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>
33 #include "i915_trace.h"
34 #include "intel_drv.h"
35 #include <linux/dma_remapping.h>
37 struct change_domains {
38 uint32_t invalidate_domains;
39 uint32_t flush_domains;
45 * Set the next domain for the specified object. This
46 * may not actually perform the necessary flushing/invaliding though,
47 * as that may want to be batched with other set_domain operations
49 * This is (we hope) the only really tricky part of gem. The goal
50 * is fairly simple -- track which caches hold bits of the object
51 * and make sure they remain coherent. A few concrete examples may
52 * help to explain how it works. For shorthand, we use the notation
53 * (read_domains, write_domain), e.g. (CPU, CPU) to indicate the
54 * a pair of read and write domain masks.
56 * Case 1: the batch buffer
62 * 5. Unmapped from GTT
65 * Let's take these a step at a time
68 * Pages allocated from the kernel may still have
69 * cache contents, so we set them to (CPU, CPU) always.
70 * 2. Written by CPU (using pwrite)
71 * The pwrite function calls set_domain (CPU, CPU) and
72 * this function does nothing (as nothing changes)
74 * This function asserts that the object is not
75 * currently in any GPU-based read or write domains
77 * i915_gem_execbuffer calls set_domain (COMMAND, 0).
78 * As write_domain is zero, this function adds in the
79 * current read domains (CPU+COMMAND, 0).
80 * flush_domains is set to CPU.
81 * invalidate_domains is set to COMMAND
82 * clflush is run to get data out of the CPU caches
83 * then i915_dev_set_domain calls i915_gem_flush to
84 * emit an MI_FLUSH and drm_agp_chipset_flush
85 * 5. Unmapped from GTT
86 * i915_gem_object_unbind calls set_domain (CPU, CPU)
87 * flush_domains and invalidate_domains end up both zero
88 * so no flushing/invalidating happens
92 * Case 2: The shared render buffer
96 * 3. Read/written by GPU
97 * 4. set_domain to (CPU,CPU)
98 * 5. Read/written by CPU
99 * 6. Read/written by GPU
102 * Same as last example, (CPU, CPU)
104 * Nothing changes (assertions find that it is not in the GPU)
105 * 3. Read/written by GPU
106 * execbuffer calls set_domain (RENDER, RENDER)
107 * flush_domains gets CPU
108 * invalidate_domains gets GPU
110 * MI_FLUSH and drm_agp_chipset_flush
111 * 4. set_domain (CPU, CPU)
112 * flush_domains gets GPU
113 * invalidate_domains gets CPU
114 * wait_rendering (obj) to make sure all drawing is complete.
115 * This will include an MI_FLUSH to get the data from GPU
117 * clflush (obj) to invalidate the CPU cache
118 * Another MI_FLUSH in i915_gem_flush (eliminate this somehow?)
119 * 5. Read/written by CPU
120 * cache lines are loaded and dirtied
121 * 6. Read written by GPU
122 * Same as last GPU access
124 * Case 3: The constant buffer
129 * 4. Updated (written) by CPU again
138 * flush_domains = CPU
139 * invalidate_domains = RENDER
142 * drm_agp_chipset_flush
143 * 4. Updated (written) by CPU again
145 * flush_domains = 0 (no previous write domain)
146 * invalidate_domains = 0 (no new read domains)
149 * flush_domains = CPU
150 * invalidate_domains = RENDER
153 * drm_agp_chipset_flush
156 i915_gem_object_set_to_gpu_domain(struct drm_i915_gem_object *obj,
157 struct intel_ring_buffer *ring,
158 struct change_domains *cd)
160 uint32_t invalidate_domains = 0, flush_domains = 0;
163 * If the object isn't moving to a new write domain,
164 * let the object stay in multiple read domains
166 if (obj->base.pending_write_domain == 0)
167 obj->base.pending_read_domains |= obj->base.read_domains;
170 * Flush the current write domain if
171 * the new read domains don't match. Invalidate
172 * any read domains which differ from the old
175 if (obj->base.write_domain &&
176 (((obj->base.write_domain != obj->base.pending_read_domains ||
177 obj->ring != ring)) ||
178 (obj->fenced_gpu_access && !obj->pending_fenced_gpu_access))) {
179 flush_domains |= obj->base.write_domain;
180 invalidate_domains |=
181 obj->base.pending_read_domains & ~obj->base.write_domain;
184 * Invalidate any read caches which may have
185 * stale data. That is, any new read domains.
187 invalidate_domains |= obj->base.pending_read_domains & ~obj->base.read_domains;
188 if ((flush_domains | invalidate_domains) & I915_GEM_DOMAIN_CPU)
189 i915_gem_clflush_object(obj);
191 if (obj->base.pending_write_domain)
192 cd->flips |= atomic_read(&obj->pending_flip);
194 /* The actual obj->write_domain will be updated with
195 * pending_write_domain after we emit the accumulated flush for all
196 * of our domain changes in execbuffers (which clears objects'
197 * write_domains). So if we have a current write domain that we
198 * aren't changing, set pending_write_domain to that.
200 if (flush_domains == 0 && obj->base.pending_write_domain == 0)
201 obj->base.pending_write_domain = obj->base.write_domain;
203 cd->invalidate_domains |= invalidate_domains;
204 cd->flush_domains |= flush_domains;
205 if (flush_domains & I915_GEM_GPU_DOMAINS)
206 cd->flush_rings |= intel_ring_flag(obj->ring);
207 if (invalidate_domains & I915_GEM_GPU_DOMAINS)
208 cd->flush_rings |= intel_ring_flag(ring);
213 struct hlist_head buckets[0];
216 static struct eb_objects *
219 struct eb_objects *eb;
220 int count = PAGE_SIZE / sizeof(struct hlist_head) / 2;
223 eb = kzalloc(count*sizeof(struct hlist_head) +
224 sizeof(struct eb_objects),
234 eb_reset(struct eb_objects *eb)
236 memset(eb->buckets, 0, (eb->and+1)*sizeof(struct hlist_head));
240 eb_add_object(struct eb_objects *eb, struct drm_i915_gem_object *obj)
242 hlist_add_head(&obj->exec_node,
243 &eb->buckets[obj->exec_handle & eb->and]);
246 static struct drm_i915_gem_object *
247 eb_get_object(struct eb_objects *eb, unsigned long handle)
249 struct hlist_head *head;
250 struct hlist_node *node;
251 struct drm_i915_gem_object *obj;
253 head = &eb->buckets[handle & eb->and];
254 hlist_for_each(node, head) {
255 obj = hlist_entry(node, struct drm_i915_gem_object, exec_node);
256 if (obj->exec_handle == handle)
264 eb_destroy(struct eb_objects *eb)
269 static inline int use_cpu_reloc(struct drm_i915_gem_object *obj)
271 return (obj->base.write_domain == I915_GEM_DOMAIN_CPU ||
272 obj->cache_level != I915_CACHE_NONE);
276 i915_gem_execbuffer_relocate_entry(struct drm_i915_gem_object *obj,
277 struct eb_objects *eb,
278 struct drm_i915_gem_relocation_entry *reloc)
280 struct drm_device *dev = obj->base.dev;
281 struct drm_gem_object *target_obj;
282 struct drm_i915_gem_object *target_i915_obj;
283 uint32_t target_offset;
286 /* we've already hold a reference to all valid objects */
287 target_obj = &eb_get_object(eb, reloc->target_handle)->base;
288 if (unlikely(target_obj == NULL))
291 target_i915_obj = to_intel_bo(target_obj);
292 target_offset = target_i915_obj->gtt_offset;
294 /* The target buffer should have appeared before us in the
295 * exec_object list, so it should have a GTT space bound by now.
297 if (unlikely(target_offset == 0)) {
298 DRM_DEBUG("No GTT space found for object %d\n",
299 reloc->target_handle);
303 /* Validate that the target is in a valid r/w GPU domain */
304 if (unlikely(reloc->write_domain & (reloc->write_domain - 1))) {
305 DRM_DEBUG("reloc with multiple write domains: "
306 "obj %p target %d offset %d "
307 "read %08x write %08x",
308 obj, reloc->target_handle,
311 reloc->write_domain);
314 if (unlikely((reloc->write_domain | reloc->read_domains)
315 & ~I915_GEM_GPU_DOMAINS)) {
316 DRM_DEBUG("reloc with read/write non-GPU domains: "
317 "obj %p target %d offset %d "
318 "read %08x write %08x",
319 obj, reloc->target_handle,
322 reloc->write_domain);
325 if (unlikely(reloc->write_domain && target_obj->pending_write_domain &&
326 reloc->write_domain != target_obj->pending_write_domain)) {
327 DRM_DEBUG("Write domain conflict: "
328 "obj %p target %d offset %d "
329 "new %08x old %08x\n",
330 obj, reloc->target_handle,
333 target_obj->pending_write_domain);
337 target_obj->pending_read_domains |= reloc->read_domains;
338 target_obj->pending_write_domain |= reloc->write_domain;
340 /* If the relocation already has the right value in it, no
341 * more work needs to be done.
343 if (target_offset == reloc->presumed_offset)
346 /* Check that the relocation address is valid... */
347 if (unlikely(reloc->offset > obj->base.size - 4)) {
348 DRM_DEBUG("Relocation beyond object bounds: "
349 "obj %p target %d offset %d size %d.\n",
350 obj, reloc->target_handle,
352 (int) obj->base.size);
355 if (unlikely(reloc->offset & 3)) {
356 DRM_DEBUG("Relocation not 4-byte aligned: "
357 "obj %p target %d offset %d.\n",
358 obj, reloc->target_handle,
359 (int) reloc->offset);
363 /* We can't wait for rendering with pagefaults disabled */
364 if (obj->active && in_atomic())
367 reloc->delta += target_offset;
368 if (use_cpu_reloc(obj)) {
369 uint32_t page_offset = reloc->offset & ~PAGE_MASK;
372 ret = i915_gem_object_set_to_cpu_domain(obj, 1);
376 vaddr = kmap_atomic(obj->pages[reloc->offset >> PAGE_SHIFT]);
377 *(uint32_t *)(vaddr + page_offset) = reloc->delta;
378 kunmap_atomic(vaddr);
380 struct drm_i915_private *dev_priv = dev->dev_private;
381 uint32_t __iomem *reloc_entry;
382 void __iomem *reloc_page;
384 ret = i915_gem_object_set_to_gtt_domain(obj, 1);
388 /* Map the page containing the relocation we're going to perform. */
389 reloc->offset += obj->gtt_offset;
390 reloc_page = io_mapping_map_atomic_wc(dev_priv->mm.gtt_mapping,
391 reloc->offset & PAGE_MASK);
392 reloc_entry = (uint32_t __iomem *)
393 (reloc_page + (reloc->offset & ~PAGE_MASK));
394 iowrite32(reloc->delta, reloc_entry);
395 io_mapping_unmap_atomic(reloc_page);
398 /* Sandybridge PPGTT errata: We need a global gtt mapping for MI and
399 * pipe_control writes because the gpu doesn't properly redirect them
400 * through the ppgtt for non_secure batchbuffers. */
401 if (unlikely(IS_GEN6(dev) &&
402 reloc->write_domain == I915_GEM_DOMAIN_INSTRUCTION &&
403 !target_i915_obj->has_global_gtt_mapping)) {
404 i915_gem_gtt_bind_object(target_i915_obj,
405 target_i915_obj->cache_level);
408 /* and update the user's relocation entry */
409 reloc->presumed_offset = target_offset;
415 i915_gem_execbuffer_relocate_object(struct drm_i915_gem_object *obj,
416 struct eb_objects *eb)
418 #define N_RELOC(x) ((x) / sizeof(struct drm_i915_gem_relocation_entry))
419 struct drm_i915_gem_relocation_entry stack_reloc[N_RELOC(512)];
420 struct drm_i915_gem_relocation_entry __user *user_relocs;
421 struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
424 user_relocs = (void __user *)(uintptr_t)entry->relocs_ptr;
426 remain = entry->relocation_count;
428 struct drm_i915_gem_relocation_entry *r = stack_reloc;
430 if (count > ARRAY_SIZE(stack_reloc))
431 count = ARRAY_SIZE(stack_reloc);
434 if (__copy_from_user_inatomic(r, user_relocs, count*sizeof(r[0])))
438 u64 offset = r->presumed_offset;
440 ret = i915_gem_execbuffer_relocate_entry(obj, eb, r);
444 if (r->presumed_offset != offset &&
445 __copy_to_user_inatomic(&user_relocs->presumed_offset,
447 sizeof(r->presumed_offset))) {
461 i915_gem_execbuffer_relocate_object_slow(struct drm_i915_gem_object *obj,
462 struct eb_objects *eb,
463 struct drm_i915_gem_relocation_entry *relocs)
465 const struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
468 for (i = 0; i < entry->relocation_count; i++) {
469 ret = i915_gem_execbuffer_relocate_entry(obj, eb, &relocs[i]);
478 i915_gem_execbuffer_relocate(struct drm_device *dev,
479 struct eb_objects *eb,
480 struct list_head *objects)
482 struct drm_i915_gem_object *obj;
485 /* This is the fast path and we cannot handle a pagefault whilst
486 * holding the struct mutex lest the user pass in the relocations
487 * contained within a mmaped bo. For in such a case we, the page
488 * fault handler would call i915_gem_fault() and we would try to
489 * acquire the struct mutex again. Obviously this is bad and so
490 * lockdep complains vehemently.
493 list_for_each_entry(obj, objects, exec_list) {
494 ret = i915_gem_execbuffer_relocate_object(obj, eb);
503 #define __EXEC_OBJECT_HAS_FENCE (1<<31)
506 need_reloc_mappable(struct drm_i915_gem_object *obj)
508 struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
509 return entry->relocation_count && !use_cpu_reloc(obj);
513 pin_and_fence_object(struct drm_i915_gem_object *obj,
514 struct intel_ring_buffer *ring)
516 struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
517 bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
518 bool need_fence, need_mappable;
522 has_fenced_gpu_access &&
523 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
524 obj->tiling_mode != I915_TILING_NONE;
525 need_mappable = need_fence || need_reloc_mappable(obj);
527 ret = i915_gem_object_pin(obj, entry->alignment, need_mappable);
531 if (has_fenced_gpu_access) {
532 if (entry->flags & EXEC_OBJECT_NEEDS_FENCE) {
533 if (obj->tiling_mode) {
534 ret = i915_gem_object_get_fence(obj, ring);
538 entry->flags |= __EXEC_OBJECT_HAS_FENCE;
539 i915_gem_object_pin_fence(obj);
541 ret = i915_gem_object_put_fence(obj);
545 obj->pending_fenced_gpu_access = true;
549 entry->offset = obj->gtt_offset;
553 i915_gem_object_unpin(obj);
558 i915_gem_execbuffer_reserve(struct intel_ring_buffer *ring,
559 struct drm_file *file,
560 struct list_head *objects)
562 drm_i915_private_t *dev_priv = ring->dev->dev_private;
563 struct drm_i915_gem_object *obj;
565 bool has_fenced_gpu_access = INTEL_INFO(ring->dev)->gen < 4;
566 struct list_head ordered_objects;
568 INIT_LIST_HEAD(&ordered_objects);
569 while (!list_empty(objects)) {
570 struct drm_i915_gem_exec_object2 *entry;
571 bool need_fence, need_mappable;
573 obj = list_first_entry(objects,
574 struct drm_i915_gem_object,
576 entry = obj->exec_entry;
579 has_fenced_gpu_access &&
580 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
581 obj->tiling_mode != I915_TILING_NONE;
582 need_mappable = need_fence || need_reloc_mappable(obj);
585 list_move(&obj->exec_list, &ordered_objects);
587 list_move_tail(&obj->exec_list, &ordered_objects);
589 obj->base.pending_read_domains = 0;
590 obj->base.pending_write_domain = 0;
592 list_splice(&ordered_objects, objects);
594 /* Attempt to pin all of the buffers into the GTT.
595 * This is done in 3 phases:
597 * 1a. Unbind all objects that do not match the GTT constraints for
598 * the execbuffer (fenceable, mappable, alignment etc).
599 * 1b. Increment pin count for already bound objects.
600 * 2. Bind new objects.
601 * 3. Decrement pin count.
603 * This avoid unnecessary unbinding of later objects in order to makr
604 * room for the earlier objects *unless* we need to defragment.
610 /* Unbind any ill-fitting objects or pin. */
611 list_for_each_entry(obj, objects, exec_list) {
612 struct drm_i915_gem_exec_object2 *entry = obj->exec_entry;
613 bool need_fence, need_mappable;
619 has_fenced_gpu_access &&
620 entry->flags & EXEC_OBJECT_NEEDS_FENCE &&
621 obj->tiling_mode != I915_TILING_NONE;
622 need_mappable = need_fence || need_reloc_mappable(obj);
624 if ((entry->alignment && obj->gtt_offset & (entry->alignment - 1)) ||
625 (need_mappable && !obj->map_and_fenceable))
626 ret = i915_gem_object_unbind(obj);
628 ret = pin_and_fence_object(obj, ring);
633 /* Bind fresh objects */
634 list_for_each_entry(obj, objects, exec_list) {
638 ret = pin_and_fence_object(obj, ring);
642 /* This can potentially raise a harmless
643 * -EINVAL if we failed to bind in the above
644 * call. It cannot raise -EINTR since we know
645 * that the bo is freshly bound and so will
646 * not need to be flushed or waited upon.
648 ret_ignore = i915_gem_object_unbind(obj);
650 WARN_ON(obj->gtt_space);
655 /* Decrement pin count for bound objects */
656 list_for_each_entry(obj, objects, exec_list) {
657 struct drm_i915_gem_exec_object2 *entry;
662 entry = obj->exec_entry;
663 if (entry->flags & __EXEC_OBJECT_HAS_FENCE) {
664 i915_gem_object_unpin_fence(obj);
665 entry->flags &= ~__EXEC_OBJECT_HAS_FENCE;
668 i915_gem_object_unpin(obj);
670 /* ... and ensure ppgtt mapping exist if needed. */
671 if (dev_priv->mm.aliasing_ppgtt && !obj->has_aliasing_ppgtt_mapping) {
672 i915_ppgtt_bind_object(dev_priv->mm.aliasing_ppgtt,
673 obj, obj->cache_level);
675 obj->has_aliasing_ppgtt_mapping = 1;
679 if (ret != -ENOSPC || retry > 1)
682 /* First attempt, just clear anything that is purgeable.
683 * Second attempt, clear the entire GTT.
685 ret = i915_gem_evict_everything(ring->dev, retry == 0);
693 list_for_each_entry_continue_reverse(obj, objects, exec_list) {
694 struct drm_i915_gem_exec_object2 *entry;
699 entry = obj->exec_entry;
700 if (entry->flags & __EXEC_OBJECT_HAS_FENCE) {
701 i915_gem_object_unpin_fence(obj);
702 entry->flags &= ~__EXEC_OBJECT_HAS_FENCE;
705 i915_gem_object_unpin(obj);
712 i915_gem_execbuffer_relocate_slow(struct drm_device *dev,
713 struct drm_file *file,
714 struct intel_ring_buffer *ring,
715 struct list_head *objects,
716 struct eb_objects *eb,
717 struct drm_i915_gem_exec_object2 *exec,
720 struct drm_i915_gem_relocation_entry *reloc;
721 struct drm_i915_gem_object *obj;
725 /* We may process another execbuffer during the unlock... */
726 while (!list_empty(objects)) {
727 obj = list_first_entry(objects,
728 struct drm_i915_gem_object,
730 list_del_init(&obj->exec_list);
731 drm_gem_object_unreference(&obj->base);
734 mutex_unlock(&dev->struct_mutex);
737 for (i = 0; i < count; i++)
738 total += exec[i].relocation_count;
740 reloc_offset = drm_malloc_ab(count, sizeof(*reloc_offset));
741 reloc = drm_malloc_ab(total, sizeof(*reloc));
742 if (reloc == NULL || reloc_offset == NULL) {
743 drm_free_large(reloc);
744 drm_free_large(reloc_offset);
745 mutex_lock(&dev->struct_mutex);
750 for (i = 0; i < count; i++) {
751 struct drm_i915_gem_relocation_entry __user *user_relocs;
753 user_relocs = (void __user *)(uintptr_t)exec[i].relocs_ptr;
755 if (copy_from_user(reloc+total, user_relocs,
756 exec[i].relocation_count * sizeof(*reloc))) {
758 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 for (i = 0; i < count; i++) {
775 obj = to_intel_bo(drm_gem_object_lookup(dev, file,
777 if (&obj->base == NULL) {
778 DRM_DEBUG("Invalid object handle %d at index %d\n",
784 list_add_tail(&obj->exec_list, objects);
785 obj->exec_handle = exec[i].handle;
786 obj->exec_entry = &exec[i];
787 eb_add_object(eb, obj);
790 ret = i915_gem_execbuffer_reserve(ring, file, objects);
794 list_for_each_entry(obj, objects, exec_list) {
795 int offset = obj->exec_entry - exec;
796 ret = i915_gem_execbuffer_relocate_object_slow(obj, eb,
797 reloc + reloc_offset[offset]);
802 /* Leave the user relocations as are, this is the painfully slow path,
803 * and we want to avoid the complication of dropping the lock whilst
804 * having buffers reserved in the aperture and so causing spurious
805 * ENOSPC for random operations.
809 drm_free_large(reloc);
810 drm_free_large(reloc_offset);
815 i915_gem_execbuffer_flush(struct drm_device *dev,
816 uint32_t invalidate_domains,
817 uint32_t flush_domains,
818 uint32_t flush_rings)
820 drm_i915_private_t *dev_priv = dev->dev_private;
823 if (flush_domains & I915_GEM_DOMAIN_CPU)
824 intel_gtt_chipset_flush();
826 if (flush_domains & I915_GEM_DOMAIN_GTT)
829 if ((flush_domains | invalidate_domains) & I915_GEM_GPU_DOMAINS) {
830 for (i = 0; i < I915_NUM_RINGS; i++)
831 if (flush_rings & (1 << i)) {
832 ret = i915_gem_flush_ring(&dev_priv->ring[i],
844 intel_enable_semaphores(struct drm_device *dev)
846 if (INTEL_INFO(dev)->gen < 6)
849 if (i915_semaphores >= 0)
850 return i915_semaphores;
852 /* Disable semaphores on SNB */
853 if (INTEL_INFO(dev)->gen == 6)
860 i915_gem_execbuffer_sync_rings(struct drm_i915_gem_object *obj,
861 struct intel_ring_buffer *to)
863 struct intel_ring_buffer *from = obj->ring;
867 if (from == NULL || to == from)
870 /* XXX gpu semaphores are implicated in various hard hangs on SNB */
871 if (!intel_enable_semaphores(obj->base.dev))
872 return i915_gem_object_wait_rendering(obj);
874 idx = intel_ring_sync_index(from, to);
876 seqno = obj->last_rendering_seqno;
877 if (seqno <= from->sync_seqno[idx])
880 if (seqno == from->outstanding_lazy_request) {
881 struct drm_i915_gem_request *request;
883 request = kzalloc(sizeof(*request), GFP_KERNEL);
887 ret = i915_add_request(from, NULL, request);
893 seqno = request->seqno;
896 from->sync_seqno[idx] = seqno;
898 return to->sync_to(to, from, seqno - 1);
902 i915_gem_execbuffer_wait_for_flips(struct intel_ring_buffer *ring, u32 flips)
904 u32 plane, flip_mask;
907 /* Check for any pending flips. As we only maintain a flip queue depth
908 * of 1, we can simply insert a WAIT for the next display flip prior
909 * to executing the batch and avoid stalling the CPU.
912 for (plane = 0; flips >> plane; plane++) {
913 if (((flips >> plane) & 1) == 0)
917 flip_mask = MI_WAIT_FOR_PLANE_B_FLIP;
919 flip_mask = MI_WAIT_FOR_PLANE_A_FLIP;
921 ret = intel_ring_begin(ring, 2);
925 intel_ring_emit(ring, MI_WAIT_FOR_EVENT | flip_mask);
926 intel_ring_emit(ring, MI_NOOP);
927 intel_ring_advance(ring);
935 i915_gem_execbuffer_move_to_gpu(struct intel_ring_buffer *ring,
936 struct list_head *objects)
938 struct drm_i915_gem_object *obj;
939 struct change_domains cd;
942 memset(&cd, 0, sizeof(cd));
943 list_for_each_entry(obj, objects, exec_list)
944 i915_gem_object_set_to_gpu_domain(obj, ring, &cd);
946 if (cd.invalidate_domains | cd.flush_domains) {
947 ret = i915_gem_execbuffer_flush(ring->dev,
948 cd.invalidate_domains,
956 ret = i915_gem_execbuffer_wait_for_flips(ring, cd.flips);
961 list_for_each_entry(obj, objects, exec_list) {
962 ret = i915_gem_execbuffer_sync_rings(obj, ring);
971 i915_gem_check_execbuffer(struct drm_i915_gem_execbuffer2 *exec)
973 return ((exec->batch_start_offset | exec->batch_len) & 0x7) == 0;
977 validate_exec_list(struct drm_i915_gem_exec_object2 *exec,
982 for (i = 0; i < count; i++) {
983 char __user *ptr = (char __user *)(uintptr_t)exec[i].relocs_ptr;
984 int length; /* limited by fault_in_pages_readable() */
986 /* First check for malicious input causing overflow */
987 if (exec[i].relocation_count >
988 INT_MAX / sizeof(struct drm_i915_gem_relocation_entry))
991 length = exec[i].relocation_count *
992 sizeof(struct drm_i915_gem_relocation_entry);
993 if (!access_ok(VERIFY_READ, ptr, length))
996 /* we may also need to update the presumed offsets */
997 if (!access_ok(VERIFY_WRITE, ptr, length))
1000 if (fault_in_multipages_readable(ptr, length))
1008 i915_gem_execbuffer_move_to_active(struct list_head *objects,
1009 struct intel_ring_buffer *ring,
1012 struct drm_i915_gem_object *obj;
1014 list_for_each_entry(obj, objects, exec_list) {
1015 u32 old_read = obj->base.read_domains;
1016 u32 old_write = obj->base.write_domain;
1019 obj->base.read_domains = obj->base.pending_read_domains;
1020 obj->base.write_domain = obj->base.pending_write_domain;
1021 obj->fenced_gpu_access = obj->pending_fenced_gpu_access;
1023 i915_gem_object_move_to_active(obj, ring, seqno);
1024 if (obj->base.write_domain) {
1026 obj->pending_gpu_write = true;
1027 list_move_tail(&obj->gpu_write_list,
1028 &ring->gpu_write_list);
1029 intel_mark_busy(ring->dev, obj);
1032 trace_i915_gem_object_change_domain(obj, old_read, old_write);
1037 i915_gem_execbuffer_retire_commands(struct drm_device *dev,
1038 struct drm_file *file,
1039 struct intel_ring_buffer *ring)
1041 struct drm_i915_gem_request *request;
1045 * Ensure that the commands in the batch buffer are
1046 * finished before the interrupt fires.
1048 * The sampler always gets flushed on i965 (sigh).
1050 invalidate = I915_GEM_DOMAIN_COMMAND;
1051 if (INTEL_INFO(dev)->gen >= 4)
1052 invalidate |= I915_GEM_DOMAIN_SAMPLER;
1053 if (ring->flush(ring, invalidate, 0)) {
1054 i915_gem_next_request_seqno(ring);
1058 /* Add a breadcrumb for the completion of the batch buffer */
1059 request = kzalloc(sizeof(*request), GFP_KERNEL);
1060 if (request == NULL || i915_add_request(ring, file, request)) {
1061 i915_gem_next_request_seqno(ring);
1067 i915_reset_gen7_sol_offsets(struct drm_device *dev,
1068 struct intel_ring_buffer *ring)
1070 drm_i915_private_t *dev_priv = dev->dev_private;
1073 if (!IS_GEN7(dev) || ring != &dev_priv->ring[RCS])
1076 ret = intel_ring_begin(ring, 4 * 3);
1080 for (i = 0; i < 4; i++) {
1081 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1082 intel_ring_emit(ring, GEN7_SO_WRITE_OFFSET(i));
1083 intel_ring_emit(ring, 0);
1086 intel_ring_advance(ring);
1092 i915_gem_do_execbuffer(struct drm_device *dev, void *data,
1093 struct drm_file *file,
1094 struct drm_i915_gem_execbuffer2 *args,
1095 struct drm_i915_gem_exec_object2 *exec)
1097 drm_i915_private_t *dev_priv = dev->dev_private;
1098 struct list_head objects;
1099 struct eb_objects *eb;
1100 struct drm_i915_gem_object *batch_obj;
1101 struct drm_clip_rect *cliprects = NULL;
1102 struct intel_ring_buffer *ring;
1103 u32 exec_start, exec_len;
1108 if (!i915_gem_check_execbuffer(args)) {
1109 DRM_DEBUG("execbuf with invalid offset/length\n");
1113 ret = validate_exec_list(exec, args->buffer_count);
1117 switch (args->flags & I915_EXEC_RING_MASK) {
1118 case I915_EXEC_DEFAULT:
1119 case I915_EXEC_RENDER:
1120 ring = &dev_priv->ring[RCS];
1123 if (!HAS_BSD(dev)) {
1124 DRM_DEBUG("execbuf with invalid ring (BSD)\n");
1127 ring = &dev_priv->ring[VCS];
1130 if (!HAS_BLT(dev)) {
1131 DRM_DEBUG("execbuf with invalid ring (BLT)\n");
1134 ring = &dev_priv->ring[BCS];
1137 DRM_DEBUG("execbuf with unknown ring: %d\n",
1138 (int)(args->flags & I915_EXEC_RING_MASK));
1142 mode = args->flags & I915_EXEC_CONSTANTS_MASK;
1143 mask = I915_EXEC_CONSTANTS_MASK;
1145 case I915_EXEC_CONSTANTS_REL_GENERAL:
1146 case I915_EXEC_CONSTANTS_ABSOLUTE:
1147 case I915_EXEC_CONSTANTS_REL_SURFACE:
1148 if (ring == &dev_priv->ring[RCS] &&
1149 mode != dev_priv->relative_constants_mode) {
1150 if (INTEL_INFO(dev)->gen < 4)
1153 if (INTEL_INFO(dev)->gen > 5 &&
1154 mode == I915_EXEC_CONSTANTS_REL_SURFACE)
1157 /* The HW changed the meaning on this bit on gen6 */
1158 if (INTEL_INFO(dev)->gen >= 6)
1159 mask &= ~I915_EXEC_CONSTANTS_REL_SURFACE;
1163 DRM_DEBUG("execbuf with unknown constants: %d\n", mode);
1167 if (args->buffer_count < 1) {
1168 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1172 if (args->num_cliprects != 0) {
1173 if (ring != &dev_priv->ring[RCS]) {
1174 DRM_DEBUG("clip rectangles are only valid with the render ring\n");
1178 cliprects = kmalloc(args->num_cliprects * sizeof(*cliprects),
1180 if (cliprects == NULL) {
1185 if (copy_from_user(cliprects,
1186 (struct drm_clip_rect __user *)(uintptr_t)
1187 args->cliprects_ptr,
1188 sizeof(*cliprects)*args->num_cliprects)) {
1194 ret = i915_mutex_lock_interruptible(dev);
1198 if (dev_priv->mm.suspended) {
1199 mutex_unlock(&dev->struct_mutex);
1204 eb = eb_create(args->buffer_count);
1206 mutex_unlock(&dev->struct_mutex);
1211 /* Look up object handles */
1212 INIT_LIST_HEAD(&objects);
1213 for (i = 0; i < args->buffer_count; i++) {
1214 struct drm_i915_gem_object *obj;
1216 obj = to_intel_bo(drm_gem_object_lookup(dev, file,
1218 if (&obj->base == NULL) {
1219 DRM_DEBUG("Invalid object handle %d at index %d\n",
1221 /* prevent error path from reading uninitialized data */
1226 if (!list_empty(&obj->exec_list)) {
1227 DRM_DEBUG("Object %p [handle %d, index %d] appears more than once in object list\n",
1228 obj, exec[i].handle, i);
1233 list_add_tail(&obj->exec_list, &objects);
1234 obj->exec_handle = exec[i].handle;
1235 obj->exec_entry = &exec[i];
1236 eb_add_object(eb, obj);
1239 /* take note of the batch buffer before we might reorder the lists */
1240 batch_obj = list_entry(objects.prev,
1241 struct drm_i915_gem_object,
1244 /* Move the objects en-masse into the GTT, evicting if necessary. */
1245 ret = i915_gem_execbuffer_reserve(ring, file, &objects);
1249 /* The objects are in their final locations, apply the relocations. */
1250 ret = i915_gem_execbuffer_relocate(dev, eb, &objects);
1252 if (ret == -EFAULT) {
1253 ret = i915_gem_execbuffer_relocate_slow(dev, file, ring,
1256 args->buffer_count);
1257 BUG_ON(!mutex_is_locked(&dev->struct_mutex));
1263 /* Set the pending read domains for the batch buffer to COMMAND */
1264 if (batch_obj->base.pending_write_domain) {
1265 DRM_DEBUG("Attempting to use self-modifying batch buffer\n");
1269 batch_obj->base.pending_read_domains |= I915_GEM_DOMAIN_COMMAND;
1271 ret = i915_gem_execbuffer_move_to_gpu(ring, &objects);
1275 seqno = i915_gem_next_request_seqno(ring);
1276 for (i = 0; i < ARRAY_SIZE(ring->sync_seqno); i++) {
1277 if (seqno < ring->sync_seqno[i]) {
1278 /* The GPU can not handle its semaphore value wrapping,
1279 * so every billion or so execbuffers, we need to stall
1280 * the GPU in order to reset the counters.
1282 ret = i915_gpu_idle(dev, true);
1286 BUG_ON(ring->sync_seqno[i]);
1290 if (ring == &dev_priv->ring[RCS] &&
1291 mode != dev_priv->relative_constants_mode) {
1292 ret = intel_ring_begin(ring, 4);
1296 intel_ring_emit(ring, MI_NOOP);
1297 intel_ring_emit(ring, MI_LOAD_REGISTER_IMM(1));
1298 intel_ring_emit(ring, INSTPM);
1299 intel_ring_emit(ring, mask << 16 | mode);
1300 intel_ring_advance(ring);
1302 dev_priv->relative_constants_mode = mode;
1305 if (args->flags & I915_EXEC_GEN7_SOL_RESET) {
1306 ret = i915_reset_gen7_sol_offsets(dev, ring);
1311 trace_i915_gem_ring_dispatch(ring, seqno);
1313 exec_start = batch_obj->gtt_offset + args->batch_start_offset;
1314 exec_len = args->batch_len;
1316 for (i = 0; i < args->num_cliprects; i++) {
1317 ret = i915_emit_box(dev, &cliprects[i],
1318 args->DR1, args->DR4);
1322 ret = ring->dispatch_execbuffer(ring,
1323 exec_start, exec_len);
1328 ret = ring->dispatch_execbuffer(ring, exec_start, exec_len);
1333 i915_gem_execbuffer_move_to_active(&objects, ring, seqno);
1334 i915_gem_execbuffer_retire_commands(dev, file, ring);
1338 while (!list_empty(&objects)) {
1339 struct drm_i915_gem_object *obj;
1341 obj = list_first_entry(&objects,
1342 struct drm_i915_gem_object,
1344 list_del_init(&obj->exec_list);
1345 drm_gem_object_unreference(&obj->base);
1348 mutex_unlock(&dev->struct_mutex);
1356 * Legacy execbuffer just creates an exec2 list from the original exec object
1357 * list array and passes it to the real function.
1360 i915_gem_execbuffer(struct drm_device *dev, void *data,
1361 struct drm_file *file)
1363 struct drm_i915_gem_execbuffer *args = data;
1364 struct drm_i915_gem_execbuffer2 exec2;
1365 struct drm_i915_gem_exec_object *exec_list = NULL;
1366 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1369 if (args->buffer_count < 1) {
1370 DRM_DEBUG("execbuf with %d buffers\n", args->buffer_count);
1374 /* Copy in the exec list from userland */
1375 exec_list = drm_malloc_ab(sizeof(*exec_list), args->buffer_count);
1376 exec2_list = drm_malloc_ab(sizeof(*exec2_list), args->buffer_count);
1377 if (exec_list == NULL || exec2_list == NULL) {
1378 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1379 args->buffer_count);
1380 drm_free_large(exec_list);
1381 drm_free_large(exec2_list);
1384 ret = copy_from_user(exec_list,
1385 (struct drm_i915_relocation_entry __user *)
1386 (uintptr_t) args->buffers_ptr,
1387 sizeof(*exec_list) * args->buffer_count);
1389 DRM_DEBUG("copy %d exec entries failed %d\n",
1390 args->buffer_count, ret);
1391 drm_free_large(exec_list);
1392 drm_free_large(exec2_list);
1396 for (i = 0; i < args->buffer_count; i++) {
1397 exec2_list[i].handle = exec_list[i].handle;
1398 exec2_list[i].relocation_count = exec_list[i].relocation_count;
1399 exec2_list[i].relocs_ptr = exec_list[i].relocs_ptr;
1400 exec2_list[i].alignment = exec_list[i].alignment;
1401 exec2_list[i].offset = exec_list[i].offset;
1402 if (INTEL_INFO(dev)->gen < 4)
1403 exec2_list[i].flags = EXEC_OBJECT_NEEDS_FENCE;
1405 exec2_list[i].flags = 0;
1408 exec2.buffers_ptr = args->buffers_ptr;
1409 exec2.buffer_count = args->buffer_count;
1410 exec2.batch_start_offset = args->batch_start_offset;
1411 exec2.batch_len = args->batch_len;
1412 exec2.DR1 = args->DR1;
1413 exec2.DR4 = args->DR4;
1414 exec2.num_cliprects = args->num_cliprects;
1415 exec2.cliprects_ptr = args->cliprects_ptr;
1416 exec2.flags = I915_EXEC_RENDER;
1418 ret = i915_gem_do_execbuffer(dev, data, file, &exec2, exec2_list);
1420 /* Copy the new buffer offsets back to the user's exec list. */
1421 for (i = 0; i < args->buffer_count; i++)
1422 exec_list[i].offset = exec2_list[i].offset;
1423 /* ... and back out to userspace */
1424 ret = copy_to_user((struct drm_i915_relocation_entry __user *)
1425 (uintptr_t) args->buffers_ptr,
1427 sizeof(*exec_list) * args->buffer_count);
1430 DRM_DEBUG("failed to copy %d exec entries "
1431 "back to user (%d)\n",
1432 args->buffer_count, ret);
1436 drm_free_large(exec_list);
1437 drm_free_large(exec2_list);
1442 i915_gem_execbuffer2(struct drm_device *dev, void *data,
1443 struct drm_file *file)
1445 struct drm_i915_gem_execbuffer2 *args = data;
1446 struct drm_i915_gem_exec_object2 *exec2_list = NULL;
1449 if (args->buffer_count < 1) {
1450 DRM_DEBUG("execbuf2 with %d buffers\n", args->buffer_count);
1454 exec2_list = kmalloc(sizeof(*exec2_list)*args->buffer_count,
1455 GFP_KERNEL | __GFP_NOWARN | __GFP_NORETRY);
1456 if (exec2_list == NULL)
1457 exec2_list = drm_malloc_ab(sizeof(*exec2_list),
1458 args->buffer_count);
1459 if (exec2_list == NULL) {
1460 DRM_DEBUG("Failed to allocate exec list for %d buffers\n",
1461 args->buffer_count);
1464 ret = copy_from_user(exec2_list,
1465 (struct drm_i915_relocation_entry __user *)
1466 (uintptr_t) args->buffers_ptr,
1467 sizeof(*exec2_list) * args->buffer_count);
1469 DRM_DEBUG("copy %d exec entries failed %d\n",
1470 args->buffer_count, ret);
1471 drm_free_large(exec2_list);
1475 ret = i915_gem_do_execbuffer(dev, data, file, args, exec2_list);
1477 /* Copy the new buffer offsets back to the user's exec list. */
1478 ret = copy_to_user((struct drm_i915_relocation_entry __user *)
1479 (uintptr_t) args->buffers_ptr,
1481 sizeof(*exec2_list) * args->buffer_count);
1484 DRM_DEBUG("failed to copy %d exec entries "
1485 "back to user (%d)\n",
1486 args->buffer_count, ret);
1490 drm_free_large(exec2_list);