file->driver_priv, ctx_id);
1027         if (IS_ERR(ctx))
1028                 return ctx;
1029
1030         hs = &ctx->hang_stats;
1031         if (hs->banned) {
1032                 DRM_DEBUG("Context %u tried to submit while banned\n", ctx_id);
1033                 return ERR_PTR(-EIO);
1034         }
1035
1036         if (i915.enable_execlists && !ctx->engine[ring->id].state) {
1037                 int ret = intel_lr_context_deferred_alloc(ctx, ring);
1038                 if (ret) {
1039                         DRM_DEBUG("Could not create LRC %u: %d\n", ctx_id, ret);
1040                         return ERR_PTR(ret);
1041                 }
1042         }
1043
1044         return ctx;
1045 }
1046
1047 void
1048 i915_gem_execbuffer_move_to_active(struct list_head *vmas,
1049                                    struct drm_i915_gem_request *req)
1050 {
1051         struct intel_engine_cs *ring = i915_gem_request_get_ring(req);
1052         struct i915_vma *vma;
1053
1054         list_for_each_entry(vma, vmas, exec_list) {
1055                 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
1056                 struct drm_i915_gem_object *obj = vma->obj;
1057                 u32 old_read = obj->base.read_domains;
1058                 u32 old_write = obj->base.write_domain;
1059
1060                 obj->dirty = 1; /* be paranoid  */
1061                 obj->base.write_domain = obj->base.pending_write_domain;
1062                 if (obj->base.write_domain == 0)
1063                         obj->base.pending_read_domains |= obj->base.read_domains;
1064                 obj->base.read_domains = obj->base.pending_read_domains;
1065
1066                 i915_vma_move_to_active(vma, req);
1067                 if (obj->base.write_domain) {
1068                         i915_gem_request_assign(&obj->last_write_req, req);
1069
1070                         intel_fb_obj_invalidate(obj, ORIGIN_CS);
1071
1072                         /* update for the implicit flush after a batch */
1073                         obj->base.write_domain &= ~I915_GEM_GPU_DOMAINS;
1074                 }
1075                 if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
1076                         i915_gem_request_assign(&obj->last_fenced_req, req);
1077                         if (entry->flags & __EXEC_OBJECT_HAS_FENCE) {
1078                                 struct drm_i915_private *dev_priv = to_i915(ring->dev);
1079                                 list_move_tail(&dev_priv->fence_regs[obj->fence_reg].lru_list,
1080                                                &dev_priv->mm.fence_list);
1081                         }
1082                 }
1083
1084                 trace_i915_gem_object_change_domain(obj, old_read, old_write);
1085         }
1086 }
1087
1088 void
1089 i915_gem_execbuffer_retire_commands(struct i915_execbuffer_params *params)
1090 {
1091         /* Unconditionally force add_request to emit a full flush. */
1092         params->ring->gpu_caches_dirty = true;
1093
1094         /* Add a breadcrumb for the completion of the batch buffer */
1095         __i915_add_request(params->request, params->batch_obj, true);
1096 }
1097
1098 static int
1099 i915_reset_gen7_sol_offsets(struct drm_device *dev,
1100                             struct drm_i915_gem_request *req)
1101 {
1102         struct intel_engine_cs *ring = req->ring;
1103         struct drm_i915_private *dev_priv = dev->dev_private;
1104         int ret, i;
1105
1106         if (!IS_GEN7(dev) || ring != &dev_priv->ring[RCS]) {
1107                 DRM_DEBUG("sol reset is gen7/rcs only\n");
1108                 return -EINVAL;
1109         }
1110
1111         ret = intel_ring_begin(req, 4 * 3);
1112         if (ret)
1113                 return ret;
1114
1115         for (i = 0; i < 4; i++) {
1116                 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1117                 intel_ring_emit(ring, GEN7_SO_WRITE_OFFSET(i));
1118                 intel_ring_emit(ring, 0);
1119         }
1120
1121         intel_ring_advance(ring);
1122
1123         return 0;
1124 }
1125
1126 static struct i915_vma*
1127 shadow_batch_pin(struct drm_i915_gem_object *obj, struct i915_address_space *vm)
1128 {
1129         struct drm_i915_private *dev_priv = to_i915(obj->base.dev);
1130         struct i915_address_space *pin_vm = vm;
1131         u64 flags;
1132         int ret;
1133
1134         /*
1135          * PPGTT backed shadow buffers must be mapped RO, to prevent
1136          * post-scan tampering
1137          */
1138         if (CMDPARSER_USES_GGTT(dev_priv)) {
1139                 flags = PIN_GLOBAL;
1140                 pin_vm = &dev_priv->gtt.base;
1141         } else if (vm->has_read_only) {
1142                 flags = PIN_USER;
1143                 obj->gt_ro = 1;
1144         } else {
1145                 DRM_DEBUG("Cannot prevent post-scan tampering without RO capable vm\n");
1146                 return ERR_PTR(-EINVAL);
1147         }
1148
1149         ret = i915_gem_object_pin(obj, pin_vm, 0, flags);
1150         if (ret)
1151                 return ERR_PTR(ret);
1152         else
1153                 return i915_gem_obj_to_vma(obj, pin_vm);
1154 }
1155
1156 static struct drm_i915_gem_object*
1157 i915_gem_execbuffer_parse(struct intel_context *ctx,
1158                           struct intel_engine_cs *ring,
1159                           struct drm_i915_gem_exec_object2 *shadow_exec_entry,
1160                           struct eb_vmas *eb,
1161                           struct i915_address_space *vm,
1162                           struct drm_i915_gem_object *batch_obj,
1163                           u32 batch_start_offset,
1164                           u32 batch_len)
1165 {
1166         struct drm_i915_gem_object *shadow_batch_obj;
1167         struct i915_vma *vma;
1168         struct i915_vma *user_vma = list_entry(eb->vmas.prev,
1169                                         typeof(*user_vma), exec_list);
1170         u64 batch_start;
1171         u64 shadow_batch_start;
1172         int ret;
1173
1174         shadow_batch_obj = i915_gem_batch_pool_get(&ring->batch_pool,
1175                                                    PAGE_ALIGN(batch_len));
1176         if (IS_ERR(shadow_batch_obj))
1177                 return shadow_batch_obj;
1178
1179         vma = shadow_batch_pin(shadow_batch_obj, vm);
1180         if (IS_ERR(vma)) {
1181                 ret = PTR_ERR(vma);
1182                 goto err;
1183         }
1184
1185         batch_start = user_vma->node.start + batch_start_offset;
1186
1187         shadow_batch_start = vma->node.start;
1188
1189         ret = i915_parse_cmds(ctx,
1190                               ring,
1191                               batch_obj,
1192                               batch_start,
1193                               batch_start_offset,
1194                               batch_len,
1195                               shadow_batch_obj,
1196                               shadow_batch_start);
1197         if (ret) {
1198                 WARN_ON(vma->pin_count == 0);
1199                 vma->pin_count--;
1200                 goto err;
1201         }
1202
1203         i915_gem_object_unpin_pages(shadow_batch_obj);
1204
1205         memset(shadow_exec_entry, 0, sizeof(*shadow_exec_entry));
1206
1207         vma->exec_entry = shadow_exec_entry;
1208         vma->exec_entry->flags = __EXEC_OBJECT_HAS_PIN;
1209         drm_gem_object_reference(&shadow_batch_obj->base);
1210         list_add_tail(&vma->exec_list, &eb->vmas);
1211
1212         shadow_batch_obj->base.pending_read_domains = I915_GEM_DOMAIN_COMMAND;
1213
1214         return shadow_batch_obj;
1215
1216 err:
1217         i915_gem_object_unpin_pages(shadow_batch_obj);
1218
1219         /*
1220          * Unsafe GGTT-backed buffers can still be submitted safely
1221          * as non-secure.
1222          * For PPGTT backing however, we have no choice but to forcibly
1223          * reject unsafe buffers
1224          */
1225         if (CMDPARSER_USES_GGTT(batch_obj->base.dev) && (ret == -EACCES))
1226                 return batch_obj;
1227         else
1228                 return ERR_PTR(ret);
1229 }
1230
1231 int
1232 i915_gem_ringbuffer_submission(struct i915_execbuffer_params *params,
1233                                struct drm_i915_gem_execbuffer2 *args,
1234                                struct list_head *vmas)
1235 {
1236         struct drm_device *dev = params->dev;
1237         struct intel_engine_cs *ring = params->ring;
1238         struct drm_i915_private *dev_priv = dev->dev_private;
1239         u64 exec_start, exec_len;
1240         int instp_mode;
1241         u32 instp_mask;
1242         int ret;
1243
1244         ret = i915_gem_execbuffer_move_to_gpu(params->request, vmas);
1245         if (ret)
1246                 return ret;
1247
1248         ret = i915_switch_context(params->request);
1249         if (ret)
1250                 return ret;
1251
1252         WARN(params->ctx->ppgtt && params->ctx->ppgtt->pd_dirty_rings & (1<<ring->id),
1253              "%s didn't clear reload\n", ring->name);
1254
1255         instp_mode = args->flags & I915_EXEC_CONSTANTS_MASK;
1256         instp_mask = I915_EXEC_CONSTANTS_MASK;
1257         switch (instp_mode) {
1258         case I915_EXEC_CONSTANTS_REL_GENERAL:
1259         case I915_EXEC_CONSTANTS_ABSOLUTE:
1260         case I915_EXEC_CONSTANTS_REL_SURFACE:
1261                 if (instp_mode != 0 && ring != &dev_priv->ring[RCS]) {
1262                         DRM_DEBUG("non-0 rel constants mode on non-RCS\n");
1263                         return -EINVAL;
1264                 }
1265
1266                 if (instp_mode != dev_priv->relative_constants_mode) {
1267                         if (INTEL_INFO(dev)->gen < 4) {
1268                                 DRM_DEBUG("no rel constants on pre-gen4\n");
1269                                 return -EINVAL;
1270                         }
1271
1272                         if (INTEL_INFO(dev)->gen > 5 &&
1273                             instp_mode == I915_EXEC_CONSTANTS_REL_SURFACE) {
1274                                 DRM_DEBUG("rel surface constants mode invalid on gen5+\n");
1275                                 return -EINVAL;
1276                         }
1277
1278                         /* The HW changed the meaning on this bit on gen6 */
1279                         if (INTEL_INFO(dev)->gen >= 6)
1280                                 instp_mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
1281                 }
1282                 break;
1283         default:
1284                 DRM_DEBUG("execbuf with unknown constants: %d\n", instp_mode);
1285                 return -EINVAL;
1286         }
1287
1288         if (ring == &dev_priv->ring[RCS] &&
1289             instp_mode != dev_priv->relative_constants_mode) {
1290                 ret = intel_ring_begin(params->request, 4);
1291                 if (ret)
1292                         return ret;
1293
1294                 intel_ring_emit(ring, MI_NOOP);
1295                 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1296                 intel_ring_emit(ring, INSTPM);
1297                 intel_ring_emit(ring, instp_mask << 16 | instp_mode);
1298                 intel_ring_advance(ring);
1299
1300                 dev_priv->relative_constants_mode = instp_mode;
1301         }
1302
1303         if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
1304                 ret = i915_reset_gen7_sol_offsets(dev, params->request);
1305                 if (ret)
1306                         return ret;
1307         }
1308
1309         exec_len   = args->batch_len;
1310         exec_start = params->batch_obj_vm_offset +
1311                      params->args_batch_start_offset;
1312
1313         ret = ring->dispatch_execbuffer(params->request,
1314                                         exec_start, exec_len,
1315                                         params->dispatch_flags);
1316         if (ret)
1317                 return ret;
1318
1319         trace_i915_gem_ring_dispatch(params->request, params->dispatch_flags);
1320
1321         i915_gem_execbuffer_move_to_active(vmas, params->request);
1322         i915_gem_execbuffer_retire_commands(params);
1323
1324         return 0;
1325 }
1326
1327 /**
1328  * Find one BSD ring to dispatch the corresponding BSD command.
1329  * The Ring ID is returned.
1330  */
1331 static int gen8_dispatch_bsd_ring(struct drm_device *dev,
1332                                   struct drm_file *file)
1333 {
1334         struct drm_i915_private *dev_priv = dev->dev_private;
1335         struct drm_i915_file_private *file_priv = file->driver_priv;
1336
1337         /* Check whether the file_priv is using one ring */
1338         if (file_priv->bsd_ring)
1339                 return file_priv->bsd_ring->id;
1340         else {
1341                 /* If no, use the ping-pong mechanism to select one ring */
1342                 int ring_id;
1343
1344                 mutex_lock(&dev->struct_mutex);
1345                 if (dev_priv->mm.bsd_ring_dispatch_index == 0) {
1346                         ring_id = VCS;
1347                         dev_priv->mm.bsd_ring_dispatch_index = 1;
1348                 } else {
1349                         ring_id = VCS2;
1350                         dev_priv->mm.bsd_ring_dispatch_index = 0;
1351                 }
1352                 file_priv->bsd_ring = &dev_priv->ring[ring_id];
1353                 mutex_unlock(&dev->struct_mutex);
1354                 return ring_id;
1355         }
1356 }
1357
1358 static struct drm_i915_gem_object *
1359 eb_get_batch(struct eb_vmas *eb)
1360 {
1361         struct i915_vma *vma = list_entry(eb->vmas.prev, typeof(*vma), exec_list);
1362
1363         /*
1364          * SNA is doing fancy tricks with compressing batch buffers, which leads
1365          * to negative relocation deltas. Usually that works out ok since the
1366          * relocate address is still positive, except when the batch is placed
1367          * very low in the GTT. Ensure this doesn't happen.
1368          *
1369          * Note that actual hangs have only been observed on gen7, but for
1370          * paranoia do it everywhere.
1371          */
1372         vma->exec_entry->flags |= __EXEC_OBJECT_NEEDS_BIAS;
1373
1374         return vma->obj;
1375 }
1376
1377 static inline bool use_cmdparser(const struct intel_engine_cs *ring,
1378                                  u32 batch_len)
1379 {
1380         return ring->requires_cmd_parser ||
1381                 (ring->using_cmd_parser && batch_len && USES_PPGTT(ring->dev));
1382 }
1383
1384 static int
1385 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1386                        struct drm_file *file,
1387                        struct drm_i915_gem_execbuffer2 *args,
1388                        struct drm_i915_gem_exec_object2 *exec)
1389 {
1390         struct drm_i915_private *dev_priv = dev->dev_private;
1391         struct eb_vmas *eb;
1392         struct drm_i915_gem_object *batch_obj;
1393         struct drm_i915_gem_exec_object2 shadow_exec_entry;
1394         struct intel_engine_cs *ring;
1395         struct intel_context *ctx;
1396         struct i915_address_space *vm;
1397         struct i915_execbuffer_params params_master; /* XXX: will be removed later */
1398         struct i915_execbuffer_params *params = &params_master;
1399         const u32 ctx_id = i915_execbuffer2_get_context_id(*args);
1400         u32 dispatch_flags;
1401         int ret;
1402         bool need_relocs;
1403
1404         if (!i915_gem_check_execbuffer(args))
1405                 return -EINVAL;
1406
1407         ret = validate_exec_list(dev, exec, args->buffer_count);
1408         if (ret)
1409                 return ret;
1410
1411         dispatch_flags = 0;
1412         if (args->flags & I915_EXEC_SECURE) {
1413                 /* Return -EPERM to trigger fallback code on old binaries. */
1414                 if (!HAS_SECURE_BATCHES(dev_priv))
1415                         return -EPERM;
1416
1417                 if (!file->is_master || !capable(CAP_SYS_ADMIN))
1418                     return -EPERM;
1419
1420                 dispatch_flags |= I915_DISPATCH_SECURE;
1421         }
1422         if (args->flags & I915_EXEC_IS_PINNED)
1423                 dispatch_flags |= I915_DISPATCH_PINNED;
1424
1425         if ((args->flags & I915_EXEC_RING_MASK) > LAST_USER_RING) {
1426                 DRM_DEBUG("execbuf with unknown ring: %d\n",
1427                           (int)(args->flags & I915_EXEC_RING_MASK));
1428                 return -EINVAL;
1429         }
1430
1431         if (((args->flags & I915_EXEC_RING_MASK) != I915_EXEC_BSD) &&
1432             ((args->flags & I915_EXEC_BSD_MASK) != 0)) {
1433                 DRM_DEBUG("execbuf with non bsd ring but with invalid "
1434                         "bsd dispatch flags: %d\n", (int)(args->flags));
1435                 return -EINVAL;
1436         }
1437
1438         if ((args->flags & I915_EXEC_RING_MASK) == I915_EXEC_DEFAULT)
1439                 ring = &dev_priv->ring[RCS];
1440         else if ((args->flags & I915_EXEC_RING_MASK) == I915_EXEC_BSD) {
1441                 if (HAS_BSD2(dev)) {
1442                         int ring_id;
1443
1444                         switch (args->flags & I915_EXEC_BSD_MASK) {
1445                         case I915_EXEC_BSD_DEFAULT:
1446                                 ring_id = gen8_dispatch_bsd_ring(dev, file);
1447                                 ring = &dev_priv->ring[ring_id];
1448                                 break;
1449                         case I915_EXEC_BSD_RING1:
1450                                 ring = &dev_priv->ring[VCS];
1451                                 break;
1452                         case I915_EXEC_BSD_RING2:
1453                                 ring = &dev_priv->ring[VCS2];
1454                                 break;
1455                         default:
1456                                 DRM_DEBUG("execbuf with unknown bsd ring: %d\n",
1457                                           (int)(args->flags & I915_EXEC_BSD_MASK));
1458                                 return -EINVAL;
1459                         }
1460                 } else
1461                         ring = &dev_priv->ring[VCS];
1462         } else
1463                 ring = &dev_priv->ring[(args->flags & I915_EXEC_RING_MASK) - 1];
1464
1465         if (!intel_ring_initialized(ring)) {
1466                 DRM_DEBUG("execbuf with invalid ring: %d\n",
1467                           (int)(args->flags & I915_EXEC_RING_MASK));
1468                 return -EINVAL;
1469         }
1470
1471         if (args->buffer_count < 1) {
1472                 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1473                 return -EINVAL;
1474         }
1475
1476         if (args->flags & I915_EXEC_RESOURCE_STREAMER) {
1477                 if (!HAS_RESOURCE_STREAMER(dev)) {
1478                         DRM_DEBUG("RS is only allowed for Haswell, Gen8 and above\n");
1479                         return -EINVAL;
1480                 }
1481                 if (ring->id != RCS) {
1482                         DRM_DEBUG("RS is not available on %s\n",
1483                                  ring->name);
1484                         return -EINVAL;
1485                 }
1486
1487                 dispatch_flags |= I915_DISPATCH_RS;
1488         }
1489
1490         intel_runtime_pm_get(dev_priv);
1491
1492         ret = i915_mutex_lock_interruptible(dev);
1493         if (ret)
1494                 goto pre_mutex_err;
1495
1496         ctx = i915_gem_validate_context(dev, file, ring, ctx_id);
1497         if (IS_ERR(ctx)) {
1498                 mutex_unlock(&dev->struct_mutex);
1499                 ret = PTR_ERR(ctx);
1500                 goto pre_mutex_err;
1501         }
1502
1503         i915_gem_context_reference(ctx);
1504
1505         if (ctx->ppgtt)
1506                 vm = &ctx->ppgtt->base;
1507         else
1508                 vm = &dev_priv->gtt.base;
1509
1510         memset(&params_master, 0x00, sizeof(params_master));
1511
1512         eb = eb_create(args);
1513         if (eb == NULL) {
1514                 i915_gem_context_unreference(ctx);
1515                 mutex_unlock(&dev->struct_mutex);
1516                 ret = -ENOMEM;
1517                 goto pre_mutex_err;
1518         }
1519
1520         /* Look up object handles */
1521         ret = eb_lookup_vmas(eb, exec, args, vm, file);
1522         if (ret)
1523                 goto err;
1524
1525         /* take note of the batch buffer before we might reorder the lists */
1526         batch_obj = eb_get_batch(eb);
1527
1528         /* Move the objects en-masse into the GTT, evicting if necessary. */
1529         need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
1530         ret = i915_gem_execbuffer_reserve(ring, &eb->vmas, ctx, &need_relocs);
1531         if (ret)
1532                 goto err;
1533
1534         /* The objects are in their final locations, apply the relocations. */
1535         if (need_relocs)
1536                 ret = i915_gem_execbuffer_relocate(eb);
1537         if (ret) {
1538                 if (ret == -EFAULT) {
1539                         ret = i915_gem_execbuffer_relocate_slow(dev, args, file, ring,
1540                                                                 eb, exec, ctx);
1541                         BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1542                 }
1543                 if (ret)
1544                         goto err;
1545         }
1546
1547         /* Set the pending read domains for the batch buffer to COMMAND */
1548         if (batch_obj->base.pending_write_domain) {
1549                 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1550                 ret = -EINVAL;
1551                 goto err;
1552         }
1553
1554         params->args_batch_start_offset = args->batch_start_offset;
1555         if (use_cmdparser(ring, args->batch_len)) {
1556                 struct drm_i915_gem_object *parsed_batch_obj;
1557
1558                 u32 batch_off = args->batch_start_offset;
1559                 u32 batch_len = args->batch_len;
1560                 if (batch_len == 0)
1561                         batch_len = batch_obj->base.size - batch_off;
1562
1563                 parsed_batch_obj = i915_gem_execbuffer_parse(ctx, ring,
1564                                                       &shadow_exec_entry,
1565                                                       eb, vm,
1566                                                       batch_obj,
1567                                                       batch_off,
1568                                                       batch_len);
1569                 if (IS_ERR(parsed_batch_obj)) {
1570                         ret = PTR_ERR(parsed_batch_obj);
1571                         goto err;
1572                 }
1573
1574                 /*
1575                  * parsed_batch_obj == batch_obj means batch not fully parsed:
1576                  * Accept, but don't promote to secure.
1577                  */
1578                 if (parsed_batch_obj != batch_obj) {
1579                         if (CMDPARSER_USES_GGTT(dev_priv))
1580                                 dispatch_flags |= I915_DISPATCH_SECURE;
1581                         params->args_batch_start_offset = 0;
1582                         batch_obj = parsed_batch_obj;
1583                 }
1584         }
1585
1586         batch_obj->base.pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
1587
1588         /* snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
1589          * batch" bit. Hence we need to pin secure batches into the global gtt.
1590          * hsw should have this fixed, but bdw mucks it up again. */
1591         if (dispatch_flags & I915_DISPATCH_SECURE) {
1592                 /*
1593                  * So on first glance it looks freaky that we pin the batch here
1594                  * outside of the reservation loop. But:
1595                  * - The batch is already pinned into the relevant ppgtt, so we
1596                  *   already have the backing storage fully allocated.
1597                  * - No other BO uses the global gtt (well contexts, but meh),
1598                  *   so we don't really have issues with multiple objects not
1599                  *   fitting due to fragmentation.
1600                  * So this is actually safe.
1601                  */
1602                 ret = i915_gem_obj_ggtt_pin(batch_obj, 0, 0);
1603                 if (ret)
1604                         goto err;
1605
1606                 params->batch_obj_vm_offset = i915_gem_obj_ggtt_offset(batch_obj);
1607         } else
1608                 params->batch_obj_vm_offset = i915_gem_obj_offset(batch_obj, vm);
1609
1610         /* Allocate a request for this batch buffer nice and early. */
1611         ret = i915_gem_request_alloc(ring, ctx, &params->request);
1612         if (ret)
1613                 goto err_batch_unpin;
1614
1615         ret = i915_gem_request_add_to_client(params->request, file);
1616         if (ret)
1617                 goto err_batch_unpin;
1618
1619         /*
1620          * Save assorted stuff away to pass through to *_submission().
1621          * NB: This data should be 'persistent' and not local as it will
1622          * kept around beyond the duration of the IOCTL once the GPU
1623          * scheduler arrives.
1624          */
1625         params->dev                     = dev;
1626         params->file                    = file;
1627         params->ring                    = ring;
1628         params->dispatch_flags          = dispatch_flags;
1629         params->batch_obj               = batch_obj;
1630         params->ctx                     = ctx;
1631
1632         ret = dev_priv->gt.execbuf_submit(params, args, &eb->vmas);
1633
1634 err_batch_unpin:
1635         /*
1636          * FIXME: We crucially rely upon the active tracking for the (ppgtt)
1637          * batch vma for correctness. For less ugly and less fragility this
1638          * needs to be adjusted to also track the ggtt batch vma properly as
1639          * active.
1640          */
1641         if (dispatch_flags & I915_DISPATCH_SECURE)
1642                 i915_gem_object_ggtt_unpin(batch_obj);
1643
1644 err:
1645         /* the request owns the ref now */
1646         i915_gem_context_unreference(ctx);
1647         eb_destroy(eb);
1648
1649         /*
1650          * If the request was created but not successfully submitted then it
1651          * must be freed again. If it was submitted then it is being tracked
1652          * on the active request list and no clean up is required here.
1653          */
1654         if (ret && params->request)
1655                 i915_gem_request_cancel(params->request);
1656
1657         mutex_unlock(&dev->struct_mutex);
1658
1659 pre_mutex_err:
1660         /* intel_gpu_busy should also get a ref, so it will free when the device
1661          * is really idle. */
1662         intel_runtime_pm_put(dev_priv);
1663         return ret;
1664 }
1665
1666 /*
1667  * Legacy execbuffer just creates an exec2 list from the original exec object
1668  * list array and passes it to the real function.
1669  */
1670 int
1671 i915_gem_execbuffer(struct drm_device *dev, void *data,
1672                     struct drm_file *file)
1673 {
1674         struct drm_i915_gem_execbuffer *args = data;
1675         struct drm_i915_gem_execbuffer2 exec2;
1676         struct drm_i915_gem_exec_object *exec_list = NULL;
1677         struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1678         int ret, i;
1679
1680         if (args->buffer_count < 1) {
1681                 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1682                 return -EINVAL;
1683         }
1684
1685         /* Copy in the exec list from userland */
1686         exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
1687         exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
1688         if (exec_list == NULL || exec2_list == NULL) {
1689                 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1690                           args->buffer_count);
1691                 drm_free_large(exec_list);
1692                 drm_free_large(exec2_list);
1693                 return -ENOMEM;
1694         }
1695         ret = copy_from_user(exec_list,
1696                              to_user_ptr(args->buffers_ptr),
1697                              sizeof(*exec_list) * args->buffer_count);
1698         if (ret != 0) {
1699                 DRM_DEBUG("copy %d exec entries failed %d\n",
1700                           args->buffer_count, ret);
1701                 drm_free_large(exec_list);
1702                 drm_free_large(exec2_list);
1703                 return -EFAULT;
1704         }
1705
1706         for (i = 0; i < args->buffer_count; i++) {
1707                 exec2_list[i].handle = exec_list[i].handle;
1708                 exec2_list[i].relocation_count = exec_list[i].relocation_count;
1709                 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
1710                 exec2_list[i].alignment = exec_list[i].alignment;
1711                 exec2_list[i].offset = exec_list[i].offset;
1712                 if (INTEL_INFO(dev)->gen < 4)
1713                         exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
1714                 else
1715                         exec2_list[i].flags = 0;
1716         }
1717
1718         exec2.buffers_ptr = args->buffers_ptr;
1719         exec2.buffer_count = args->buffer_count;
1720         exec2.batch_start_offset = args->batch_start_offset;
1721         exec2.batch_len = args->batch_len;
1722         exec2.DR1 = args->DR1;
1723         exec2.DR4 = args->DR4;
1724         exec2.num_cliprects = args->num_cliprects;
1725         exec2.cliprects_ptr = args->cliprects_ptr;
1726         exec2.flags = I915_EXEC_RENDER;
1727         i915_execbuffer2_set_context_id(exec2, 0);
1728
1729         ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list);
1730         if (!ret) {
1731                 struct drm_i915_gem_exec_object __user *user_exec_list =
1732                         to_user_ptr(args->buffers_ptr);
1733
1734                 /* Copy the new buffer offsets back to the user's exec list. */
1735                 for (i = 0; i < args->buffer_count; i++) {
1736                         ret = __copy_to_user(&user_exec_list[i].offset,
1737                                              &exec2_list[i].offset,
1738                                              sizeof(user_exec_list[i].offset));
1739                         if (ret) {
1740                                 ret = -EFAULT;
1741                                 DRM_DEBUG("failed to copy %d exec entries "
1742                                           "back to user (%d)\n",
1743                                           args->buffer_count, ret);
1744                                 break;
1745                         }
1746                 }
1747         }
1748
1749         drm_free_large(exec_list);
1750         drm_free_large(exec2_list);
1751         return ret;
1752 }
1753
1754 int
1755 i915_gem_execbuffer2(struct drm_device *dev, void *data,
1756                      struct drm_file *file)
1757 {
1758         struct drm_i915_gem_execbuffer2 *args = data;
1759         struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1760         int ret;
1761
1762         if (args->buffer_count < 1 ||
1763             args->buffer_count > UINT_MAX / sizeof(*exec2_list)) {
1764                 DRM_DEBUG("execbuf2 with %d buffers\n", args->buffer_count);
1765                 return -EINVAL;
1766         }
1767
1768         if (args->rsvd2 != 0) {
1769                 DRM_DEBUG("dirty rvsd2 field\n");
1770                 return -EINVAL;
1771         }
1772
1773         exec2_list = kmalloc(sizeof(*exec2_list)*args->buffer_count,
1774                              GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
1775         if (exec2_list == NULL)
1776                 exec2_list = drm_malloc_ab(sizeof(*exec2_list),
1777                                            args->buffer_count);
1778         if (exec2_list == NULL) {
1779                 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1780                           args->buffer_count);
1781                 return -ENOMEM;
1782         }
1783         ret = copy_from_user(exec2_list,
1784                              to_user_ptr(args->buffers_ptr),
1785                              sizeof(*exec2_list) * args->buffer_count);
1786         if (ret != 0) {
1787                 DRM_DEBUG("copy %d exec entries failed %d\n",
1788                           args->buffer_count, ret);
1789                 drm_free_large(exec2_list);
1790                 return -EFAULT;
1791         }
1792
1793         ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list);
1794         if (!ret) {
1795                 /* Copy the new buffer offsets back to the user's exec list. */
1796                 struct drm_i915_gem_exec_object2 __user *user_exec_list =
1797                                    to_user_ptr(args->buffers_ptr);
1798                 int i;
1799
1800                 for (i = 0; i < args->buffer_count; i++) {
1801                         ret = __copy_to_user(&user_exec_list[i].offset,
1802                                              &exec2_list[i].offset,
1803                                              sizeof(user_exec_list[i].offset));
1804                         if (ret) {
1805                                 ret = -EFAULT;
1806                                 DRM_DEBUG("failed to copy %d exec entries "
1807                                           "back to user\n",
1808                                           args->buffer_count);
1809                                 break;
1810                         }
1811                 }
1812         }
1813
1814         drm_free_large(exec2_list);
1815         return ret;
1816 }