2 * Copyright © 2008,2010 Intel Corporation
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:
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
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
24 * Eric Anholt <eric@anholt.net>
25 * Chris Wilson <chris@chris-wilson.co.uk>
30 #include <drm/i915_drm.h>
32 #include "i915_trace.h"
33 #include "intel_drv.h"
34 #include <linux/dma_remapping.h>
37 struct list_head vmas;
40 struct i915_vma *lut[0];
41 struct hlist_head buckets[0];
45 static struct eb_vmas *
46 eb_create(struct drm_i915_gem_execbuffer2 *args)
48 struct eb_vmas *eb = NULL;
50 if (args->flags & I915_EXEC_HANDLE_LUT) {
51 unsigned size = args->buffer_count;
52 size *= sizeof(struct i915_vma *);
53 size += sizeof(struct eb_vmas);
54 eb = kmalloc(size, GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
58 unsigned size = args->buffer_count;
59 unsigned count = PAGE_SIZE / sizeof(struct hlist_head) / 2;
60 BUILD_BUG_ON_NOT_POWER_OF_2(PAGE_SIZE / sizeof(struct hlist_head));
61 while (count > 2*size)
63 eb = kzalloc(count*sizeof(struct hlist_head) +
64 sizeof(struct eb_vmas),
71 eb->and = -args->buffer_count;
73 INIT_LIST_HEAD(&eb->vmas);
78 eb_reset(struct eb_vmas *eb)
81 memset(eb->buckets, 0, (eb->and+1)*sizeof(struct hlist_head));
85 eb_lookup_vmas(struct eb_vmas *eb,
86 struct drm_i915_gem_exec_object2 *exec,
87 const struct drm_i915_gem_execbuffer2 *args,
88 struct i915_address_space *vm,
89 struct drm_file *file)
91 struct drm_i915_gem_object *obj;
92 struct list_head objects;
95 INIT_LIST_HEAD(&objects);
96 spin_lock(&file->table_lock);
97 /* Grab a reference to the object and release the lock so we can lookup
98 * or create the VMA without using GFP_ATOMIC */
99 for (i = 0; i < args->buffer_count; i++) {
100 obj = to_intel_bo(idr_find(&file->object_idr, exec[i].handle));
102 spin_unlock(&file->table_lock);
103 DRM_DEBUG("Invalid object handle %d at index %d\n",
109 if (!list_empty(&obj->obj_exec_link)) {
110 spin_unlock(&file->table_lock);
111 DRM_DEBUG("Object %p [handle %d, index %d] appears more than once in object list\n",
112 obj, exec[i].handle, i);
117 drm_gem_object_reference(&obj->base);
118 list_add_tail(&obj->obj_exec_link, &objects);
120 spin_unlock(&file->table_lock);
123 list_for_each_entry(obj, &objects, obj_exec_link) {
124 struct i915_vma *vma;
127 * NOTE: We can leak any vmas created here when something fails
128 * later on. But that's no issue since vma_unbind can deal with
129 * vmas which are not actually bound. And since only
130 * lookup_or_create exists as an interface to get at the vma
131 * from the (obj, vm) we don't run the risk of creating
132 * duplicated vmas for the same vm.
134 vma = i915_gem_obj_lookup_or_create_vma(obj, vm);
136 DRM_DEBUG("Failed to lookup VMA\n");
141 list_add_tail(&vma->exec_list, &eb->vmas);
143 vma->exec_entry = &exec[i];
147 uint32_t handle = args->flags & I915_EXEC_HANDLE_LUT ? i : exec[i].handle;
148 vma->exec_handle = handle;
149 hlist_add_head(&vma->exec_node,
150 &eb->buckets[handle & eb->and]);
157 while (!list_empty(&objects)) {
158 obj = list_first_entry(&objects,
159 struct drm_i915_gem_object,
161 list_del_init(&obj->obj_exec_link);
163 drm_gem_object_unreference(&obj->base);
168 static struct i915_vma *eb_get_vma(struct eb_vmas *eb, unsigned long handle)
171 if (handle >= -eb->and)
173 return eb->lut[handle];
175 struct hlist_head *head;
176 struct hlist_node *node;
178 head = &eb->buckets[handle & eb->and];
179 hlist_for_each(node, head) {
180 struct i915_vma *vma;
182 vma = hlist_entry(node, struct i915_vma, exec_node);
183 if (vma->exec_handle == handle)
190 static void eb_destroy(struct eb_vmas *eb) {
191 while (!list_empty(&eb->vmas)) {
192 struct i915_vma *vma;
194 vma = list_first_entry(&eb->vmas,
197 list_del_init(&vma->exec_list);
198 drm_gem_object_unreference(&vma->obj->base);
203 static inline int use_cpu_reloc(struct drm_i915_gem_object *obj)
205 return (HAS_LLC(obj->base.dev) ||
206 obj->base.write_domain == I915_GEM_DOMAIN_CPU ||
207 !obj->map_and_fenceable ||
208 obj->cache_level != I915_CACHE_NONE);
212 relocate_entry_cpu(struct drm_i915_gem_object *obj,
213 struct drm_i915_gem_relocation_entry *reloc)
215 struct drm_device *dev = obj->base.dev;
216 uint32_t page_offset = offset_in_page(reloc->offset);
220 ret = i915_gem_object_set_to_cpu_domain(obj, true);
224 vaddr = kmap_atomic(i915_gem_object_get_page(obj,
225 reloc->offset >> PAGE_SHIFT));
226 *(uint32_t *)(vaddr + page_offset) = reloc->delta;
228 if (INTEL_INFO(dev)->gen >= 8) {
229 page_offset = offset_in_page(page_offset + sizeof(uint32_t));
231 if (page_offset == 0) {
232 kunmap_atomic(vaddr);
233 vaddr = kmap_atomic(i915_gem_object_get_page(obj,
234 (reloc->offset + sizeof(uint32_t)) >> PAGE_SHIFT));
237 *(uint32_t *)(vaddr + page_offset) = 0;
240 kunmap_atomic(vaddr);
246 relocate_entry_gtt(struct drm_i915_gem_object *obj,
247 struct drm_i915_gem_relocation_entry *reloc)
249 struct drm_device *dev = obj->base.dev;
250 struct drm_i915_private *dev_priv = dev->dev_private;
251 uint32_t __iomem *reloc_entry;
252 void __iomem *reloc_page;
255 ret = i915_gem_object_set_to_gtt_domain(obj, true);
259 ret = i915_gem_object_put_fence(obj);
263 /* Map the page containing the relocation we're going to perform. */
264 reloc->offset += i915_gem_obj_ggtt_offset(obj);
265 reloc_page = io_mapping_map_atomic_wc(dev_priv->gtt.mappable,
266 reloc->offset & PAGE_MASK);
267 reloc_entry = (uint32_t __iomem *)
268 (reloc_page + offset_in_page(reloc->offset));
269 iowrite32(reloc->delta, reloc_entry);
271 if (INTEL_INFO(dev)->gen >= 8) {
274 if (offset_in_page(reloc->offset + sizeof(uint32_t)) == 0) {
275 io_mapping_unmap_atomic(reloc_page);
276 reloc_page = io_mapping_map_atomic_wc(
277 dev_priv->gtt.mappable,
278 reloc->offset + sizeof(uint32_t));
279 reloc_entry = reloc_page;
282 iowrite32(0, reloc_entry);
285 io_mapping_unmap_atomic(reloc_page);
291 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
293 struct drm_i915_gem_relocation_entry *reloc,
294 struct i915_address_space *vm)
296 struct drm_device *dev = obj->base.dev;
297 struct drm_gem_object *target_obj;
298 struct drm_i915_gem_object *target_i915_obj;
299 struct i915_vma *target_vma;
300 uint32_t target_offset;
303 /* we've already hold a reference to all valid objects */
304 target_vma = eb_get_vma(eb, reloc->target_handle);
305 if (unlikely(target_vma == NULL))
307 target_i915_obj = target_vma->obj;
308 target_obj = &target_vma->obj->base;
310 target_offset = target_vma->node.start;
312 /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
313 * pipe_control writes because the gpu doesn't properly redirect them
314 * through the ppgtt for non_secure batchbuffers. */
315 if (unlikely(IS_GEN6(dev) &&
316 reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION &&
317 !target_i915_obj->has_global_gtt_mapping)) {
318 i915_gem_gtt_bind_object(target_i915_obj,
319 target_i915_obj->cache_level);
322 /* Validate that the target is in a valid r/w GPU domain */
323 if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
324 DRM_DEBUG("reloc with multiple write domains: "
325 "obj %p target %d offset %d "
326 "read %08x write %08x",
327 obj, reloc->target_handle,
330 reloc->write_domain);
333 if (unlikely((reloc->write_domain | reloc->read_domains)
334 & ~I915_GEM_GPU_DOMAINS)) {
335 DRM_DEBUG("reloc with read/write non-GPU domains: "
336 "obj %p target %d offset %d "
337 "read %08x write %08x",
338 obj, reloc->target_handle,
341 reloc->write_domain);
345 target_obj->pending_read_domains |= reloc->read_domains;
346 target_obj->pending_write_domain |= reloc->write_domain;
348 /* If the relocation already has the right value in it, no
349 * more work needs to be done.
351 if (target_offset == reloc->presumed_offset)
354 /* Check that the relocation address is valid... */
355 if (unlikely(reloc->offset >
356 obj->base.size - (INTEL_INFO(dev)->gen >= 8 ? 8 : 4))) {
357 DRM_DEBUG("Relocation beyond object bounds: "
358 "obj %p target %d offset %d size %d.\n",
359 obj, reloc->target_handle,
361 (int) obj->base.size);
364 if (unlikely(reloc->offset & 3)) {
365 DRM_DEBUG("Relocation not 4-byte aligned: "
366 "obj %p target %d offset %d.\n",
367 obj, reloc->target_handle,
368 (int) reloc->offset);
372 /* We can't wait for rendering with pagefaults disabled */
373 if (obj->active && in_atomic())
376 reloc->delta += target_offset;
377 if (use_cpu_reloc(obj))
378 ret = relocate_entry_cpu(obj, reloc);
380 ret = relocate_entry_gtt(obj, reloc);
385 /* and update the user's relocation entry */
386 reloc->presumed_offset = target_offset;
392 i915_gem_execbuffer_relocate_vma(struct i915_vma *vma,
395 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
396 struct drm_i915_gem_relocation_entry stack_reloc[N_RELOC(512)];
397 struct drm_i915_gem_relocation_entry __user *user_relocs;
398 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
401 user_relocs = to_user_ptr(entry->relocs_ptr);
403 remain = entry->relocation_count;
405 struct drm_i915_gem_relocation_entry *r = stack_reloc;
407 if (count > ARRAY_SIZE(stack_reloc))
408 count = ARRAY_SIZE(stack_reloc);
411 if (__copy_from_user_inatomic(r, user_relocs, count*sizeof(r[0])))
415 u64 offset = r->presumed_offset;
417 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, r,
422 if (r->presumed_offset != offset &&
423 __copy_to_user_inatomic(&user_relocs->presumed_offset,
425 sizeof(r->presumed_offset))) {
439 i915_gem_execbuffer_relocate_vma_slow(struct i915_vma *vma,
441 struct drm_i915_gem_relocation_entry *relocs)
443 const struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
446 for (i = 0; i < entry->relocation_count; i++) {
447 ret = i915_gem_execbuffer_relocate_entry(vma->obj, eb, &relocs[i],
457 i915_gem_execbuffer_relocate(struct eb_vmas *eb)
459 struct i915_vma *vma;
462 /* This is the fast path and we cannot handle a pagefault whilst
463 * holding the struct mutex lest the user pass in the relocations
464 * contained within a mmaped bo. For in such a case we, the page
465 * fault handler would call i915_gem_fault() and we would try to
466 * acquire the struct mutex again. Obviously this is bad and so
467 * lockdep complains vehemently.
470 list_for_each_entry(vma, &eb->vmas, exec_list) {
471 ret = i915_gem_execbuffer_relocate_vma(vma, eb);
480 #define __EXEC_OBJECT_HAS_PIN (1<<31)
481 #define __EXEC_OBJECT_HAS_FENCE (1<<30)
484 need_reloc_mappable(struct i915_vma *vma)
486 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
487 return entry->relocation_count && !use_cpu_reloc(vma->obj) &&
488 i915_is_ggtt(vma->vm);
492 i915_gem_execbuffer_reserve_vma(struct i915_vma *vma,
493 struct intel_ring_buffer *ring,
496 struct drm_i915_private *dev_priv = ring->dev->dev_private;
497 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
498 bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
499 bool need_fence, need_mappable;
500 struct drm_i915_gem_object *obj = vma->obj;
504 has_fenced_gpu_access &&
505 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
506 obj->tiling_mode != I915_TILING_NONE;
507 need_mappable = need_fence || need_reloc_mappable(vma);
509 ret = i915_gem_object_pin(obj, vma->vm, entry->alignment, need_mappable,
514 entry->flags |= __EXEC_OBJECT_HAS_PIN;
516 if (has_fenced_gpu_access) {
517 if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
518 ret = i915_gem_object_get_fence(obj);
522 if (i915_gem_object_pin_fence(obj))
523 entry->flags |= __EXEC_OBJECT_HAS_FENCE;
525 obj->pending_fenced_gpu_access = true;
529 /* Ensure ppgtt mapping exists if needed */
530 if (dev_priv->mm.aliasing_ppgtt && !obj->has_aliasing_ppgtt_mapping) {
531 i915_ppgtt_bind_object(dev_priv->mm.aliasing_ppgtt,
532 obj, obj->cache_level);
534 obj->has_aliasing_ppgtt_mapping = 1;
537 if (entry->offset != vma->node.start) {
538 entry->offset = vma->node.start;
542 if (entry->flags & EXEC_OBJECT_WRITE) {
543 obj->base.pending_read_domains = I915_GEM_DOMAIN_RENDER;
544 obj->base.pending_write_domain = I915_GEM_DOMAIN_RENDER;
547 if (entry->flags & EXEC_OBJECT_NEEDS_GTT &&
548 !obj->has_global_gtt_mapping)
549 i915_gem_gtt_bind_object(obj, obj->cache_level);
555 i915_gem_execbuffer_unreserve_vma(struct i915_vma *vma)
557 struct drm_i915_gem_exec_object2 *entry;
558 struct drm_i915_gem_object *obj = vma->obj;
560 if (!drm_mm_node_allocated(&vma->node))
563 entry = vma->exec_entry;
565 if (entry->flags & __EXEC_OBJECT_HAS_FENCE)
566 i915_gem_object_unpin_fence(obj);
568 if (entry->flags & __EXEC_OBJECT_HAS_PIN)
569 i915_gem_object_unpin(obj);
571 entry->flags &= ~(__EXEC_OBJECT_HAS_FENCE | __EXEC_OBJECT_HAS_PIN);
575 i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring,
576 struct list_head *vmas,
579 struct drm_i915_gem_object *obj;
580 struct i915_vma *vma;
581 struct i915_address_space *vm;
582 struct list_head ordered_vmas;
583 bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
586 if (list_empty(vmas))
589 vm = list_first_entry(vmas, struct i915_vma, exec_list)->vm;
591 INIT_LIST_HEAD(&ordered_vmas);
592 while (!list_empty(vmas)) {
593 struct drm_i915_gem_exec_object2 *entry;
594 bool need_fence, need_mappable;
596 vma = list_first_entry(vmas, struct i915_vma, exec_list);
598 entry = vma->exec_entry;
601 has_fenced_gpu_access &&
602 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
603 obj->tiling_mode != I915_TILING_NONE;
604 need_mappable = need_fence || need_reloc_mappable(vma);
607 list_move(&vma->exec_list, &ordered_vmas);
609 list_move_tail(&vma->exec_list, &ordered_vmas);
611 obj->base.pending_read_domains = I915_GEM_GPU_DOMAINS & ~I915_GEM_DOMAIN_COMMAND;
612 obj->base.pending_write_domain = 0;
613 obj->pending_fenced_gpu_access = false;
615 list_splice(&ordered_vmas, vmas);
617 /* Attempt to pin all of the buffers into the GTT.
618 * This is done in 3 phases:
620 * 1a. Unbind all objects that do not match the GTT constraints for
621 * the execbuffer (fenceable, mappable, alignment etc).
622 * 1b. Increment pin count for already bound objects.
623 * 2. Bind new objects.
624 * 3. Decrement pin count.
626 * This avoid unnecessary unbinding of later objects in order to make
627 * room for the earlier objects *unless* we need to defragment.
633 /* Unbind any ill-fitting objects or pin. */
634 list_for_each_entry(vma, vmas, exec_list) {
635 struct drm_i915_gem_exec_object2 *entry = vma->exec_entry;
636 bool need_fence, need_mappable;
640 if (!drm_mm_node_allocated(&vma->node))
644 has_fenced_gpu_access &&
645 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
646 obj->tiling_mode != I915_TILING_NONE;
647 need_mappable = need_fence || need_reloc_mappable(vma);
649 WARN_ON((need_mappable || need_fence) &&
650 !i915_is_ggtt(vma->vm));
652 if ((entry->alignment &&
653 vma->node.start & (entry->alignment - 1)) ||
654 (need_mappable && !obj->map_and_fenceable))
655 ret = i915_vma_unbind(vma);
657 ret = i915_gem_execbuffer_reserve_vma(vma, ring, need_relocs);
662 /* Bind fresh objects */
663 list_for_each_entry(vma, vmas, exec_list) {
664 if (drm_mm_node_allocated(&vma->node))
667 ret = i915_gem_execbuffer_reserve_vma(vma, ring, need_relocs);
672 err: /* Decrement pin count for bound objects */
673 list_for_each_entry(vma, vmas, exec_list)
674 i915_gem_execbuffer_unreserve_vma(vma);
676 if (ret != -ENOSPC || retry++)
679 ret = i915_gem_evict_vm(vm, true);
686 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
687 struct drm_i915_gem_execbuffer2 *args,
688 struct drm_file *file,
689 struct intel_ring_buffer *ring,
691 struct drm_i915_gem_exec_object2 *exec)
693 struct drm_i915_gem_relocation_entry *reloc;
694 struct i915_address_space *vm;
695 struct i915_vma *vma;
699 unsigned count = args->buffer_count;
701 if (WARN_ON(list_empty(&eb->vmas)))
704 vm = list_first_entry(&eb->vmas, struct i915_vma, exec_list)->vm;
706 /* We may process another execbuffer during the unlock... */
707 while (!list_empty(&eb->vmas)) {
708 vma = list_first_entry(&eb->vmas, struct i915_vma, exec_list);
709 list_del_init(&vma->exec_list);
710 drm_gem_object_unreference(&vma->obj->base);
713 mutex_unlock(&dev->struct_mutex);
716 for (i = 0; i < count; i++)
717 total += exec[i].relocation_count;
719 reloc_offset = drm_malloc_ab(count, sizeof(*reloc_offset));
720 reloc = drm_malloc_ab(total, sizeof(*reloc));
721 if (reloc == NULL || reloc_offset == NULL) {
722 drm_free_large(reloc);
723 drm_free_large(reloc_offset);
724 mutex_lock(&dev->struct_mutex);
729 for (i = 0; i < count; i++) {
730 struct drm_i915_gem_relocation_entry __user *user_relocs;
731 u64 invalid_offset = (u64)-1;
734 user_relocs = to_user_ptr(exec[i].relocs_ptr);
736 if (copy_from_user(reloc+total, user_relocs,
737 exec[i].relocation_count * sizeof(*reloc))) {
739 mutex_lock(&dev->struct_mutex);
743 /* As we do not update the known relocation offsets after
744 * relocating (due to the complexities in lock handling),
745 * we need to mark them as invalid now so that we force the
746 * relocation processing next time. Just in case the target
747 * object is evicted and then rebound into its old
748 * presumed_offset before the next execbuffer - if that
749 * happened we would make the mistake of assuming that the
750 * relocations were valid.
752 for (j = 0; j < exec[i].relocation_count; j++) {
753 if (copy_to_user(&user_relocs[j].presumed_offset,
755 sizeof(invalid_offset))) {
757 mutex_lock(&dev->struct_mutex);
762 reloc_offset[i] = total;
763 total += exec[i].relocation_count;
766 ret = i915_mutex_lock_interruptible(dev);
768 mutex_lock(&dev->struct_mutex);
772 /* reacquire the objects */
774 ret = eb_lookup_vmas(eb, exec, args, vm, file);
778 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
779 ret = i915_gem_execbuffer_reserve(ring, &eb->vmas, &need_relocs);
783 list_for_each_entry(vma, &eb->vmas, exec_list) {
784 int offset = vma->exec_entry - exec;
785 ret = i915_gem_execbuffer_relocate_vma_slow(vma, eb,
786 reloc + reloc_offset[offset]);
791 /* Leave the user relocations as are, this is the painfully slow path,
792 * and we want to avoid the complication of dropping the lock whilst
793 * having buffers reserved in the aperture and so causing spurious
794 * ENOSPC for random operations.
798 drm_free_large(reloc);
799 drm_free_large(reloc_offset);
804 i915_gem_execbuffer_move_to_gpu(struct intel_ring_buffer *ring,
805 struct list_head *vmas)
807 struct i915_vma *vma;
808 uint32_t flush_domains = 0;
809 bool flush_chipset = false;
812 list_for_each_entry(vma, vmas, exec_list) {
813 struct drm_i915_gem_object *obj = vma->obj;
814 ret = i915_gem_object_sync(obj, ring);
818 if (obj->base.write_domain & I915_GEM_DOMAIN_CPU)
819 flush_chipset |= i915_gem_clflush_object(obj, false);
821 flush_domains |= obj->base.write_domain;
825 i915_gem_chipset_flush(ring->dev);
827 if (flush_domains & I915_GEM_DOMAIN_GTT)
830 /* Unconditionally invalidate gpu caches and ensure that we do flush
831 * any residual writes from the previous batch.
833 return intel_ring_invalidate_all_caches(ring);
837 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
839 if (exec->flags & __I915_EXEC_UNKNOWN_FLAGS)
842 return ((exec->batch_start_offset | exec->batch_len) & 0x7) == 0;
846 validate_exec_list(struct drm_i915_gem_exec_object2 *exec,
850 unsigned relocs_total = 0;
851 unsigned relocs_max = UINT_MAX / sizeof(struct drm_i915_gem_relocation_entry);
853 for (i = 0; i < count; i++) {
854 char __user *ptr = to_user_ptr(exec[i].relocs_ptr);
855 int length; /* limited by fault_in_pages_readable() */
857 if (exec[i].flags & __EXEC_OBJECT_UNKNOWN_FLAGS)
860 /* First check for malicious input causing overflow in
861 * the worst case where we need to allocate the entire
862 * relocation tree as a single array.
864 if (exec[i].relocation_count > relocs_max - relocs_total)
866 relocs_total += exec[i].relocation_count;
868 length = exec[i].relocation_count *
869 sizeof(struct drm_i915_gem_relocation_entry);
871 * We must check that the entire relocation array is safe
872 * to read, but since we may need to update the presumed
873 * offsets during execution, check for full write access.
875 if (!access_ok(VERIFY_WRITE, ptr, length))
878 if (likely(!i915_prefault_disable)) {
879 if (fault_in_multipages_readable(ptr, length))
888 i915_gem_execbuffer_move_to_active(struct list_head *vmas,
889 struct intel_ring_buffer *ring)
891 struct i915_vma *vma;
893 list_for_each_entry(vma, vmas, exec_list) {
894 struct drm_i915_gem_object *obj = vma->obj;
895 u32 old_read = obj->base.read_domains;
896 u32 old_write = obj->base.write_domain;
898 obj->base.write_domain = obj->base.pending_write_domain;
899 if (obj->base.write_domain == 0)
900 obj->base.pending_read_domains |= obj->base.read_domains;
901 obj->base.read_domains = obj->base.pending_read_domains;
902 obj->fenced_gpu_access = obj->pending_fenced_gpu_access;
904 i915_vma_move_to_active(vma, ring);
905 if (obj->base.write_domain) {
907 obj->last_write_seqno = intel_ring_get_seqno(ring);
908 if (obj->pin_count) /* check for potential scanout */
909 intel_mark_fb_busy(obj, ring);
912 trace_i915_gem_object_change_domain(obj, old_read, old_write);
917 i915_gem_execbuffer_retire_commands(struct drm_device *dev,
918 struct drm_file *file,
919 struct intel_ring_buffer *ring,
920 struct drm_i915_gem_object *obj)
922 /* Unconditionally force add_request to emit a full flush. */
923 ring->gpu_caches_dirty = true;
925 /* Add a breadcrumb for the completion of the batch buffer */
926 (void)__i915_add_request(ring, file, obj, NULL);
930 i915_reset_gen7_sol_offsets(struct drm_device *dev,
931 struct intel_ring_buffer *ring)
933 drm_i915_private_t *dev_priv = dev->dev_private;
936 if (!IS_GEN7(dev) || ring != &dev_priv->ring[RCS])
939 ret = intel_ring_begin(ring, 4 * 3);
943 for (i = 0; i < 4; i++) {
944 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
945 intel_ring_emit(ring, GEN7_SO_WRITE_OFFSET(i));
946 intel_ring_emit(ring, 0);
949 intel_ring_advance(ring);
955 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
956 struct drm_file *file,
957 struct drm_i915_gem_execbuffer2 *args,
958 struct drm_i915_gem_exec_object2 *exec,
959 struct i915_address_space *vm)
961 drm_i915_private_t *dev_priv = dev->dev_private;
963 struct drm_i915_gem_object *batch_obj;
964 struct drm_clip_rect *cliprects = NULL;
965 struct intel_ring_buffer *ring;
966 struct i915_ctx_hang_stats *hs;
967 u32 ctx_id = i915_execbuffer2_get_context_id(*args);
968 u32 exec_start, exec_len;
973 if (!i915_gem_check_execbuffer(args))
976 ret = validate_exec_list(exec, args->buffer_count);
981 if (args->flags & I915_EXEC_SECURE) {
982 if (!file->is_master || !capable(CAP_SYS_ADMIN))
985 flags |= I915_DISPATCH_SECURE;
987 if (args->flags & I915_EXEC_IS_PINNED)
988 flags |= I915_DISPATCH_PINNED;
990 switch (args->flags & I915_EXEC_RING_MASK) {
991 case I915_EXEC_DEFAULT:
992 case I915_EXEC_RENDER:
993 ring = &dev_priv->ring[RCS];
996 ring = &dev_priv->ring[VCS];
997 if (ctx_id != DEFAULT_CONTEXT_ID) {
998 DRM_DEBUG("Ring %s doesn't support contexts\n",
1004 ring = &dev_priv->ring[BCS];
1005 if (ctx_id != DEFAULT_CONTEXT_ID) {
1006 DRM_DEBUG("Ring %s doesn't support contexts\n",
1011 case I915_EXEC_VEBOX:
1012 ring = &dev_priv->ring[VECS];
1013 if (ctx_id != DEFAULT_CONTEXT_ID) {
1014 DRM_DEBUG("Ring %s doesn't support contexts\n",
1021 DRM_DEBUG("execbuf with unknown ring: %d\n",
1022 (int)(args->flags & I915_EXEC_RING_MASK));
1025 if (!intel_ring_initialized(ring)) {
1026 DRM_DEBUG("execbuf with invalid ring: %d\n",
1027 (int)(args->flags & I915_EXEC_RING_MASK));
1031 mode = args->flags & I915_EXEC_CONSTANTS_MASK;
1032 mask = I915_EXEC_CONSTANTS_MASK;
1034 case I915_EXEC_CONSTANTS_REL_GENERAL:
1035 case I915_EXEC_CONSTANTS_ABSOLUTE:
1036 case I915_EXEC_CONSTANTS_REL_SURFACE:
1037 if (ring == &dev_priv->ring[RCS] &&
1038 mode != dev_priv->relative_constants_mode) {
1039 if (INTEL_INFO(dev)->gen < 4)
1042 if (INTEL_INFO(dev)->gen > 5 &&
1043 mode == I915_EXEC_CONSTANTS_REL_SURFACE)
1046 /* The HW changed the meaning on this bit on gen6 */
1047 if (INTEL_INFO(dev)->gen >= 6)
1048 mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
1052 DRM_DEBUG("execbuf with unknown constants: %d\n", mode);
1056 if (args->buffer_count < 1) {
1057 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1061 if (args->num_cliprects != 0) {
1062 if (ring != &dev_priv->ring[RCS]) {
1063 DRM_DEBUG("clip rectangles are only valid with the render ring\n");
1067 if (INTEL_INFO(dev)->gen >= 5) {
1068 DRM_DEBUG("clip rectangles are only valid on pre-gen5\n");
1072 if (args->num_cliprects > UINT_MAX / sizeof(*cliprects)) {
1073 DRM_DEBUG("execbuf with %u cliprects\n",
1074 args->num_cliprects);
1078 cliprects = kcalloc(args->num_cliprects,
1081 if (cliprects == NULL) {
1086 if (copy_from_user(cliprects,
1087 to_user_ptr(args->cliprects_ptr),
1088 sizeof(*cliprects)*args->num_cliprects)) {
1094 ret = i915_mutex_lock_interruptible(dev);
1098 if (dev_priv->ums.mm_suspended) {
1099 mutex_unlock(&dev->struct_mutex);
1104 eb = eb_create(args);
1106 mutex_unlock(&dev->struct_mutex);
1111 /* Look up object handles */
1112 ret = eb_lookup_vmas(eb, exec, args, vm, file);
1116 /* take note of the batch buffer before we might reorder the lists */
1117 batch_obj = list_entry(eb->vmas.prev, struct i915_vma, exec_list)->obj;
1119 /* Move the objects en-masse into the GTT, evicting if necessary. */
1120 need_relocs = (args->flags & I915_EXEC_NO_RELOC) == 0;
1121 ret = i915_gem_execbuffer_reserve(ring, &eb->vmas, &need_relocs);
1125 /* The objects are in their final locations, apply the relocations. */
1127 ret = i915_gem_execbuffer_relocate(eb);
1129 if (ret == -EFAULT) {
1130 ret = i915_gem_execbuffer_relocate_slow(dev, args, file, ring,
1132 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1138 /* Set the pending read domains for the batch buffer to COMMAND */
1139 if (batch_obj->base.pending_write_domain) {
1140 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1144 batch_obj->base.pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
1146 /* snb/ivb/vlv conflate the "batch in ppgtt" bit with the "non-secure
1147 * batch" bit. Hence we need to pin secure batches into the global gtt.
1148 * hsw should have this fixed, but bdw mucks it up again. */
1149 if (flags & I915_DISPATCH_SECURE && !batch_obj->has_global_gtt_mapping)
1150 i915_gem_gtt_bind_object(batch_obj, batch_obj->cache_level);
1152 ret = i915_gem_execbuffer_move_to_gpu(ring, &eb->vmas);
1156 hs = i915_gem_context_get_hang_stats(dev, file, ctx_id);
1167 ret = i915_switch_context(ring, file, ctx_id);
1171 if (ring == &dev_priv->ring[RCS] &&
1172 mode != dev_priv->relative_constants_mode) {
1173 ret = intel_ring_begin(ring, 4);
1177 intel_ring_emit(ring, MI_NOOP);
1178 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1179 intel_ring_emit(ring, INSTPM);
1180 intel_ring_emit(ring, mask << 16 | mode);
1181 intel_ring_advance(ring);
1183 dev_priv->relative_constants_mode = mode;
1186 if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
1187 ret = i915_reset_gen7_sol_offsets(dev, ring);
1192 exec_start = i915_gem_obj_offset(batch_obj, vm) +
1193 args->batch_start_offset;
1194 exec_len = args->batch_len;
1196 for (i = 0; i < args->num_cliprects; i++) {
1197 ret = i915_emit_box(dev, &cliprects[i],
1198 args->DR1, args->DR4);
1202 ret = ring->dispatch_execbuffer(ring,
1203 exec_start, exec_len,
1209 ret = ring->dispatch_execbuffer(ring,
1210 exec_start, exec_len,
1216 trace_i915_gem_ring_dispatch(ring, intel_ring_get_seqno(ring), flags);
1218 i915_gem_execbuffer_move_to_active(&eb->vmas, ring);
1219 i915_gem_execbuffer_retire_commands(dev, file, ring, batch_obj);
1224 mutex_unlock(&dev->struct_mutex);
1232 * Legacy execbuffer just creates an exec2 list from the original exec object
1233 * list array and passes it to the real function.
1236 i915_gem_execbuffer(struct drm_device *dev, void *data,
1237 struct drm_file *file)
1239 struct drm_i915_private *dev_priv = dev->dev_private;
1240 struct drm_i915_gem_execbuffer *args = data;
1241 struct drm_i915_gem_execbuffer2 exec2;
1242 struct drm_i915_gem_exec_object *exec_list = NULL;
1243 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1246 if (args->buffer_count < 1) {
1247 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1251 /* Copy in the exec list from userland */
1252 exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
1253 exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
1254 if (exec_list == NULL || exec2_list == NULL) {
1255 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1256 args->buffer_count);
1257 drm_free_large(exec_list);
1258 drm_free_large(exec2_list);
1261 ret = copy_from_user(exec_list,
1262 to_user_ptr(args->buffers_ptr),
1263 sizeof(*exec_list) * args->buffer_count);
1265 DRM_DEBUG("copy %d exec entries failed %d\n",
1266 args->buffer_count, ret);
1267 drm_free_large(exec_list);
1268 drm_free_large(exec2_list);
1272 for (i = 0; i < args->buffer_count; i++) {
1273 exec2_list[i].handle = exec_list[i].handle;
1274 exec2_list[i].relocation_count = exec_list[i].relocation_count;
1275 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
1276 exec2_list[i].alignment = exec_list[i].alignment;
1277 exec2_list[i].offset = exec_list[i].offset;
1278 if (INTEL_INFO(dev)->gen < 4)
1279 exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
1281 exec2_list[i].flags = 0;
1284 exec2.buffers_ptr = args->buffers_ptr;
1285 exec2.buffer_count = args->buffer_count;
1286 exec2.batch_start_offset = args->batch_start_offset;
1287 exec2.batch_len = args->batch_len;
1288 exec2.DR1 = args->DR1;
1289 exec2.DR4 = args->DR4;
1290 exec2.num_cliprects = args->num_cliprects;
1291 exec2.cliprects_ptr = args->cliprects_ptr;
1292 exec2.flags = I915_EXEC_RENDER;
1293 i915_execbuffer2_set_context_id(exec2, 0);
1295 ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list,
1296 &dev_priv->gtt.base);
1298 /* Copy the new buffer offsets back to the user's exec list. */
1299 for (i = 0; i < args->buffer_count; i++)
1300 exec_list[i].offset = exec2_list[i].offset;
1301 /* ... and back out to userspace */
1302 ret = copy_to_user(to_user_ptr(args->buffers_ptr),
1304 sizeof(*exec_list) * args->buffer_count);
1307 DRM_DEBUG("failed to copy %d exec entries "
1308 "back to user (%d)\n",
1309 args->buffer_count, ret);
1313 drm_free_large(exec_list);
1314 drm_free_large(exec2_list);
1319 i915_gem_execbuffer2(struct drm_device *dev, void *data,
1320 struct drm_file *file)
1322 struct drm_i915_private *dev_priv = dev->dev_private;
1323 struct drm_i915_gem_execbuffer2 *args = data;
1324 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1327 if (args->buffer_count < 1 ||
1328 args->buffer_count > UINT_MAX / sizeof(*exec2_list)) {
1329 DRM_DEBUG("execbuf2 with %d buffers\n", args->buffer_count);
1333 exec2_list = kmalloc(sizeof(*exec2_list)*args->buffer_count,
1334 GFP_TEMPORARY | __GFP_NOWARN | __GFP_NORETRY);
1335 if (exec2_list == NULL)
1336 exec2_list = drm_malloc_ab(sizeof(*exec2_list),
1337 args->buffer_count);
1338 if (exec2_list == NULL) {
1339 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1340 args->buffer_count);
1343 ret = copy_from_user(exec2_list,
1344 to_user_ptr(args->buffers_ptr),
1345 sizeof(*exec2_list) * args->buffer_count);
1347 DRM_DEBUG("copy %d exec entries failed %d\n",
1348 args->buffer_count, ret);
1349 drm_free_large(exec2_list);
1353 ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list,
1354 &dev_priv->gtt.base);
1356 /* Copy the new buffer offsets back to the user's exec list. */
1357 ret = copy_to_user(to_user_ptr(args->buffers_ptr),
1359 sizeof(*exec2_list) * args->buffer_count);
1362 DRM_DEBUG("failed to copy %d exec entries "
1363 "back to user (%d)\n",
1364 args->buffer_count, ret);
1368 drm_free_large(exec2_list);